Communications

in Computer and Information Science 204

Dhinaharan Nagamalai Eric Renault
Murugan Dhanuskodi (Eds.)

Trends in Computer Science,

Engineering and
Information Technology

First International Conference on Computer Science,

Engineering and Information Technology, CCSEIT 2011
Tirunelveli, Tamil Nadu, India, September 23-25, 2011
Proceedings

13
Volume Editors

Dhinaharan Nagamalai
Wireilla Net Solutions PTY Ltd
Melbourne, VIC, Australia
E-mail: [email protected]

Eric Renault
Institut Telecom/Telecom SudParis (ex. GET-INT)
Departement Reseaux et Services Multimedia Mobiles (RS2M)
Samovar UMR INT-CNRS 5157
9, rue Charles Fourier, 91011 Evry Cedex, France
E-mail: [email protected]

Murugan Dhanuskodi
Manonmaniam Sundaranar University
Department of Computer Science and Engineering
Tirunelveli, Tamil Nadu, India
E-mail: [email protected]

ISSN 1865-0929 e-ISSN 1865-0937

ISBN 978-3-642-24042-3 e-ISBN 978-3-642-24043-0
DOI 10.1007/978-3-642-24043-0
Springer Heidelberg Dordrecht London New York

Library of Congress Control Number: Applied for

CR Subject Classification (1998): I.2, C.2, H.4, H.3, D.2, J.1

© Springer-Verlag Berlin Heidelberg 2011

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer. Violations are liable
to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
 Preface

The First International Conference on Computer Science, Engineering and In-

formation Technology (CCSEIT-2011), The First International Conference on
Parallel, Distributed Computing Technologies and Applications (PDCTA-2011),
and The First International Conference on Digital Image Processing and Pat-
tern Recognition (DPPR-2011) were held in Tirunelveli - Tamil Nadu, India,
during September 23–25, 2011. The events attracted many local and interna-
tional delegates, presenting a balanced mixture of intellects from all over the
world. The goal of this conference series is to bring together researchers and
practitioners from academia and industry to focus on understanding parallel,
distributed computing technologies, digital image processing and pattern recog-
nition and all areas of computer science, information technology, and to establish
new collaborations in these areas.
The CCSEIT 2011, PDCTA 2011 and DPPR 2011 committees invited original
submissions from researchers, scientists, engineers, and students that illustrate
research results, projects, survey works, and industrial experiences describing
significant advances in the areas related to the relevant themes and tracks of the
conferences. This effort guaranteed submissions from an unparalleled number of
internationally recognized top-level researchers. All the submissions underwent
a strenuous peer-review process which comprised expert reviewers. Besides the
members of the Technical Program Committee, external reviewers were invited
on the basis of their specialization and expertise. The papers were reviewed based
on their technical content, originality, and clarity. The entire process, which in-
cludes the submission, review, and acceptance processes, was done electronically.
All these efforts undertaken by the Organizing and Technical Program Commit-
tees led to an exciting, rich, and high-quality technical conference program, which
featured high-impact presentations for all attendees to enjoy and to expand their
expertise in the latest developments in this field.
There were a total 1,256 submissions to the conference, and the Technical
Program Committee selected 185 papers for presentation at the conference and
subsequent publication in the proceedings. This small introduction would be
incomplete without expressing our gratitude and thanks to the General and
Program Chairs, members of the Technical Program Committees, and external
reviewers for their excellent and diligent work. Thanks to Springer for the strong
support. Finally, we thank all the authors who contributed to the success of the
conference. We also sincerely wish that all attendees benefited academically from
the conference and wish them every success in their research.

Dhinaharan Nagamalai
Eric Renault
Murugan Dhanushkodi
 Organization

General Chairs
David C. Wyld Southeastern Louisiana University, USA
Michal Wozniak Wroclaw University of Technology, Poland

Steering Committee Chairs

Murugan Dhanuskodi Manonmaniam Sundaranar University,
India
Jan Zizka SoNet/DI, FBE, Mendel University in
Brno, Czech Republic
John Karamitsos University of the Aegean, Samos, Greece
Khoa N. Le University of Western Sydney, Australia
Nabendu Chaki University of Calcutta, India
Salah S. Al-Majeed University of Essex, UK
Dhinaharan Nagamalai Wireilla Net Solutions, Australia

Publicity Chairs
Roberts Masillamani Hindustan University, India
Chih-Lin Hu National Central University, Taiwan

Program Committee
A. Arokiasamy Eastern Mediterranean University, Cyprus
A.P. Sathish Kumar PSG Institute of Advanced Studies, India
Abdul Aziz University of Central Punjab, Pakistan
Abdul Kadir Ozcan The American University, Cyprus
Al-Sakib Khan Pathan Kyung Hee University, South Korea
Andreas Riener Johannes Kepler University Linz, Austria
Andy Seddon Asia Pacific Institute of Information
Technology, Malaysia
Antelin Vijila Manonmaniam Sundaranar University,
India
Arvinth Kumar M.S. University, India
Athanasios Vasilakos University of Western Macedonia, Greece
Atilla Elci Eastern Mediterranean University, Cyprus
B. Srinivasan Monash University, Australia
VIII Organization

Balasubramaniam Manonmaniam Sundaranar University,

India
Bong-Han Kim Chongju University , South Korea
Boo-Hyung Lee KongJu National University, South Korea
Brajesh Kumar Kaushik Indian Institute of Technology, India
Charalampos Z. Patrikakis National Technical University of
Athens, Greece
Chih-Lin Hu National Central University, Taiwan
Chin-Chih Chang Chung Hua University, Taiwan
Cho Han Jin Far East University, South Korea
Cynthia Dhinakaran Hannam University, South Korea
Danda B. Rawat Old Dominion University, USA
David W. Deeds Shingu College, South Korea
Debasis Giri Haldia Institute of Technology, India
Deepak Garg Thapar University, India
Dhinaharan Nagamalai Wireilla Net Solutions Pty Ltd., Australia
Dimitris Kotzinos Technical Educational Institution of
Serres, Greece
Dong Seong Kim Duke University, USA
Emmanuel Bouix iKlax Media, France
Eric Renault Institut Telecom – Telecom SudParis,
France
Farhat Anwar International Islamic University , Malaysia
Firkhan Ali Bin Hamid Ali Universiti Tun Hussein Onn Malaysia,
Malaysia
Ford Lumban Gaol University of Indonesia
Geuk Lee Hannam University, South Korea
Girija Chetty University of Canberra, Australia
H.V. Ramakrishnan Dr. MGR University, India
Henrique Joao Lopes Domingos University of Lisbon, Portugal
Ho Dac Tu Waseda University, Japan
Hoang, Huu Hanh Hue University, Vietnam
Hwangjun Song Pohang University of Science and
Technology, South Korea
J.Arunadevi Thiagarajar College, Madurai, India
Jacques Demerjian Communication & Systems, Homeland
Security, France
Jae Kwang Lee Hannam University, South Korea
Jan Zizka SoNet/DI, FBE, Mendel University in
Brno, Czech Republic
Jansirani Manonmaniam Sundaranar University,
India
Jeong-Hyun Park Electronics Telecommunication Research
Institute, South Korea
Jivesh Govil Cisco Systems Inc. - CA, USA
Johann Groschd University of Bristol, UK
 Organization IX

John Karamitsos University of the Aegean, Samos, Greece

Johnson Kuruvila Dalhousie University, Halifax, Canada
Jose Enrique Armendariz-Inigo Universidad Publica de Navarra, Spain
Jung-Gil Song Hannam University, South Korea
Jungwook Song Konkuk University, South Korea
K.P. Thooyamani Bharath University, India
Kamaljit I Lakhtaria Atmiya Institute of Technology & Science,
India
Kannan Anna University, Chennai, India
Khamish Malhotra University of Glamorgan, UK
Khoa N. Le Griffith University, Australia
Krzysztof Walkowiak Wroclaw University of Technology, Poland
L. Ganesan Alagappa University, India
Lu S. Veiga Technical University of Lisbon, Portugal
Lu Yan University of Hertfordshire, UK
Maode Ma Nanyang Technological University,
Singapore
Marco Roccetti Universty of Bologna, Italy
Michael Peterson University of Hawaii at Hilo, USA
Michal Wozniak Wroclaw University of Technology, Poland
Mohsen Sharifi Iran University of Science and Technology,
Iran
Murugan D. Manonmaniam Sundaranar University,
India
Murugeswari M.S. University, India
Muthulakshmi M.S. University, India
N. Krishnan Manonmaniam Sundaranar University,
India
Nabendu Chaki University of Calcutta, India
Natarajan Meghanathan Jackson State University, USA
Neerajkumar Madha Vaisnavi Devi University, India
Nicolas Sklavos Technological Educational Institute of
Patras, Greece
Nidaa Abdual Muhsin Abbas University of Babylon, Iraq
Paul D. Manuel Kuwait University, Kuwait
Phan Cong Vinh London South Bank University, UK
Ponpit Wongthongtham Curtin University of Technology, Australia
Prabu Dorairaj NetApp Inc., India
Rajalakshmi Manonmaniam Sundaranar University,
India
Rajendra Akerkar Technomathematics Research Foundation,
India
Rajesh Kumar P. The Best International, Australia
Rajesh Manonmaniam Sundaranar University,
India
Rajesh Bawa Punjabi University, India
X Organization

Rajkumar Kannan Bishop Heber College, India

Rakhesh Singh Kshetrimayum Indian Institute of Technology-Guwahati,
India
Ramayah Thurasamy Universiti Sains Malaysia, Malaysia
Rituparna Chaki West Bengal University of Technology,
India
Roberts Masillamani Hindustan University, India
S. Arumugam Nandha Engineering College, India
S. Hariharan B.S.Abdur Rahman University, India
Sadasivam Manonmaniam Sundaranar University,
India
Sagarmay Deb Central Queensland University, Australia
Sajid Hussain Acadia University, Canada
Sajid Hussain Fisk University, USA
Salah S. Al-Majeed University of Essex, UK
Sanguthevar Rajasekaran University of Connecticut, USA
Sarmistha Neogy Jadavpur University, India
Sattar B. Sadkhan University of Babylon, Iraq
Seemabawa Thapar University, India
Sergio Ilarri University of Zaragoza, Spain
Serguei A. Mokhov Concordia University, Canada
Seungmin Rho Carnegie Mellon University, USA
Shivan Haran Arizona State University, USA
Somitra Sanadhya IIT-Delhi, India
Soodamani, ramalingam Uuniversity of Hertfordshire,UK
Sriman Narayana Iyengar VIT University, India
Subha M.S. University, India
Sudip Misra Indian Institute of Technology-Kharagpur,
India
Sundarapandian Viadyanathan Vel Tech Dr.RR & Dr.SR Technical
University, India
Sundaresan Manonmaniam Sundaranar University,
India
SunYoung Han Konkuk University, South Korea
Suruliandi Manonmaniam Sundaranar University,
India
Susana Sargento University of Aveiro, Portugal
Syed Rahman University of Hawaii-Hilo, USA
Syed Rizvi University of Bridgeport, USA
Taruna S. Banasthali University, India
Thamaraiselvi Madras Institute of Technology,India
Thambidurai Pondicherry University, India
Velmurugan Ayyadurai Center for Communication Systems, UK
Vishal Sharma Metanoia Inc., USA
Wei Jie University of Manchester, UK
 Organization XI

Yan Luo University of Massachusetts Lowell, USA

Yannick Le Moullec Aalborg University, Denmark
Yao-Nan Lien National Chengchi University, Taiwan
Yeong Deok Kim Woosong University, South Korea
Yuh-Shyan Chen National Taipei University, Taiwan
Yung-Fa Huang Chaoyang University of Technology,
Taiwan

External Reviewers
Alejandro Regalado Mendez Universidad del Mar, Mexico
Alireza Mahini Islamic Azad University-Gorgan, Iran
Amandeep Singh Thethi Guru Nanak Dev University Amritsar,
India
Ashok Kumar Sharma YMCA Institute of Engineering, India
Ayman Khalil Institute of Electronics and
Telecommunications of Rennes (IETR),
France
Buket Barkana University of Bridgeport, USA
Christos Politis Kingston University, UK
Hao Shi Victoria University, Australia
Indrajit Bhattacharya Kalyani Government Engineering College,
India
Jyotirmay Gadewadikar Alcorn State University, USA
Khoa N. Le University of Western Sydney, Australia
Laiali Almazaydeh University of Bridgeport, USA
Lakshmi Rajamani Osmania University, India
Michel Owayjan American University of Science &
Technology - AUST, Lebanon
Mohamed Hassan American University of Sharjah, UAE
Monika Verma Punjab Technical University, India
N.K. Choudhari Bhagwati Chaturvedi College of
Engineering, India
Nitiket N. Mhala B.D. College of Engineering - Sewagram,
India
Nour Eldin Elmadany Arab Acadmy for Science and
Technology, Egypt
Premanand K. Kadbe Vidya Pratishthan’s College of Engineering,
India
R. Murali Dr. Ambedkar Institute of Technology,
Bangalore, India
Raman Maini Punjabi University, India
Rushed Kanawati LIPN - Université Paris 13, France
S.A.V. Satyamurty Indira Gandhi Centre for Atomic Research,
India
Shrikant K. Bodhe Bosh Technologies, India
XII Organization

Sridharan CEG Campus - Anna University, India

Utpal Biswas University of Kalyani, India
Wichian Sittiprapaporn Mahasarakham University, Thailand

Technically Supported By
Networks & Communications Community (NCC)
Digital Signal & Image Processing Community (DSIPC)
Computer Science & Information Technology Community (CSITC)

Organized By

ACADEMY & INDUSTRY RESEARCH COLLABORATION CENTER (AIRCC)

 Table of Contents

Computer Science, Engineering and Information

Technology
Hybrid Chaos Synchronization of Liu and Lü Systems by Active
Nonlinear Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Sundarapandian Vaidyanathan

Methods for Preventing Search Engine-Based Web Server Attacks . . . . . . 11

Keerthiram Murugesan and Mukesh Singhal

Sea Object Detection Using Shape and Hybrid Color Texture

Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
M. Uma Selvi and S. Suresh Kumar

Multi-layer Logon Verification System: A Case Study of Indian

Banks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Mohammad Asim

Image Compression Using Two Dimensional Spherical Coder in Wavelet

Lifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
S. Panimalar, R. Anisha, and M. Gomathi

An Efficient Visible Watermarking for Copyright Protection Using

Discrete Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
R. Anisha, S. Panimalar, and M. Gomathi

Evolving Test Patterns for Use Case Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 65

P.G. Sapna, Hrushikesha Mohanty, and Arunkumar Balakrishnan

A Novel Approach for Compression of Encrypted Grayscale Images

Using Huffman Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
S. Saravana Kumar and S. Panimalar

Global Chaos Synchronization of Hyperchaotic Pang and Wang Systems

by Active Nonlinear Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Sundarapandian Vaidyanathan and Karthikeyan Rajagopal

Mining Web Path Traversals Based on Generation of FP Tree with

Utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
E. Poovammal and Pillai Cigith

Evaluating Degree of Dependency from Domain Knowledge Using

Fuzzy Inference System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Vibha Gaur and Anuja Soni
XIV Table of Contents

Novel Methodologies to Avert Hacker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

Paul Bharath Bhushan Petlu, N.L. Kumar Anantapalli,
Devaki Pendlimarri, and Dr. M. Muralidhara Rao

Multi-Domain Meta Search Engine with an Intelligent Interface for

Efficient Information Retrieval on the Web . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Minnie D. and Srinivasan S.

Improved Video Watermarking Using Nested Watermark, Wavelets and

Geometric Warping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
T. Jayamalar and V. Radha

Robust Fault Tolerant AOMDV Routing Mechanism in MANET . . . . . . . 142

Jayalakhsmi Vaithiyanathan and B. SenthilRajaManokar

ZKIP and Formal System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

M. Thiyagarajan and S. Samundeeswari

Personalized Web Based Collaborative Learning in Web 3.0:

A Technical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
S. Padma and Ananthi Seshasaayee

Rule Acquisition in Data Mining Using a Self Adaptive Genetic

Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
K. Indira, S. Kanmani, D. Gaurav Sethia, S. Kumaran, and
J. Prabhakar

Dynamic Multi Dimensional Matchmaking Model for Resource

Allocation in Grid Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Japhynth Jacob, Elijah Blessing Rajsingh, and
Isaac Balasingh Jesudasan

A Comparative Study of CSO and PSO Trained Artificial Neural

Network for Stock Market Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Suresh Chittineni, Vabbilisetty Mounica, Kaligotla Abhilash,
Suresh Chandra Satapathy, and P.V.G.D. Prasad Reddy

Scaled Conjugate Gradient Algorithm in Neural Network Based

Approach for Handwritten Text Recognition . . . . . . . . . . . . . . . . . . . . . . . . . 196
Haradhan Chel, Aurpan Majumder, and Debashis Nandi

A Llinear Framework for Identification of Ship and Its Velocity

Estimation Using SAR Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
P. Subashini and M. Krishnaveni

Creational Patterns to Create Decorator Pattern Objects in Web

Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Vijay K. Kerji
 Table of Contents XV

Scalable Geographic Based Multicasting (SGBM) Protocol . . . . . . . . . . . . 233

H. Venkateswaran and A. John Prakash
Median Adjusted Constrained PDF Based Histogram Equalization for
Image Contrast Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
P. Shanmugavadivu, K. Balasubramanian, and K. Somasundaram
Isolated Word Recognition System Using Back Propagation Network
for Tamil Spoken Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
V. Radha, C. Vimala, and M. Krishnaveni
Enhanced Hybrid Compression Models for Compound Images . . . . . . . . . 265
D. Maheswari and Dr. V. Radha
Index Based Ordered Broadcast with Status (IOBS) Algorithm for
Consistent Data Broadcast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
K. Chitra, K. Senthamarai Kannan, and S. Abirami
Capturing High-Level Semantics of Images in Web Documents Using
Strength Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
P. Shanmugavadivu, P. Sumathy, and A. Vadivel
Multiple Access Scheme Using MSE-OFDM with Pseudo Random
Cyclic Prefix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Biswajit Datta, Anupam Karmakar, and Mrinal K. Naskar
Elimination of Irrelevancy during Semantic Service Discovery Using
Clustering Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Chellammal Surianarayanan and Gopinath Ganapathy
Leaf and Flower Recognition Using Preferential Image Segmentation
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
N. Valliammal and S.N. Geethalakshmi
Analysis of Next Hop Selection for Geocasting in VANET . . . . . . . . . . . . . 326
Sanjoy Das and D.K. Lobiyal
Multi Objective Genetic Approach for Solving Vehicle Routing Problem
with Time Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
Padmabati Chand and J.R. Mohanty
Application of Modified NSGA-II Algorithm to Reactive Power
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
S. Ramesh, S. Kannan, and S. Baskar
A Revised TESOR Algorithm for Perfect Routing . . . . . . . . . . . . . . . . . . . . 355
P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani
An Accomplishment of New Level Register File Management
Architecture for Montgomery Algorithm in Elliptic Curve . . . . . . . . . . . . . 367
M. Prabu and R. Shanmugalakshmi
XVI Table of Contents

Modified Partial Differential Equations Based Adaptive Two-Stage

Median Filter for Images Corrupted with High Density Fixed-Value
Impulse Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
P. Shanmugavadivu and P.S. Eliahim Jeevaraj

Enhanced Classification Performance Using Computational

Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Indrajit Mandal and N. Sairam

Feature Modeling of the Evolving Access Control Requirements . . . . . . . . 392

K. Shantha Kumari and T. Chithralekha

Fault Diagnosis of Pneumatic Valve Using PCA and ANN

Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
P. Subbaraj and B. Kannapiran

A Generalized Classification Approach for Indian Sign Language

Datasets Using Instance Based Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
M. Krishnaveni and V. Radha

Agent Based Adaptive Learning System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423

Gowri R., Kanmani S., and Sampath Kumar T.T.

GPGPU Implementation of Parallel Memetic Algorithm for VLSI

Floorplanning Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
Subbaraj Potti and Sivakumar Pothiraj

Improving Cloud Security through Virtualization . . . . . . . . . . . . . . . . . . . . . 442

B. Loganayagi and S. Sujatha

Selection of Software Testing Technique: A Multi Criteria Decision

Making Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
Monisha Victor and Nitin Upadhyay

A Hybrid Evolutionary Algorithm for the Page Number Minimization

Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Dharna Satsangi, Kamal Srivastava, and Gursaran

Trust-Enhanced Recommendation of Friends in Web Based Social

Networks Using Genetic Algorithms to Learn User Preferences . . . . . . . . . 476
Vinti Agarwal and Kamal K. Bharadwaj

Modeling and Verification of Server Aided Verification Protocol Using

NuSMV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
Vikram Saralaya, Kishore J.K., Sateesh Reddy, and Sanjay Singh

Modeling and Verification of Inter Realm Authentication in Kerberos

Using Symbolic Model Verifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
Punit Mundra, Madhavi Sharma, Shobhit Shukla, and Sanjay Singh
 Table of Contents XVII

A New Bluetooth Security Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507

Mintu Philip and Asha Das
A New Trace Backing Algorithm for Maximizing Streaming Data
Join . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
G. Hemalatha and K. Thanushkodi
A Comparative Analysis of Demand Assignment Multiple Access
Protocols for Wireless ATM Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
C. Shanmuganathan and Dr. P. Raviraj
Perceptual Linear Predictive Cepstral Coefficient for Malayalam
Isolated Digit Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
Cini Kurian and Kannan Balakrishnan
Agent Based Modeling of Individual Voting Preferences with Social
Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 542
Vivek Kumar Singh, Swati Basak, and Neelam Modanwal
Texture Analysis of Brain MRI and Classification with BPN for the
Diagnosis of Dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
T.R. Sivapriya, V. Saravanan, and P. Ranjit Jeba Thangaiah
A Host Based Kernel Level Rootkit Detection Mechanism Using
Clustering Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
Jestin Joy and Anita John
Fuzzy Clustering for Effective Customer Relationship Management in
Telecom Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
Gayathri Asokan and S. Mohanavalli
Dematerialized Deposits Using XSI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
V. Radha, Surabhi Upender, N.V.K.D. Ramesh Gorrila, and
B. Vijay Kumar
Classifying Gender from Faces Using Independent Components . . . . . . . . 589
Sunita Kumari and Banshidhar Majhi
Experimental Evaluation of Fast Solution Algorithms for the Simple
Motif Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
Tarek El Falah, Thierry Lecroq, and Mourad Elloumi
Secure Web Services Negotiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
Vakul Mohanty and Chittaranjan Hota
A Recommender System for Rural and Urban Learners . . . . . . . . . . . . . . . 619
L. Arockiam, S. Charles, V. Arul Kumar, and P. Cijo
Study of Organizational Factors Affecting Usability of Online Helps
with Programming Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 628
Nilesh Dhannaseth and Ganesh Bhutkar
XVIII Table of Contents

Finding Reliable Path for Peer to Peer Application in Mobile

Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
D. Jayashree and P. Yogesh

Text Mining for Interpreting Gene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647

K. Prabavathy and P. Sumathi

Improved Input-Delayed Dynamic Neural Network for GPS/INS

Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

Application of Artificial Neural Network for the Prediction of

Groundwater Level in Hard Rock Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
M. Kavitha Mayilvaganan and K.B. Naidu

Energy Efficient and Fault Tolerant GPSR in Ad Hoc Wireless

Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683
Jyotsana Jaiswal and Pabitra Mohan Khilar

An IDE for Developing Mobile Applications for Smart Phones . . . . . . . . . 693

Sakila Banu Ali and Nisar Hundewale

Function Approximation Using SVM with FCM and Slope Based

Partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704
Amera Almas and Nisar Hundewale

A PIC-Controller Based Doubly Fed Induction Generator (DFIG) for

Wind Energy Conversion Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
K. Sureshkumar, P. Vijaya kumar, R. Dhandayuthabani, and
A. Sakthivel

Packet Scheduling Scheme with Enhanced Quality of Service for Mobile

WiMAX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735

 Hybrid Chaos Synchronization of Liu and Lü Systems
by Active Nonlinear Control

Sundarapandian Vaidyanathan

Research and Development Centre,

Vel Tech Dr. RR & Dr. SR Technical University,
Avadi-Alamathi Road, Avadi, Chennai-600 062, India
[email protected]
http://www.vel-tech.org/

Abstract. This paper investigates the hybrid chaos synchronization of identical

Liu systems, identical Lü systems, and non-identical Liu and Lü systems by ac-
tive nonlinear control. Liu system (Liu et al. 2004) and Lü system (Lü and Chen,
2002) are important models of three-dimensional chaotic systems. Hybrid syn-
chronization of the three-dimensional chaotic systems considered in this paper
are achieved through the synchronization of the first and last pairs of states and
anti-synchronization of the middle pairs of the two systems. Sufficient conditions
for hybrid synchronization of identical Liu, identical Lü, and non-identical Liu
and Lü systems are derived using active nonlinear control and Lyapunov stabil-
ity theory. Since the Lyapunov exponents are not needed for these calculations,
the active nonlinear control is an effective and convenient method for the hybrid
synchronization of the chaotic systems addressed in this paper. Numerical simu-
lations are shown to illustrate the effectiveness of the proposed synchronization
schemes.

Keywords: Chaos, hybrid synchronization, Liu system, Lü system, nonlinear

control.

1 Introduction

Chaotic systems are dynamical systems that are highly sensitive to initial conditions.
This sensitivity is popularly referred to as the butterfly effect [1].
Chaos is an interesting nonlinear phenomenon and has been extensively and inten-
sively studied in the last two decades ([1]-[17]). In 1990, Pecora and Carroll [2] intro-
duced a method to synchronize two identical chaotic systems and showed that it was
possible for some chaotic systems to be completely synchronized. From then on, chaos
synchronization has been widely explored in a variety of fields including physical [3],
chemical [4], ecological [5] systems, secure communications ([6]-[7]) etc.
In most of the chaos synchronization approaches, the master-slave or drive-response
formalism is used. If a particular chaotic system is called the master or drive system
and another chaotic system is called the slave or response system, then the idea of
synchronization is to use the output of the master system to control the slave system so
that the output of the slave system tracks the output of the master system asymptotically.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 1–10, 2011.

c Springer-Verlag Berlin Heidelberg 2011
2 S. Vaidyanathan

Since the seminal work by Pecora and Carroll [2], a variety of impressive approaches
have been proposed for the synchronization of the chaotic systems such as the PC
method [2], sampled-data feedback synchronization method [8], OGY method [9], time-
delay feedback method [10], backstepping method [11], adaptive design method [12],
sliding mode control method [13], etc.
So far, many types of synchronization phenomenon have been studied such as com-
plete synchronization [2], phase synchronization [5], generalized synchronization [7],
anti-synchronization [14], hybrid synchronization [15] etc.
Complete synchronization (CS) is characterized by the equality of state variables
evolving in time, while anti-synchronization (AS) is characterized by the disappearance
of the sum of relevant state variables evolving in time.
In hybrid synchronization of chaotic systems [15], one part of the systems is synchro-
nized and the other part is anti-synchronized so that complete synchronization (CS) and
anti-synchronization (AS) co-exist in the systems. The co-existence of CS and AS is
very useful in secure communication and chaotic encryptation schemes.
This paper is organized as follows. In Section 2, we derive results for the hybrid
synchronization of identical Liu systems ([16], 2004). In Section 3, we derive results
for the hybrid synchronization of two identical Lü systems ([17], 2002). In Section 4,
we derive results for the hybrid synchronization of non-identical Liu and Lü systems.
In Section 5, we present the conclusions of this paper.

2 Hybrid Synchronization of Identical Liu Systems

In this section, we consider the hybrid synchronization of identical Liu systems [16].
Thus, we consider the master system as the Liu dynamics, which is described by the
dynamics
ẋ1 = a(x2 − x1 )
ẋ2 = bx1 − x1 x3 (1)
ẋ3 = −cx3 + dx21
where xi (i = 1, 2, 3) are the state variables and a, b, c, d are positive parameters.
We consider the slave system as the controlled Liu system, which is described by the
dynamics
ẏ1 = a(y2 − y1 ) + u1
ẏ2 = by1 − y1 y3 + u2 (2)
ẏ3 = −cy3 + dy12 + u3
where yi (i = 1, 2, 3) are the states of the system and ui (i = 1, 2, 3) are the active
controls.
The Liu system ([16], 2004) is one of the paradigms of the three dimensional chaotic
systems and this system is chaotic when the parameters take the values
a = 10, b = 40, c = 2.5 and d = 4.
Figure 1 illustrates the chaotic portrait of the Liu system (1).
For the hybrid synchronization of the identical Liu systems (1) and (2), the errors
are defined as
 Hybrid Synchronization of Liu and Lü Systems 3

140

120

100

80
x3

60

40

20

0
40

20 20
15
0 10
5
0
x −20 −5
2 −10 x
−40 −15
1

Fig. 1. State Orbits of the Liu Chaotic System

e1 = y1 − x1
e2 = y2 + x2 (3)
e3 = y3 − x3
A simple calculation yields the error dynamics as

ė1 = a(e2 − e1 ) − 2ax2 + u1

ė2 = be1 + 2bx1 − (y1 y3 + x1 x3 ) + u2 (4)
ė3 = −ce3 + d(y12 − x21 ) + u3

We consider the nonlinear controller defined by

u1 = −(a + b)e2 + 2ax2

u2 = −e2 − 2bx1 + y1 y3 + x1 x3 (5)
u3 = −d(y12 − x21 )

Substitution of (5) into (4) yields the linear error dynamics

ė1 = −ae1 − be2

ė2 = be1 − e2 (6)
ė3 = −ce3

We consider the candidate Lyapunov function defined by

1 T 1  2 
V (e) = e e= e + e22 + e23 (7)
2 2 1
Differentiating (7) along the trajectories of the system (6), we get
4 S. Vaidyanathan

V̇ (e) = −ae21 − e22 − ce23 (8)

3
which is a negative definite function on IR , since a and c are positive constants.
Thus, by Lyapunov stability theory [18], the error dynamics (6) is globally exponen-
tially stable. Hence, we obtain the following result.

Theorem 1. The identical hyperchaotic Liu systems (1) and (2) are globally and expo-
nentially hybrid synchronized with the active nonlinear controller (5).

Numerical Simulations. For the numerical simulations, the fourth order Runge-Kutta
method with initial step h = 10−6 is used to solve the two systems of differential
equations (1) and (2) with the nonlinear controller (5).

20

x
10 1

0 y
1
−10
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

40
x2
20

−20
y2
−40
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

100

50
x
3
y3
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec)

Fig. 2. Hybrid Synchronization of Identical Liu Systems

The parameters of the identical Liu systems (1) and (2) are taken as those which
result in the chaotic behaviour of the systems, viz. a = 10, b = 40, c = 2.5 and d = 4.
The initial values for the master system (1) are taken as

x1 (0) = 6, x2 (0) = 3, x3 (0) = 7

and the initial values for the slave system (2) are taken as

y1 (0) = 10, y2 (0) = 8, y3 (0) = 3

Figure 2 depicts the hybrid synchronization of the identical Liu systems (1) and (2).
 Hybrid Synchronization of Liu and Lü Systems 5

3 Hybrid Synchronization of Identical Lü Systems

In this section, we consider the hybrid synchronization of identical Lü systems [17].
Thus, we consider the master system as the Lü dynamics, which is described by the
dynamics
ẋ1 = α(x2 − x1 )
ẋ2 = γx2 − x1 x3 (9)
ẋ3 = −βx3 + x1 x2

where xi (i = 1, 2, 3) are the state variables and α, β, γ are positive constants.

We consider the slave system as the controlled Lü system, which is described by the
dynamics
ẏ1 = α(y2 − y1 ) + u1
ẏ2 = γy2 − y1 y3 + u2 (10)
ẏ3 = −βy3 + y1 y2 + u3

where yi (i = 1, 2, 3) are the states of the system and ui (i = 1, 2, 3) are the active
controls.
The Lü system ([17], 2002) is one of the paradigms of the three dimensional chaotic
systems and this system is chaotic when the parameters take the values

α = 36, β = 3 and γ = 20.

Figure 3 illustrates the chaotic portrait of the Lü system (9).

50

40

30
x3

20

10

0
15
10
5 15
10
0 5
−5 0
−5
x2 −10
−10 x1
−15 −15

Fig. 3. State Orbits of the Lü Chaotic System

6 S. Vaidyanathan

For the hybrid synchronization of the identical Lü systems (9) and (10), the errors
are defined as
e1 = y1 − x1
e2 = y2 + x2 (11)
e3 = y3 − x3

A simple calculation yields the error dynamics as

ė1 = α(e2 − e1 ) − 2αx2 + u1

ė2 = γe2 − (y1 y3 + x1 x3 ) + u2 (12)
ė3 = −βe3 + y1 y2 − x1 x2 + u3

We consider the nonlinear controller defined by

u1 = −αe2 + 2αx2
u2 = −(γ + 1)e2 + y1 y3 + x1 x3 (13)
u3 = −y1 y2 + x1 x2

Substitution of (13) into (12) yields the linear error dynamics

ė1 = −αe1
ė2 = −e2 (14)
ė3 = −βe3

We consider the candidate Lyapunov function defined by

1 T 1  2 
V (e) = e e= e + e22 + e23 (15)
2 2 1
Differentiating (15) along the trajectories of the system (14), we get

V̇ (e) = −αe21 − e22 − βe23 (16)

which is a negative definite function on IR3 , since α and β are positive constants.
Thus, by Lyapunov stability theory [18], the error dynamics (14) is globally expo-
nentially stable. Hence, we obtain the following result.

Theorem 2. The identical Lü systems (9) and (10) are globally and exponentially hy-
brid synchronized with the active nonlinear controller (13).

Numerical Simulations. For the numerical simulations, the fourth order Runge-Kutta
method with initial step h = 10−6 is used to solve the two systems of differential
equations (9) and (10) with the nonlinear controller (13).
The parameters of the identical Lü systems (9) and (10) are taken as those which
result in the chaotic behaviour of the systems, viz. α = 36, β = 3 and γ = 20.
 Hybrid Synchronization of Liu and Lü Systems 7

20
x
10
1

0 y1

−10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

40
x
2
20

0
y
2
−20
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

30 y3
20

10 x3
0
0 0.5 1 1.5 2 2.5 3
Time (sec)

Fig. 4. Hybrid Synchronization of Identical Lü Systems

The initial values for the master system (9) are taken as
x1 (0) = 12, x2 (0) = 5, x3 (0) = 10
and the initial values for the slave system (10) are taken as
y1 (0) = 5, y2 (0) = 14, y3 (0) = 16
Figure 4 depicts the hybrid synchronization of the identical Lü systems (9) and (10).

4 Hybrid Synchronization of Liu and Lü Systems

In this section, we consider the hybrid synchronization of the non-identical Liu and Lü
systems.
Thus, we consider the master system as the Lü system (2002), which is described by
the dynamics
ẋ1 = α(x2 − x1 )
ẋ2 = γx2 − x1 x3 (17)
ẋ3 = −βx3 + x1 x2
where xi (i = 1, 2, 3) are the states of the system and α, β, γ are positive parameters of
the system.
Next, we consider the slave system as the controlled Liu system (2004), which is
described by the dynamics
ẏ1 = a(y2 − y1 ) + u1
ẏ2 = by1 − y1 y3 + u2 (18)
ẏ3 = −cy3 + dy12 + u3
8 S. Vaidyanathan

where yi (i = 1, 2, 3) are the states of the system and ui (i = 1, 2, 3) are the active
controls.
For the hybrid synchronization of the Lü system (17) and Liu system (18), the errors
are defined as
e1 = y1 − x1
e2 = y2 + x2 (19)
e3 = y3 − x3

A simple calculation yields the error dynamics as

ė1 = a(e2 − e1 ) − (a − α)x1 − (a + α)x2 + u1

ė2 = be1 + bx1 + γx2 − (y1 y3 + x1 x3 ) + u2 (20)
ė3 = −ce3 + (β − c)x3 − x1 x2 + dy12 + u3

We consider the nonlinear controller defined by

u1 = −(a + b)e2 + (a − α)x1 + (a + α)x2

u2 = −e2 − bx1 − γx2 + y1 y3 + x1 x3 (21)
u3 = −(β − c)x3 + x1 x2 − dy12

Substitution of (21) into (20) yields the linear error dynamics

ė1 = −ae1 − be2

ė2 = be1 − e2 (22)
ė3 = −ce3

We consider the candidate Lyapunov function defined by

1 T 1  2 
V (e) = e e= e1 + e22 + e23 (23)
2 2
Differentiating (23) along the trajectories of the system (22), we get

V̇ (e) = −ae21 − e22 − ce23 (24)

which is a negative definite function on IR3 , since a and c are positive constants.
Thus, by Lyapunov stability theory [18], the error dynamics (22) is globally expo-
nentially stable. Hence, we obtain the following result.

Theorem 3. The Lü system (17) and the Liu system (18) are globally and exponentially
hybrid synchronized with the active nonlinear controller (21).

Numerical Simulations. For the numerical simulations, the fourth order Runge-Kutta
method with initial step h = 10−6 is used to solve the two systems of differential
equations (17) and (18) with the nonlinear controller (21).
 Hybrid Synchronization of Liu and Lü Systems 9

40
y
1
20

x1
0

−20
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

40
x2
20

−20 y
2
−40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

40

30 y3
20

10 x3
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (sec)

Fig. 5. Hybrid Synchronization of the Liu and Lü Systems

For the Lü and Liu systems, the parameter values are taken so that the systems are
chaotic. Hence, we take
a = 10, b = 40, c = 2.5, d = 4, α = 36, β = 3 and γ = 20.
The initial values for the master system (17) are taken as
x1 (0) = 2, x2 (0) = 4, x3 (0) = 9
and the initial values for the slave system (18) are taken as
y1 (0) = 5, y2 (0) = 8, y3 (0) = 15
Figure 5 depicts the hybrid synchronization of the non-identical Lü system (17) and
Liu system (18).

5 Conclusions
In this paper, nonlinear control method based on Lyapunov stability theory has been de-
ployed to globally and exponentially hybrid synchronize identical Liu systems (2004),
identical Lü systems (2002) and non-identical Liu and Lü systems. Numerical simu-
lations are also shown to illustrate the effectiveness of the proposed hybrid synchro-
nization schemes. Since Lyapunov exponents are not required for these calculations,
the proposed active control method is effective and convenient to achieve global hybrid
chaos synchronization of the chaotic systems addressed in this paper.
10 S. Vaidyanathan

References
1. Alligood, K.T., Sauer, T., Yorke, J.A.: Chaos: An Introduction to Dynamical Systems.
Springer, New York (1997)
2. Pecora, L.M., Carroll, T.L.: Synchronization in chaotic systems. Phys. Rev. Lett. 64, 821–824
(1990)
3. Lakshmanan, M., Murali, K.: Chaos in Nonlinear Oscillators: Controlling and Synchroniza-
tion. World Scientific, Singapore (1996)
4. Han, S.K., Kerrer, C., Kuramoto, Y.: Dephasing and bursting in coupled neural oscillators.
Phys. Rev. Lett. 75, 3190–3193 (1995)
5. Blasius, B., Huppert, A., Stone, L.: Complex dynamics and phase synchronization in spatially
extended ecological system. Nature 399, 354–359 (1999)
6. Feki, M.: An adaptive chaos synchronization scheme applied to secure communication.
Chaos, Solit. Fract. 18, 141–148 (2003)
7. Murali, K., Lakshmanan, M.: Secure communication using a compound signal from gener-
alized synchronizable chaotic systems. Phys. Rev. Lett. A 241, 303–310 (1998)
8. Yang, T., Chua, L.O.: Control of chaos using sampled-data feedback control. Internat. J.
Bifurcat. Chaos. 9, 215–219 (1999)
9. Ott, E., Grebogi, C., Yorke, J.A.: Controlling chaos. Phys. Rev. Lett. 64, 1196–1199 (1990)
10. Park, J.H., Kwon, O.M.: A novel criterion for delayed feedback control of time-delay chaotic
systems. Chaos, Solit. Fract. 17, 709–716 (2003)
11. Yu, Y.G., Zhang, S.C.: Adaptive backstepping synchronization of uncertain chaotic systems.
Chaos, Solit. Fract. 27, 1369–1375 (2006)
12. Liao, T.L., Tsai, S.H.: Adaptive synchronization of chaotic systems and its applications to
secure communications. Chaos, Solit. Fract. 11, 1387–1396 (2000)
13. Konishi, K., Hirai, M., Kokame, H.: Sliding mode control for a class of chaotic systems.
Phys. Lett. A. 245, 511–517 (1998)
14. Jian, X.: Anti-synchronization of Lienard chaotic systems via feedback control. Internat. J.
Phys. Sci. 5, 2757–2761 (2010)
15. Sudheer, K.S., Sabir, M.: Hybrid synchronization of hyperchaotic Lü system. Pramana 73,
1387–1396 (2000)
16. Liu, C., Liu, T., Liu, L., Liu, K.: A new chaotic attractor. Chaos, Solit. Fract. 22, 1031–1038
(2004)
17. Lü, J., Chen, G.: A new chaotic attractor coined. Internat. J. Bifur. Chaos. 12, 659–661 (2002)
18. Hahn, W.: The Stability of Motion. Springer, New York (1967)
 Methods for Preventing Search Engine-Based
Web Server Attacks

Keerthiram Murugesan and Mukesh Singhal

Department of Computer Science,

University of Kentucky,
Lexington, Kentucky - 40508, USA

Abstract. Many Internet users make use of search engines to locate a

web page or information they want. These search engines have become a
fundamental tool in the World Wide Web. Since the capability of search
engines to reach any part of the web has increased, security loopholes of
any websites face the risk of being exposed globally. This poses serious
security threats on the web. In this paper, we discuss various methods
to prevent web server attacks through search engines.

Keywords: Search engine security, information security.

1 Introduction

The number of websites in the Internet is growing at a fast rate. This increase in
number of websites forces Internet users to turn to search engines to find specific
information. A user can search for any content available in the Internet with the
help of search engines.
Due to the power of search engines to reach any part of the web, vulner-
abilities in a website is accessible to anyone who actively searches for com-
promised websites. These vulnerabilities occur due to known and/or unknown
software bugs, improper or insufficient security measures, etc. in the websites.
If a website is unaware of these vulnerabilities, search engine might accidently
reveal these information in their search results. Attackers can now use search en-
gines as a tool for identifying websites with potential vulnerabilities and search
for the queries that expose these vulnerabilities in search results without re-
vealing the attackers identity. This results in certain users having an unautho-
rized access to restricted information, posing a serious threat to web security.
There are several cases reported on web server attacks using search engine as a
tool [1,2].
The rest of the paper is organized as follows: The next section provides a
background on search engine architecture. In Section 3, we examine the usage
of malicious queries in search engines. Section 4 presents web server attacks
that have been mounted successfully using search engines. We propose several
methods in Section 5 to prevent web server attacks through search engines.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 11–18, 2011.

c Springer-Verlag Berlin Heidelberg 2011
12 K. Murugesan and M. Singhal

2 General Search Engine Architecture

In this section, we provide a brief description of general search engine architecture

[3]. We focus on three main components of search engine architecture; namely,
crawler, indexer and query engine that help us to understand various security
attacks through search engine. Figure 1 shows the general search architecture
with all the main components.
The crawler is an important component of the search engine architecture and
runs with a basic set of URLs to start with. There are one or more crawlers run
concurrently to maximize the searching for information. Crawlers hop from one
web page to another in the Internet in search of any useful information.
When a crawler visits a web page, it extracts all the URLs in the web page
and gives them to the crawler controller. The crawler controller organizes the
URLs received from each crawlers based on the feedback information received
from the indexer and query engine components. The web page is then added to
the queue, for further processing, by the Indexer component. Optionally, a copy
of the web page is stored in the page repository for future uses.
The indexer uses a parser to extract all the important keywords or terms
from each page in the queue. These keywords are stored along with their link
information in the indexes using the index controller. The index controller links
these indexed web pages together based on their contents. These indexes, which
may be distributed, form the basic data structure of the search engine and can
be used for further processing like classification, text summarization, etc.
When a user sends a query to the query engine, it extracts all the important
terms from the query using the query parser. This query parser is similar in
functionality to the index parser. The parsed query is then sent to the index
controller, which uses the indexes to generate the list of URLs that are rele-
vant to the user query. The query engine may use a specific ranking algorithm
(PageRank, HITS, etc.) to sort the retrieved list.

Fig. 1. General Search Engine Architecture

 Methods for Preventing Search Engine-Based Web Server Attacks 13

As mentioned earlier, the indexer and query engine components generate feed-
back information such as usage patterns, index last modified details and ranking
and importance of a web page, with respect to the search engine model and
users. These feedbacks are used by the crawler controller for computing refresh
policies for each web page in the Internet. Refresh policies help the crawlers to
identify the next important web page to crawl and the frequency with which
they need to fetch each web page in the Internet.
Although there are many other supporting modules in a search engine archi-
tecture, these three components determine the search result returned for a user-
supplied query. In short, crawlers work together to extract information about a
web page, whereas the indexer and query engine (along with ranking module)
work together to provide list of URLs relevant to a user query.
There are two main metrics (cost functions) that determine the efficiency and
quality of search engines:

– Freshness: This metric tells whether the local copy is accurate or not.
– Age: This metric tells how outdated the local copy is.

A good search engine should maintain the freshness of a collection of web pages
as high as possible and the age as low as possible. Readers are referred to [3] for
the complete description of the search engine architecture.

3 Malicious Queries

A search for vulnerabilities is easier if you know what you are looking for. An at-
tacker can search queries for website vulnerabilities that expose this information
in the search results, without revealing the attackers identity.

Fig. 2. Search engine query statistics for queries such as index of/, 500 server error
and @yahoo.com from the Google search engine since 2004 (Google Trends [4]). The
data is scaled based on the average search traffic of the given terms over the period
(Search Volume Index).
14 K. Murugesan and M. Singhal

For example, searching index of / can show the list of websites with a directory
listing enabled. Sometimes, it could be easy to find sensitive files or even system
password files like /etc/passwd, .htpasswd, sam password files, etc. with this
search, if the websites have accidently enabled a directory listing. A search for the
query 500 server error shows a list of vulnerable websites that can provide useful
information like operating system version, port numbers, sql queries, source code,
etc. for an attacker. Similarly, a query @yahoo.com reveals a list of yahoo mail
users, which can be used by a spammer for phishing or advertising. Figure 2
shows the Google search engine statistics for the above queries.

4 Search Engine-Based Web Server Attacks

In this section, we will discuss about web server attacks that involves search
engines to find vulnerable servers and to propagate its attack.
In July 2004, a variant of MYDOOM worm used SMTP mail service to spread
and steal email addresses from the address book of target machines [5]. When
an email address ([email protected]) is found, the worm uses the domain name
(example.com) as a query in search engines such as Google, Yahoo, Lycos, and
AltaVista to find email addresses of that domain in the search result and prop-
agates further.
In December 2004, PhpBB Bulletin Board System, an open source php-based
web forum application, has been reported with a remote PHP script injection
and input validation vulnerabilities due to an improper use of urldecode() on
user-supplied input for highlight parameter in viewtopic.php phpBB script file
[6]. The urldecode() function, which is used to filter potentially malicious con-
tent from user-supplied input, failed to properly translate input. Using these
vulnerabilities, a remote attacker could have executed arbitrary commands on a
vulnerable server without sufficient privileges, which could have been exploited
to trigger SQL injection attacks.
After this issue was posted on the Internet, a Santy.A worm was created to
exploit these vulnerabilities using the Google search engine as a tool to search for
the keyword viewtopic.php in order to find the potentially vulnerable servers [7].
When a victim site was found, the worm would attack the system by exploiting
the highlight parameter vulnerability in the phpBB software. This worm would
execute itself by using the PERL environment of the compromised server and
propagate further. When Google stopped the worm from using its search engine
for finding the vulnerable sites, variants of the worm (Santy.B ) started using
AOL and Yahoo search engines to locate the vulnerable websites [8]. Thus, if
one of the several search engines is able to stop a threat, there are several other
search engines that attackers can use to perform their malicious actions.
In May 2008, Asprox trojan, a spam botnet for sending phishing emails, with
Danmec, a password-stealing trojan, searches Google for web servers that use
asp, aspx, or php scripts to dynamically generate web pages from SQL databases
[9,10]. With the list of URLs from search engine, it launches SQL-Injection at-
tack to inject an iframe into the website source which will redirect visitors of that
 Methods for Preventing Search Engine-Based Web Server Attacks 15

website to direct84.com to download a malicious javascript file. This javascript

file may redirects the visitor to another website for downloading additional ma-
licious javascript files which tries one or more exploits to crack the visitor’s
computer. When successful, it builds a data stream directly into the visitor’s file
system and installs this trojan on the computer, which in turn repeats the above
steps on other set of potential URLs returned by a search engine.
In August 2008, a directory listing vulnerability was reported in TWiki,
an open source enterprise wiki application. Due to a failure in filter-
ing the malicious shell meta-characters from the user-supplied string for
the TWiki query image parameter of the open() function, an attacker
could access any files if the twiki/bin/configure script was not properly se-
cured and the web server had sufficient permissions to view them. For ex-
ample, to get the /etc/passwd file of the vulnerable server, an attacker
could go to: http://www.example.org/twiki/bin/configure?action=image;
image=../../../../../../etc/passwd;type=text/plain [11]. This could
also be used to execute arbitrary commands via shell meta-characters. Since
the TWiki application uses predictable URLs, the vulnerable servers could be
found using the search engines.
Although these attacks were fixed, many web security threats are still exposed
through search engines everyday.

5 Proposed Methods
In this section, we propose several methods to address the growing problem
of web attacks through search engines. The main purpose of this section is to
stimulate research ideas in this problem domain.

5.1 IN:SITE Crawlers

IN:SITE crawlers or local crawlers are a simple extension to the crawler com-
ponent shown in Figure 1. This crawler may be specific to a web server or a
network of servers that belongs to an organization. Unlike a crawler in Figure 1,
this local crawler is deployed in each web server in the Internet and is completely
under the control of web administrator. Together they provide a distributed ap-
proach to the search engine architecture as shown in Figure 3. The effort needed
to generate a powerful but secured search engine are distributed throughout the
Internet. Each web server in the Internet now acts as a resource pool, which
constantly provides information to the search engine to make it effective and
secure.
IN:SITE crawler can crawl all the web pages or publicly accessible contents
in a web server or network and send this information to a search engine. This
crawler can keep track of last modified information of all publicly accessible
web pages of that web server. When a web page is modified, it notifies the
crawler controller about the modified content. Crawler controller updates the
refresh policy of the web page based on its importance and its freshness and age
metrics.
16 K. Murugesan and M. Singhal

Fig. 3. IN:SITE crawler. All the information exchanged between a search engine and
web servers are encrypted.

Since a crawler is local to a web server or a network, it may have an access to

some restricted information that cannot be accessed from outside their computer
or network. For instance, employee details of an organization, which can be
accessed within the organization network, may be open to a local crawler of that
network. This can be avoided by placing this local crawler in a proxy server of
that network or web server or by providing a set of judgments or rules on what
to access, to a local crawler, specified by the web administrator.
IN:SITE crawler has several advantages:
– These local crawlers reduce the unnecessary check performed by a search
engine crawler to identify any recent modifications in a web page since its
last crawl. This reduces the bandwidth wastages caused by a search engine
in the Internet.
– This increases the freshness metric and considerably reduces the age metric
of each web page in the search engine.
– The complete control of information sent to the crawler controller is with
the web administrator.
– When sensitive information is accidently released into the Internet, this
crawler can be configured to inform a search engine to block an access to a
specific web page or an entire web site immediately.

5.2 Intelligent Crawlers

Intelligent crawlers measure the amount of security issues exposed in a website.
From this knowledge, intelligent crawlers can now avoid indexing vulnerable
websites or block an access to those websites from search results to reduce further
security risks.
 Methods for Preventing Search Engine-Based Web Server Attacks 17

In general, an intelligent crawler performs two steps: fetching a web page

from the Internet and risk analysis. The risk analysis step analyzes each web
page for any vulnerabilities and sensitive information before adding it to the
queue. The risk analysis step uses patterns to identify the sensitive information.
These patterns are based on simple regular expressions. For instance, patterns
for SSN and email address can be given as:
SSN: ^(?!000)([0-6]\d{2}|7([0-6]\d|7[012]))
([-]?)(?!00)\d\d\3(?!0000)\d{4}$

Email Address: ^([\w\-\.]+)@((\[([0-9]{1,3}

\.){3}[0-9]{1,3}\])|(([\w\-]+\.)+)([a-zA-Z]{2,4}))$
In addition, the risk analysis step scans a web page for any known vulnerabili-
ties such as SQL injection, directory listing, etc. using vulnerability signatures
[12,13]. If this step finds any security issues in a web page, the search engine
will blacklist the web page or reduce the ranking based on the concentration of
security issues in a web site.

5.3 Feedback Information

Search engines can use feedback information to improve search results returned
by the query engine. In addition, this can be used to identify vulnerable websites
to avoid listing them in a search result [14].
Explicit feedback is the feedback information given by a user to search engine
about vulnerabilities in a web site and update refresh policies or rank web pages.
These feedbacks are based on ratings given by users for a set of web pages. Since
these feedbacks are based on the users judgment, it may have false positive or
false negative errors. Modules in query engine and indexer components should
analyze this feedback information again to reduce these errors. These feedbacks
are more effective if security experts rate web pages listed in search results for
popular queries. In order to ensure the security of the web, search engines can
use these feedbacks to avoid listing vulnerable websites or reduce their ranking.
This feedback mechanism can be used by search engine to prevent web attacks
through search engine in the long run.

5.4 Query Filters

When a user sends a query to the query engine, the query filter checks for any
malicious content in the user-supplied query using known regular expressions,
keywords and/or phrases. In addition, malicious queries can be identified based
on the percentage of blacklisted web pages listed in search results returned for
a user-supplied query.

6 Conclusion
Search engines can be used as a tool in locating vulnerable servers in the Internet
anonymously. Vulnerabilities, such as SQL injection, Cross-Site Scripting, etc. in
18 K. Murugesan and M. Singhal

a website can be found easily if an attacker knows what to look for. In this paper,
we discussed several methods to address the problem of search engine-based web
server attacks. These proposed methods can only prevent an attacker from using
a search engine to locate a vulnerable server. A web administrator should take
certain precautionary measures such as keeping the server up-to-date, blocking
unwanted ports and services, using firewall and antivirus programs, etc. to protect
a web server completely from these attacks.

References
1. Hernandez, J.C., Sierra, J.M., Ribagorda, A., Ramos, B.: Search engines as a se-
curity threat. Computer 34, 25–30 (2001)
2. Newman, A.C.: Search Engines used to attack databases. White Paper, Application
Security, Inc.
3. Arasu, A., Cho, J., Molina, H.G., Paepcke, A., Raghavan, S.: Searching the Web.
ACM Trans. Internet Technol. 1, 2–43 (2001)
4. Google Trends, https://www.google.com/trends
5. Symantec Threats and Risks, http://www.symantec.com/security_response/
writeup.jsp?docid=2004-072615-3527-99
6. United States Computer Emergency Readiness Team, http://www.us-cert.gov/
cas/techalerts/TA04-356A.html
7. McAfee Virus Profile and Definition, http://home.mcafee.com/VirusInfo/
VirusProfile.aspx?key=130471
8. Symantec Threats and Risks, http://www.symantec.com/security_response/
writeup.jsp?docid=2004-122610-5904-99
9. Dell SecureWorks,
http://www.secureworks.com/research/threats/danmecasprox
10. Shin, Y., Myers, S., Gupta, M.: A Case Study on Asprox Infection Dynamics. In:
Flegel, U., Bruschi, D. (eds.) DIMVA 2009. LNCS, vol. 5587, pp. 1–20. Springer,
Heidelberg (2009)
11. TWiki Security Alert,
http://twiki.org/cgi-bin/view/Codev/SecurityAlert-CVE-2008-3195
12. Xie, Y., Yu, F., Achan, K., Panigrahy, R., Hulten, G., Osipkov, I.: Spamming
botnets: signatures and characteristics. In: ACM SIGCOMM Conference on Data
Communication (SIGCOMM 2008), pp. 171–182. ACM, New York (2008)
13. Brumley, D., Newsome, J., Song, D., Wang, H., Jha, S.: Towards Automatic Gen-
eration of Vulnerability-Based Signatures. In: IEEE Symposium on Security and
Privacy (SP 2006), pp. 2–16. IEEE Computer Society, Washington, DC (2006)
14. White, R., Jose, J., Ruthven, I.: Comparing explicit and implicit feedback tech-
niques for web retrieval: Trec-10 interactive track report. In: Text Retrieval Con-
ference, TREC (2001)
 Sea Object Detection Using Shape and Hybrid Color
Texture Classification

M. Uma Selvi and S. Suresh Kumar

Vivekanadha College of Engineering for Women,

{Mumaselvi,nice.ssk}@gmail.com

Abstract. Sea target detection from remote sensing imagery is very important,
with a wide array of applications in areas such as fishery management, vessel
traffic services, and naval warfare. This paper focuses on the issue of ship
detection from spaceborne optical images (SDSOI). Although advantages of
synthetic aperture radar (SAR) result in that most of current ship detection
approaches are based on SAR images. But disadvantages of SAR still exist.
Such as the limited number of SAR sensors, the relatively long revisit cycle,
and the relatively lower resolution. To overcome these disadvantages a new
classification algorithm using color and texture is introduced for Ship detection.
Color information is computationally cheap to learn and process. However in
many cases, color alone does not provide enough information for classification.
Texture information also can improve classification performance. This
algorithm uses both color and texture features. In this approach for the
construction of a hybrid color-texture space we are using mutual information
and three aspects: 1) Classifies ship candidates 2) The relevant classes are
automatically built by the samples’ Appearances and 3) Shape and Texture
features. Experimental results of SDSOI on a large image set captured by
optical sensors from multiple satellites show that our approach is effective in
distinguishing between ships and non ships, and obtains a satisfactory ship
detection performance.. Feature extraction is done by the co-occurrence matrix
with SVM (Support Vectors Machine) as a classifier. Therefore this algorithm
may attain a very good classification rate.

Keywords: Ship detection from spaceborne optical images, synthetic aperture

radar, Support vector machines classifier, Co- Occurrence Matrix, Hybrid
Color, Hybrid color-texture.We would like to encourage you to list your
keywords in this section.

1 Introduction

Ship detection from remote sensing imagery is very important and has a wide array of
applications such as fishery management, vessel traffic services, and naval warfare. In
particular, in recent years, because of the decrease in fishery resources in the world,
ship detection has become much more important for effective and efficient ship
monitoring to prohibit illegal fishing activities in time.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 19–31, 2011.
© Springer-Verlag Berlin Heidelberg 2011
20 M. Uma Selvi and S. Suresh Kumar

However, ship detection based on SAR has limitations. First, with a limited
number of SAR satellites, the revisit cycle is relatively long and, then, cannot meet
the needs of the application of real-time ship monitoring. Second, the resolution of
most satellite SAR images is often not high enough to extract detailed ship
information.
Ship detection based on satellite optical images can partly overcome the
aforementioned shortcomings of ship detection based on SAR and is complementary
to SAR-based ship detection. Thus, it is advantageous to investigate SDSOI to better
satisfy the requirements of ship monitoring. There are not many researches about
target detection in sea through direct image processing and among them papers that
utilize color features are even fewer. For this reason almost all of the existing
approaches suffer from one or more drawbacks like sensitivity to different sensors or
invalidity for low SNR images . Many papers have worked on ship detection based on
SAR2 images [5]. A new method based on combinatorial improved PNN3 model for
ship detection in SAR imagery and [4] proposes a method to reduce speckle noise for
SAR images and to improve the detected ratio for SAR ship targets from the SAR
imaging mechanism. A specific technique for automatic spot detection, based on the
Wavelet Transform is presented and justified in [10]. Marivi et al [11] proposed a ship
target detection method of SAR images by means of the discrete wavelet transform,
taking advantage of the difference of statistical behavior of the vessels and the sea.
All above methods are dependent on SAR images; they expend largely and can only
obtain target points, which cannot be used to recognize targets. There are also some
papers that use remote sensing images for ship detection [6]. They present a method
based on cumulative projection curve (CPC) to estimate the number of ships of small
size, which is only efficient on especial images from stationary ships along coastline
in a harbor. One of the few researches that uses color feature, from Lab color
coordinate system, for sea target detection is [9]. They presented a definition on the
degree of overlap between two clusters and developed an algorithm for calculating the
overlap rate. Using this theory, they also developed a new hierarchical cluster
merging algorithm for image segmentation and applied it to the ship detection in high
resolution images is one of the several papers that worked on IR images. They used
PCA, Bayes classification and wavelet-denoising to classify the sea targets, but in
several papers, limitations and disadvantages of the methods based on statistical
analysis are pointed out [8]. One of the papers that have more superiority to previous
works in visible images domain is [7]. Their work is based on calculating different
chaos by obtaining largest Lyapunov exponent of target and sea background which is
not appropriate for images that contain seaside or some low chaos objects. Also the
authors have proposed [8] based on the natural measure feature. Although the results
of the method are considerable for some images but it still suffers from previously
mentioned imperfection and needs analyzing several frames for exact results. So it is
still important to find new methods of detecting target from sea background. Except
[9] all above researches have worked on grey level images and for sure our method is
one of the few one that utilizes color feature for sea target detection.
We want to further investigate how we can resolve issues in SDSOI. SDSOI
includes two stages: 1) sea detection and 2) ship detection in sea. This paper mainly
focuses on how we can detect a ship in sea, assuming that sea regions have been
detected by prior geographic information or algorithms. In addition, in this paper, we
 Sea Object Detection Using Shape and Hybrid Color Texture Classification 21

are only interested in ship detection on a panchromatic band or on a band of a

multispectral image and not on multispectral images.In SDSOI, several factors such
as clouds, ocean waves, and small islands are often detected as false ship candidates
due to similar characteristics, which affect the performance of ship detection.
In this paper, a novel hierarchical complete and operational SDSOI approach with
multiple features.
Stage 1–Extraction of Ship Candidates: This stage includes image segmentation
and simple shape analysis. On the one hand, image segmentation with global and
local information, including the characteristics of gray and edge features, is applied to
obtain possible ship candidate regions. On the other hand, because ships are generally
thin and long, simple shape features are extracted to eliminate obvious false
candidates. In this stage, the main aim is to extract ship candidates with missing
alarms as low as possible, and false alarms that are due to objects such as small clouds
and islands may be extracted as ship candidates. Stage 2–Classification of Ship
Candidates: In this stage, a novel semisupervised hierarchical classification of ship
candidates based on shape and texture features is further conducted to remove most
false alarms and obtain the results of ship detection.
Some differences between the feature distributions of ships and nonships can be
noticed; therefore, we can remove nonship candidates and detect ships by pattern
classification. Furthermore, a significant difference usually exists between the feature
distributions of different samples within one class, which often has a negative
influence on the classification performance. Thus, all candidates are divided into
several subclasses according to their appearances, and the subclasses are divided into
miniclasses by feature clustering (so-called semisupervised) to improve the
performance. In addition, a hierarchical ensemble classification approach based on the
miniclass distributions, rather than on the directed extracted features, is presented for
candidate recognition. Moreover, in feature extraction, besides commonly used shape
and gray distribution features, a new texture operator, local multiple patterns (LMP),
for enhancing the representation ability of the feature set is first introduced. The basic
idea of LMP is that it extends binary patterns to multiple patterns. In addition, the
LMP preserves more structural information and is more suitable for image analysis
than local binary patterns (LBP).

2 Learning and Extraction of Ship Candidates

2.1 Image Segmentation

In spaceborne optical images, a stationary gray distribution with a slow variation in

the sea region usually exists, whereas usually, an obvious edge between the ship and
sea due to a high ship hull is observed. Thus, image segmentation with edge
information can be applied to extract possible candidate regions, which may include
many false alarms such as islands, clouds, and ocean waves due to their similar edge
characteristics with those of ships. On the other hand, there is usually a strong wake
round steering ship, and it is sometimes difficult to distinguish the ship and the wake
in optical images with low resolutions. When the ship and its wake are segmented into
one region, the area is expanded, which often helps us find the small ships. Thus, they
are considered a whole part in the following processing. In the experiments that
22 M. Uma Selvi and S. Suresh Kumar

involve extraction of ship candidates, only the original regions of more than ten pixels
in size and more than four pixels in length are considered for analysis. The algorithm
is described as follows.

Image preprocessing: There is sometimes a dark gray distribution of ships, which

has a negative effect on ship segmentation. However, there is usually obvious edge
information around the ship. To enhance the distinction between the ship and sea to
extract the whole ship, a new mixed image is proposed. The new mixed image is
computed with the original gray value and edge magnitude as follows. Suppose that
f(i, j) is the gray value of a point (i, j) in an image. Then, the Sobel operator is applied
to compute the edge magnitude Mag(i, j). The mixed image is defined as
Mixf(i, j) = f(i, j) + a × Mag(i, j) (1)
where a is a coefficient that controls the proportion of the edge magnitude to gray
value, which was experientially set to equal to 1 in our experiments.

Coarse image segmentation: The whole ship region is relatively salient in the mixed
image, ship candidates can be obtained by coarse image segmentation of the mixed
image with a proper threshold. Here, a simple and effective segmentation method
based on the optimal principle of maximum between-class variance and minimum
within-class variance [14] is used to compute a global adaptive threshold.
Furthermore, a region-filling algorithm [15] is adopted to delete the holes within
regions, and morphological open and close operators with a three-pixel-diameter
circle element [15] are adopted to eliminate very thin lines and random noise regions,
which are perhaps generated by ocean waves for example.

Refined image segmentation: Due to the usage of only a global threshold, the coarse
segmentation results may have edge localization errors. This processing aims at
refining every region’s contour with the local characteristics for the following feature
extraction. A level set by the Chan–Vese model [16] is adopted to refine image
segmentation. Chan and Vese (2001) proposed an algorithm based on the Mumford–
Shah model, which can provide an optimal partition of two classes. The simple energy
function is given as follows:
)& )R&)E& Ȓ_X[\íFR_G[G\
LQVLGH&
 (2)
Ȓ_X[\íFE_G[G\
RXWVLGH&

where co is the mean value inside of the curve C, and cb is the mean value outside of
the curve C. The minimum energy given by the Chan–Vese model will be an optimal
piecewise smooth approximation of the edge. A level set is done on the corresponding
subimages that are cut from the original images, with ranges depending on the ranges
of segmented regions with adaptive thresholds. The initial curve C is set as the edge
of the coarse segmentation.
 Sea Object Detection Using Shape and Hybrid Color Texture Classification 23

2.2 Simple Shape Analysis

After image segmentation, simple shape analyses can be applied to eliminate obvious
false candidates. First, ships have a limited area, length, and width range. According
to this constraint, false candidates such as very large or very small islands and clouds
can be eliminated with proper thresholds, which should take the wake imaging into
account.
Second, ships are commonly long and thin. Thus, the ratio of the length to the
width of the region ounding rectangle [15], [17] is larger than a given threshold.
According to this condition, obvious false alarms, including islands and clouds with
very small ratios, are eliminated. The relatively low threshold aims at keeping us from
eliminating ships that may be either amid a curved wake or too small to exactly
extract the ratio value. It aims at having a low missing alarm rate. Ship candidates are
detected by image segmentation and simple shape analysis, as previously discussed.
In our experiments, to reduce the time consumption, the level set is done after the
simple shape analysis: only ship candidates need refined image segmentation with
level set. Fig. 3 gives the intermediate results of one typical image sample. It is easily
concluded that the refined segmentation results with level sets are closer to the true
ship edges than those with an adaptive threshold due to the use of local gray
characteristics.

3 Feature Extraction of Ship Candidates

It is important to extract effective features to distinguish ships from other objects,
which mainly comprise clouds and ocean waves. Both shape and appearance
attributes are crucial for each object class. In our approach, various features, including
shape and texture, are extracted from a ship candidate and are concatenated as a
feature vector for classification.

3.1 Shape Feature Extraction of Ship Candidates

Binary images, which are obtained from image segmentation, readily provide simple
geometric properties such as perimeter and area. In our approach, the following shape
descriptors are adopted.

Compactness: Compactness is given as [17] Compactness = (Perimeter)2 /Area The

most compact region in a Euclidean space is a circle.

Convexness: Let S represent a set of contour points obtained from level set and CH(S)
[15], [17] be defined as its convex hull. The convexity measure is defined as
CM(S) = Area(S)/Area(CH(S)) [15], [17].
Rectangularity and Eccentricity: The simplest eccentricity [17] is the ratio of the
major to the minor axes of an object approximated by its best fit ellipse. Ships are
commonly long and thin; therefore, it can be adopted as a shape feature.
Rectangularity [17] is the maximum ratio of region area to the area of a bounding
rectangle according to its different directions. It assumes values in the interval (0, 1],
with 1 representing a perfectly rectangular region.
24 M. Uma Selvi and S. Suresh Kumar

Moment invariants: Moments are extensively used for shape representation, pattern
recognition [18], and image reconstruction, which makes them a very useful feature
set to include. Here, we adopt the first seven moment invariants introduced by Hu
[19].

3.2 Texture Feature Extraction of Ship Candidates

After taking a close look at these candidates that have been expanded to more than
100 pixels, we consider that most candidates are large enough so that some stable
features can be computed from the gray distribution. Furthermore, little difference
between the gray distribution of a ship and that of a nonship can be observed. Texture
analysis can be applied to eliminate false alarms.
In our approach, besides commonly used features such as simple texture features,
wavelet-based features, and multiscale Gaussian differential features (MGDFs), a new
texture operator, LMP, is introduced to enhance the representation ability of the
feature set. Simple texture features include mean, variance, moments, and entropy of
gray value . Other features are described briefly as follows.

Wavelet-Based Features: Wavelet-based features aim at extracting the information in

the frequency domain, and their capability of undergoing texture analysis has been
shown [20]. By applying the wavelet transform to the candidate image, a number of
subbands are generated. For each subband I(l,s), the following features are calculated
[20]:
@
OTOT
H $™™Ň,LMŇ
MM
OTOTOT
H ORJ$™™Ň,LMŇDQRUPORJŇ,LMŇDQRUP (3)
MM


OT

:KHUHDQRUP ™Ň,LMŇ
LM

A is the area of each subband, l is the decomposition level, and q represents the
subband number within the decomposition level l. The feature e1(l,q) shows the amount
of signal energy at a specific resolution, whereas e2(l,q) shows the nonuniformity of the
subband values.

4 Learning Colour and Texture Features

4.1 Color Feature

Color information is readily available as input from a color camera so no extra

processing is required. Color of a pixel is represented as a vector in the Red, Green,
Blue (RGB) color space or Hue, Saturation, Value (HSV). Real world surfaces often
have more than a single color. They are either white, pink, gray, or black. To learn the
color appearance of a real world surface, the histogram method is often used. A
 Sea Object Detection Using Shape and Hybrid Color Texture Classification 25

normalized histogram H approximates the distribution of colors in an image patch P,

each bin in H is calculated by
Hi = ni/N (1)
where ni is the number of pixels whose color falls into the region defined by bin i and
N is the total number of pixels in P. When histogram is used for classification, during
the training stage, a set of model histograms of a class is learned. In classification
stage, a sample histogram is compared with the model histograms. The sample’s class
is assigned with the best matched model’s class.

4.2 Texture Feature

Although there is no formal definition of texture, the texture of a visual surface can be
thought of as spatial distribution of intensity/color variations. In the case of co-
occurrence matrix, the performance depends on the number of gray levels,
displacement sets are used. Texture is used to improve the performance of a color
based classifier but with a consideration of computational speed and memory
requirement. A good tradeoff between performance improvement and added
complexity is desired. For this reason, two texture learning methods were chosen: the
Local Binary Pattern (LBP) method and a statistical method using 8 histograms of the
intensity difference between two pixels at 2, 4, 6, and 8 pixels apart at horizontal and
vertical directions. These two methods are significantly less complex than other
texture methods. The LBP contains local intensity pattern of a pixel. The LBP of pixel
n with intensity x is defined as:
lbp (n) =∑ 2bi(x-x1) (2)
i =1...8
bi(x) = {1 if x ≥ 0,0 if x < 0. (3)
where x1...8 are the intensity values of the 8 pixels neighboring n.

5 Combining Colour and Texture Features

5.1 Partitioning Training Set Using Color Histograms

From initial tests of the color based classifier, it was found that the average
performance was relatively good. The Color information is rich enough in most cases.
Classification errors arise when the sample images were taken under bad illumination
condition when there were targets with very similar color appearance to the path
(even to a human eye). It is noted that extracting texture features is several magnitude
more complex than extracting color feature. From these observations, algorithm was
built that uses color as the primary feature and only consider texture feature in special
cases where it is difficult to use color to discriminate. Instead of having one threshold
for each model color histogram, two thresholds are learned. One high threshold is
used to determine if a sample is matched with a model, one low threshold used to
determine if it is not. If the intersection value falls in between these two thresholds,
26 M. Uma Selvi and S. Suresh Kumar

the texture feature is used. The samples that fall in this region are the difficult cases
for the color classifier.

5.2 Classification Using Color and Texture

During classification stage, the algorithm maps a pixel patch of unknown class label
to either targets or non-targets. A patch is labeled as path if it’s color histogram is
matched with any of the h color histograms saved from the learning stage using the
high threshold. If none of the model histograms matches with the sample, the
algorithm iterates through the h models again. This time the intersection value
between the sample and the model color histogram is compared with the low
threshold. If a matched model is found, the sample’s texture histogram is compared
with the model’s texture histograms. A pseudo-code of the algorithm is listed below

for i = 1 to h do
if M(sco,mco,i) > high thresholdi then
s is path
return
end if
end for
for i = 1 to h do
if M(sco,mco,i) > low thresholdi AND
M(stx,mtx, i) > texture thresholdi then
s is path
return
end if
end for
s is non path
sco, stx are the sample’s color and texture histogram, mco,i and mtx,i are color and
texture histogram of the ith model.

6 Hybrid Color-Texture Space Construction

6.1 Color Spaces

The construction of the hybrid color-texture space starts with transforming a classified
image, initially represented in the RGB standard system, into different color spaces.
Color can be divided into four families [4] namely:
1. The primary spaces which are based on the trichromatic theory assuming that it
is possible to match any color by mixing appropriate amounts of three primary colors.
2. The perceptual spaces which try to quantify the subjective human color
perception using the intensity, the hue and the saturation.
3. The luminance–chrominance spaces where one component represents the
luminance and the two others the chrominance.
4. The independent axis spaces resulting from different statistical methods which
provide as less correlated components as possible.
 Sea Object Detection Using Shape and Hybrid Color Texture Classification 27

a global color space Cs gathers the components chosen for our approach, with NCs
=30 vectors:
Cs={R,G,B,r,g,b,X,Y1,Z,x,y,z,L,a,b,u1,v1,c,h,H,S,V,Y,i,q,u2,v2,I1,I2,I3}
The color study is associated with texture analysis, by calculating 2 statistical
parameters and 5 parameters of Haralick of the co-occurrence matrix.

6.2 Gray-Level Co-occurrence Matrices

Let E denote the space of candidates attributes. The number of attributes of E is 7 x 30

=210, where 7 represent the five parameters of Haralick + the two statistical
parameters, and 30 is the number of colors components. From E we select the
attributes that will Compose the hybrid color-texture space using mutual information.

6.3 Mutual Information

Mutual information (MI) is the measure of the dependence between random variables.
The MI of two discrete variables x and y is defined based on their joint probabilistic
distribution p(x, y) and the respective marginal probabilities p(x) and p(y):
MI (x,y)=∑alog(a/b) (4)
i,j
a= p(xi,yj), b=p(xi)p(yj)
The estimation of the joint probabilistic distribution p(x, y) represents the principal
difficulty for the use of MI. Various methods have been studied in the literature to
estimate such joint distribution (histograms and kernel-based methods). Authors in
propose a fast method to estimate the MI. One of the advantages of this method is that
it based in the estimation of the entropy. The relation between MI and the entropy can
be defined as :
MI(x, y) = H(x) − H( y | x) (5)
Using the properties of the entropy, the mutual information can be rewritten into
MI(x, y) = H(x) + H( y) − H(x, y) (6)
This entropy is calculated using k-nearest neighbor method. This method has a
second advantage. It allows the calculation between a set of features x and the output
variable y.

6.4 Hybrid Color-Texture Space

In this part, an algorithm will select the attributes composing the hybrid color-texture
space. These attributes are chosen from the space of candidates attributes E. The
algorithm based on the mutual information is developed. It works in two steps. In the
first step, the forward phase, attributes are adding one by one. At each iteration, the
attribute selected to integrate the current subspace is the one that most increases the
mutual information with the output variable. The forward phase is stopped when
adding any new attribute decreases the mutual information. The second step is the
backward phase. Attributes are eliminated one at time. The attribute that most
increase the mutual information when it is discarded is eliminated from the subset of
28 M. Uma Selvi and S. Suresh Kumar

features. As in the forward step, the backward phase is stopped when discarding any
other attributes decreases the mutual information of the subspace with the output
variable. The final subspace is the hybrid color-texture space. It contains the attributes
that are the most discriminate.

7 Support Machine Vector

Among the methods with kernels based on the statistical learning theory of Vladimir
Vapnik, SVM are the most known. SVM is a method of binary classification by
supervised learning, it was introduced by Vapnik in 1995 . This method is an
alternative to the recent classification. This method relies on the existence of a linear
classifier in an appropriate space.

7.1 SVM Classifier

Support Vector Machine Algorithm are attractive, especially for non-linearly

separable problems, where the problem is high-dimensional. A classification task
usually involves with training and testing data which consists of some data instances.
Each instance in training set contains one “target value”(class lables) and several
“attributes”(features).The goal of SVM is to produce a model which predicts target
value of data instances in the test set which are given only the attributes. It provides a
Novel means of classification using the principles of structural risk minimization
(SRM). It is one of the most sophisticated nonparametric supervised classifiers
available today, with many different configurations depending upon the functions
used for generating the transform space in which the decision surface is constructed.
Here the classification problem is restricted to consideration of the two-class problem
without loss of generality. In a binary classification task, the aim is to find an optimal
separating hyper-plane. Considering the example shown in fig.1,there are many
possible linear classifiers that can separate the data, but there is only one that
maximizes the margin(maximizes the distance between the hyper-plane and the
nearest data of every other class).
Only those vectors that are on the perimeter of the data cloud are used to determine
the boundary. In other words, those data vectors, which are closest to the separating
margins in the transformed space, are called the support vectors(SV).
Since this is a problem of classification of two classes, this method uses a set of
learning data to learn the parameters of the model. It is based on the use of so-called
kernel function that allows an optimal separation of the data. It was successfully
evaluated on pattern recognition problem . For example in two classes
problem(positive and negative sets of simples), the basic form of linear SVM
classifier try to find an optimal hyperplane that separates the set of positive samples
from the set of negative samples. The LIBSVM package that support multi-class
problem available on to classify the different textures is used in our work.

8 Experiments
Our experiments were conducted using a PC with a Pentium 4 CPU 1.8 G with 1-GB
memory, and they involve the following two image data sets.
 Sea Object Detection Using Shape and Hybrid Color Texture Classification 29

Fig. 1. Typical sea images. (a) Images with large clouds. (b) Images withsmall clouds. (c)
Images with strong ocean waves. (d) Images with small islands.

Fig. 2. Typical samples in Data set

Data Set 1: This data set consists of spaceborne optical images, about 2000 ∗ 2000
pixels in size, with a total of 232 images and a resolution of 5–20 m. Typical samples
are shown in Fig. 1. It is used to test our proposed approach of ship detection. For
simplicity, all of the images were scaled to the spatial resolution of 10 m to train only
one classifier in our experiments.
Data Set 2: This data set includes more than 600 typical ship candidate sub images
obtained by our ship candidate extraction from the spaceborne optical images in Data
set 1. Some typical samples are shown in Fig. 2. The aim is to test the performance of
the presented hierarchical classification approach, which is a very important section
of the ship detection approach.

9 Conclusion
Because of the security problems sea target databases, unlike face and skin images,
are not published. Lack of famous color images for sea target detection, forced us to
create our own database. I am going to test the method on 3 different databases. First
database is about 2000 frames of the sea targets from several mainstreams movies.
The second database is collected from different military websites and the third one is
database from low quality movies filmed by one of famous military companies for sea
target detection purposes. It is obvious that because of observing framing rules the
third database has the best result.
30 M. Uma Selvi and S. Suresh Kumar

An efficient classification algorithm for ship detection that combines color and
texture features were presented to detect ship in the sea . The Hybrid color-texture
space was constructed for classified image using mutual information. Thus the
algorithm may achieve good classification rate when compared to existing algorithms.

References
1. Zhu, C., Zhou, H., Wang, R., Guo, J.: A Novel Hierarchical Method of Ship Detection
from Spaceborne Optical Image Based on Shape and Texture Features. IEEE Transactions
On Geoscience And Remote Sensing 48, 3446–3456 (2010)
2. Mirghasemi, S., Yazdi, H.S., Lotfizad, M.: Linear and Quadratic PSO Based Color Space
Conversion for Sea Target Detection. International Journal of Computer and Electrical
Engineering 2, 1793–8163 (2010)
3. Mirghasemi, S.: Sea Target Detection Based on SVM Method Using HSV Color Space. In:
IEEE Students Conference on Research and Development (November 2009)
4. Huang, S.-q., Liu, D.-z., Gao, G.-q., Guo, X.-j.: A novel method for speckle noise
reduction and ship target detection in SAR images. In: Science Direct, Pattern
Recognition, pp. 1533–1542 (October 2010)
5. Du, Z., Liu, R., Liu, N., Chen, P.: A New Method for Ship Detection in SAR Imagery
based on Combinatorial PNN model. In: IEEE First Int. Conf. on Intelligent Networks and
Intelligent Systems, pp. 531–534 (2008)
6. Hu, Y., Wu, Y.: Number Estimation of Small-Sized Ships in Remote Sensing Image Based
on Cumulative Projection Curve. In: IEEE Int. Conf. on Audio, Language and Image
Processing (ICALIP), pp. 1522–1526 (July 2008)
7. Yang, S., He, S., Lin, H.: Video Image Targets Detection Based on the Largest Lyapunov
Exponent. In: IEEE The 9th Int. Conf. for Young Computer Scientists, pp. 2973–2977
(November 2008)
8. He, S., Yang, S., Shi, A., Li, T.: A Novel Image Moving Sea Targets Detection Method
Based on Natural Measure Feature. In: IEEE Int. Symposium on Information Science and
Engineering, pp. 397–400 (December 2008)
9. Hong, Z., Jiang, Q., Guan, H., Weng, F.: Measuring overlap-rate in hierarchical cluster
merging for image segmentation and ship detection. In: Fourth Int. Conf. on Fuzzy
Systems and Knowledge Discovery (FSKD), vol. 4, pp. 420–425 (2007)
10. Tello, M., Lopez-Martinez, C., Mallorqui, J.J., Bonastre, R.: Automatic Detection of Spots
and Extraction of Frontiers in SAR Images by Means of the Wavelet Transform:
Application to Ship and Coastline Detection. In: IEEE Int. Con. on Geoscience and
Remote Sensing Symposium, IGARSS, pp. 383–386 (August 2006)
11. Marivi, T., Carlos, L.M., Jordi, J.M.: A novel algorithm for ship detection in SAR imagery
based on the wavelet transform. IEEE Geoscience And Remote Sensing Letters 2, 201–205
(2005)
12. Maia, H.E., Aboutajdine, D., Hammouch, A.: Color-texture analysis by mutual
information for multispectral image classification. In: IEEE Pacific Rim Conference on
Communications, Computers and Signal Processing, pp. 359–364 (August 2009)
13. El Maia, H., Hammouch, A., Aboutajdine, D.: Hybridization of the color information with
the texture by using the mutual information for the classification of satellite images. In: Int.
Conference on Multimedia Computing and Systems (ICMCS), pp. 372–376 (April 2009)
14. Otsu, N.: A threshold selection method from gray-level histogram. IEEE Trans. Syst.,
Man, Cybern. SMC-9(1), 62–66 (1979)
 Sea Object Detection Using Shape and Hybrid Color Texture Classification 31

15. Gonzalez, R.C., Woods, R.E.: Digital Image Processing, 2nd edn. Prentice-Hall, Upper
Saddle River (2002)
16. Chan, F.T., Vese, L.: Active contours without edges. IEEE Trans. Image Process. 10(2),
266–277 (2001)
17. Sonka, M., Hlavac, V., Boyle, R.: Image Processing, Analysis, and Machine Vision, 2nd
edn. Brooks/Cole, Pacific Grove (2002)
18. Dudani, S.A., Breeding, K., McGhee, R.: Aircraft identification by moment invariants.
IEEE Trans. Comput. C-26(1), 39–46 (1977)
19. Hu, M.: Visual pattern recognition by moment invariants. IEEE Trans. Inf. Theory IT-8(2),
179–187 (1962)
20. Kourosh, J.K., Hamid, S.Z.: Radon transform orientation estimation for rotation invariant
texture analysis. IEEE Trans. Pattern Anal. Mach. Intell. 27(6), 1004–1008 (2005)
 Multi-layer Logon Verification System:
A Case Study of Indian Banks

Mohammad Asim

Dept.of CSSE
University of Hail, KSA
[email protected]

Abstract. Internet is only the medium which connects billions of people

globally and finally has been removed all physical barriers. Bank is also one of
the sector which greatly influenced with the invent of Internet. However,
Internet banking is one of the service of the bank which provides many services
online to the customers without being physically visit the bank.On the one end
Internet banking bring many services to the customer but on the contrary also
pose a risk of fraud, theft and vandalism.
The key aspect of any online services to authenticate the customer without
any face-to-face communication.which ultimately challenge for the technology
to validate them without any compromise. In the whole scenario security can be
compromised at three level ;at the client ,in transit and at the server.Presently
there are number of authentications methods proposed by the researcher
worldwide for the financial institutions but my study is particularly focus on
Indian Banks. Moreover, a special focus given on online services providing by
the banks to their customers. Study reveals the authentication methods is a
layered approach to achieve high degree of privacy and security in online
services and which is a combination of two approach such as mobile and
Token based authentication methods.
I have conducted a study of 25 banks of India to collect the data regarding
present Authentication methods adopted presently. The paper describe in detail
working of the said authentication methods and future implementation of these
technologies.

Keywords: Internet banking, Protection and Security, Mobile Authentication,

Token Authentication.

1 Introduction
In the current scenario, the Internet has become extensively used gizmo for corporate
organizations, education and bank is using Internet Technology to provide various
online services to serve their customers with ease and comfort. However, with the
growing application of the Internet, security has become an important area of
attention for various levels of users and however, a big challenge for the organization
to implement security methods for safe and hassle free services for their customers to
gain their confidence. There are several factors that may be affecting the adoption of
internet banking. However, an important factor is the level of security and the risk
associated with it[1].

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 32–41, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Multi-layer Logon Verification System: A Case Study of Indian Banks 33

The paper mainly discusses and analysis the current procedures adopted by major
banks in India.however, it is important to note that whenever there is trust, there is
risk. Thus, trust is accepting risk. Trust, risk, security and reliability are fundamental
to internet banking.I investigated the security of the internet banking systems in
Banks in India including Public,private and foreign banks through visiting their sites
and security guidelines/manual they have online.
The rest of paper is outlined as follows. In section 2, I describe current Internet
Banking Authentication System adopted by Indian Banks. In section 3, I have
purposed a approach for authentication, In section 4 is concern the case study of the
banks.In section 5,There are some recommendations for the banks and finally I
concluded in section 6.

2 Existing Internet Banking Authentication Methods in India

I have conducted the study of 25 major banks in India through visit their official sites
to obtain the information for their security methods adopted for authentication
process:
SSL + HTTP + Virtual Keyboard

Authentication is a method to verify the credentials of the customer before using any
online services without any paper work or physical identification system. However,
now the question arises here for validating the credentials of the customer[2].After a
thoroughly studied of the banks sites, I have found that all the banks are using method
of authentication based on HTTPS. Which allow a secure session between the clients
machine and the bank’s server.
Presently the authentication system of the Internet banking in India is based on
public key cryptography system and the digital signature technology for
authenticating, thus the safe and smooth electronic transaction can be ensured. The
Certificate based authentication system is a certificate center system based on PKI.
The security authentication system of internet banking is composed of Central Bank
Certification Authority (CA) and the Commercial Bank internal management CA.
These two systems work independently and cooperatively. The tasks are divided as
follows:
The Central Bank CA issues the digital certificates to the clients of the internet
banking, and the internal management CA issues management digital certificates to
commercial bank employees. These digital management certificates are mainly used
in the Registration Authority (RA) system. The main purpose of Commercial bank
employees to use digital certificate to authenticate ID, encrypt transmission and
resist repudiation, also audit clients' certificate requests and create the clients'
certificates.
Security proxy permit a web-based transaction interface embedded with
functionalities such as authentication, integrity, non-repudiation, encryption,
authorization and security enhancements. The main function of Digital signature is to
ensure the integrity and non-repudiation of any transaction [3].
34 M. Asim

Database

Web Server
CA Admin
Signature
System

CA Server

Safe Agent
Server
RA Server User
Database

Internet Intranet

USER USER Administrator Terminal user

Fig. 1. Internet Banking Security Architecture (Source: IEEE 2006,p.1570 )

2.1 Virtual Keyboard for Internet Banking

Virtual Keyboard is an online application to enter password with the help of a mouse.

2.2 Advantage of a Virtual Keyboard

The Virtual Keyboard is designed to protect your password from malicious

“Spyware” and “Trojan Programs”. Use of Virtual keyboard will reduce the risk of
password theft.
Process to Use Virtual Keyboard
Steps to use Virtual Keyboard are as follows:
• Enter your Login Id using Physical Keyboard.
• Select the check box “Use Virtual Keyboard”.
• Use the Virtual Keyboard to enter the login password.
• Once you have entered your password, click “Log-in”[4].
 Multi-layer Logon Verification System: A Case Study of Indian Banks 35

3 Proposed Approach

In today's Internet-driven society, most financial institutions offer online banking

services to their customers. However, with the fast growth of online fraud, Internet
banking comes with risks. A majority of users are worried about the improper use of
their bank account details on the Internet.
Therefore, banks are looking for ways to protect their brand image and preserve the
trust they have with their customers [5].
My approach is based on the concept that encryption of the user name/password is
not enough that is why we introduce a second layer of security which would be highly
safe and secure for online banking services. That is called multi-layer logon
verification system.
1. Layer is concern with the level and type of security and authentication
method.In this layer customer is suppose to enter pair of information
username/password like traditional method which would be secure and
protected by using HTTPS technology.
2. Second layer allow authenticating the user further to enter some sort of code
generated by bank server and send it to the registered mobile number of the
user or the token generated by hardware devices. The two methods are:

a. Mobile Authentication and

b. Token Authentication

I will discuss in detail, the working of these two methods in the next two sections.

3.1 Mobile Authentication

The First Method: Using the mobile phone

One of new security enhancements is the new authentication method that has been
developed for our individual customers. This method allows the customer to use
his/her mobile phone as an authentication tool.

How does the authentication using the mobile phone work?

• This relates to when the user chooses to authenticate some
transactions/payments through his/her mobile phone, while using Internet
Banking.
• When the customer attempts to make a transaction/payment, the banking system
generates a special password for this transaction, and sends it as an SMS
message to their mobile phone.
• The user then enters this password in order to authenticate the requested
transaction, thereby adding another layer of security.
• The banking system checks the entered password against the system-generated
password, and if it matches the transaction is completed.
36 M. Asim

Code send by the bank

CUSTOMERS’ MOBILE

BANK

Fig. 2. Architecture of Mobile Authentication

How to register a mobile phone number of the customer on the banks site to
allow it to be used as an authentication tool?
The user enters the mobile phone number, on Internet banking service after which the
system will give the user the following options for authenticating the mobile phone
number:

Through the ATM:

• When using this method the system creates an application code and sends it
as an SMS message to the entered mobile number.
• The user needs to authenticate his/her ID and confirm the entered mobile
number using the application code. This can be done by visiting a any Bank
ATM and entering the Order Number in the appropriate service menu.
• Upon entry, the system checks the entered application code against the
system-generated application code, and if it matches, the system then
registers the mobile number for use as an authentication tool.

Through Internet Banking Module – Security Token

• When the user chooses this method to authenticate the mobile phone number,
the system then asks for a security token code to be entered.
• When successfully entered the system then generates an application code and
then sends it as an SMS message to the entered mobile number.
• The user then need to authenticate his/her ID and confirm the entered mobile
number using the application code received, by entering the application code
on Internet banking module.
• Upon entry, the system checks the application code against the system-
generated application code, and if it matches, the system then registers the
mobile number for use as an authentication tool.
 Multi-layer Logon Verification System: A Case Study of Indian Banks 37

3.2 Token Authentication

This is a method for authentication that will be expanded to cover all third party
transactions as well as any Internet banking services. This option applies using a
small electronic device supplied by the Bank to the customers; the device generates a
one off password, for every transaction. To date this security token device has only
been used for some transactions (i.e. transactions exceeding a certain amounts).
However, my idea to introduce this method of transaction authentication with all
transactions/payments made from the customer’s account to third party, as well when
updating certain essential information.
The small devices are typically of the size of small key chains, calculators or credit
cards. Usually an authentication token has the following features:
1. Processors
2. Liquid Crystal Display (LCD) for displaying outputs.
3. Battery
4. (Optionally) a small keypad for entering information.
5. (Optionally) a real time clock.
Each authentication token is pre-programmed with a unique number, called as a
random seed. The seed forms the basis for ensuring the uniqueness of the output
produced by the token.
Let we understand. How does it work step by step?
Step1: Creation of a Token: Whenever a authentication token is created, the
corresponding random seed is generated for the token by the authentication server.
This seed is stored or pre-programmed inside the token, as well as its entry is made

ID=ASIM
Seed=8976654

User Record Creation ID Seed

ASIM 8976654
Humaid 7665439
Server Zarrin 5113324

User Database

Seed

Authentication Token

Fig. 3. Random seed storage in the database and the authentication token
38 M. Asim

against that user’s record in the user database. Conceptually it’s like a password but
not exactly a password because the user know about the password but the user does
not know about the value of the seed unlike password.

Step 2: Use of Token

An authentication token automatically generates pseudorandom numbers ,called as

one time passwords.One time passwords are generated randomly by an authentication
token, based on the seed value that they are programmed with.They are one-time
because they are generated ,used once ,and discarded forever. When a user wants to
be authenticated, the user will get a screen to enter the user id and the latest one time
password obtained from the authentication token.The user id and password travel to
the server as a part of the login request.The server obtains the seed corresponding to
the user id from the user database,using a seed retrieval program.It then calls another
program called as password validation program,to which the server gives the seed and
the one time password .This program knows how to establish the relationship between
the seed and the one time password.Now we are explaining in simple way,the
program uses synchronization techniques,to generate the same one-time password as
was done by the authentication token.However,the main point to be noted here is that
the authentication server can use this program to determine if a particular seed value
relates to a particular one-time password or not.This is shown in the figure 4.A
question at this stage could be ,what would happen if a user loses an authentication
token? Can another user simply grab it and use it?To deal such situations,the
authentication token is generally protected by a password or a 4 digit PIN. Only when
this PIN is entered can the one-time password be generated. This is also the basis for
what is called as multi-layer logon verification system.What are these factors?There
are three most common factors:
• Something that you know,e.g a password or PIN
• Something you have,e.g a credit card or an identity card
• Something you are,e.g you voice or finger print
Based on these principles,we can see that a password is a single layer
authentication,because it is only something that you know.In contrast ,authentication
token are examples of second layer of authentication,here you must have something
(the authentication token itself) and you must also know something(the PIN is used to
protect it).Someone only knowing the PIN or only having the token cannot use it.
Both the layers of authentication are required that is why it is called multi -layer
authentication.

Step 3: Server return an appropriate message back to the user

Finally ,the server sends an appropriate message back to the user,depending
on whether the previous operations yielded success or failure[6].This is shown in
figure 5:
 Multi-layer Logon Verification System: A Case Study of Indian Banks 39

Step 1: the user’s id and the one-time password obtained from the authentication token are
sent to the server.

1. Login request

Client ID=Asim
Password=5432216 Server

5432216

Authentication Token

Step 2: The server’s seed retrieval program now retrieves the seed for this user from the user database

Id = Asim
Seed=?

Seed Retrieval Program ID Seed

ASIM 8976654
Humaid 7665439
Zarrin 5113324
Seed Found

Seed=8976654
User database

Step 3: The server’s password validation program calculates the one-time password and checks the seed against
the one-time password.

Password validation Program Password

Validation
Program

Password is correct

Fig. 4. Server validate the one-time password

40 M. Asim

Login Successful

Application Menu
1. View Menu
2. Transfer Money
3. ………

Fig. 5. Server sends an appropriate message back to the user

4 Case Study of Indian Banks

The following table shows the current methods by the banks in India and I have
collected and analysis the data after visiting the official site of the banks.

S.No Name of the Bank URL Security Methods for online Banking
1 State Bank of India https://www.onlinesbi.com PKI/Digital Signature Virtual Keyboard
2 Union Bank of India http://www.unionbankonline.co.in PKI/Digital Signature Virtual Keyboard
3 Bank of Baroda http://bankofbaroda.com PKI/Digital Signature Virtual Keyboard
4 Andhra Bank https://www.onlineandhrabank.net.in PKI/Digital Signature Virtual Keyboard
5 Allahabad Bank https://www.allbankonline.in PKI/Digital Signature Virtual Keyboard
6 ICICI Bank http://icicibank.com PKI/Digital Signature Virtual Keyboard
7 HDFC Bank http://www.hdfcbank.com PKI/Digital Signature Virtual Keyboard
8 Bank of India http://www.bankofindia.com PKI/Digital Signature Virtual Keyboard
9 Axis Bank http://www.axisbank.com PKI/Digital Signature NO
10 Karnataka Bank http://www.karnatakabank.com PKI/Digital Signature Virtual Keyboard
11 Kotak Mahindra Bank http://www.kotak.com PKI/Digital Signature Virtual Keyboard
12 Yes bank http://www.yesbank.co.in PKI/Digital Signature NO
13 American Express https://home.americanexpress.com PKI/Digital Signature NO
14 Corporation Bank http://www.corpbank.com/ PKI/Digital Signature Virtual Keyboard
15 Indian bank https://www.indianbank.net.in PKI/Digital Signature
NO
16 Punjab National Bank http://www.pnbindia.in PKI/Digital Signature Virtual Keyboard
17 Syndicate Bank http://www.syndicatebank.in PKI/Digital Signature Virtual Keyboard
18 Vijaya bank http://vijayabank.com/vijaya/vijaya PKI/Digital Signature Virtual Keyboard
19 UCO Bank http://www.ucobank.com/default.htm PKI/Digital Signature Virtual Keyboard
20 Induslnd bank http://www.indusind.com/indusind/wcms/en/ho PKI/Digital Signature Virtual Keyboard
me/personal-banking
21 INGVysya http://www.ingvysyabank.com PKI/Digital Signature Virtual Keyboard
22 State Bank of http://www.sbhyd.com/default.asp PKI/Digital Signature Virtual Keyboard
Hyderabad
23 Oriental Bank of https://www.obcindia.co.in/obcnew/site/index.as PKI/Digital Signature Virtual Keyboard
Commerce px
24 State Bank of Bikaner & http://www.sbbjbank.com PKI/Digital Signature Virtual Keyboard
Jaipur
25 Indian Overseas Bank http://www.iob.in PKI/Digital Signature Virtual Keyboard
 Multi-layer Logon Verification System: A Case Study of Indian Banks 41

5 Recommendations for the Banks

1. Winning the faith of the customers is the primary goal for any financial
institutions. But customers are not merely looking towards the services as
well as they are worried about the protection and security of their personal
and financial data.
2. Study revealed, that most of the banks in India, they are using traditional
approach to authenticate the customers which is not enough. However,
increasing cyber attacks demands more sophisticated and advance methods.
3. Study proposed a method of authentication. Which is basically broken into
two separate layers and each one is performing individual function to
achieve high degree of privacy and security for online services.

6 Conclusion
This paper is based on the case study of Indian banks and particularly focuses on
authentication methods in Internet banking system. I have also explored the current
method of authentications by the banks and proposing the new techniques for
authentication addition to the existing methods. The implementation of these new
techniques would be greatly influence the security and create confidence in customers
to adopt Internet banking. Still Internet banking is less popular due to lack of
knowledge of using Internet banking, privacy issues and risks involved among the
customers. My research definitely proved to be good idea for banking people to
understand the role of technology to win the faith of the customer. The recognition of
any banks depends on two factors, the services they are providing to their customers
and the technology implementation especially at security side.

References
[1] Sanayei, A., Noroozi, A.: Security of Internet Banking Services and its linkage with Users’
Trust. In: International Conference on Information Management and Engineering 2009, pp.
3–7 (2009)
[2] Xing, F., Zhan, J.: Online banking Authentication Using Mobile Phones.Future
Information Technology (FutureTech). In: 2010 5th International Conference, pp. 1–5
(2010)
[3] Lin, S., Liu, Y.: An Improved Security Authentication Description and Implementation in
Internet Banking System. In: Proceedings of International Conference in Communications,
Circuits and Systems, vol. 3, pp. 1570–1572 (2006)
[4] http://www.icicibank.com/pfsuser/webnews/virtualkeyboad.htm
[5] http://www.gemalto.com/brochures/download/ezio_mobile_auth.pdf
[6] Kahate, A.: Cryptography and Network Security. Tata McGraww-Hill Publishing
Company Limited, New York
 Image Compression Using Two Dimensional Spherical
Coder in Wavelet Lifting

S. Panimalar, R. Anisha, and M. Gomathi

Manonmaniam Sundaranar University,

Tirunelveli
[email protected]
[email protected]

Abstract. In recent years, many wavelet coders that use various spatially
adaptive coding technique to compress the image. Level of flexibility and the
coding efficiency are two crucial issues in spatially adaptive methods. So in this
paper “spherical coder” is introduced. The objective of this paper is to combine
the spherical tree with the wavelet lifting technique and compare the
performance of the spherical tree between the different coding technique such
as arithmetic, Huffman and run length. The Comparison is made by using
PSNR and Compression Ratio (CR). It is shown the Spherical tree in wavelet
lifting with the arithmetic coder gives high CR value.

Keywords: Discrete Wavelet Transform, Wavelet Lifting, Shperical Coder,

Arithmetic Coding, Run length Coding, Huffman Coding.

1 Introduction
Image compression is the application of Data Compression on Digital Images. When
a digital image is transmitted through a communication medium like a cable or a
simple telephone line, the cost of the transmission depends on the size of the file and
is a very important factor. To reduce the transmission cost the file has to be
compressed. The objective of the compression is to reduce redundancy of the image
data in order to be able to store or data in an efficient form.
Image compression may be lossy or lossless. Lossless compression is mostly
preferred for artificial images such as icons, comics and technical drawings. Because
of using low bit rate lossy compression produce some compression artifacts. Lossy
methods are suitable for natural images. This is because the minor loss of information
is acceptable to achieve a substantial reduction in bit rate.
Lossy compression is most commonly used to compress the multimedia data such
as audio, video, still images especially in applications such as internet telephony and
streaming media. Lossless compression is most commonly used to compress the text
and data files such as bank records, text articles, etc.
To compress the image any one of the coder is used. All image coders are based
on some statistical model for natural images. The coder is explicitly or implicitly
optimized for the specific model and applied to sample images. Wavelet domain
provides a good match to the space-frequency characteristics of natural images.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 42–54, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Image Compression Using Two Dimensional Spherical Coder in Wavelet Lifting 43

Hence, it is much easier to build a realistic image model in the wavelet domain than in
the pixel domain. That is because a simple coder in the wavelet domain could
outperform a complex coder in pixel domain.
Wavelet subbands are best characterized by spatially varying nonhomogeneous
processes. The “spherical representation” that has been introduced in [1] as a novel
way of representing image information in wavelet domain. Based on the spherical
representation, we develop a coding algorithm that handles this nonhomogeneity in an
effective and nonparametric way.
Successful coding schemes used in DCT-based algorithms, such as run-length
coding, and vector quantization [2], [3] were tried in wavelet image coding, but
demonstrated modest coding gains over standard transform-based algorithms. The
breakthrough in wavelet image coding arrived with coders using hierarchical wavelet
trees, such as EZW [4], SPIHT [5], SFQ [6]–[9]. Grouping wavelet coefficients that
belong to the same spatial region under a tree-structure, these coders were able to
adapt to the properties of different regions in the image. Other coders, such as EQ
[10], classification-based algorithms [11], EBCOT [12], etc., achieved improved
coding efficiency by introducing more spatial adaptivity in modeling the subbands.
The rest of this paper is organized as follows: in Section II we review the Discrete
Wavelet Transform (DWT). Wavelet Lifting Scheme and features are discussed in
Section 3. Section 4 describes the details of the spherical representation and spherical
coding algorithm. In Section 5 the different coding techniques are discussed. The
performance of the spherical representation with various coders and various wavelet
domain are compared in Section 6 in a series of computational experiments. In
Section 7 we summarise the results of the study . The conclusion of the paper is in
Section 8.

2 Discrete Wavelet Transform

The discrete wavelet transform (DWT) is an implementation of the wavelet transform
using a discrete set of the wavelet scales and translations obeying some defined rules.
In other words, this transform decomposes the signal into mutually orthogonal set of
wavelets, which is the main difference from the continuous wavelet transform (CWT),
or its implementation for the discrete time series sometimes called discrete-time
continuous wavelet transform (DT-CWT).
The discrete wavelet transform is a hierarchical subband technique. The subbands
are created by applying decomposition on original image. To start the decomposition,
the image is filtered in both the horizontal and vertical directions using separable
filters. This creates four subbands as follows: Subband LL1 represents the horizontal
and vertical low frequency components of the image. Subband HH1 represents the
horizontal and vertical high frequency components of the image. Subband LH1
represents the horizontal low and vertical high frequency components. Subband HL1
represents the horizontal high and vertical low frequency components.
To obtain the next level in the decomposition the two filters are again applied, but
only to the LL1 subband. Again we obtain four subbands labeled LL2, HH2, LH2,
and HL2 with representations similar to the first level subbands. The following
example illustrates the discrete wavelet transform.
44 S. Panimalar, R. Anissha, and M. Gomathi

Fig. 1. Original Lena Image

The two level decompo osition image and corresponding decomposition mapp of
above figure is show below
w.

Fiig. 2. Two Level Decomposed Image

Fiig. 3. Two Level Decomposition Map

3 Wavelet Lifting Scceme

The wavelet lifting schemee divides the wavelet transform into a set of steps. Onee of
the elegant qualities of wav
velet algorithms expressed via the listing scheme is the ffact
that the inverse transform m is a mirror of the forward transform. lifting scheeme
wavelets the predict waveleet transform starts with a split step, which divides the ddata
 Image Compression Using Two Dimensional Spherical Coder in Wavelet Lifting 45

set into odd and even elements. The predict step uses a function that approximates the
data set. The difference between the approximation and the actual data replaces the
odd elements of the data set. The even elements are left unchanged and become the
input for the next step in the transform. The predict step, where the odd value is
"predicted" from the even value is described by the equation

oddj+1,i = oddj,i - P( evenj,i ) (1)

The inverse predict transform adds the prediction value to the odd element (reversing
the prediction step of the forward transform). In the inverse transform the predict step
is followed by a merge step which interleaves the odd and even elements back into a
single data stream.
The simple predict wavelets are not useful for most wavelet applications. The even
elements that are used to "predict" the odd elements result from sampling the original
data set by powers of two (e.g., 2, 4, 8...). Viewed from a compression point of view
this can result in large changes in the differences stored in the odd elements (and less
compression). One of the most powerful applications for wavelets is in the
construction of filters. The "down sampling" in the predict wavelets does not provide
an approximation of the data at each step, which is one of the requirements for filter
design.
The update step replaces the even elements with an average. This results in a
smoother input for the next step of the next step of the wavelet transform. The odd
elements also represent an approximation of the original data set, which allows filters
to be constructed. A simple lifting scheme forward transform is diagrammed in below
figure.

Fig. 4. Lifting Scheme Forward Wavelet Transform

The update phase follows the predict phase. The original value of the odd elements
has been overwritten by the difference between the odd element and its even
"predictor". So in calculating an average the update phase must operate on the
differences that are stored in the odd elements:

evenj+1,i = evenj,i + U( oddj+1,i ) (2)

In the lifting scheme version of the Haar transform, the prediction step predicts that
the odd element will be equal to the even element. The difference between the
predicted value (the even element) and the actual value of the odd element replaces
46 S. Panimalar, R. Anisha, and M. Gomathi

the odd element. For the forward transform iteration j and element i, the new odd
element, j+1,i would be
oddj+1,i = oddj,i - evenj,I (3)
In the lifting scheme version of the Haar transform the update step replaces an even
element with the average of the even/odd pair (e.g., the even element si and its odd
successor, si+1):
evenj+1,i=(evenji+oddji)/2 (4)
The original value of the oddj,i element has been replaced by the difference between
this element and its even predecessor. Simple algebra lets us recover the original
value:
oddj,i = evenj,i + oddj+1,I (5)
Substituting this into the average, we get

evenj+1,i=(evenj,i+evenj,i+oddj+1,i)/2
evenj+1,i= evenj,i+(oddj+1,i)/2 (6)

The averages (even elements) become the input for the next recursive step of the
forward transform. This is shown in Figure 5, below

Fig. 5. Two Steps in Wavelet Lifting Scheme Forward Scheme

One of the elegent features of the lifting scheme is that the inverse transform is a
mirror of the forward transform. In the case of the Haar transform, additions are
substituted for subtractions and subtractions for additions. The merge step replaces
the split step. This is shown in Figure 6, below

Fig. 6. Lifting Scheme Inverse Transform

 Image Compression Using Two Dimensional Spherical Coder in Wavelet Lifting 47

4 Spherical Representation
The “spherical representation” is a hierarchical description of how total coefficient
energy gets distributed within each wavelet subband. A hierarchical tree of subband
energy is formed by summing up the squared coefficients. Phase variables are defined
that describe how the energy in a given region is split into energies of two sub-
regions. Phase variables are coded based on a simple and effective model. The
nonhomogeneity of wavelet subbands is handled through this nonparametric model of
the hierarchy.
We use the following hierarchical structure to represent a random process X (see
Fig.7): In 1-D for X={xi}1≤i≤2k and for 0<m ≤ k, 0 ≤ n < 2k-m.

Em(n)=∑ (7)

(8)

where 0 ≤ Ψm(n) ≤ Π/2. Here, X could be seen as one of the wavelet subbands of a
1-D signal. In the next section, this formulation is easily extended to 2-D subbands.
The variables Em(n) provide local energy information at different resolution levels
m. The phase variables Ψm(n) indicate how the local energy gets split between the
two neighboring regions

Em-1(2n) = Em(n)sin(ψm(n))2 (9)

Em-1(2n+1) = Em(n)cos(ψm(n))2 (10)

The phase variables in a sense represent the difference in information content between
the two regions. Going from the top level (m=k) to the bottom level (m=1) of the
hierarchy, the phase values provide a refinement of the available information in each
region of the subband.
When the total energy, Ek(0)=∑ and 2k-1the phases at all levels of the
hierarchy are given, the coefficients are easily determined up to a sign bit; i.e.,
. The sign bits could also be defined as part of the epresentation if the
phase values at the bottom of the hierarchy cover full range;i.e., -π<ψ1(i)≤π, and

X2i = (i) sin(φ(i)) (11)

X2i+1 = i cos (12)

In this type of representation, we are able to use local energy not only to differentiate
between statistically distinct parts of the process but also to provide direct information
about the underlying coefficient values. Coding Ek(0) and Ψm(n) can be seen as an
alternative to coding xi. We might say that, instead of cartesian coordinates, spherical
coordinate system is used in representing the process; hence, the name spherical
representation.
48 S. Panimalar, R. Anisha, and M. Gomathi

Fig. 7. Spherical Representation in 1-D

Spherical representation could be easily extended to 2-D to be used in wavelet

image coding. Partial squared sums need to be defined in both vertical and horizontal
directions in an alternating fashion (see Fig. 8). Let us represent the phase variables
ψu,v(s,t) and the local energies Гu,v(s,t) as and respectively.

Fig. 8. Spherical Representation in 2-D

5 Encoding
Encoding is the process of transforming information from one format into another.
A. Run Length Encoding(RLE)
One relatively simple way to compress an image is called Run Length Encoding
(RLE), which describes the image as a list of "runs", where a run is a sequence of
horizontally adjacent pixels of the same color. It codes the data by measuring the
length of runs of the values
The simplest form of compression technique which is widely supported by most
bitmap file formats such as TIFF, BMP, and PCX. RLE performs compression
regardless of the type of information stored, but the content of the information does
affect its efficiency in compressing the information.
 Image Compression Using Two Dimensional Spherical Coder in Wavelet Lifting 49

B. Huffman Encoding
We use this coding to encode/compress a file. Steps:
1. Read the file.
2. Calculate the probability of each symbol (since we use
an ASCII file, there are 256 possible symbols). Instead of actual
probability we can use instance count.
3. Use the Huffman coding algorithm to find the coding
for each symbol. Need to apply only on subset of symbols that actually
appear in the file.
4. Encode the file, and write it out.
The out file must also include the encoding table, so as to permit decoding.
C. Arithmetic Encoding
The principles of arithmetic coding describe an arithmetic coding engine that will
produce a compliant bit stream when used in conjunction with the correct methods for
binarization and context selection (described below).
Arithmetic coding is employed on integer-valued wavelet coefficients and uses a
two-stage process as shown in figure 9. First, integer values are binarized into a
sequence of bits or boolean values. At the same time a “context” is generated for each
of these bits. These boolean values, together with the corresponding context, are then
coded by the binary arithmetic coding engine.

Fig. 9. Arithmetic Encoding Process

Arithmetic coding is a lossless coding technique. In arithmetic coding, a data is

encoded as a real number in an interval from one to zero. Arithmetic coding typically
has a better compression ratio than Huffman coding. Arithmetic coding produces a
single symbol rather than several separate codeword.

6 Spherical Coding Scheme

The spherical coding algorithm is given (for each wavelet subband at different scales
and in different orientations) as follows
50 S. Panimalar, R. Anisha, and M. Gomathi

1. Apply Wavelet Lifting to input image

2. Use soft-thresholding to estimate zero-quantized wavelet coefficients

C^(m,n)= c(m,n) if |c(m,n)| > T

C^(m,n)= 0 if |c(m,n)| < T (13)

3. Create a spherical representation

u
(s+1)-1 2v(t+1)-1
Γu,v(s,t) = Σ Σ (m,n)2
m=2us n=2vt

0≤ S<2(J-u) , 0≤2(J-v) and for,0≤u<J

, ,
Ψu+1, u+1(s,t)=arctan
, ,

0≤s,t<2(J-u-1)
, ,
Ψu+1, u(s,t)=arctan (14)
, , ,

0≤s<2(J-u-1),0≤t<2(J-u)

4. Find the Lagrangian cost by using the below forumla

where I(m,n) represents the sign bit for coefficient c(m,n)

L0,0(m,n)=(c(m,n)-c(m,n))2+I(m,n) (15)

where I(m,n) represents the sign bit for coefficient c(m,n), i.e.,

I(m,n)=0 if c(m,n)=0

I(m,n)=1 else (16)

5. Compare the Lagrangian cost to previous cost. If the latter is smaller then
quantize to zero
6. Coding the local energies and phase variables using coder(arithmetic coder,
runlength coder, Huffman coder)
7. Decode the local energies and phase variables using arithmetic decoder
8. Decode the wavelet coefficients

, , Г , , (17)

Apply inverse Wavelet Lifting finally we get compressed image

 Image Compression Usin
ng Two Dimensional Spherical Coder in Wavelet Lifting 51

The block diagram of thee proposed spherical coding scheme is shown below

Fig. 10. Proposed Spherical Coding Process

7 Performance Analysis of Spherical Coding Scheme

In this performance analyssis of proposed method we have to analyse the spherrical
coding with various waveleet domain such as Discrete Wavelet Domain and Wavvelet
Lifting as well as with various coder such as arithmetic coder, run length coder and
the Huffman coder and find d out which one is the best. We carried out computatioonal
experiments to conform thee effectiveness of the above techniques.
To evaluate the performmance of the compression techniques several performaance
metrics are available. we usse the PSNR and CR to analyses the performance.
PSNR means Peak Sign nal to Noise Ratio (PSNR). This performance metric giives
the signal to noise ratio between the original image and the compressed image. PSSNR
ollowing formula.
is calculated by using the fo
2 1
20

∑ 2

CR means Compression Raatio. This performance metric gives compression ratio. T This
ratio gives an indication of how much compression is achieved for a particular image.
The performance of the algorithm is evaluated on Lena, Peppergray and Sailbboat
All the images are of 512x512 pixels. These pictures are the most widely uused
standard test images used for compression algorithms. The image contains a nnice
mixture of detail, flat regiions, shading, and texture that do a good job of testting
various image processing allgorithms.
We can evaluate these immages by the CR value and PSNR value.

(a) (b) (c)

Fig. 11. Expeerimental Images (a) Lena (b) Pepper (c) Sailboat
52 S. Panimalar, R. Anisha, and M. Gomathi

The performance analysis of spherical coding on PSNR value is given below

Table 1. Performance analysis on PSNR value

Image Wavelet Lifting

RLC HMC AMC
Lena 34.468 34.468 34.468
Pepper 35.046 35.046 35.046
Sailboat 32.848 32.848 32.848

Performance Analysis on PSNR value for different coder

35.5

35

34.5

34
PSNR Value

Lena
33.5 Pepper
Sailboat
33

32.5

32

31.5
Runlength Huffman Arithmetic
Coding Method

Fig. 12. Performance Analysis on PSNR Value of different coder

In the above graph that comparing is made with different coding techniques such
as run length , Huffman and arithmetic for lena, pepper gray and sailboat images.
From that graph the psnr value for Pepper image in all the coding technique is very
high.

Table 2. Performance analysis on CR value

Image Wavelet Lifting

RLC HMC AMC
Lena 16.20 33.33 66.12
Pepper 12.54 34.44 66.13
Boat 26.84 33.14 66.72
 Image Compression Using Two Dimensional Spherical Coder in Wavelet Lifting 53

Performance Analysis on Compression Ratio value of

different coder

80

70

60

50
CR value

Runlength
40 Huffman
Arithmetic
30

20

10

0
Lena Pepper Sailboat
Images

Fig. 13. Performance Analysis on CR Value of different coder

From the above graph we found that comparing with arithmetic coder, run length
coder and Huffman coder on lena pepper gray and sailboat images the arithmetic
coder gives higher CR value than the other coders.

8 Conclusions
In this paper the Spherical coding in the wavelet lifting with various wavelets domain
with various encoders have discussed. From that discussion the wavelet lifting
scheme and arithmetic coding gives best reconstructed image and also better CR and
PSNR is concluded. Arithmetic coder gives better CR as well as gives better PSNR.
Wavelet compression is a lossy method. To make it lossless and more efficient, we
go in for adaptive schemes in various factors for the transform. For example, we have
different algorithms like directional adaptive algorithm, Directional Qucinix Lifting,
Adaptive filters etc for future enhancement.

References
[1] Ates, H., Orchard, M.: Wavelet image coding using the spherical representation. In: Proc.
IEEE Int. Conf. Image Processing, Genova, Italy, vol. 1, pp. 89–92 (September 2005)
[2] Antonini, M., Barlaud, M., Mathieu, P., Daubechies, I.: Image coding using wavelet
transform. IEEE Trans. Image Process. 1(2), 205–220 (1992)
[3] Antonini, M., Barlaud, M., Mathieu, P., Daubechies, I.: Image coding using vector
quantization in the wavelet transform domain. In: Proc. IEEE Int. Conf. Acoustics,
Speech, Signal Processing, Albuquerque, NM, vol. 4, pp. 2297–2300 (April 1990)
[4] Shapiro, J.: Embedded image coding using zerotrees of wavelet coefficients. IEEE Trans.
Signal Process. 41(12), 3445–3462 (1993)
54 S. Panimalar, R. Anisha, and M. Gomathi

[5] Said, A., Pearlman, W.: A new fast and efficient image codec based on set partitioning in
hierarchical trees. IEEE Trans. Circuit Syst. Video Technol. 6(3), 243–250 (1996)
[6] Xiong, Z., Ramchandran, K., Orchard, M.: Space-frequency quantization for wavelet
image coding. IEEE Trans. Image Process. 6(5), 677–693 (1997)
[7] Xiong, Z., Ramchandran, K., Orchard, M.: Joint optimization of scalar and tree-structured
quantization of wavelet image decomposition. In: Proc. Conf. Rec. 33th Asilomar, Pacific
Grove, CA, vol. 2, pp. 891–895 (November 1993)
[8] Ramchandran, K., Orchard, M.T.: An investigation of waveletbased image coding using
an entropy-constrained quantization framework. IEEE Trans. Signal Process. 46(2), 342–
353 (1998)
[9] Xiong, Z., Galatsanos, N., Orchard, M.: Marginal analysis prioritization for image
compression based on a hierarchical wavelet decomposition. In: Proc. IEEE Int. Conf.
Acoustics, Speech, Signal Processing, Minneapolis, MN, vol. 5, pp. 546–549 (April 1993)
[10] LoPresto, S.M., Ramchandran, K., Orchard, M.T.: Image coding based on mixture
modeling of wavelet coefficients and a fast estimation-quantization framework. In: Proc.
Data Compression Conf., Snowbird, UT, pp. 221–230 (March 1997)
[11] Joshi, R., Jafarkhani, H., et al.: Comparison of different methods of classification in
subband coding of images. IEEE Trans. Image Process. 6(11), 1473–1486 (1997)
[12] Taubman, D.: High performance scalable image compression with EBCOT. IEEE Trans.
Image Process. 9(7), 1219–1235
 An Efficient Visible Watermarking for Copyright
Protection Using Discrete Wavelet Transform

R. Anisha, S. Panimalar, and M. Gomathi

Manonmaniam Sundaranar University,

Tirunelveli
[email protected],
[email protected]

Abstract. The objective of this project is to use a novel method for generic
visible watermarking with a capability of lossless image recovery. This method
is based on one-to-one compound mapping algorithm. The compound
mappings are proved to be reversible, which allows for lossless recovery of
original images from watermarked images. The mappings may be adjusted to
yield pixel values close to those of desired visible watermarks. Different types
of visible watermarks, including opaque monochrome and translucent full color
ones, are embedded as applications of the proposed generic approach. To
increase the visual perception Discrete Wavelet Transforms is used.
Experimental results show that our method provides a higher embedding
capacity compared to the other algorithms proposed in the literature.

Index Terms: Lossless reversible visible, watermarking, one-to-one compound

mapping, Discrete Wavelet Transforms, Translucent watermark.

1 Introduction

Digital watermarking is defined as a process of embedding data (watermark) into a

multimedia object to help to protect the owner's right to that object. The embedded
data (watermark) may be either visible or invisible .In visible watermarking , a
secondary image ie. a watermark is embedded in a primary (source) image such that
watermark is intentionally visible to a human observer whereas in the case of invisible
watermarking the embedded data is not visible, but may be extracted/detected by a
computer program.
Some of the desired characteristics of visible watermarks are listed below [23][24].

1. A visible watermark should be obvious in both color and monochrome

images.

2. The watermark should be spread in a large or important area of the image in

order to prevent its deletion by clipping.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 55–64, 2011.
© Springer-Verlag Berlin Heidelberg 2011
56 R. Anisha, S. Panimalar, and M. Gomathi

3. The watermark should be visible yet must not significantly obscure the
image details beneath it.

4. The watermark must be difficult to remove; removing a watermark should be

more costly and labor intensive than purchasing the image from the owner.

5. The watermark should be applied automatically with little human

intervention and labor.
Embedding of watermarks, either visible or invisible, affect the quality of the source
image in general. A group of techniques, named reversible watermarking [8]–[19],
allow users to remove the embedded watermark and restore the original content as
needed. However, not all reversible watermarking techniques guarantee lossless
image recovery, which means that the recovered image is identical to the original,
pixel by pixel. Lossless recovery is important in many applications where serious
concerns about image quality arise. Some examples include forensics, medical image
analysis, historical art imaging, or military applications.
Compared with their invisible counterparts, there are relatively few mentions of
lossless visible watermarking in the literature. Several lossless invisible watermarking
techniques have been proposed in the past. The most common approach is to
compress a portion of the original host and then embed the compressed data together
with the intended payload into the host [5], [13]–[15]. Another approach is to
superimpose the spread-spectrum signal of the payload on the host so that the signal is
detectable and removable [3]. A third approach is to manipulate a group of pixels as a
unit to embed a bit of information [16], [17]. Although one may use lossless invisible
techniques to embed removable visible watermarks [11], [18], the low embedding
capacities of these techniques hinder the possibility of implanting large-sized visible
watermarks into host media.
As to lossless visible watermarking, the most common approach is to embed a
monochrome watermark using deterministic and reversible mappings of pixel values
or DCT coefficients in the watermark region [6], [9], [19]. Another approach is to
rotate consecutive watermark pixels to embed a visible watermark [19]. One
advantage of these approaches is that watermarks of arbitrary sizes can be embedded
into any host image. However, only binary visible watermarks can be embedded
using these approaches, which is too restrictive since most company logos are
colorful.
In this paper, a method for lossless visible watermarking using an appropriate
compound mappings that allow mapped values to be controllable is used. The
mappings are proved to be reversible for lossless original image recovery . The
approach is generic, leading to the possibility of embedding different types of visible
watermarks into cover images. Two applications of the proposed method are
demonstrated, where opaque monochrome watermarks and non uniformly translucent
full-color ones are respectively embedded into color images. More specific compound
mappings are also created and proved to be able to yield visually more distinctive
 An Efficient
E Visible Watermarking for Copyright Protection 57

visible watermarks in the watermarked

w image. Embedding removable translucent ffull-
color watermarks provide better advertising effects than traditional monochroome
ones.
In the remainder of this paper, DWT is described in SectionII. The one-to-one
compound mappings is described in Section III. Related lemmas and theorems are
also proved and security prrotection measures described. Applications of the propoosed
method for embedding opaaque monochrome and translucent color watermarks iinto
color images are described in Sections 3 and 4, respectively. In Section 5, the speccific
compound mapping for yieelding more distinctive visible watermarks is describedd. In
Section 6, experimental ressults are presented to demonstrate the effectiveness of the
proposed method. Finally, a conclusion with some suggestions for future workk is
included in Section 7.

2 Discrete Wavelet Transform

T

The discrete wavelet transfo orm (DWT) is an implementation of the wavelet transfoorm
using a discrete set of the wavelet
w scales and translations obeying some defined ruules.
In other words, this transfoorm decomposes the signal into mutually orthogonal seet of
wavelets, which is the main n difference from the continuous wavelet transform (CWWT),
or its implementation for the discrete time series sometimes called discrete-tiime
continuous wavelet transforrm (DT-CWT).
The discrete wavelet transform is a hierarchical subband technique. The subbaands
are created by applying deccomposition on original image. To start the decompositiion,
the image is filtered in bo oth the horizontal and vertical directions using separaable
filters. This creates four sub
bbands as follows: Subband LL1 represents the horizonntal
and vertical low frequency y components of the image. Subband HH1 represents the
horizontal and vertical hig gh frequency components of the image. Subband L LH1
represents the horizontal loow and vertical high frequency components. Subband H HL1
represents the horizontal higgh and vertical low frequency components.
To obtain the next level in the decomposition the two filters are again applied, but
only to the LL1 subband. Again we obtain four subbands labeled LL2, HH2, LH2,
and HL2 with representattions similar to the first level subbands. The follow wing
example illustrates the discrrete wavelet transform.

Fig. 1. Original Lena Image

58 R. Anisha, S. Panimaalar, and M. Gomathi

The two level decompo osition image and corresponding decomposition mapp of
above figure is show below
w.

Fiig. 2. Two Level Decomposed Image

3 Reversible One-to--One Compound Mapping

First, we propose a generiic one-to-one compound mapping for converting a sett of
numerical values 1,, 2 … to another set Q={q1,q2...qM}, such that the
respective mapping from pi to qi for all i=1,2...,M is reversible. Here, for the
copyright protection applicaations investigated in this study, all the values pi and qi are
image pixel values (grayscaale or color values). The compound mapping f is goverrned
by a one-to-one function Fx with one parameter x=a or b in the following way:

q=f(p)= Fb-1(Fa(p)) (1)

where Fx-1 is the inverse of Fx which, by the one-to-one property, leads to the fact tthat
if Fa(p)=p' , then Fa-1(p')=pp for all values of a and b. On the other hand, Fa(p) and
Fb(p) generally are set to bee unequal if a≠b.
The compound mapping g described by (1) is indeed reversible, that is, p cann be
derived exactly from q usin ng the following formula:

p= f-1(q)=Fa-1(Fb(q)) (2)
 An Efficient Visible Watermarking for Copyright Protection 59

4 Generic Visible Watermark Embedding and Extraction

Theorem 1 (Lossless Reversible Visible Watermarking):
There exist one-to-one compound mappings for use to embed into a given imageI a
visible watermark Q whose pixel values are close to those of a given watermark L ,
such that the original image can be recovered from Q losslessly.

Proof: This is a consequence of Lemmas 1 and 2 after regarding the individual pixel
values in I ,L and Q respectively as those of p,l and q mentioned in Lemma 2. And it
is clear by Lemma 1 that the value p can be recovered losslessly from the mapped
value q which is derived in Lemma 2.
The above discussions are valid for embedding a watermark in a grayscale image.
If color images are used both as the cover image and the watermark, we can apply the
mappings to each of the color channels to get multiple independent results. The
resulting visible watermark is the composite result of the color channels.
Based on Theorem 1, the proposed generic lossless reversible visible watermarking
scheme with a given image and a watermark as input is described as an algorithm as
follows.

--------------------------------------------------------------
Algorithm 1: Generic Visible Watermark Embedding
--------------------------------------------------------------
Input: an image and a watermark L .

Output: watermarked image W.

Steps:
1) Select a set P of pixels from I where L is to be embedded, and call P a
watermarking area.
2) Denote the set of pixels corresponding to P in W by Q in by .
3) For each pixel X with value p in P , denote the corresponding pixel in Q as Z as
and the value of the corresponding pixel Y in L as l , and conduct the following
steps.
a) Apply an estimation technique to derive a to be a value close to p , using the
values of the neighboring pixels of X (excluding X itself).
b) Set b to be the value l .
c) Map p to a new value p=Fb-1(Fa(p)) .
d) Set the value of Z to be q .
4) Set the value of each remaining pixel in W, which is outside the region P, to be
equal to that of the corresponding pixel in I .
60 R. Anisha, S. Panimalar, and M. Gomathi

The corresponding watermark removal process for a watermarked Image W

generated by Algorithm 1 is described as an algorithm as follows.

Algorithm 2: Generic Watermark Removal for Lossless Image Recovery

--------------------------------------------------------------
Input: a watermarked image W and a watermark L.

Output: the original image R recovered from W .

Steps:
1) Select the same watermarking area Q in W as that selected in Algorithm 1.
2) Set the value of each pixel in R , which is outside the region Q , to be equal to
that of the corresponding pixel In W .
3) For each pixel Z with value q inQ , denote the corresponding pixel in the
recovered image R as X and the value of the corresponding pixel Y in L as l,
and conduct the following steps.
a) Obtain the same value a as that derived in Step 3a of Algorithm 1 by applying
the same estimatdion technique used there.
b) Set b to be the value l .
c) Restore p from q by setting p=Fa-1(Fb(q)) .
d) Set the value of X to be p .
 An Efficient Visible Watermarking for Copyright Protection 61

5 Experimental Results
The following four images are taken for this experiment.

To quantitatively measure the effectiveness of this method, we define a set of

performance metrics here. First, the quality of a watermarked Image W is measured
by the peak signal-to-noise ratio (PSNR) of W with respect to the non recoverable
watermarked image B in the following way:

PSNRW = 20 × log10 ﴾255⁄

∑ ∑ , , ﴿

Also, the quality of a recovered image R is measured by the PSNR of R with respect
to the original image I in a similar way

PSNRR = 20 × log10 ﴾255⁄

∑ ∑ , , ﴿
62 R. Anisha, S. Panimalar, and M. Gomathi

PSNR analysis after watermark embedding

IMAGES PSNR
Leena 45.02
Pepper 43.01
Boat 42.23
Baboon 47.22

PSNR
48
46
44
42
40 PSNR
38

Leena Pepper Boat Baboon

PSNR analysis after watermark extraction

IMAGES PSNR
Leena 48.34
Pepper 46.24
Boat 45.01
Baboon 50.02

6 Conclusions
In this paper, a new method for reversible visible watermarking with lossless image
recovery capability has been developed. The method uses one-to-one compound
 An Efficient Visible Watermarking for Copyright Protection 63

mappings that can map image pixel values to those of the desired visible watermarks.
The compound mappings allow different types of visible watermarks to be embedded.
To increase the visual perception Discrete Wavelet Transform is used. PSNR values
are increased.

References
[1] Petitcolas, F.A.P., Anderson, R.J., Kuhn, M.G.: Information hiding—A survey. Proc.
IEEE 87(7), 1062–1078 (1999)
[2] Johnson, N.F., Duric, Z., Jajodia, S.: Information Hiding. In: Steganography and
Watermarking—Attacks and Countermeasures. Kluwer, Boston (2001)
[3] Cox, I.J., Kilian, J., Leighton, F.T., Shamoon, T.: Secure spread spectrum watermarking
for multimedia. IEEE Trans. Image Process. 6(12), 1673–1687 (1997)
[4] Kankanhalli, M.S., Rajmohan, Ramakrishnan, K.R.: Adaptive visible watermarking of
images. In: Proc. IEEE Int. Conf. Multimedia Computing and Systems, vol. 1, pp. 568–573
(1999)
[5] Hu, Y., Kwong, S.: Wavelet domain adaptive visiblewatermarking. Electron. Lett. 37(20),
1219–1220 (2001)
[6] Mohanty, S.P., Ramakrishnan, K.R., Kankanhalli, M.S.: A DCT domain visible
watermarking technique for images. In: Proc. IEEE Int. Conf. Multimedia and Expo,
vol. 2, pp. 1029–1032 (July 2000)
[7] Braudaway, G., Magerlein, K.A., Mintzer, F.: Protecting publicly available images with
a visible image watermark. In: Proc. SPIE Int. Conf. Electronic Imaging, vol. 2659,
pp. 126–133 (February 1996)
[8] Cheng, Y.J., Tsai, W.H.: A new method for copyright and integrity protection for bitmap
images by removable visible watermarks and irremovable invisible watermarks. Presented
at the Int. Computer Symp.—Workshop on Cryptology and Information Security,
Hualien, Taiwan, R.O.C. (December 2002)
[9] Huang, P.M., Tsai, W.H.: Copyright protection and authentication of grayscale images by
removable visible watermarking and invisible signal embedding techniques: A new
approach. Presented at the Conf. Computer Vision, Graphics and Image Processing,
Kinmen, Taiwan, R.O.C. (August 2003)
[10] Hu, Y., Kwong, S., Huang, J.: An algorithm for removable visible watermarking. IEEE
Trans. Circuits Syst. Video Technol. 16(1), 129–133 (2006)
[11] Hu, Y., Jeon, B.: Reversible visible watermarking and lossless recovery of original
images. IEEE Trans. Circuits Syst. Video Technol. 16(11), 1423–1429 (2006)
[12] Macq, B.: Lossless multiresolution transform for image authenticating watermarking.
Presented at the European Signal Processing Conf., Tampere, Finland (September 2000)
[13] Fridrich, J., Goljan, M., Du, R.: Lossless data embedding—New paradigm in digital
watermarking. J. Appl. Signal Process. (2), 185–196 (2002)
[14] Awrangjeb, M., Kankanhalli, M.S.: Lossless watermarking considering the human visual
system. Presented at the Int. Workshop on Digital Watermarking, Seoul, Korea (October
2003)
[15] Awrangjeb, M., Kankanhalli, M.S.: Reversible watermarking using a perceptual model. J.
Electron. Imag. 14(013014) (March 2005)
[16] de Vleeschouwer, C., Delaigle, J.F., Macq, B.: Circular interpretation of bijective
transformations in lossless watermarking for media asset management. IEEE Trans.
Multimedia 5(1), 97–105 (2003)
64 R. Anisha, S. Panimalar, and M. Gomathi

[17] Tian, J.: Reversible data embedding using a difference expansion. IEEE Trans. Circuits
Syst. Video Technol. 13(8), 890–896 (2003)
[18] Tsai, H.M., Chang, L.W.: A high secure reversible visible watermarking scheme. In: Proc.
IEEE Int. Conf. Multimedia and Expo, Beijing, China, pp. 2106–2109 (July 2007)
[19] Yip, S.K., Au, O.C., Ho, C.W., Wong, H.M.: Lossless visible watermarking. In: Proc.
IEEE Int. Conf. Multimedia and Expo, pp. 853–856 (July 2006)
[20] Lumini, A., Maio, D.: Adaptive positioning of a visible watermark in a digital image. In:
Proc. Int. Conf. Multimedia and Expo, Taipei, Taiwan, R.O.C., pp. 967–970 (June 2004)
[21] Li, X., Orchard, M.T.: Edge-directed prediction for lossless compression of natural
images. IEEE Trans. Image Process. 10(6), 813–817 (2001)
[22] Liu, T.-Y., Tsai, W.-H.: Generic Lossless Visible Watermarking—A New Approach.
IEEE Transactions on Image Processing 19(5) (May 2010)
[23] Yeung, M.M., et al.: Digital Watermarking for High- Quality Imaging. In: Proc. of IEEE
First Workshop on Multimedia Signal Processing, Princeton, NJ, pp. 357–362 (June
1997)
[24] Mintzer, F., et al.: Effective and Ineffective Digital Watermarks. In: Proc. of IEEE
International Conference on Image Processing, ICIP 1997, vol. 3, pp. 9–12 (1997)
 Evolving Test Patterns for Use Case Maps

P.G. Sapna1 , Hrushikesha Mohanty2 , and Arunkumar Balakrishnan1

1
Coimbatore Institute of Technology, Coimbatore, India
2
University of Hyderabad, Hyderabad, India
[email protected], {mohanty.hcu,arunkumar.cta.cit}@gmail.com

Abstract. An important aspect of software testing today is the abil-

ity to seamlessly move from analysis models to generating test scenarios
and test cases. Use Case Maps(UCMs) are a scenario-based software en-
gineering technique used at early stages of software development. The
notation is most suited to capturing and eliciting use cases as well as
high-level architectural design and for test case generation. Test pat-
terns provide established solutions for designing tests or for supporting
the testing process. In this work, we define test patterns based on the
the primitives of the Use Case Maps to aid the process of test scenario
generation.

1 Introduction
There is a lot of effort to standardize the User Requirements Notation(URN).
Use Case Maps(UCMs) are used in this direction as it is a scenario based notation
that helps in capturing functional requirements of a system. Also, UCMS bridge
the modeling gap between requirements(represented as use cases) and design by
linking behaviour and structure in an explicit way. UCMs elaborate one or more
paths describing the causal relationships between responsibilities of one or more
use cases i.e. a UCM shows the progression of scenarios along use cases. Thus,
UCMs integrate many scenarios enabling reasoning about interactions between
scenarios.
A pattern is a named description of a problem and solution that can be
applied to new contexts. It suggests how to apply in varying novel circumstances,
considering the trade-offs [6]. Patterns exists that address design and architecture
issues. Many well-known patterns address design, architecture, or process issues.
Testing patterns help give solutions for designing tests. For example, Binder [4]
suggests a collection of test design patterns for object-oriented systems. Patterns
for artifacts like classes, methods and scenarios as well as for test automation
and test oracles are discussed. Testing patterns help bridge the divide between
manual test generation and automated test case generation.
Individual scenarios help in the understanding of functionalities as well as
guide the definition of details scenarios for design and testing phase. Amyot and
Eberlein [7] in their work handle the issue of extracting individual scenarios from
UCM specifications thereby enabling the validation of requirements and facili-
tating the transition from the requirements to the design phase. The objective of

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 65–74, 2011.

c Springer-Verlag Berlin Heidelberg 2011
66 P.G. Sapna, H. Mohanty, and A. Balakrishnan

this work is to develop test patterns from Use Case Map patterns. The number
of patterns and their unstructured nature(e.g. OR-forks and OR-joins in UCM
need not be nested) require that test patterns be defined to aid in the process
of automated test case generation. For this, we consider that UCMs consisting
of multiple scenarios embedded in one have been converted to individual sce-
narios. The work of [2] looks at building test patterns for basic UCM patterns,
namely, alternatives, concurrency, loops and stubs. Mussbacher’s [5] work uses
new notational elements and defines 18 additional patterns of UCMs in addition
to work in [2]. This work studies [5] and aims at building test patterns using
basic testing techniques, namely, equivalence class partitioning, boundary value
testing, loop testing and orthogonal array testing techniques.
Section 2 introduces UCMS and discusses related work in the area. Section 3
details the approach used and then discusses the UCM patterns. Test patterns
corresponding to each UCM patterns is given. Section 4 gives concludes the work
and identifies areas for future work.

2 Survey
2.1 Use Case Maps
Use Case Maps (UCMs) are used as a visual notation for describing causal rela-
tionships between responsibilities of one or more use cases [1]. Responsibilities
relate to processing like actions, activities, operations and tasks. Software enti-
ties like objects, processes and databases are represented using components. The
relationships in UCMs are called causal as they link causes represented using pre-
conditions and triggering events to effects represented using postconditions and

Fig. 1. Basic Elements of UCM Notation

 Evolving Test Patterns for Use Case Maps 67

Fig. 2. Example Use Case Map

resulting events. Thus, UCMs show use cases and use case maps scenarios model
functional requirements of a system.
The basic elements of the UCM notation are shown in Figure 1. A map
can contain any number of paths and structural elements(components). Ac-
tions to fulfil a scenario are represented using responsibilities. Paths show causal
sequences. Alternatives are represented using OR-forks and OR-joins whereas
concurrency is represented using AND-forks and AND-joins. Loops can be mod-
eled using OR-forks and OR-joins. UCM models can be decomposed using stubs
which contain sub-maps called plug-ins, which are reusable units of behaviour
and structure. There are two types of stubs: static and dynamic. A static stub
can have atmost one plug-in, whereas a dynamic stub may have many plug-ins
that are selected during runtime. The choice of a plug-in in the case of dynamic
stub is dependent on a selection policy. Figure 2 shows a sample use case map.

2.2 Defining Scenarios Using UCM

UCMs are useful for describing multiple scenarios abstractly in a single view. The
advantage of such a representation is that it helps understanding and reasoning
about the system as a whole as well as early detection of conflicting or inconsis-
tent scenarios [7]. However, this same advantage causes difficulty in extracting
individual scenarios.
Scenarios begin at the start point representing some precondition and trigger-
ing event and finish at end points, which describe postconditions and resulting
events [8].
For example, the scenarios extracted from the sample UCM in Figure 2 in-
clude:
1. <start point, responsibility, alternate branch1, parallel branch1, parallel branch2,
end point>
2. <start point, responsibility, alternate branch2, parallel branch1, parallel branch2,
end point>
3. <start point, responsibility, alternate branch1, parallel branch2, parallel branch1,
end point>
4. <start point, responsibility, alternate branch2, parallel branch2, parallel branch1,
end point>
68 P.G. Sapna, H. Mohanty, and A. Balakrishnan

2.3 Related Work

Amyot et al [1] handle the problem of automated scenario generation from UCMs
in their work. Scenario definitions and traversal algorithms are used to extract
individual scenarios from UCMs, which are stored as XML files. Transformation
to UML diagrams(viz. sequence diagrams) is achieved through the use of XSLT.
Roy et al [10] in their work built a new Eclipse-based editing tool for UCMs,
jUCMNav. The tool helps in creating, maintaining, analyzing and transform-
ing UCM models. Kealey et al [11] in their work propose enhancement to the
jUCMNav Eclipse plug-in by introducing new scenario traversal semantics as well
as transformation to Message Sequence Charts(MSC). The objective is to use
UCMs for scenario generation as well as translating UCMs to design diagrams
like MSC.
Mussbacher [5] study the applicability of UCMs for workflow description with
reference to scenarios and study how to evolve the UCM notation. They com-
pare with standards for workflow, business process design and business process
execution languages like Business Process Modeling Notation(BPMN), Business
Process Execution Language for Web Services(BPEL4WS) and UML Activity
diagrams. With the simple traversal mechanism as the basis, Mussbacher shows
that UCMs support only eight of the possible twenty one workfloww patterns
which is relatively lesser than for the other notations. However, with new nota-
tional elements and a new traversal mechanism, the number of patterns UCM
supports raises to eighteen. In our work, we concentrate our efforts on developing
test patterns for the UCM patterns defined in [5].
In [4], testing patterns are developed that target the coverage of scenarios de-
scribed in terms of Use Case Maps. The objective is to cover functional scenarios
at different levels of completeness, like all results, all causes and all results, all
path segments, all end-to-end paths, all plug-ins. The idea is that covering paths
in the UCMs leads to coverage of events and responsibilities that together form
scenarios. The basis for the patterns is the use of white-box techniques for test
case generation applied at the level of scenarios.

3 Test Patterns for Use Case Maps

3.1 The Approach
UCMs as mentioned before, is a scenario based notation that helps in capturing
functional requirements of a system. The number of scenarios generated for even
a small system is very large which makes exhaustive testing impossible. Hence,
there is a need to test software in an effective way considering cost, time, coverage
and customer requirements. Path testing is considered as a criteria to determine
adequacy of testing with reference to scenarios. The number of paths tested can
be used as a measure to determine the level of testing. Questions that can be
asked are [9]:
 Evolving Test Patterns for Use Case Maps 69

– What is the maximum number of different paths possible ?

– What is the fewest number of paths possible ?
– How many different paths can be deduced ?
Also, equivalence class partitioning and boundary value analysis is used in case
of preconditions and loops respectively. In this work, for each UCM pattern,
we define corresponding test patterns. The patterns taken into consideration
include:
a. Basic Control patterns - includes, sequence, parallel split, synchronization,
exclusive choice and simple merge
b. Advanced branching and synchronization patterns - includes, multiple choice,
synchronizing merge, multiple merge.

3.2 Basic Control Patterns

Consider two segments ’a’ and ’b’. Here, a path that consists of two successive
path segments is expressed by the path product(a*b) of the segment names. In
case of parallel paths between two nodes, path sum is taken(a+b). In case of
loops, an infinite sum is denoted by b* and a loop taken atleast once is denoted
by b+ . Here, we look at developing test cases for basic control patterns. Test
cases for alternatives, concurrency, loops, multiple start points, single stub and
plugins and causal stubs have been discussed in [5]. In this work, we look at
obtaining test cases from other patterns found in UCMs.
Sequence: Figure 3a shows the sequence pattern consisting of a sequence of
responsibilities from start to end. The test case is a path product of all the se-
quences contributing to the path.

Parallel Split: Figure 3b indicates that activities can be performed in paral-

lel or in any order. An AND-fork is used to represent parallel split. Hence, we
consider that the activities may happen in sequence or parallel.

Synchronization: Figure 3c indicates a case where concurrent branches con-

verge at an AND-join. The assumption here is that each concurrent branch is
executed only once, and all incoming branches should be executed before control
moves from the AND-join forwards.

Exclusive Choice: Figure 3d shows an exclusive choice represented using an

OR-fork. Here, the branches have to be mutually exclusive.

Simple Merge: Figure 3e shows a case where branches merge together into
one represented using an OR-join.

3.3 Advanced Branching and Synchronization Patterns

Multiple Choice: Figure 4a shows a case where more than one alternative
branch can be selected and executed at the same time. Here, equivalence class
70 P.G. Sapna, H. Mohanty, and A. Balakrishnan

Fig. 3. Testing patterns for Basic control patterns

partitioning(ECP) is used to to partition the input domain, thereby eliciting all

possible input classes for testing.
Multiple Merge : Figure 4b shows the case where multiple branches are
merged without synchronization. The assumption is that activities following the
merge will be performed once for each active branch. Here, each path from the
start to the end is considered.

Static Stub: Figure 4c shows the case of static stub which contains only one
plug-in and therefore having hierarchical decomposition of the UCM.

Synchronizing Merge: Figure 4d shows the case of dynamic stubs represented

as dashed diamond containing several plug-ins. Here, the selection of the plug-in
can be determined at run-time based on the selection policy defined(otherwise
called pre-condition). The traversal waits until all plug-ins have completed their
 Evolving Test Patterns for Use Case Maps 71

Fig. 4. Testing patterns for Advanced control patterns

72 P.G. Sapna, H. Mohanty, and A. Balakrishnan

activities before continuing along the path after the stub. Here, ECP is used
taking into consideration valid and invalid input classes, thereby generating test
cases where at least one branch is executed and at most all branches are executed.

N-out-of-M Join: Figure 4e is a type of merge where multiple concurrent

incoming branches merge. Here, the continuation of the workflow happens when
nth out of m incoming branches complete and thereby trigger the continuation
of the activity. The other plug-ins may complete but do not trigger the continu-
ation of the workflow. In this case, equivalence class partitioning and boundary
value analysis is used. Equivalence class partitioning is used for defining the in-
put domain and boundary value analysis used to determine the bounds groups
for n out of m branches.

Discriminator: Figure 4f is a special case of n-out-of-m join and test cases

generated similarly.

3.4 Structural Patterns

Unstructured cycles: Figure 5 shows an unstructured loop. With UCMs it
is possible to create unstructured loops with OR-forks and OR-joins as nesting
of the same is not a requirement. Beizer [9] defines four difference classes of
loops and a methodology to design test cases for both simple and nested loops
which is followed in this work. In cases of unstructured loops, Beizer recommends
that unstructured loops be redesigned as structured loops which is one solution.
Otherwise, in such a case, a condition that loops must be executed such that a
responsibility is executed a maximum of two times can be applied[Linzhang04].
For the UCM in Figure 4, a total of three test cases will ensure that all paths
are tested if the maximum condition is taken.

Fig. 5. Testing pattern for UCM Structural pattern

3.5 Patterns Involving Multiple Instances

Figure 6 shows a pattern involving multiple instances. Three such patterns have
been defined by [5]. However, in all these cases, boundary value analysis with
 Evolving Test Patterns for Use Case Maps 73

Fig. 6. Testing pattern for Multiple Instances

reference to the number of times the instances need to be executed is adopted.

Valid test cases for ends of the range and invalid test cases for just above and
below the range is included.

4 Conclusions and Future Work

The focus of this paper was to create scenarios for testing based on Use Case
Maps. Use Case Maps have the advantage of integrating scenarios in a single
view which helps designers and test engineers to understand and analyse re-
quirements better than when they are represented individually. However, given
the varied patterns possible when using UCMs, there is need to formulate pro-
cedures for testing. Currently, UCMs support only eight workflow patterns for
which test scenarios have been defined. These patterns are basic in nature. This
work looks at additional patterns defined in [5] and defines test patterns for the
same. Future work involves building test patterns for ’communication patterns’
defined in [5] which is not handled in this work as well as looking at a reduction
procedure to reduce the number of scenarios generated.

References
1. Amyot, D.: Specification and Validation of Telecommunications Systems with
Use Case Maps and LOTOS, Ph.D. thesis, SITE, University of Ottawa, Canada
(September 2001)
2. Amyot, D., Logrippo, L., Weiss, M.: Generation of test purposes from Use Case
Maps. Computer Networks (2005)
3. Wang, L., Yuan, J., Yu, X., Hu, J., Li, X., Zheng, G.: Generating Test Cases from
UML Activity Diagram based on Gray-Box Method. In: Proceedings of the 11th
Asia-Pacific Software Engineering Conference (APSEC). IEEE Computer Society,
Los Alamitos (2004)
4. Binder, R.V.: Testing Object-Oriented Systems Models, Patterns, and Tools. Ad-
dison Wesley, Reading (1999)
74 P.G. Sapna, H. Mohanty, and A. Balakrishnan

5. Mussbacher, G.: Evolving Use Case Maps as a Scenario and Workflow Description
Language. In: 10th Workshop of Requirement Engineering (WER 2007), pp. 56–67
(2007)
6. Larman, C.: Applying UML and Patterns: An Introduction to Object-Oriented
Analysis and Design and the Unified Process, 2nd edn. Prentice Hall, Englewood
Cliffs (2001)
7. Amyot, D., Eberlein, A.: An Evaluation of Scenario Notations and Construction
Approaches for Telecommunication Systems Development. Telecommunications
Systems Journal 24(1), 61–94 (2003)
8. Amyot, D., He, X., He, Y., Cho, D.Y.: Generating Scenarios from Use Case Map
Specifications. In: 3rd International Conference on Quality Software (QSIC), pp.
108–115 (2003)
9. Beizer, B.: Software Testing Techniques. Dreamtech Press (2002)
10. Roy, J.-F., Kealey, J., Amyot, D.: Towards Integrated Tool Support for the User Re-
quirements Notation. In: Gotzhein, R., Reed, R. (eds.) SAM 2006. LNCS, vol. 4320,
pp. 198–215. Springer, Heidelberg (2006)
11. Kealey, J., Amyot, D.: Enhanced Use Case Map Traversal Semantics. In: Gaudin,
E., Najm, E., Reed, R. (eds.) SDL 2007. LNCS, vol. 4745, pp. 133–149. Springer,
Heidelberg (2007)
 A Novel Approach for Compression of Encrypted
Grayscale Images Using Huffman Coding

S. Saravana Kumar and S. Panimalar

Manonmaniam Sundaranar University, Tirunelveli

[email protected],
[email protected]

Abstract. Currently information security is becoming more important in data

storage and transmission. Images are broadly used in several processes.
Therefore, the protection of image data from unauthorized access is essential.
Image encryption plays a major role in the field of information hiding. The
objective of this paper is to provide an efficient coding technique for compress
the encrypted gray scale images without any loss. This is made possible by
using Huffman coding. This paper uses resolution-progressive compression
(RPC) technique to achieve better result. RPC compresses an encrypted image
progressively in resolution, such that the decoder can observe a low-resolution
version of the image, study local statistics based on it, and use the statistics to
decode the next resolution level. The experimental results show that the method
used in this paper for compression gives good result.

Keywords: Compression of encrypted images, resolution progressive

compression, Huffman coding.

1 Introduction
In many convenient scenarios multimedia contents need both to be compressed and
protected. Whereas the conventional way to compress and protect data requires that
data are first compressed and then encrypted, in some cases it may be desirable to
operate the other way around. This is the case, for example of an encrypted content
that is transmitted without compression through the internet, but at a certain point
needs to be compressed in order to match the characteristics of the transmission link.
If the node is not trusted it may not be provided with the encryption key, then
compression of the encrypted content is required.
In the literature, solutions have been proposed to treat image and video data as
Markov sources and to exploit the Markovian property in the Slepian-Wolf decoder
[1], [12].A similarwork is also found in [13] for nonencrypted colored sources. Some
good results have been reported for binary images. However, there are some
limitations with this approach: first, Markov decoding in a Slepian-Wolf decoder is
expensive, especially in dealing with sources with nonbinary alphabets; second, bit-
plane based Markov decoding certainly reduces the complexity, but the source
dependency that originally defined in the symbol domain is usually not fully utilized
when translated to bit-planes; third, since image and video data are known to be

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 75–83, 2011.
© Springer-Verlag Berlin Heidelberg 2011
76 S. Saravana Kumar and
a S. Panimalar

highly nonstationary, a gllobal Markov model cannot describe its local statisstics
precisely. As reported in [12],
[ for 8-bit grayscale images, only the first two m most
significant bit-planes (MSB B) are compressible by employing a 2-D Markov modeel in
bit-planes. The challengin ng issue still remains is, how to effectively exploit the
correlation of encrypted image data.
In this paper, lossless co
ompression techniques are discussed and experimentedd on
encrypted gray scale imag ges. This paper provides an efficient way to comprress
encrypted images through resolution-progressive compression (RPC). The encooder
starts by sending a down nsampled version of the ciphertext. At the decoder, the
corresponding low-resolutio on image is decoded and decrypted, from which a highher-
resolution image is obtaineed by intraframe prediction. The predicted image, togetther
with the secret encryption key,
k is used as the side information (SI) to decode the nnext
resolution level. This proceess is iterated until the whole image is decoded. Huffm man
technique is used as coder and
a decoder.
The rest of paper is orgaanized as follows. The next section II briefly describes the
resolution-progressive com mpression technique. Section III describes the conttext
adaptive technique. Section n IV describes the Huffman coding technique. Sectionn V
briefly describes the comp pression algorithm by step wise. Experimental setupp is
discussed in Section V.Finaally the conclusion is presented in Section VI.

2 Resolution Progresssive Compression of Encrypted Images

The input grayscale imag ge I is encrypted by using encryption algorithm. A After
applying the encryption alg gorithm the input image is converted into the ciphertextt Y.
This is shown in below Fig..1.
The encoder takes the ciphertext
c Y and decomposes it into four subimages. IIt is
called as 00, 01, 10, and 11
1 sub-images. Each sub-image is a downsampled-by-ttwo
version of the encrypted im mage Y. The name of a sub-image denotes the horizonntal
and vertical offsets of the downsampling.
d The 00 sub-image is further downsamppled
to create multiple resolutionn levels. An example of the decomposition is illustratedd in
Fig. 2.

(a) (b)

Fig. 1. Encryptio
on Process (a) Pepper Image (b) Encrypted Image Y
 A Novel Approach for Compression of Encrypted Grayscale Images 77

(a) (b)
Fig. 2. (a) Illustration of a three-level decomposition of the unencrypted “Pepper” image. (b)
Layout of the sub-images.

After the downsampling, each sub-image is encoded independently using Huffman

codes, and the resulting bits are transmitted from the lowest resolution to the highest.
Decoding starts from the 00 sub-image of the lowest-resolution level, say, level 00n.
This paper suggests transmitting the uncompressed 00n sub-image as the doped bits].
Thus, the 00n sub-image can be known by the decoder without ambiguity, and
knowledge about the local statistics will be derived based on it. Next, other sub-images
of the same resolution level are interpolated from the decrypted 00n sub-image. Then the
Context Adaptive Interpolation technique is applied for SI generation.

3 Context Adaptive Interpolation

The SI is generated through interpolation. For simplicity, for any pixel in the target
sub-image, this technique only uses the four horizontal and vertical neighbors or the
four diagonal neighbors in the known sub-images for the interpolation. For this
purpose a simple and effective context adaptive interpolator (CAI) technique is used
in this paper. Due to the isomorphism, this technique only describes the horizontal-
vertical interpolator illustrated in Fig. 3.

Fig. 3. Illustration of the two-step interpolation at the decoder side. The dashed arrows denote
the first step interpolation, and the solid arrows denote the second step.
78 S. Saravana Kumar and S. Panimalar

Let s be the pixel value to be interpolated, t=[t1,t2,t3,t4]be the vector of neighboring

pixels. The interpolator classifies the local region into four types: smooth,
horizontally-edged, vertically-edged and other. In smooth regions, a mean filter is
applied; in horizontally/vertically edged regions, the interpolation is done along the
edge; otherwise use a median filter. CAI is formulated as
, max min 20
| | | | 20
(1)
| |– – | 20
, .

4 Huffman Encoding
We use this coding to encode/compress a file. Steps:
1. Read the file.
2. Calculate the probability of each symbol (since we use an ASCII file, there
are 256 possible symbols). Instead of actual probability we can use instance
count.
3. Use the Huffman coding algorithm to find the coding
for each symbol. Need to apply only on subset of symbols that actually
appear in the file.
4. Encode the file, and write it out.
The out file must also include the encoding table, so as to permit decoding.

5 Compression Algorithm
1. Get the input gray scale image
2. Apply the encryption algorithm on to the input image to get the cipher text
3. Then apply the downsampling technique on to the cipher text

001,011,101,111,…,
Original Cipher 00n,01n,10n,11n
Image Image Y
Encrypt the Image Down Sampling

Huffman Coding

Context Adaptive
Compressed Interpolation
Image
Decrypt the Image

Fig. 4. Overall Process Flow Diagram of Compression of Encrypted Gray Scale Images
 A Novel Approach for Compression of Encrypted Grayscale Images 79

4. Apply the Huffman cod ding algorithm on to the each downsampled images.
5. Then the context adapttive interpolation technique to create the Side Informatioon
6. Then apply the decrypttion algorithm to get the compressed image.

The overall process flow diagram is shown in Fig.4

6 Expriements and Performance

P Analysis
The images Pepper, Lena, Barbara
B and Plane and are used for the experiments. Thiis is
shown in Fig.5.

(a) (b)

(c) (d)

Fig. 5. Experimen
ntal Images (a) Pepper (b) Lena (c) Barbara (d) Plane

First the experimental im

mages are encrypted to create the cipher text. The encryppted
images of the experimental images are shown in Fig.6. Then the encrypted images are
downsampled to create the sub images. Then the each sub images are encoded ussing
the Huffman techniques.
Then the context adaptivve interpolation technique is used for predicts the unknoown
sub images. Finally the inteerpolated images are decrypt to get the compressed imaage.
The decrypted images are shown in Fig.7.
80 S. Saravana Kumar and
a S. Panimalar

Fig. 6. Encrypteed Images (a) Pepper (b) Lena (c) Barbara (d) Plane

Fig. 7. Decrypteed Images (a) Pepper (b) Lena (c) Barbara (d) Plane

A. Performance Analysis
The Peak-Signal-to-Noise--Ratio (PSNR) and Compression Ratio (CR) are ttwo
metrics used for analysis thhe performance of the above algorithm. The PSNR and CR
values for the images are taabulated in Table1.
PSNR means Peak Signal to Noise Ratio (PSNR). This performance metric giives
the signal to noise ratio between the original image and the compressed image. PS
SNR
is calculated by using the fo
ollowing formula.
 A Novel Approach for Compression of Encrypted Grayscale Images 81

2 1
20

∑ 2

CR means Compression Ratio. This performance metric gives compression ratio. This
ratio gives an indication of how much compression is achieved for a particular image.
The PSNR & CR analysis of experimental images is shown in Fig.8.

Table 1. PSNR & CR Value For expirental images

Image PSNR CR
Lena 34.468 33.33
Pepper 35.046 34.44
Boat 32.848 33.14

Barbara 33.235 32.15

PSNR & CR Analysis

35.5
35
34.5
PSNR & CR Values

34
33.5
PSNR
33
CR
32.5
32
31.5
31
30.5
Lena Pepper Boat Barbara
Images

Fig. 8. PSNR & CR Analysis of Experimental Images

7 Conclusions
This paper provides an efficient coding technique for compress the encrypted gray
scale images without any loss. This is made possible by using Huffman coding. This
paper uses resolution-progressive compression (RPC) technique to achieve better
result. RPC compresses an encrypted image progressively in resolution, such that the
decoder can observe a low-resolution version of the image, study local statistics based
on it, and use the statistics to decode the next resolution level. The experimental
results show that the method used in this paper for compression gives good result.
82 S. Saravana Kumar and S. Panimalar

References
[1] Johnson, M., Ishwar, P., Prabhakaran, V.M., Schonberg, D., Ramchandran, K.:
On compressing encrypted data. IEEE Trans. Signal Process. 52(10), 2992–3006
(2004)
[2] Stallings, W.: Cryptography and Network Security: Principles and Practice, 3rd edn.
Prentice-Hall, Upper Saddle River (2003)
[3] Pradhan, S.S., Ramchandran, K.: Distributed source coding using syndromes (DISCUS).
IEEE Trans. Inf. Theory 49(3), 626–643 (2003)
[4] Slepian, J.D., Wolf, J.K.: Noiseless coding of correlated information sources. IEEE Trans.
Inf. Theory IT-19, 471–480 (1973)
[5] García-Frías, J., Zhao, Y.: Compression of correlated binary sources using turbo codes.
IEEE Commun. Lett. 5(10), 417–419 (2001)
[6] Bajcsy, J., Mitran, P.: Coding for the Slepian-Wolf problem with turbo codes. In: Proc.
IEEE Global Telecommun. Conf., San Antonio, TX, pp. 1400–1404 (November 2001)
[7] Aaron, A., Girod, B.: Compression with side information using turbo codes. In: Proc.
IEEE Data Compression Conf., Snowbird, UT, pp. 252–261 (April 2002)
[8] Liveris, A., Xiong, Z., Georghiades, C.: Compression of binary sources with side
information at the decoder using LDPC codes. IEEE Commun. Lett. 6(10), 440–442
(2002)
[9] Varodayan, D., Aaron, A., Girod, B.: Rate-adaptive distributed source coding using low-
density parity-check codes. In: Proc. Asilomar Conf. Signals, Systems and Computers,
Pacific Grove, CA, pp. 1203–1207 (November 2005)
[10] Xiong, Z., Liveris, A., Cheng, S.: Distributed source coding for sensor networks. IEEE
Signal Process. Mag. 21, 80–94 (2004)
[11] Yang, Y., Stankovic, V., Xiong, Z.: Image encryption and data hiding: Duality and code
designs. In: Proc. Inf. Theory Workshop, Lake Tahoe, CA, pp. 295–300 (September
2007)
[12] Schonberg, D.: Practical Distributed Source Coding and its Application to the
Compression of Encrypted Data. Ph.D. dissertation, Univ. California, Berkeley (2007)
[13] Varodayan, D., Aaron, A., Girod, B.: Exploiting spatial correlation in pixel-domain
distributed image compression. In: Proc. Picture Coding Symposium, Beijing, China
(April 2006)
[14] Weinberger, M., Seroussi, G., Sapiro, G.: The LOCO-I lossless image compression
algorithm: principles and standardization into JPEG-LS. IEEE Trans. Image Process. 9(8),
1309–1324 (2000)
[15] Wu, X., Memon, N.: Context-based adaptive lossless image coding. IEEE Trans.
Commun. 45(4), 437–444 (1997)
[16] Aaron, A., Rane, S., Setton, E., Girod, B.: Transform-domain Wyner-Ziv codec for video.
In: Proc. SPIE Visual Commun. Image Process., San Jose, CA, pp. 520–528 (January
2004)
[17] Berger, T.: Rate Distortion Theory: A Mathematical Basis for Data Compression.
Prentice-Hall, Upper Saddle River (1971)
[18] Gersho, A., Gray, R.: Vector Quantization and Signal Compression. Kluwer, Boston
(1992)
[19] Jayant, N., Noll, P.: Digital Coding of Waveforms: Principles and Applications to Speech
and Video. Prentice-Hall, Upper Saddle River (1984)
 A Novel Approach for Compression of Encrypted Grayscale Images 83

[20] Rowitch, D.N., Milstein, L.B.: On the performance of hybrid FEC/ARQ systems using
rate-compatible punctured turbo (RCPT) codes. IEEE Trans. Commun. 48, 948–959
(2000)
[21] Yao, Q., Zeng, W., Liu, W.: Multi-resolution based hybrid spatiotemporal compression of
encrypted videos. In: Proc. IEEE Int. Conf. Acous., Speech and Sig. Process., Taipei,
Taiwan, R.O.C., pp. 725–728 (April 2009)
[22] Liu, W., Zeng, W., Dong, L., Yao, Q.: Efficient Compression of Encrypted Grayscale
Images. IEEE Transactions on Image Processing 19(4) (April 2010)
 Global Chaos Synchronization of Hyperchaotic Pang
and Wang Systems by Active Nonlinear Control

Sundarapandian Vaidyanathan1 and Karthikeyan Rajagopal2

1
R & D Centre, Vel Tech Dr. RR & Dr. SR Technical University
Avadi-Alamathi Road, Avadi, Chennai-600 062, India
[email protected]
http://www.vel-tech.org/
2
School of Electronics and Electrical Engineering, Singhania University
Dist. Jhunjhunu, Rajasthan-333 515, India
[email protected]
http://www.singhaniauniversity.co.in/

Abstract. This paper investigates the global chaos synchronization of identical

hyperchaotic Pang systems (Pang and Liu, 2011) and synchronization of non-
identical hyperchaotic Pang system and Wang system (Wang and Liu, 2006). Ac-
tive nonlinear feedback control is the method used to achieve the synchronization
of the identical and different hyperchaotic Pang and Wang systems addressed in
this paper and our results are established using Lyapunov stability theory. Since
the Lyapunov exponents are not required for these calculations, the active control
method is effective and convenient to synchronize identical and different hyper-
chaotic Pang and Wang systems. Numerical simulations are given to illustrate
the effectiveness of the proposed synchronization schemes for the global chaos
synchronization of hyperchaotic systems addressed in this paper.

Keywords: Active control, chaos synchronization, nonlinear control, Pang sys-

tem, Wang system.

1 Introduction

Chaos theory has been developed and extensively studied over the past two decades.
A chaotic system is a nonlinear dynamical system that is highly sensitive to initial
conditions. This sensitivity is popularly referred to as the butterfly effect [1]. There has
been active research done in the past three decades on the control and synchronization
of nonlinear chaotic systems.
Chaos synchronization problem was first described by Fujisaka and Yemada [2] in
1983. This problem did not receive great attention until Pecora and Carroll ([3]-[4])
published their results on chaos synchronization in early 1990s. From then on, chaos
synchronization has been extensively and intensively studied in the last three decades
([3]-[21]). Chaos theory has been explored in a variety of fields including physical [5],
chemical [6], ecological [7] systems, secure communications ([8]-[10]) etc.
In most of the chaos synchronization approaches, the master-slave or drive-response
formalism is used. If a particular chaotic system is called the master or drive system

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 84–93, 2011.

c Springer-Verlag Berlin Heidelberg 2011
 Global Chaos Synchronization of Hyperchaotic Pang and Wang Systems 85

and another chaotic system is called the slave or response system, then the idea of the
synchronization is to use the output of the master system to control the slave system so
that the output of the slave system tracks the output of the master system asymptotically.
Since the seminal work by Pecora and Carroll ([3]-[4]), a variety of impressive ap-
proaches have been proposed for the synchronization for the chaotic systems such as
PC method ([3]-[4]), the sampled-data feedback synchronization method ([10]-[11]),
OGY method [12], time-delay feedback approach [13], backstepping design method
[14], adaptive design method ([15]-[18]), sliding mode control method [19], active con-
trol method ([20]-[21]), etc.
In this paper, we derive new results for the global chaos synchronization for iden-
tical and different hyperchaotic Pang and Wang systems. Explicitly, using the active
nonlinear control method, we achieve global chaos synchronization for identical hy-
perchaotic Pang systems ([22], 2011) and non-identical hyperchaotic Pang system and
Wang system ([23], 2006).
This paper has been organized as follows. In Section 2, we give the problem state-
ment and our methodology. In Section 3, we derive results for the chaos synchronization
of two identical hyperchaotic Pang systems. In Section 4, we discuss the synchroniza-
tion of non-identical hyperchaotic Pang and Wang systems. In Section 5, we present the
conclusions of this paper.

2 Problem Statement and Our Methodology

Consider the chaotic system described by the dynamics
ẋ = Ax + f (x) (1)
where x ∈ IRn is the state of the system, A is the n × n matrix of the system parameters
and f is the nonlinear part of the system. We consider the system (1) as the master or
drive system.
As the slave or response system, we consider the following chaotic system described
by the dynamics
ẏ = By + g(y) + u (2)
where y ∈ IRn is the state of the slave system, B is the n × n matrix of the system
parameters, g is the nonlinear part of the system and u is the controller of the slave
system.
If A = B and f = g, then x and y are the states of two identical chaotic systems. If
A = B or f = g, then x and y are the states of two different chaotic systems. In the ac-
tive nonlinear control approach, we design a feedback controller u which synchronizes
the states of the master system (1) and the slave system (2) for all initial conditions
x(0), y(0) ∈ IRn .
If we define the synchronization error as
e = y − x, (3)
then the synchronization error dynamics is obtained as
ė = By − Ax + g(y) − f (x) + u (4)
86 S. Vaidyanathan and K. Rajagopal

Thus, the global synchronization problem is essentially to find a feedback controller

(active control) u so as to stabilize the error dynamics (4) for all initial conditions, i.e.

lim e(t) = 0, ∀ e(0) ∈ IRn (5)

t→∞

We use the Lyapunov stability theory as our methodology. We take as a candidate Lya-
punov function
V (e) = eT P e,
where P is a positive definite matrix. Note that V is a positive definite function by
construction. We assume that the parameters of the master and slave systems are known
and that the states of both systems (1) and (2) are available for measurement.
If we find a feedback controller u so that

V̇ (e) = −eT Qe

where Q is a positive definite matrix, then V is a negative definite function on IRn .

Thus, by Lyapunov stability theory [24], the error dynamics (4) is globally exponen-
tially stable and hence the states of the master system (1) and slave system (2) are
globally exponentially synchronized.

3 Synchronization of Identical Hyperchaotic Pang Systems

In this section, we apply the active nonlinear control method for the synchronization of
two identical hyperchaotic Pang systems ([22], 2011).
Thus, the master system is described by the Pang dynamics

ẋ1 = a(x2 − x1 )
ẋ2 = cx2 − x1 x3 + x4
(6)
ẋ3 = −bx3 + x1 x2
ẋ4 = −k1 x1 − k2 x2

where x1 , x2 , x3 , x4 are the state variables and a, b, c, d, k1 , k2 are positive real con-
stants.
When a = 36, b = 3, c = 20, k1 = 2 and k2 = 2, the Pang system (6) is hyper-
chaotic as illustrated in Figure 1.
The slave system is also described by the Pang dynamics

ẏ1 = a(y2 − y1 ) + u1
ẏ2 = cy2 − y1 y3 + y4 + u2
(7)
ẏ3 = −by3 + y1 y2 + u3
ẏ4 = −k1 y1 − k2 y2 + u4

where y1 , y2 , y3 , y4 are the state variables and u1 , u2 , u3 , u4 are the nonlinear con-
trollers to be designed.
 Global Chaos Synchronization of Hyperchaotic Pang and Wang Systems 87

60 20

40 0
x3

x4
20 −20

0 −40
50 50
50 50
0 0
x2 0 0
x x2
−50 −50 1 −50 −50 x
1

20 20

0 0
x4

−20 x4
−20

−40 −40
50 50
50 50
0 0
0 −50 x x3 0 −50 x
x3 1 2

Fig. 1. The Hyperchaotic Pang System

The synchronization error is defined by

ei = yi − xi , (i = 1, 2, 3, 4) (8)
A simple calculation yields the error dynamics as
ė1 = a(e2 − e1 ) + u1
ė2 = ce2 + e4 − (y1 y3 − x1 x3 ) + u2
(9)
ė3 = −be3 + y1 y2 − x1 x2 + u3
ė4 = −k1 e1 − k2 e2 + u4
We choose the nonlinear controller as
u1 = −ae2
u2 = −(c + 1)e2 − e4 + y1 y3 − x1 x3
(10)
u3 = −(y1 y2 − x1 x2 )
u4 = k1 e1 + k2 e2 − e4
Substituting the controller u defined by (10) into (9), we get
ė1 = −ae1 , ė2 = −e2 , ė3 = −be3 , e˙4 = −e4 (11)
We consider the candidate Lyapunov function
1 1  2 
V (e) = eT e = e1 + e22 + e23 + e24 (12)
2 2
which is a positive definite function on IR4 .
88 S. Vaidyanathan and K. Rajagopal

Differentiating V along the trajectories of (11), we find that

V̇ (e) = −ae21 − e22 − be23 − e24 (13)

which is a negative definite function on IR4 since a and b are positive constants.
Thus, by Lyapunov stability theory [24], the error dynamics (11) is globally expo-
nentially stable. Hence, we have proved the following result.

Theorem 1. The identical hyperchaotic Pang systems (6) and (7) are exponentially and
globally synchronized with the active nonlinear controller u defined by (10).

Numerical Results. For the numerical simulations, the fourth order Runge-Kutta
method with time-step 10−6 is used to solve the two systems of differential equations
(6) and (7). We take the parameter values as a = 36, b = 3, c = 20, k1 = 2, k2 = 2
so that the two Pang systems (6) and (7) are hyperchaotic. The active controller u is
defined by (10).
The initial values of the master system (6) are taken as

x1 (0) = 18, x2 (0) = 12, x3 (0) = 20, x4 = 10

and the initial values of the slave system (7) are taken as

y1 (0) = 26, y2 (0) = 30, y3 (0) = 10, y4 = 20.

Figure 2 shows that synchronization between the states of the master system (6) and
the slave system (7) occur in 4 seconds.

40
y1
20

0 x1
−20
0 1 2 3 4 5 6 7 8
50
y
2

0
x2

−50
0 1 2 3 4 5 6 7 8
40

x3
20
y3
0
0 0.5 1 1.5 2 2.5 3
40

20
y4

0 x4
−20
0 1 2 3 4 5 6
Time (sec)

Fig. 2. Synchronization of the Identical Hyperchaotic Pang Systems

 Global Chaos Synchronization of Hyperchaotic Pang and Wang Systems 89

4 Synchronization of Hyperchaotic Pang and Wang Systems

In this section, the nonlinear control method is applied for the synchronization of the
non-identical hyperchaotic systems, viz. Pang system (2011) and Wang system (2006).
The dynamics of the Wang system (2006) taken as the master system, is described
by
ẋ1 = α(x2 − x1 )
ẋ2 = βx1 − ηx1 x3 + x4
(14)
ẋ3 = −γx3 + x21
ẋ4 = −δx1
where x1 , x2 , x3 , x4 are the state variables and α, β, γ, δ, , η are positive real
constants.
When α = 10, β = 40, γ = 25, δ = 10.6,  = 4 and η = 1, the Wang dynamics (14)
is hyperchaotic as illustrated in Figure 3.

150 50

100
x4

0
x3

50

0 −50
50 50
20 20
x2 0 0 x2
0
0
−50 −20 x1 −50 −20 x1

50 50
x4

0 0
x4

−50 −50
200 200
20 50
100 100
x3 0 0
0 −20 x1 x 0 −50
3 x2

Fig. 3. The Hyperchaotic Wang System

90 S. Vaidyanathan and K. Rajagopal

The slave system is described by the Wang dynamics

ẏ1 = a(y2 − y1 ) + u1
ẏ2 = cy2 − y1 y3 + y4 + u2
(15)
ẏ3 = −by3 + y1 y2 + u3
ẏ4 = −k1 y1 − k2 y2 + u4

where y1 , y2 , y3 , y4 are the state variables and u1 , u2 , u3 , u4 are the nonlinear con-
trollers to be designed.
A simple calculation yields the error dynamics as

ė1 = a(e2 − e1 ) + (a − α)(x2 − x1 ) + u1

ė2 = ce2 + e4 + cx2 − βx1 − (y1 y3 − ηx1 x3 ) + u2
(16)
ė3 = −be3 + (γ − b)x3 + y1 y2 − x21 + u3
ė4 = −k1 e1 − k2 e2 + (δ − k1 )x1 − k2 x2 + u4

We choose the nonlinear controller as

u1 = −ae2 − (a − α)(x2 − x1 )
u2 = −(c + 1)e2 − e4 − cx2 + βx1 + (y1 y3 − ηx1 x3 )
(17)
u3 = −(γ − b)x3 − y1 y2 + x21
u4 = k1 e1 + k2 e2 − e4 − (δ − k1 )x1 + k2 x2

Substituting the controller u defined by (17) into (16), we get

ė1 = −ae1
ė2 = −e2
(18)
ė3 = −be3
e˙4 = −e4

We consider the candidate Lyapunov function

1 T 1  2 
V (e) = e e= e1 + e22 + e23 + e24 (19)
2 2
which is a positive definite function on IR4 .
Differentiating V along the trajectories of (18), we find that

V̇ (e) = −ae21 − e22 − be23 − e24 (20)

which is a negative definite function on IR4 since a and b are positive constants.
Thus, by Lyapunov stability theory [24], the error dynamics (18) is globally expo-
nentially stable. Hence, we have proved the following result.

Theorem 2. The non-identical hyperchaotic Wang system (14) and hyperchaotic Pang
system (15) are exponentially and globally synchronized with the active nonlinear con-
troller u defined by (17).
 Global Chaos Synchronization of Hyperchaotic Pang and Wang Systems 91

Numerical Results. For the numerical simulations, the fourth order Runge-Kutta method
with time-step 10−6 is used to solve the two systems of differential equations (14) and
(15) with the active nonlinear controller u defined by (17). The parameter values are
chosen so that the master and slave systems exhibit hyperchaotic behaviour, i.e.

a = 36, b = 3, c = 20, k1 = 2, k2 = 2

α = 10, β = 40, γ = 25, δ = 10.6,  = 4 and η = 1

The initial values of the master system (14) are taken as

x1 (0) = 15, x2 (0) = 20, x3 (0) = 10, x4 (0) = 40

and the initial values of the slave system (15) are taken as

y1 (0) = 25, y2 (0) = 10, y3 (0) = 26, y4 (0) = 18

Figure 4 shows that synchronization between the states of the master system (14)
and the slave system (15) occur in 3 seconds.

50
y
1

0 x1

−50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
100

x2
0
y2
−100
0 0.5 1 1.5 2 2.5 3 3.5 4
150

100
y3

50 x3

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
50
x
4
0
y4
−50
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (sec)

Fig. 4. Synchronization of the Hyperchaotic Wang and Pang Systems

92 S. Vaidyanathan and K. Rajagopal

5 Conclusions

In this paper, we have used nonlinear control method based on Lyapunov stability the-
ory to achieve global chaos synchronization for the identical hyperchaotic Pang systems
(2011) and non-identical hyperchaotic systems, viz. hyperchaotic Pang system (Pang
and Liu, 2011) and hyperchaotic Wang system (Wang and Liu, 2006). Numerical sim-
ulations are also given to validate all the synchronization results derived in this paper.
Since the Lyapunov exponents are not required for these calculations, the nonlinear con-
trol method is very effective and convenient to achieve global chaos synchronization for
the hyperchaotic systems addressed in this paper.

References
1. Alligood, K.T., Sauer, T., Yorke, J.A.: Chaos: An Introduction to Dynamical Systems.
Springer, New York (1997)
2. Fujikasa, H., Yamada, T.: Stability theory of synchronized motion in coupled-oscillator sys-
tems. Progr. Theoret. Phys. 69, 32–47 (1983)
3. Pecora, L.M., Carroll, T.L.: Synchronization in chaotic systems. Phys. Rev. Lett. 64, 821–824
(1990)
4. Pecora, L.M., Carroll, T.L.: Synchronizing chaotic circuits. IEEE Trans. Circ. Sys. 38,
453–456 (1991)
5. Lakshmanan, M., Murali, K.: Chaos in Nonlinear Oscillators: Controlling and Synchroniza-
tion. World Scientific, Singapore (1996)
6. Han, S.K., Kerrer, C., Kuramoto, Y.: Dephasing and burstling in coupled neural oscillators.
Phys. Rev. Lett. 75, 3190–3193 (1995)
7. Blasius, B., Huppert, A., Stone, L.: Complex dynamics and phase synchronization in spatially
extended ecological system. Nature 399, 354–359 (1999)
8. Kwok, H.S., Wallace, K., Tang, S., Man, K.F.: Online secure communication system using
chaotic map. Internat. J. Bifurcat. Chaos. 14, 285–292 (2004)
9. Kocarev, L., Parlitz, U.: General approach for chaos synchronization with applications to
communications. Phys. Rev. Lett. 74, 5028–5030 (1995)
10. Murali, K., Lakshmanan, M.: Secure communication using a compound signal using
sampled-data feedback. Applied Math. Mech. 11, 1309–1315 (2003)
11. Yang, T., Chua, L.O.: Control of chaos using sampled-data feedback control. Internat. J.
Bifurcat. Chaos. 9, 215–219 (1999)
12. Ott, E., Grebogi, C., Yorke, J.A.: Controlling chaos. Phys. Rev. Lett. 64, 1196–1199 (1990)
13. Park, J.H., Kwon, O.M.: A novel criterion for delayed feedback control of time-delay chaotic
systems. Chaos, Solit. Fract. 17, 709–716 (2003)
14. Wu, X., Lü, J.: Parameter identification and backstepping control of uncertain Lü system.
Chaos, Solit. Fract. 18, 721–729 (2003)
15. Samuel, B.: Adaptive synchronization between two different chaotic dynamical systems.
Adaptive Commun. Nonlinear Sci. Num. Simul. 12, 976–985 (2007)
16. Yu, Y.G., Zhang, S.C.: Adaptive backstepping synchronization of uncertain chaotic systems.
Chaos, Solit. Fract. 27, 1369–1375 (2006)
17. Park, J.H., Lee, S.M., Kwon, O.M.: Adaptive synchronization of Genesio-Tesi system via a
novel feedback control. Physics Lett. A. 371, 263–270 (2007)
18. Park, J.H.: Adaptive control for modified projective synchronization of a four-dimensional
chaotic sytem with uncertain parameters. J. Comput. Applied Math. 213, 288–293 (2008)
 Global Chaos Synchronization of Hyperchaotic Pang and Wang Systems 93

19. Yau, H.T.: Design of adaptive sliding mode controller for chaos synchronization with uncer-
tainties. Chaos, Solit. Fract. 22, 341–347 (2004)
20. Chen, H.K.: Global chaos synchronization of new chaotic systems via nonlinear control.
Chaos, Solit. Fract. 23, 1245–1251 (2005)
21. Sundarapandian, V., Suresh, R.: Global chaos synchronization for Rössler and Arneodo
chaotic systems. Far East J. Math. Sci. 44, 137–148 (2010)
22. Pang, S., Liu, Y.: A new hyperchaotic system from the Lü system and its control. J. Comput.
Applied Math. 235, 2775–2789 (2011)
23. Wang, F.Q., Liu, C.X.: Hyperchaos evolved from the Liu chaotic system. Chin. Phys. 15,
963–968 (2006)
24. Hahn, W.: The Stability of Motion. Springer, New York (1967)
 Mining Web Path Traversals Based on Generation of FP
Tree with Utility

E. Poovammal and Pillai Cigith

Department of Computer Science, SRM University,

Kattankulathur, Chennai 603203, Tamil Nadu, India
[email protected],[email protected]

Abstract. Web Mining is mining information from the Web. Web Usage
Mining extracts user accessing patterns and gives the desired web page quickly.
While accessing web, the user activities gets recorded in a web server log.
Using data mining algorithm, if mined properly, the web log acts as the gateway
to user’s interests. But, merely considering presence or absence of web page in
web log does not gauge its importance to user. Hence, in this work two
parameters namely frequency of the access pattern and utility value are used to
find better access patterns of the user and succeeded. This paper uses the
method of Frequent Pattern (FP) Tree generation to find frequency of access
pattern and the time spent by user on each web page is considered as the utility
value.

Keywords: FP Tree, Utility, Usage patterns, Web Usage Mining, frequent user
access patterns.

1 Introduction
The World Wide Web (WWW) is a huge galaxy of information which acts as a source
for rich and dynamic collection of hyperlink data with usage characteristics. Data
mining techniques are widely applied to obtain useful and understandable patterns
from these massive sources of data. All the data mining techniques exploited on the
Web comes under the category of Web mining. Web mining can be broadly classified
into three parts, i.e. content mining, usage mining, and link structure mining. Web
Content Mining describes the discovery of useful information from the web
content/data/documents such as text, image, audio and video. Web Structure Mining
is the process of discovering knowledge from the World Wide Web organization and
links across the webs. Web Usage Mining is the process of extracting interesting
patterns or knowledge from various web access log records.
A Web server usually registers a log entry for every access of a Web page in a web
log. Web Usage Mining deals with performing data mining techniques on web log
entries to find association patterns, sequential patterns, and trends of Web accessing.
Analyzing and exploring regularities in web log entries can identify potential
customers for electronic commerce, improve the quality and performance of Web
services system, and optimize the site architecture to cater to the preference of end
users.

D. Nagamalai, E. Renault, and M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 94–100, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Mining Web Path Traversals Based on Generation of FP Tree with Utility 95

When it comes to mining frequent patterns, most of the studies carried out
previously adopted an Apriori [1]-like approach based on the anti-monotone property:
if any length k pattern is not frequent in the database, its length (k + 1) super-pattern
can never be frequent.
But an Apriori-like algorithm can suffer from two nontrivial costs:
─ It is costly to handle a huge number of candidate sets.
─ It is tedious to repeatedly scan the database and check a large set of candidates
by pattern matching
Hence, we use another data mining algorithm known as FP Tree [2],[3] generation
which have the following advantages:
─ It does not involve creation of candidate sets
─ It makes use of a compact data structure called as FP Tree which is an extended
prefix tree structure storing important information about frequent patterns.
─ It makes use of partitioning based divide- and- conquer approach instead of
Apriori’s level wise approach.
However, considering only the binary frequency for each website of a web traversal
path is not a good measure for finding interesting web traversal paths because each
user may spend different time period for each website.
This paper uses FP tree with utility value to give more relevant results to the user.
This approach first pre-processes the web server log and with that data as input to the
FP Tree algorithm mines for finding out frequent patterns. After mining the utility
value is calculated which results in useful patterns to the user and helps in decision
making.
The remainder of this paper is divided as follows. In section 2 we describe related
work. In section 3 we describe the problem definition. In section 4 we present the
experiments and their results. Finally conclusions are presented in section 5.

2 Related Work
Chen et al. [4] developed two efficient algorithms for determining large reference
sequences for web traversal paths. The first one, called full-scan (FS) algorithm,
solved discrepancy between traversal patterns and association rules. The second one,
called selective-scan (SS) algorithm is able to avoid database scans in some passes so
as to reduce the disk I/O cost involved. But both these algorithms had some
limitations like they required large number of database scans, candidate generation for
large databases where difficult.
Sequential pattern mining [5] discovered sequential patterns from customer
sequence databases. But sequential pattern mining suffered from mining closed and
maximal sequential patterns as well as mining approximate sequential patterns. The
Two-Phase algorithm [6], [7] was developed, based on the definitions of [8] to find
high utility itemsets using the downward closure property of Apriori. The authors [6],
[7] have defined the transaction weighted utilization (twu) and proved that it is
possible to maintain the downward closure property. This algorithm suffers from the
same problem of the level-wise candidate generation-and-test methodology.
96 E. Poovammal and P. Cigith

A novel approach of finding frequent patterns without generating candidate sets

was proposed in the FP Tree generation algorithm. This method uses a data structure
called FP Tree which stores the database in a compressed form. A brief comparison
between Apriori algorithm and FP Tree algorithm with its advantages and
disadvantages can be studied in [9].
Zhou et al. [10] introduced the concept of utility in web path traversal mining
model. They adopted the definitions of utility from the high utility pattern mining
model [6], [7], [8], [11]. They expressed the browsing time of a user as a utility of a
website. Their algorithm is based on the Two-Phase [6], [7] high utility mining
algorithm which suffers from the level-wise candidate generation-and-test
methodology of the Apriori algorithm. Hence, they also suffered from generating too
many candidate patterns and need several database scans to find out the frequent web
traversal paths. Also combining both frequency and utility to find web path traversals
was still not explored.
So, a more effective algorithm EUWPTM (Effective Utility based Web Path
Traversal Mining) [12] was introduced which was a modification of the previous
work and is based on pattern sequential approach.

3 Problem Definition
We start with the basics of all the important terms that lead to a formal definition of
web path traversal mining using both FP Tree generation and Utility value.
A Web server log file contains requests made to the Web server, recorded in a
chronological order. The most popular log file formats are the Common Log Format
(CLF) and an extended version of the CLF, the Extended CLF. A line in the ECLF is
as shown in Figure 1. Some of the fields it contains are as follows:
─ The client's host name or its IP address
─ The date and time of the request
─ The operation type (GET, POST, HEAD, etc.)
─ The status code of the request (200, 404 etc.)
─ Referrer Link, type of OS and the type of browser.
A sample weblog data is as follows

192.168.0.1 - - [29/Sep/2004:17:10:31 +0200] \GET /axis/people.shtml

HTTP/1.1"200 8289 \http://www-sop.inria.fr/axis/table.html" \Mozilla/4.0
(compatible; MSIE6.0; Windows NT 5.2; .NET CLR 1.1.4322)"

Fig. 1. A web request from INRIA’s Web Server Log in the ECLF Format

When a user starts his session he moves from one web page to a different web
page. He moves around the Web and leaves back a trace of his activities in the web
server log. Pre-processing the web server log is the most important step in using the
log for mining. After pre-processing is done all the vital information needs to be
stored in a database. This database can be a data structure also.
A web path traversal is the set of links accessed by the user in his one particular
session. For e.g. if ‘a’,’ b’, ‘c’, ‘d’ are four websites and the user first goes to web page ‘a’
followed by ‘c’, ‘b’ and ‘d’ in one session then the path traversal for the user is a-c-b-d.
 Mining Web Path Traversals Based on Generation of FP Tree with Utility 97

Let I = {i1, i2 …im} be a set of links, and a Weblog W = {T1, T2… Tn} where Ti
is a traversal path which contains a set of links in I.
The support (or occurrence frequency) of a pattern A, which is a set of links, is the
number of traversals containing A in W. A, is a frequent pattern if A’s support is no
less than a predefined minimum support threshold S. Given a Weblog W and a
minimum support threshold, S, the problem of finding the complete set of frequent
path traversals is called the frequent path traversal mining problem. These web path
traversals are provided as the input to the FP Tree algorithm [2].
In this algorithm, first a data structure called frequent-pattern tree is constructed
which stores significant quantitative information about the frequent patterns.
Secondly, an FP-tree based pattern-fragment growth mining is developed starting
from a length-1 pattern (initial suffix pattern) examining only its conditional-pattern
base (a sub database which consists of the set of frequent items co-occurring with the
suffix pattern), constructs its (conditional) FP Tree and performs mining recursively
with such a tree.
The pattern growth is obtained by concatenating the suffix pattern with the newly
generated from the conditional FP Tree. And finally, the search techniques used is a
partitioning-based, divide-and conquer method rather than Apriori-like level-wise
generation of the combinations of frequent itemsets. It employs the least frequent
items as suffix, which offers good selectivity. All these techniques contribute to
substantial reduction of search costs.
Merely considering the presence or absence of a web page does not really judge the
importance of a web page. Different web pages mean different to different users.
Hence some parameter is chosen which acts as a measure of the importance to the
particular user. Here the parameter taken is the time spent by each user on the web
page known as the utility [10] value.

3.1 The Proposed Approach

Pre processing the web server log [13] results in the removal of useless data like
searches made by bots, spiders, Page Not Found error(status code:404) etc. This
cleaned data is the input to the FP Growth algorithm. The utility value is calculated
using the following steps:
i. The identification of users is done based on their ip addresses. Here the
assumption made is that each individual user uses the same machine.
ii. The utility value of a web page is calculated using the timestamp values of
consecutive requests of the user. For e.g.

Table 1. A sample server log

Users Links Traversed

174.129.130.230 A(04:23:56)
83.109.145.234 P(04:24:07)
174.129.130.230 C(04:24:11)
83.109.145.234 S(04:24:13)
98 E. Poovammal and P. Cigith

Table 2. Transformed one

Users Links Traversed

174.129.130.230 A(04:23:56), C(04:24:11)

Consider the sample of the pre-processed server log shown in Table 1. Table 1 is pre
processed and the result is shown in Table 2.
And finally the utility value for page A is calculated as follows:
Utility (A) = Timestamp (C) – Timestamp (A) (1)
So, in this example Utility (A) = 15 secs.
iii. At the end, those links are selected whose utility value satisfies a minimum
utility threshold (Ω) value.

4 Experimental Results
To evaluate the performance of the proposed method, experiments were conducted on
the log file access.log of web server of www.vinrcorp.com which had 389 entries.
For description of the entries in detail refer section 3 of this paper. Out of these entries
there were requests from spiders and robots which are not useful in mining. So we
removed them and the data was cleaned. This reduced the number of entries on which
the algorithm is applied. After cleaning the data 362 entries were obtained. These
entries were logged for 81 different users who searched for 138 unique web pages.
Figure 2 shows the snapshot of the output of the parsed server log. The programs
were written in Java and the IDE used was NetBeans 6.9.1. The programs were run on
Windows 7 operating system on an Intel Core i3 processor with 2.26 Ghz CPU and
4GB RAM.

Fig. 2. shows a snapshot of the output of the parsed web server log. The values are separated by
tab. To the extreme left is the ip address, in the middle is the time of access and at the end the
link accessed.

The BMS-WebView-1 contains several months’ worth of click-stream data from

two e- commerce web sites. Each transaction in these datasets is a web session
 Mining Web Path Traversals Based on Generation of FP Tree with Utility 99

Fig. 3. shows the runtime performance curve for the BMS-Webview-1 dataset. It is measured
against minimum utility threshold on the x-axis and runtime (in secs) on y-axis.

Fig. 4. shows the runtime performance curve for kosarak dataset. It is measured against
minimum utility threshold on the x-axis and runtime (in secs) on y-axis.

consisting of all the product detail pages viewed in that session. The goal for the
dataset is to find associations between products viewed by visitors in a single visit to
the web site. This dataset contain 59,602 transactions (with 497 distinct items). Figure
3 shows the runtime performance curve for BMS-WebView-1 dataset.
Another dataset, the kosarak dataset contains web click-stream data of a Hungarian
on-line news portal. It is a big dataset containing almost one million transactions
(990,002) and 41,270 distinct items. Figure 4 shows the runtime performance curve
for the kosarak dataset.

5 Conclusion
Discovering frequent Web accessing sequences from Weblog databases is not only
useful in improving the website design but also lead to better marketing decisions.
But considering only the frequency of the web pages does not reflect the importance
of the web page for the user. It is because the user might not visit an important web
page as frequently as say the user visiting web pages for daily news or even checking
his personal e-mails. So taking into account web pages only on the basis of frequency
will easily neglect a web page which is not frequent but important to the user. The
quality of the web pages thus obtained makes a big impact on the time taken for
searching effectively.
100 E. Poovammal and P. Cigith

In this paper, an attempt has been made to combine two methods namely frequency
of the access pattern and the Utility value. FP Growth algorithm finds the frequently
accessed patterns while Utility mining technique incorporates the ‘utility value’ which
measures the significance of a web page for the user.
The proposed method efficiently reduces the performance time. For a particular
user this method gives a set of links which satisfies a minimum support threshold and
also arranges all the links traversed by the user in descending order of the utility
value.
The resulting links obtained from the proposed method are checked for both
‘frequency’ and ‘interestingness’ and it is found that the results are significant for
decision making process. The links which were previously discarded because of their
low frequency but were important were also included in the final result.

References
1. Agrawal, R., Srikant, R.: Fast Algorithms for Mining Association Rules in Large
Databases. In: 20th International Conference on Very Large Data Bases, pp. 487–499
(1994)
2. Han, J., Pei, J., Yin, J.: Mining Frequent Patterns without Candidate Generation. In: ACM-
SIGMOD International Conference Management of Data, Dallas, TX, pp. 1–12 (2000)
3. Aiman, M.S., Dr. Dominic, P.D.D., Dr. Azween, B.A.: A Comparative Study of FP-
growth Variations. J. Computer Science and Network Security 09 (2009)
4. Chen, M., Park, J., Yu, P.: Efficient data mining for path traversal patterns. IEEE
Transactions on Knowledge and Data Engineering, 209–221 (1998)
5. Yao, H., Hamilton, H.: Mining itemset utilities from transaction databases. Data &
Knowledge Engineering 59, 603–626 (2006)
6. Liu, Y., Liao, W., Choudhary, A.: A fast high utility itemsets mining algorithm. In: 1st
International Conference on Utility-Based Data Mining, pp. 90–99 (2005)
7. Liu, Y., Liao, W., Choudhary, A.: A Two Phase algorithm for fast discovery of High
Utility of Itemsets. In: Ninth Pacific-Asia Conference on Knowledge Discovery and Data
Mining, pp. 689–695 (2005)
8. Yao, H., Hamilton, H., Butz, C.: A Foundational Approach to Mining Itemset Utilities
from Databases. In: Third SIAM International Conference on Data Mining, pp. 482–486
(2004)
9. Santosh Kumar, B., Rukmani, K.: Implementation of Web Usage Mining Using APRIORI
and FP Growth Algorithms. J. Advanced Networking and Applications 400(01), 400–404
(2010)
10. Zhou, L., Liu, Y., Wang, J., Shi, Y.: Utility-based Web Path Traversal Pattern Mining. In:
Seventh IEEE International conference on Data Mining Workshops, pp. 373–378 (2007)
11. Pei, J., Han, J., Asl, B., Wang, J., Pinto, H., Chen, Q., Dayal, U., Hsu, M.: Mining
sequential patterns by pattern-growth: The PrefixSpan Approach. IEEE Transactions on
Knowledge and Data Engineering 16, 1424–1440 (2004)
12. Chowdhury, F., Syed, T., Byeong-Soo, J., Young-Koo, L.: Efficient Mining of Utility
Based Web Path Traversals. In: International Conference on Advanced Communication
Technology 2009 (2009) ISBN 978-89-5519-139-4
13. Mohd, W., Mohd, H.M., Hafizul, F.H., Mohamad, F.M.M.: Data Pre-processing on Web
Server Logs for Generalized Association Rules Mining Algorithm. World Academy of
Science, Engineering and Technology 48 (2008)
 Evaluating Degree of Dependency from Domain
Knowledge Using Fuzzy Inference System

Vibha Gaur and Anuja Soni

Department of Computer Science

University of Delhi
{3.vibha,30.anuja}@gmail.com

Abstract. Inter-agent communication is one of the main concerns of agent

oriented requirements engineering that is delineated as managing inter-
dependencies and interaction among various agents performing collaborative
activities. To carry out cooperative activities, tasks are disseminated and
delegated to other agents with a rationale of sharing mutual expertise and
potential. So an agent may be dependent on other agent for accomplishing
a goal or for want of a resource to achieve that goal. This requires an agent
to quantify the dependency needs termed as Degree of Dependency (DoD) to
realize whether it should delegate the task to another agent, if yes, then to
whom so that overall quality of Multi-Agent System (MAS) is not
compromised. To evaluate DoD, this work employs domain knowledge that is
an inclusive knowledge of an environment containing the business rules,
credentials and reports to understand the business needs precisely. As the
domain knowledge may be fuzzy and uncertain, Fuzzy Inference system is
utilized to evaluate DoD from domain knowledge. This will assist the developer
to address inter-agent coordination issues without squandering resources and
hence in building MAS of high quality.

Keywords: Fuzzy Inference System, Degree of Dependency, MAS, Domain

Knowledge, PARETO’s rule.

1 Introduction
In software engineering community, there is an ever increasing demand to amend and
streamline software projects in terms of Multi-Agent Systems (MAS). MAS involve a
large number of agents playing different roles, interacting with each other to achieve
personal and common goals [1], [2]. To achieve common goals, tasks are distributed
and entrusted to other agents with purpose to share proficiency and potential; to work
in parallel or sequence on common problems. By depending on others, an agent may
be able to accomplish goals that are difficult or infeasible to achieve otherwise.
Requirements analysis necessitates supporting ways for identifying and analyzing
dependencies in inter-agent coordination.
In literature various requirements frameworks such as Formal Tropos [1], [5],
i*[6], REF [7] have been recommended for inter-agent coordination. Tropos [1] is
well known; being one of the most cited methodologies and is suitable for the very

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 101–111, 2011.
© Springer-Verlag Berlin Heidelberg 2011
102 V. Gaur and A. Soni

early phases of requirements analysis of Agent-Oriented Systems. In it dependency

requirements are modeled using UML. Tropos exhibits diverse kinds of dependencies
viz. goal, task and resource by employing agent interaction diagarms and dependency
UML diagrams. i* [6] recommends a modeling framework that focuses on strategic
actor relationships. It is employed for obtaining a better understanding of the
organizational relationships among the various organizational agents. REF [7]
involves novel requirements engineering techniques integrated with software quality
modeling approaches to provide an environment in which the agents can freely
collaborate to obtain goals, tasks and resources.
However all these frameworks of early phase of requirements engineering focus
more on capturing various kinds of dependencies instead of their degree and do not
support the ways to examine dependency needs of an agent. This work facilitates to
quantify dependency needs of an agent entitled as Degree of Dependency (DoD) that
would assist the developer in taking decision of delegating a task to another agent in
MAS so that its quality may not compromised. As an information system is implanted
in the domain knowledge of a specific business or activities in an organization, the
acquaintance with target business area is indispensable[3], [4].
As the domain knowledge may be fuzzy, incomplete and uncertain, this work
employs the Fuzzy Inference System with a number of inference rules to fuzzify
domain characteristics and result in a single crisp output value entitled as Degree of
Dependency. In addition, PARETO’s rule is applied to identify critical tasks of the
goals with high degree of dependency to overcome the delay in pursuing these goals.
The organization of this paper is as follows. The Section 2 introduces Fuzzy
Inference System and PARETO’s rule. Section 3 presents Fuzzy Inference Model for
DoD using Domain Knowledge. Section 4 illustrates the proposed methodology with
an experimental study and analyses the results. Finally section 5 concludes the paper.

2 Fuzzy Inference System

Fuzzy logic is widely used due to its ability in representing the vagueness and
imprecise information. Mamdani’s Fuzzy Inference System (FIS) is widely known
and used in developing fuzzy models. It consists of rules of the form “IF x1 is very
less AND x2 is less THEN y is high,’’ where very less, less and high are linguistic
terms with functional forms like Gaussian etc., also known as membership
functions[8],[9]. Implication is applied for each rule. Generally min operator
representing the AND method is used for implication.Aggregation is used to unify the
outputs of all the rules resulting into a single fuzzy set. The aggregated output
function is defuzzified in a single crisp number using defuzzification method viz.
Centroid method [8].The framework of deriving DoD from domain knowledge is
shown in fig. 1 that is explained in detail in experimental study in section 4.
Pareto’s rule: Dr. Juran's study of the "vital few and trivial many", the principle that
20 percent of something always are accountable for 80 percent of the outcome, was
entitled as Pareto's principle or the 80/20 rule[10]. This Pareto’s rule would be
utilized in deciding DoD at task level as explained in section 4.
 Evaluating Degree of Dependency from Domain Knowledge 103

Fuzzy Domain Attribute1

Fuzzy Domain Attribute2

... ... ∑
... ...

Rule Base
Degree of
.... .... Rules aggregation
& defuzzification
Dependency
... ...
..
Fuzzy Domain Attributen

Fig. 1. Structure of Mamdani’s Fuzzy Inference System

3 Fuzzy Inference Model for DoD Using Domain Knowledge

Advances in networking technology have revitalized agent technology for engineering
complex distributed software systems. However, the success of agent-based system is
subject to the adoption of effective agents’ cooperation that would be required to
share expertise for achieving goals of MAS [2].

A goal with a
number of
Problem with a tasks
Problem
Domain number of goals
Fuzzy Inference System

DoD at goal level

Domain Requirements Analyst

Knowledge
(Goal)
Domains)
Goals with high DoD
PARETO’S Significant
rule tasks with
DoD at task level actual need of
dependency

Fig. 2. Fuzzy Inference Model for DoD using Domain Knowledge

Degree of Dependency is defined as a parameter to quantify an agent’s dependency

needs for delegating a goal in a distributed environment. For designating a goal to
other agents, an agent requires to exercise its own dependency requirements which
may originate from domain knowledge. The focal business area is entitled as problem
domain and supportive detailed information of problem is characterized as domain
knowledge [3], [4]. This study is aimed to obtain DoD of an agent w.r.t. a goal
exercised over domain knowledge of business area. The fig. 2 illustrates a frame work
to derive DoD from domain knowledge of the problem domain which can be
described by the following subsections.

3.1 Degree of Dependency

Degree of Dependency signifies the quantification of dependency needs of a depender

agent for accomplishing a goal. A dependency describes an agreement between two
actors, the depender and the dependee [5]. Depender agent necessitates identifying the
104 V. Gaur and A. Soni

critical goals with very high or high DoD from the list of goals. The following section
describes the domain knowledge required to compute DoD.

3.2 Domain Knowledge

In software engineering, domain knowledge is knowledge about the environment in

which the target system functions [3], [4]. Domain knowledge is essential to be
learned from software users in the problem domain viz. domain experts. Domain
Knowledge utilized in this work consists of a number of goal domains that specify the
problem. Goal domain is an extensive term which incorporates all attributes along
their domain that characterize the problem. It is an area of expertise of an application
that needs to be investigated and converted in a set of rules in knowledge base, by
knowledge engineers to achieve a goal. An example of goal domain identified from
procurement data is shown in fig. 3 which further is described in detail in section 4.

Goal
Goal Domain

Product Cost Transaction Nature

Time Constraint
Transaction Feasibility
Product Quality
Domain Characteristics

Fig. 3. Domain characteristics in contributing DoD

As the domain knowledge is fuzzy, uncertain and ambiguous, this work applies FIS
on domain knowledge to evaluate DoD in the following sub-section.

3.3 Methodology to Evaluate Degree of Dependency from Domain Knowledge

In a MAS, to carry out collaborative activities, tasks are delegated to other agents
with a rationale of contributing mutual potential and expertise[1], [5], [6], [7]. For
quantifying the dependency requirements of a goal, this work defines DoD which is a
parameter in [0, 1] exercised over domain knowledge. Following steps are performed
to evaluate DoD of an agent at goal and task level from domain knowledge:
i. Obtain domain knowledge for a specific goal from domain expert or real data set
of business environment.
ii. Derive domain characteristics of the goals from domain knowledge with the
assistance of domain experts.
iii. Fuzzify all inputs formulated on domain characteristics in the form of fuzzy sets
via Gaussian membership function.
iv. Formulate a fuzzy set represented by a membership function on output variable.
In this work, DoD is treated as output variable.
v. Once fuzzy inputs and outputs parameters are obtained, observe varying fuzzy
nature of outputs as per fuzzy inputs leading to a number of rules.
vi. Apply FIS to obtain final aggregated crisp value DoD at goal level.
vii. Decompose goal into a number of tasks.
 Evaluating Degree of Dependency from Domain Knowledge 105

viii. If DoD at goal level, is 0.5 or more, belonging to high or very high category,
further analyze it at task level.
ix. Obtain importance of tasks with respect to a goal in human linguistic terms viz.
less, high, very high from domain experts.
x. Linguistic terms obtained are converted into corresponding crisp values using
defuzzification method mentioned in our previous work [13] using the following
equations. The crisp interval αA by α-cut operation against fuzzy number (l,m,u)
can be obtained as[9],[13]:
α
A I ,I m l α l, u m α u 1
where a α-cut is viewed as a bridge between fuzzy sets and crisp sets.
Crisp values against the crisp intervals representing the degree of optimism for
domain experts, is computed as [9], [13]:
CW µI 1 µ I , where µ 0,1 2
where CWtsk is crisp weight at task level, Iαl =lower bound of crisp interval,
Iαu=upper bound of crisp interval, µ is called as the index of optimism representing
optimistic level of domain experts as optimistic, moderate or pessimistic. Higher
value of µ represents the higher degree of optimism. Similarly value of α shows the
level of confidence in most promising crisp value.
xi. CWtsk is converted to a normalized value in [0,1] i.e. CWtsk[0,1] by dividing it by
summation of all task level weights of respective goal [13].
xii. To assess DoD at task level, distribute DoD at goal level to its corresponding
tasks using the following equation:

DoDatTskLvl = CWtsk[0,1] * DoDatgoal (3)

where DoDatgoal refers to DoD at goal level and DatTskLvl to DoD at task level.
xiii. Apply Pareto’s Principle-the 80-20 Rule in identifying most critical tasks for
which depender agent needs to be dependent on other agents.
The value of DoD obtained above would assist the developer in deciding the number
of optimized dependee agents during requirements analysis so as to obtain quality of
MAS in limited resources.

4 Experimental Study
To see the application of the method, an experiment was carried out to compute
Degree of Dependency of PurchaseHeadAgent in Agent-Mediated Materials
Procurement Environment (AMMPE). AMMPE is composed of PurchaseHeadAgent,
RawMaterialAgents, SparesAgents, PackagingAgents, ConsumableAgents and
MiscelleneousItemsAgents involved with procuring various items viz. raw materials,
spares, packaging, consumable and miscellaneous for various projects distributed over
various locations. PurchaseHeadAgent is concerned with the actual assessment of
dependency pertaining to various procurement activities.
106 V. Gaur and A. Soni

Fig. 4. Data seet formulated on quarterly report of purchase order

Fig. 5. Lead
L time of various procurement materials

The data set of procureement area of an Infrastructure industry was utilizedd to

capture the values of fuzzzy input variables characterizing the domain attribuutes
explained in detail in sectio
on 4.1. The data set consists of quarterly report of 10 yeears
record of data of procuremeent activity as shown in fig. 4 and lead time of procuremment
items illustrated in fig. 5. Lead time is the time elapsed between the initiationn of
procurement action and thee receipt into the supply system of the material procuured
[12]. The purchase order of procurement activities consists of the quarteerly
description of various materials
m viz. raw material, consumable items, spaares,
packaging, consumable and d miscellaneous items in the infrastructure industry.
The present experimentaal work in concern with the AMMP activity is classiffied
under following three categ gories:

4.1 Fuzzification of Dom

main Characteristics and DoD, Preprocessing of
Purchase Order Repo ort

Fuzzification of domain characteristics along with Degree of Dependency and

preprocessing of purchase order
o can be described in following subsections.

4.1.1 Identification of Do omain Characteristics

Fifteen domain experts fro om materials management were involved to capture and
formulate five domain chaaracteristics namely Product Cost, Transaction Feasibillity,
Time Constraint, Product Quality
Q and Transaction Nature for the procurement dataa set
of infrastructure industry which
w are further defined.
Product Cost is the summ mation of acquiring all the components required to builld a
product. The product with very
v high cost requires significant attention and hencefoorth
high DoD in obtaining that product within assigned time constraints. T The
 Evaluating Degree of Dependency from Domain Knowledge 107

Transaction Feasibility is concerned with the availability of product in the market.

Less feasible product requires high efforts in procuring it. Time Constraint is
concerned with the time restriction imposed on a transaction. For example time
constraint in obtaining raw material is always rigid as time delay may hamper the
entire production work in the industry. Hence the activity with tight time constraint
requires additional agents and efforts leading to high DoD. Product Quality is related
to the collection of features and characteristics of a product that contribute to its
ability to meet given requirements. The product with high quality acquires high cost
and henceforth requires significant concern. Transaction Nature signifies the nature
of transaction as highly critical, moderately critical or trivial. The items viz. raw
materials, which are directly required in the production process, are considered as
highly critical as their shortage would have a severe impact on the production. Highly
critical items require momentous attention in comparison to other trivial items
employed in industry.

4.1.2 Fuzzification of Domain Characteristics Utilized as Input Variables

Aforesaid domain characteristics viz. ProductCost, TimeConstraint are fuzzified
and various levels of Gaussian membership functions against the linguistic terms are
defined with the assistance of domain experts as shown in fig. 6,7,8,9 and 10.

Fig. 6. Variation of Product Cost Fig. 7. Variation of Transaction Feasibility

Fig. 8. Variation of Time Constraint Fig. 9. Variation of Product Quality

Fig. 10. Variation of Transaction Nature

Fig. 11. Output Variable or DOD

108 V. Gaur and A. Soni

4.1.3 Fuzzification of DoD Employed as Output Variable

The variation of the DoD as output variable is shown in fig. 11.The output varies from
very less to very high signifying the level of inter-agent dependency in this
experiment.
4.1.4 Preprocessing of Purchase Order to Assess Domain Characteristics
The data set of purchase order was worked out to assess the values of various domain
characteristics in the following manner. The cost of various materials was obtained by
the equation below:
Product cost M T ∑ Price M T
Qty M T
4
where n refers to total number of categories of various materials and m number of
items under category MtrlType[p], where1<=p<=n.
ProductCostMtrlType[q] is the total cost of qth material, Price M T
refers to
price of q item of MtrlType[p] and Qty
th
M T
th
is the quantity of q item of
materialtype[p]. The cost of various materials computed using equation (4) over
quarterly report of 10 years record of data is normalized in [0,1] using the equation (5)
which further exhibits the behavior very less, less, high or very high.
Product cost M T ProductCost M T /HighestCost 5

Time Constraint is derived from required date of items from purchase order in the
following manner. If (required time is less than lead time of item) then Time Constraint
is tight. If (required time is more than lead time of item) then Time Constraint is lenient
and If (required time is equal to lead time of item) then Time Constraint is medium.
Transaction Nature is directly related to Time Constraint. Tight Time Constraint
signifies the Transaction Nature as critical, lenient Time Constraint indicates
Transaction Nature as trivial and medium as moderately critical.
Various levels of Product Quality are mapped to the extent of quality standards
followed for manufacturing process such as ISO,QS9000 etc. Less : signifies that no
quality standards are adopted, Medium: refers to less adherence of quality standards,
High : refers to strict adherence of standards. Levels of quality are directly related to
the Transaction Nature. The critical Transaction Nature of product requires the high
quality, moderately critical as medium quality and trivial as less quality. The status of
Transaction Feasibility of the product as very less, less, high and very high would be
captured from the frequency of cancelled items encapsulated in data set.
4.2 Obtaining Degree of Dependency at Goal Level
To show the application of aforesaid proposed methodology, an experiment was
conducted using Mamdani’s Fuzzy Inference System [8], [9] for evaluating DoD of
PurchaseHeadAgent for obtaining raw material. Rules were formulated in the rule
base of the Fuzzy Inference System with the help of domain experts. The team of 15
domain experts from the infrastructure industry was involved to provide the assistance
in formulating the inference rules. Inference rules can be characterized by inferring
various output values by blending several input values. The inference rules in rule
base pertaining to raw material with a goal domain of less TransactionFeasibility,
 Evaluating Degree of Dependency from Domain Knowledge 109

tight TimeConstraint, very high ProductCost, high ProductQuality and highly critical
TransactionNature are represented as follows:
If (TransactionFeasibility is less) and (TimeConstraint is tight) and (ProductCost is
high) and (ProductQuality is high) and (TransactionNature is highlycritical) then
(DoD is high)
The other rules were formulated in the rule base in the similar way. Similarly, DoD
for obtaining spares, packaging, consumable and miscellaneous items was obtained as
shown in table 1.

4.3 Obtaining DoD at Task Level and Implication of PARETO’S Rule

Goals were further decomposed into a number of tasks. DoD for critical goals was
distributed at task level using equation 3. The importance of tasks in human linguistic
terms, captured from various domain experts was defuzzified into crisp values as
explained in section 3.3. Thereafter PARETO’s rule was applied to overlook trivial
tasks and pick out critical tasks to be delegated to efficient procurement agents.

4.4 Result Analysis

The decomposition of various procurement goals into a number of tasks along with
their DoD is illustrated table 1. Raw materials and spares are the critical goals with
very high and high DoD and rest of other goals viz. obtaining packaging, consumable
and miscellaneous items with DoD less than 0.5 are considered as trivial.

Table 1. Decomposition of goals into a number of tasks along with their DoD

Procurement Tasks DoD at Goal Crisp DoD Resulted

Goals goal Status Weights at task Task Status
level CWtsk[0,1] level (using PARETO’S rule)
Raw Material Planning 0.4375 0.3067 Critical
Ordering 0.7010 0.3437 0.2410
Delivery 0.0937 0.0657
Receiving Critical 0.0937 0.0657 Trivial
Issuing 0.0312 0.0219
Spares Planning 0.4375 0.2678 Critical
Ordering 0.6121 0.3437 0.2104
Delivery 0.0937 0.0574
Receiving 0.0937 0.0574 Trivial
Issuing 0.0312 0.0191
Packaging Planning 0.2913

Consumable Ordering 0.2900 Trivial

items Delivery
Receiving
Miscellaneous Issuing 0.2000
items

Fig.12 evidently segregates various goals with various levels of dependencies

formulated on a number of rules in rule base.
110 V. Gaur and A. Soni

Fig. 12. DoD with respect to various goals

As evident from fig. 13, it iss observed that in raw material activity by applying Pareeto’s
rule, 70%- 80% of DoD is derived
d from two activities PLANNING and ORDERING G.

Fig. 13. DoD at task level

Fig. 14. DOD w.r.t. (a) Transaaction Nature, Time Constraint (b) Product Cost, Product Quaality

Hence dependent agent needs to be dependent for these two critical activities and
rest of other activities viz. DELEVERY, RECEIVING, ISSUING less contributee in
DoD , henceforth dependen nt agent may manage these activities of his own to avvoid
unnecessary communication n overheads.
Fig. 14 shows the fuzzy surface plot obtained in the present experiment in ordeer to
depict the relationship betw ween input and output variables. Fig. 14 (a) exhibits the
relationship between the in nput variables Transaction Nature and Time Constraintt on
one hand and the output vaariable DoD on the other. DoD is inversely proportionaal to
Time Constraint and direcctly proportional to Transaction Nature. It signifies tthat
highly critical item viz. raww material urgently required for the production with tiight
time constraint requires higgh DoD for accomplishing it. Fig. 14 (b) exhibits that D
DoD
is directly proportional to Product Cost as well as Product Quality signifying tthat
higher product cost incurrred in obtaining a goal and higher quality requiired
necessitates higher depend dency requirements in the form of DoD. The relationsship
thus obtained between in nput variables and output variable DoD would asssist
PurchaseHeadAgent in pllanning future strategies for the adjustments of innput
variables with inappropriatee infrastructure and very high DoD.
 Evaluating Degree of Dependency from Domain Knowledge 111

5 Conclusion
The advent of ubiquitous computing has revolutionized distributed Multi-agent
systems. However, its successful application is subject to the adoption of effective
agents’ communication that would be needed to share expertise for achieving goals of
MAS. Tasks are disseminated and delegated to other agents with the intent of
propagating proficiency and capability. Before indulging in the process of
coordination, agents are required to assess their own dependency needs. To quantify
dependency needs of an agent, this work provides a dependency metric entitled as
DoD that would facilitate an agent to estimate its exertion load devised in the form of
Degree of Dependency in a distributed environment. To evaluate DoD, this work
employs Fuzzy Inference System exercised over domain knowledge that may be
fuzzy, uncertain and ambiguous. PARETO’s rule is applied to signify the inevitability
of dependency of an agent for critical tasks and ignore trivial tasks. This study would
assist developer to address dependency requirements of an agent during requirements
engineering and provide him a sound underpinning in estimating number of agents to
execute these dependency needs, which consequently may eschew unnecessary
communication overheads.

References
1. Bresciani, P., Perini, A.: Tropos: An Agent-Oriented Software Development Methodology.
Autonomous Agents and Multi-Agent Sytems 8, 203–236 (2004)
2. Bedi, P., Gaur, V.: Optimal Partner Selection Model for Cooperation. In: 3rd Indian
International Conference on Artificial Intelligence, pp. 1297–1307 (2007)
3. Osada, A.: The Role of Domain Knowledge Representation in Requirements Elicitation.
In: 25th Conference on IASTED International Multi-Conference: Software Engineering
(2007)
4. Rose, P.: A Method and Framework for Domain Knowledge Assisted Requirements
Evolution. In: International Conference on Software Engineering (2009)
5. Castro, J., Kolp, M.: Towards requirements-driven information systems engineering: the
Tropos project. Elsevier Information Systems 27, 365–389 (2002)
6. Singh, Y.: Evaluation of Agent Oriented Requirements Engineering Frameworks, vol. 2,
pp. 33–38. IEEE, Los Alamitos, ISBN: 978-0-7695-3336-0,2008
7. Donzelli, P.: A goal-driven and agent-based requirements engineering framework*.
Springer, Heidelberg (2004), doi:10.1007/s00766-003-0170-4
8. Zadeh, L.A.: Outline of a new Approach to the Analysis of Complex Systems and
Decision Processes. IEEE SMC-3, 28–44 (1973)
9. Sivarao, P.B.: GUI Based Mamdani Fuzzy Inference System Modeling To Predict Surface
Roughness in Laser Machining. IJECS 9(9)
10. Chen, Y.-S., Chong, P.P.: Mathematical and computer modelling of the Pareto principle.
Mathematical and Computer Modelling 19(9), 61–80 (1994)
11. Klir, G.J.: Fuzzy Sets and Fuzzy Logic. PHI Publications (1995) ISBN:81-203-1136-1
12. Srinivasan, S.: Multi-Agent System Supply Chain Management in Steel Pipe
Manufacturing. IJCSI 7(4) (2010) ISSN : 1694-0814
13. Gaur, V., Soni, A.: An Integrated Approach to Prioritize Requirements Using Fuzzy
Decision Making. IACSIT IJET 2(4), 320–328 (2010) ISSN: 1793-8236
 Novel Methodologies to Avert Hacker

Paul Bharath Bhushan Petlu1, N.L. Kumar Anantapalli1,

Devaki Pendlimarri2, and M. Muralidhara Rao1
1
Swarnandhra College of Engineering and Technology
Narsapur – 534 280, India
{trishipaul,kumaranantapalli,drmmrao54}@gmail.com
2
Swarnandhra Institute of Engineering and Technology
Narsapur – 534 280, India
[email protected]

Abstract. With the advent of internet the world has become very small in the
arena of communication. The information can be communicated from one place
to another place through the internet very easily and with great speed. Before
communicating with each other the user must be authenticated by the remote
system to provide security to the information. The concept of username with
password is an efficient method of maintaining shared secret information
between a user and a remote system. The information that is communicated
through the internet passes across an unsecured network. This gives security
breaches to the information of the user authentication and is being hacked by
the hacker. Even though the user credentials are encrypted into a cipher
text, still with the constant monitoring, the hacker is able to guess and decrypt
the user credentials. Here, we are presenting some novel methods to prevent the
hacker in knowing the user credentials even though it passes through the
unsecured networks.

Keywords: Attacks, hacker, authentication, network/information security,

security, time-slot based algorithms, multiple databases, cipher text.

1 Introduction
With the advent of internet the world has become very small in the arena of
communication. The use of computers has been increased due to the vast utilization of
internet for e-mails, e-commerce, social networking etc. The information that is being
communicated through internet passes through the unsecured networks, which gives
chance to the security breaches [4] and [5]. There are several techniques to provide
security to the information communicated between two users. The username along
with password is the efficient method of maintaining a shared secret between the user
and the remote system. The other way of authenticating the user is using biometrics.
This technique uses the human parts to authenticate the user, such as, thumb
impression, retinal of eye etc. To implement biometrics, we need additional hardware
and its technology support. If the hardware fails then we cannot authenticate the user.
The username along with password can be proved as an efficient technique for the
user authentication by adding additional protection and security measures. As the

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 112–120, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Novel Methodologies to Avert Hacker 113

information passes through the unsecured networks, the attacker/hacker will

attack/hack the information and do harm to the system or individual information.
One of the main threats to the data transmitted is sniffing. This involves using a
wired or wireless sniffer to listen to the network traffic. Sniffing may occur as a
passive eavesdropping or active interception, such as man-in-the-middle attack with
an attacker/hacker serving as an intermediary between the two systems through which
message passes. If the sniffer knows the packet structure, most sniffers have an ability
to decode and analyze the data gathered. Some sniffers can automatically filter out
usernames along with passwords from other observed information, providing the
attacker with an uncluttered view of captured account information and data.
Another threat against the password transmission is attacks. Classes of attacks
might include passive monitoring of communications, active network attacks, close-in
attacks, exploitation by insiders, and attacks through the service provider. Some of the
attacks are [7]:

1.1 Replay Attacks

A breach of security in which information is stored without authorization and then

retransmitted to trick the receiver into unauthorized operations such as false
identification or authentication or a duplicate transaction. For example, messages
from an authorized user who is logging into a network may be captured by an attacker
and resent (replayed) the next day. The retransmission of valid logon messages is
sufficient to gain access to the network even though the messages may be encrypted
and the hacker does not know the actual keys and passwords.

1.2 Passive Attacks

A passive attack on a cryptosystem is one in which the cryptanalyst cannot interact

with any of the parties involved, attempting to break the system solely based upon
observed data i.e. the cipher text. This can also include known plaintext attacks where
both the plain text and its corresponding cipher text are known. While most classical
ciphers are vulnerable to this form of attack, most modern ciphers are designed to
prevent this type of attack above all others.

1.3 Active Attacks

An intrusion into a computer network which attempts to delete or modify the data
stored on the computers which form part of the network. This is one of the most
serious forms of attack since the operations of many companies' critically depend on
data.

1.4 Password Attacks

Password cracking is the process of recovering passwords from data that has been
stored in or transmitted by a computer system. A common approach is to try
repeatedly in guessing for the password. The purpose of password cracking might be
to help a user to recover a forgotten password (though installing an entirely new
password is less of a security risk, but involves system administration privileges), to
114 P.B. Bhushan Petlu et al.

gain unauthorized access to a system, or as a preventive measure by system

administrators to check for easily crack-able passwords. On a file-by-file basis,
password cracking is utilized to gain access to digital evidence but access to a
particular file is restricted.

1.5 Phishing Attacks

Phishing is a way of attempting to acquire sensitive information such as usernames,

passwords and credit card details by masquerading as a trustworthy entity in an
electronic communication. Communications purporting to be from popular social web
sites, auction sites, online payment processors or IT administrators are commonly
used to lure the unsuspecting public. Phishing is typically carried out by e-mail
spoofing or instant messaging and it often directs users to enter details at a fake
website whose look and feel are almost identical to the legitimate one. Phishing is an
example of social engineering techniques used to deceive users and exploit the poor
usability of current web security technologies. Attempts to deal with the growing
number of reported phishing incidents include legislation, user training, public
awareness and technical security measures.

1.6 Hijack Attack

Session hijacking is the exploitation of a valid computer session — sometimes also

called a session key — to gain unauthorized access to information or services in a
computer system. In particular, it is used to refer to the theft of a magic cookie used to
authenticate a user to a remote server. It has particular relevance to web developers, as
the HTTP cookies used to maintain a session on many web sites can be easily stolen
by an attacker using an intermediary computer or with access to the saved cookies on
the victim's computer.

1.7 Spoof Attack

Spoofing attack is a situation in which one person or program successfully

masquerades as another by falsifying data and thereby gaining an illegitimate
advantage.

1.8 Buffer Overflow

In computer security and programming, a buffer overflow, or buffer overrun, is an

anomaly where a program, while writing data to a buffer, overruns the buffer's
boundary and overwrites adjacent memory. This is a special case of violation of
memory safety.

1.9 Exploit Attack

The attacker knows of a security problem within an operating system or a piece of

software and leverages that knowledge by exploiting the vulnerability.
 Novel Methodologies to Avert Hacker 115

Here, we suggest some novel set of methodologies to prevent the hacker or

attacker from decrypting the user credentials. The hacker/attacker tries to know the
encrypted user credentials by constant monitoring.

2 Background
The username along with the password is an efficient method of maintaining a shared
secret between the user and the remote system. The user credentials are
communicated through the unsecured network, which causes to security breach. In
order to provide security, the user credentials are encrypted into a cipher text by using
the Data Encryption Standard algorithm. But, still by constant monitoring the
attacker/hacker is able to guess and decrypt the cipher text. The username along with
the password can be proved as an efficient authentication technique by giving more
protection and security measures to the user credentials that are being communicated
through the network. Some novel methods are suggested by the authors to provide
more protection and security to the user credentials. They are: cipher text based on the
time-stamp, multiple user passwords, and multiple cipher texts. These methods can be
applied to prevent the hacker/attacker in decrypting the user credentials that are being
communicated through the unsecured network by constant monitoring.

3 Methodology
Here, we suggest some novel set of methodologies to prevent the hacker/attacker from
capturing the username and the password via an unsecured network. Most of the
services that are provided through the internet are using the username with password
mechanism for the user authentication. This username and password will be given to
the user at the time of creation of the new account. Later, whenever the user wants to
communicate with the system provided through the internet, he/she is identified and
authenticated using this username with password. As this username and password
information passes to the remote system via unsecured networks, causes a breach to
the hackers/attackers to decrypt the username and password. To prevent the
hacker/attacker from capturing the username and the password, we suggest the below
novel set of methodologies.

3.1 Cipher Text Based on the Time-Stamp

The round the clock i.e. 24 hours is divided into number of equal or unequal time-
slots. Each time-slot is associated with an algorithm using which the user credentials
are encrypted into cipher text.

Equal time-slots: The 24 hours are divided into number of time-slots. Each time-slot
t1, t 2...tn is of equal time period such as 1 hour or 1½ hour or 2 hours etc. depending
upon the value of n, total number of time-slots. For example, if n = 12 then each
time-slot is of two hour time period.
116 P.B. Bhushan Petlu et al.

Unequal time-slots: The 24 hours are divided into number of time-slots of different
1
time periods. For example, t1 = 1hr , t 2 = 2hrs, t 3 = 2hrs...tn = 1 hr .
2
In either the case the below equation must be satisfied:
n
N= t
a =1
a = 24hours (1)

Each time-slot is represented as t1, t 2...tn where n represents the number of time-
slots divided per day. Each time-slot ti is associated with an algorithm a lg orithmi ,
where 1 ≤ i ≤ n . There is an algorithm s tan dard _ a lg orithm , which is used to
encrypt the user credentials i.e., username and password at the first time of user
account creation.
There are many systems offering their services through the internet. Every user
need user account to access the services of the system. This user account holds the
username along with the password to authenticate the user uniquely. After creation of
user account, whenever the user wants to access the services of the system, the user
credentials are used to authenticate the user by the sign up page. This sign up page
will encrypt the user credentials using the a lg orithmi associated with the time-slot ti
when the user submits the credentials. This encrypted data will be sent to the server
along with the time stamp as a packet. After receiving this packet the server will
decrypt the encrypted data with the use of a lg orithmi associated with the same time-
slot ti received along with the packet. This gives the original username and the
password as entered by the user. Due to security reasons, as the username and the
passwords shouldn’t be stored as a plain text in the database, the server will again
encrypt the data using s tan dard _ a lg orithm and store this into the database as
shown in Fig 1. With this the user will have an account through which he/she can
access the services of the organization through the internet at any time.
Whenever the user wants to access the services of the system, he/she is prompted to
submit the user credentials for the user identification and authentication. After submitting
these credentials the sign in page will encrypt the user credentials using a lg orithmi
associated with the time-slot ti that the user has submitted the details. This encrypted
data will be sent to the server along with the time stamp as a packet through the network.
After receiving the packet, the server will decrypt the data using a lg orithmi associated
with the same time-slot ti which is being received along with the packet. This will give
the original username and password as submitted by the user. The server encrypts the
decrypted user credentials using the algorithm s tan dard _ a lg orithm , and verifies in the
database. Once the encrypted user credentials matches with the user credentials stored in
the database, the server will create a session through which the user can communicate
with the system for further access.
Each time-slot is associated with different algorithms; the details submitted by the
user will be encrypted into different cipher texts, which will prevent the
hacker/attacker in guessing the user credentials. More number of time-slots and/or
data encryption algorithms will lead to more prevention of hacker/attacker. By
applying the below strategies, even the hacker can be averted more. They are:
 Novel Methodologies to Avert Hacker 117

• Changing encryption algorithms: The data encryption algorithms associated with

each time-slot can be changed in regular intervals.
• Repeated algorithms: The set of data encryption algorithms can be repeated to
some time-slots.
• Changing time periods: The time period of the time-slots can be changed in
regular intervals.
• Hybrid approach: The above strategies can be applied simultaneously.

Authentication Server
Network

Client
Database

Server

Fig. 1. Generation of cipher text based on time-stamp with two tier architecture

The problem in the above methodology is that, the number of decryption and
encryption operations at the server side will increase the burden on the server. We can
reduce this burden by using the authentication server. The user identification and
authentication process will be carried out by the authentication server instead of the
main server. Once the user is authenticated as a valid user, then the Authentication
Server will communicate this to the Main Server to create a session which will further
allow the user to communicate with the server. With the introduction of authentication
server we have additional cost, space and technology.

3.2 Multiple User Passwords

The services of the system though the internet can be accessed by the users having an
account. This account has the information of the user credentials, username along
with the password. In this methodology, we propose that, at the time of account
creation, the user may select the number of passwords he would like to have. This
number may be restricted either to two or three. The user is required to choose two or
three passwords instead of one at the time of account creation. The chosen username
and the password(s) are stored in the database at the server as username, password1,
password2, and password3. Certain password selection measures are to be taken
while choosing the passwords, such as, complexity of the password in terms of size,
combination of alphabets and digits, etc., which results the password to be strong. The
user is able to access the services of the system by submitting the user credentials,
118 P.B. Bhushan Petlu et al.

username along with any one of the three passwords. The user credentials submitted
by the user are verified with the stored information to authenticate the user. Once the
user is authenticated as a valid user, the server will create a session through which the
user is able to access the system for further communication.

3.3 Multiple Databases

The methodology is as shown in Fig. 2. The round the clock, i.e. 24 hours is divided into
number of time-slots such as t1, t 2...tn , where n represent the number of time-slots. Each
time-slot ti is associated with an encryption algorithm a lg orithmi , where 1 ≤ i ≤ n .
Each distinct data encryption algorithm is associated with a database. These databases are
represented as db1, db 2...dbm where m represents the number of distinct data encryption
algorithms. The encryption algorithm s tan darda lg orithm is used to encrypt the user
credentials into cipher text at the time of creation of a user account.
Every user can access the system with a user account. At the time of creation of
user account, the submitted user credentials by the user are encrypted into a cipher
text using the algorithm s tan darda lg orithm . This encrypted cipher text will be sent
to the server along with the time stamp as a packet. After receiving this packet, the
server will decrypt the cipher text with the use of the algorithm s tan darda lg orithm .
This gives the original credentials as submitted by the user. The server will encrypt
these user credentials into different cipher texts cipher1, cipher 2...cipherm using the
algorithms a lg orithm1, a lg orithm 2...a lg orithmm respectively. These encrypted
cipher texts are stored in the respective databases db1, db 2...dbm .
After creation of a new account, whenever the user wants to communicate with the
system, the user is prompted to submit the user credentials for the user identification
and authentication. After submitting these details the sign in page will encrypt the
user credentials into a cipher text using the algorithm a lg orithmi , where i represent
the time-slot the user has submitted the details. This encrypted cipher text will be sent
to the server as a packet through the network. The server will verify this cipher text
with the databases db1, db 2...dbm simultaneously. If the user submitted credentials
matches with the previously stored user credentials in any one of the database, then
the user is authenticated as a valid user. Once the user is authenticated as a valid user,
the server will create a session through which the user can communicate with the
system for further access.
In every time-slot, as the data encryption algorithm is changing the details entered
by the user will be encrypted into different cipher texts, which will prevent the
hacker/attacker in guessing the user credentials. Hence, the more number of time-slots
and/or data encryption algorithms will lead to more restriction to the hacker/attacker.
To provide more protection and security to the user credentials from the man-in-the-
middle, the below strategies can be applied:
• Changing encryption algorithms: The data encryption algorithms associated with
each time-slot can be changed in regular intervals.
• Repeated algorithms: The set of data encryption algorithms can be repeated to
some time-slots.
 Novel Methodologies to Avert Hacker 119

• Random selection of encryption algorithms: The data encryption algorithms to

each time-slot can be selected randomly.
• Changing time periods: The time period of the time-slots can be changed in
regular intervals.
• Hybrid approach: The above strategies can be applied simultaneously.

Algorithm1

DB1

DB2
Algorithm2
Client Network Server

Algorithmm DBm

Fig. 2. Methodology for generating different forms of cipher texts for the user credentials

Note: The same set of data encryption algorithms must be used. Addition of a new
algorithm leads to the addition of a database at the server side.
The problem here is that the server is maintaining the multiple databases to
maintain the user credentials in different forms of encrypted cipher texts using
different algorithms. However, the maintenance of multiple databases can lead to an
advantage in the case of database recovery. Suppose, one database is crashed due to
some reasons, we can recover the information from the remaining databases. Hence,
more number of databases will lead to a more security in database recovery.

4 Conclusion and Future Work

Since the advent of internet there is a vast utilization of internet for e-mailing, e-
commerce etc. by the numerous users being at different places. The concept of
username and password is an efficient method of maintaining a shared secret between
the server and a remote client system. The communication between the server and the
client system is via an unsecured network. This gives breaches to the hacker/attacker
to know the username and the password of the user by sniffing, hacking etc. being a
man-in-the-middle. To restrict the man-in-the-middle from hacking/attacking, we
proposed three methodologies. Each methodology has its own advantages as well as
disadvantages. The selection of methodology may depend upon the available
technology and the cost of establishment of the methodology. Further, with the
innovative methods, techniques and algorithms the data that has been passed through
the network can be made more secured. Hence, the man-in-the-middle can be
prevented or nullified more from either hacking or attacking the packets.
120 P.B. Bhushan Petlu et al.

References

[1] Scarfone, K., Souppaya, M.: Guide to Enterprise Password Management (Draft).
Recommendations of the National Institute of Standards and Technology, Special
Publication 800-118, 38 pages (2007)
[2] Elminaam, D.S.A., Kader, H.M.A., Hadhoud, M.M.: Performance Evaluation of
Symmetric Encryption Algorithms. IJCSNS International Journal of Computer Science and
Network Security 8(12), 280–286 (2008)
[3] Identification and Authentication, NIST Computer Security Hand Book,
http://www.sos.cs.ru.nl/applications/courses/security2008/
nistiadraft.pdf
[4] Raj, A.K., Tewari, R.R., Upadhyay, S.K.: Different Types of Attacks on Integrated
MANET- Internet Communication. International Journal of Computer Science and
Security (IJCSS) 4(3), 265–274 (2010)
[5] Network security types of attack passive attack active attack,
http://www.computernetworkingnotes.com/ccna_certifications/
types_of_attacks.htm
[6] Petlu, P.B.B., Pendlimarri, D., Anantapalli, N.L.K., Naga Srinivas, P.V., Muralidhara Rao,
M.: Proceedings of the IC2IT, International Conference on Computing and Information
Technology, held on 11-12, 2011 at King Mongkut’s University of Technology, North
Bangkok, Bangkok, Thailand, pp: 116-121 (May 2011)
[7] http://en.wikipedia.org/wiki/Password_cracking
 Multi-Domain Meta Search Engine with an Intelligent
Interface for Efficient Information Retrieval on the Web

D. Minnie1 and S. Srinivasan2

1
Madras Christian College, Chennai, India
[email protected]
2
Anna University of Technology Madurai, Madurai, India
[email protected]

Abstract. A web search engine searches for information in the World Wide
Web. The user needs to search for information in various domains such as
financial and healthcare. The user is expected to remember various Vertical
Search Engine names to extract appropriate information. A Vertical Search
Engine is a domain specific search engine that provides the user with results for
specific queries on that domain. Often the user is unable to get the exact
information from the web due to the different page ranking techniques followed
by individual search engines. Meta Search Engines solve this problem to a
certain level by sending the user’s search queries to various search engines and
combining the search results. This paper proposes Multi Domain Meta Search
Engines that facilitate efficient Information Retrieval on multiple domains. User
interfaces for selecting domain, category and search engines is presented. An
intelligent User Interface is also designed to get the user query and to send it to
appropriate search engines. Few algorithms are designed to combine results
from various search engines and also to display the results.

Keywords: Information Retrieval, Web Search Engine, Vertical Search Engine,

Meta Search Engine.

1 Introduction and Related Work

The World Wide Web (WWW) has evolved into a huge repository of information.
The web resources increase everyday and that in turn increases the complexity of
accessing the web data. There is a need for efficient searching techniques to extract
appropriate information from the web, as the users require correct and complex
information from the web.

1.1 Web Search Engine

A Web Search Engine is a search engine designed to search WWW for information
about given search query and returns links to various documents in which the search
query’s key words are found. The types of search engines [1 – 3] used in this paper
are General Purpose Search Engines such as Google, Vertical Search Engines and
Meta Search Engines.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 121–129, 2011.
© Springer-Verlag Berlin Heidelberg 2011
122 Minnie D. and Srinivasan S.

1.2 Vertical Search Engine

A Vertical Search Engine is a domain specific search engine that provides the user
with results for specific queries on that domain. It indexes web pages that are related
to a particular domain. The filter component of Vertical Search Engine classifies the
web pages downloaded by the crawler into appropriate domains [4]. Medical search
engines such as “MedSearch”, facilitate the user to get information on medical
domain [5 - 7]. Search engines assume that users are capable of creating appropriate
search queries. Vertical Search Engines such as iMed creates search queries for the
user by interacting with the user [8].

1.3 Meta Search Engine

Different Search Engines follow different page ranking techniques and hence
approximately 20% of the results are same for a specific search query. The user may
fail to receive the necessary information. This gives rise to the use of Meta Search
Engines [9]. Meta Search Engine accepts a search query from the user and sends the
search query to a limited set of Search Engines. The queries may be sent as it is or it
may be modified to suit the requirements of the search engines. The results are
retrieved from the various search engines and they are combined to produce a result
set and are given to the user.

1.4 Information Retrieval

Information Retrieval is the science of searching and retrieving information. Web

search is also a type of information retrieval as the user searches the web for
information. The efficiency of a search facility is measured using two metrics
Precision and Recall. Precision specifies whether the documents retrieved are relevant
and Recall specifies whether all the relevant documents are retrieved.
Precision = Relevant and Retrieved / Retrieved (1)
Recall = Relevant and Retrieved / Relevant (2)
Precision is calculated as the ratio of the relevant web pages that are retrieved to the
number of web pages that are retrieved. Recall is not calculated as it is not possible to
find the number of relevant web pages out of the millions of web pages in the WWW.

2 Web Search Engine Architecture

A web search engine consists of a Web Crawler or Spider, Indexer, Query Parser and
Ranker. The architecture of a web search engine is shown in Fig. 1.
Web search engines work by storing information about many web pages, which
they retrieve from the WWW itself. These pages are retrieved by a Web crawler
which follows every link it sees. The contents of each page are then analyzed to
determine how it should be indexed.
Web pages are filtered using Term Frequency-Inverse Document Frequency
(TFIDF) scores for that page. TFIDF is calculated as the product of Term Frequency
(TF) of a term t in document d and Inverse Document Frequency (IDF) of a term t.
 Multi-Domain Meta Search Engine with an Intelligent Interface 123

TFIDFtd = TFtd * IDFt (3)

Term Frequency (TFtd) of document d and term t is given as the ratio of the number of
occurrence of a term in a web page to the total number of words in that page.
TFtd = No. of term t in document d / Total no. of words in document d (4)
Inverse Document Frequency (IDFt) of a term specifies the uniqueness of a document.
It is given as a ratio of the total number of documents to the number of documents
containing the term.
IDFt = log [Total no. of documents (N) / No. of documents with term t ] (5)
The higher value for IDFt specifies that the document is unique as the term is present
in few documents. The lower value specifies that the term is present in many
documents and hence the document is not unique even though the Term Frequency of
the document could be high.

Fig. 1. Architecture of Web Search Engine

Fig. 2. Meta Search Engine Architecture

124 Minnie D. and Srinivasan S.

3 Meta Search Engine Architecture

Meta Search engine is not a repository of physical data. It receives request from the
user and sends the request to various search engines. The search engines check their
indices and extract a list of web pages as links and pass the result to the Meta Search
Engine. The Meta Search Engine receives the links, applies few algorithms, ranks the
results and finally displays the result. The Meta Search Engine architecture is shown
in Fig.2.

4 Methodology
Few Web Search Engines[10] such as Google, Yahoo, MSN and Vertical Search
Engines[11] such as Yahoo Health, webMD, MSNHealth and Meta Search Engines
[12] such as DogPile, MetaCrawler, MonsterCrawler are used in this study. A
Multiple Search Engine [13] is also analyzed in which a Search Engine can be
selected from a list of search engines and queries can be sent to it.
Four topics Cancer, Diabetes, Paracetamol and Migraine are identified to be
searched in the nine search engines. The search results are classified under the
categories Relevant, Reliable, Redundant and Removed based on the retrieved web
page contents. 20 results from each of the nine Search Engines are considered for
each topic and are given in table 1. The interpretation of the results is shown in Fig.3.
The precision values are calculated using the ratio of relevant documents to the
retrieved documents and is presented in Table 2. The precision details are also plotted
in a graph and are shown in Fig. 4.
The user searches for financial information, healthcare information and so on in
the web. The user is expected to remember various search engine names to extract
necessary information. The search engines also have a biased view of presenting the
documents and the user may miss the necessary information.
Hence the facility to get efficient domain related information on multiple domains
is a need of the hour. Multi-Domain Meta Search Engine [14] sends queries to various
search engines and consolidates the results from the search engines.

Table 1. Relevance, Reliability, Redundant and Removed details for Search Engines

Search Engine Relevant Reliable Redundan Removed

DogPile 18.75 16.25 1.75 0.5
Google 19.25 19.25 0.25 0.25

Meta Crawler 17.75 14.25 4 0

Monster Crawler 19 14 1 0

MSN 18.25 12.5 1.5 1.25

MSN Health 1 1 0 0

WebMD 17 17 0.25 0

Yahoo 19.25 17.25 2.75 0

Yahoo Health 11 11 0 0
 Multi-Domain Meta Search Engine with an Intelligent Interface 125

Fig. 3. Search Engine’s Relevant, Reliable, Redundant and Removed details

Table 2. Precision value table

Search Engine Precision Search Engine Precision

DogPile 0.94 MSN Health 0.25
Google 0.96 WebMD 0.85
Meta Crawler 0.89 Yahoo 0.96
Monster Crawler 0.95 Yahoo Health 0.55
MSN 0.91

Fig. 4. Precision for various Search Engines for the given topics

5 Proposed Multi-Domain Meta Search Engine

We propose a user-friendly Multi-Domain Meta Search Engine that sends search
queries to various search engines and to retrieve results from them. Various cases of
selection of search engines are proposed in this paper.
126 Minnie D. and Srinivasan S.

Various cases of Meta Search Engines are designed and analyzed in this study. The
basic architecture of the Multi-Domain Meta Search Engine is given in Fig. 5. In the
first model the Meta Search Engine was designed to send queries to Search Engines
such as Google, Yahoo, AltaVista and AskJeeves. The queries were formed for a
specific domain by adding the domain name as part of the search query. An interface
for the Meta Search Engine was formed consisting of push buttons to select the
domain and text box to enter the query string. The query string to be sent to the search
engines are formed with the + symbol applied on the string entered and the selected
domain name.
The second model of Multi Domain Meta Search Engine aims at providing an
efficient information retrieval on a particular domain for the user by accessing various
Vertical Search Engines.
The result of the query can be shown as a list of results from each search engine in
individual windows, or as a list of results in different frames of a same window or as a
single ordered list of results.

6 Multi-Domain Meta Search Engine Operations

The Multi-Domain Meta Search Engine consists of a user interface for domain and
category selection, a user interface for search engine selection, query formulator,
dispatcher, results receiver, filter, ranker and display of results.

6.1 User Interface – Domain and Category Selection

The Domain selector presents a list of domains in this interface to the user. The
domains are listed in the alphabetical order for easier access.
The user is allowed to enter the request as a query. The user can directly move to
search or can select the domain from which the information is requested. User can
also select the category of his request such as research papers, news articles, and
homepages and so on so that the search can be performed on relevant pages.

Fig. 5. Multi-Domain Meta Search Engine Architecture

 Multi-Domain Meta Search Engine with an Intelligent Interface 127

6.2 User Interface – Search Engine Selection

The Search Engine Selection Interface presents to the user a list of Search Engines
that are appropriate for the selected domain and category. Vertical Search Engines if
available for a specific domain is also listed here. The user has the option to select one
or more of the search engines. If the user has not selected the search engines then all
the listed search engines are selected for further processing.

6.3 Query Formulator and Dispatcher

As different Search Engines follow different styles for the representation of the query
search string, the search query strings are generated for a given user input. Different
query strings are generated as per the requirement for the different search engines.
The query strings are then sent to various search engines to extract desired results
from the search engines.

6.4 Results Receiver

The user generally searches only the first and second pages of a search result and also
the most relevant results are presented in those pages. The Search Engines provide the
best results also in the top few pages. Hence 25 results from each Search Engine are
selected as the results from the Search Engines. The results are stored for further
processing.

6.5 Filtering and Ranking Results

The results from a single search engine are found to be having redundant links and
removed links during some of the searches. These links are removed from the results
set from each of the search engine results in the first step. All the results are populated
and the redundant links across the different search engine results are eliminated. The
ranking used by the different search engines is used to rank the results.

6.6 Results Generation in Different Windows

The results from the various search engines are displayed in different windows. The
user has the freedom to look into the results given in various windows. The advantage
of this method is that the user can select as many Search Engines as needed. There is
no need for the ordering and ranking of results. The disadvantage of the method is
that the user has to switch between windows for effective usage of the system.

6.7 Results Generation in Multiple Frames of Same Window

The results from various search engines are displayed in various frames of the same
window. The output window is designed to display results from 4 search engines
simultaneously. This method is faster to use but the user has to make the decision
about the result links to be visited. The result for the searches on Cancer, Diabetes,
Paracetamol and Migrane for this case is given in Fig.6 and the precision details are
given in Fig. 7. In this method also the ordering and ranking of the results is not done
as the result from the various search engines are used as it is.
128 Minnie D. and Srinivasan S.

Fig. 6. Relevant, Reliable, Redundant and Removed data of Multi-Domain Meta Search Engine

Fig. 7. Precision for MDMSE and various Search Engines for the given topics

It can be seen from the figures 6 and 7 that the Multi-Domain Meta Search
Engine’s performance is better than the performance of individual search engines.

6.8 Results Generation in Same Window

Alternatively the ordering process is to be performed and the ranked list is to be

displayed along with the rank and the name of the search engine that has produced the
link.

7 Conclusion and Future Work

This paper analyses two Multi Domain Meta Search Engines that provide an
information retrieval of information on various domains for the user using various
Search Engines that are already available. The Web Search Engine, Vertical Search
Engine and Meta Search Engine features are also presented. Few Search Engines are
tested for the relevancy, reliability, redundancy and availability of search results for
 Multi-Domain Meta Search Engine with an Intelligent Interface 129

few topics. A domain and category selection user interface is presented. A query
interface is designed to send user’s search query to the various search engines and the
resultant links are consolidated to display the results to the user.
The system can be extended to process the content of the web pages in addition to
processing of the links.

Acknowledgement. The authors wish to thank Ms.L.R.Rita, project student,

Department of Computer Science, Madras Christian College, Chennai, India for her
contribution towards development of the Multi-Domain Meta Search Engine.

References
1. Dunham, M.H., Sridar, S.: Data Mining. Pearson Education, London (2007)
2. Kosla, R., Blockeel, H.: Web Mining Research: A Survey. SIGKDD Explorations 2(1),
1–15 (2000)
3. Jeyaveeran, N., Haja Abdul Khader, A., Balasubramaniyan, R.: E-Learning and Web
Mining: An Evaluation. In: Proceedings of the 2nd International Conference on Semantic
e-Business and Enterprise Computing (2009)
4. Shettar, R., Bhuptani, R.: A Vertical Search Engine – Based on Domain Classifier.
International Journal of Computer Science and Security 2(4), 18–27
5. Luo, G., Tang, C.: On Iterative Intelligent Medical Search. In: SIGIR 2008, The 31st
Annual International ACM SIGIR Conference, Singapore, July 20-24, pp. 3–10 (2008)
6. Luo, G.: Intelligent Output Interface for Intelligent Medical Search Engine, Association for
the Advancement of Artificial Intelligence. In: Proceedings of the Twenty-Third AAAI
Conference on Artificial Intelligence, pp. 1201–1206 (2008)
7. Ilic, D., Bessell, T.L., Silagy, C.A., Green, S.: Specialized Medical search-engines are no
better than general search-engines in sourcing consumer information about androgen
deficiency. Human Reproduction 18(3), 557–561 (2003)
8. Luo, G.: Design and Evaluation of the iMed Intelligent Medical Search Engine. In:
Proceedings of the 2009 IEEE International Conference on Data Engineering, ICDE 2009,
pp. 1379–1390 (2009)
9. Chan, K.-P.: Meta search engine. Theses and dissertations. Paper 232 (2007),
http://digitalcommons.ryerson.ca/dissertations/232
10. General Search Engine web sites, http://google.com, http://google.com,
http://google.com
11. Vertical Search Engine web sites,
http://health.yahoo.net,
http://www.webmd.com, http://health.msn.com
12. Meta Search Engine web sites,
http://dogpile.com/, http://metacrawler.com,
http://monstercrawler.com
13. White, R.W., Richardson, M., Bilenko, M.: Enhancing Web Search by Promoting Multiple
Search Engine Use. In: Proceedings of the 31st Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval (SIGIR 2008),
Singapore, July 20-24 (2008)
14. Minnie, D., Srinivasan, S.: Meta Search Engines for Information Retrieval on Multiple
Domains. In: Proceedings of the International Joint Journal Conference on Engineering
and Technology (IJJCET 2011), pp. 115–118. Gopalax Publications & TCET (2011)
 Improved Video Watermarking Using Nested
Watermark, Wavelets and Geometric Warping

T. Jayamalar1 and V. Radha2

1
Research Scholar, Avinashilingam Deemed University for Women, Coimbatore, India
[email protected]
2
Associate Professor, Avinashilingam Deemed University for Women, Coimbatore, India
[email protected]

Abstract. The advancements in communication software and hardware is

increasing the amount of information in the form of text, image, audio and video
being shared. This increase arise the issue of protecting the intellectual property.
This paper studies the protection of video signals through watermarking. The
system proposed uses a combination of visual cryptography and wavelets to create
a nested watermark which is embedded into the raw uncompressed video signal
using a geometric warping technology. To improve robustness against
compression, the method embeds the watermark in relevant area using a block
selection method. The experimental results prove that the proposed system is
efficient in terms of capacity and robustness against various attacks.

Keywords: Video Watermarking, Nested Watermarking, Geometric Warping,

Wavelets.

1 Introduction
The current era of information explosion is envisaging tremendous growth in
communication medium and hardware growth. This in turn, is generating a huge
volume of digital signals in the form of text, images, videos and audio content. The
World Wide Web (WWW) technology is providing an easy-to-access medium, which
is allowing a large number of people to share information in a cost effective manner.
However, these advancements prove to be challenging while taking the intellectual
content protection into consideration. To answer this challenge, techniques like
encryption, stegnography, cryptography and watermarking are being used. Among
these techniques, digital watermarking techniques have gained more attention while
proving the integrity and authenticity of the owner ([25], [17], [18]).
The art of embedding hidden data into a host signal is termed as ‘digital
watermark’. The techniques can be used for copyright protection, fingerprinting, copy
protection, broadcasting monitoring and data authentication. The techniques of
watermarking can be text-based watermarking, image watermarking ([19], [7], [9],
[14]), video watermarking ([5],[10]), audio watermarking and 3D watermarking [26].
Among them image and video watermarking are two important areas that has attracted
several researchers ([33], [21]).

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 130–141, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Improved Video Watermarking Using Nested Watermark, Wavelets 131

All these techniques have the common objective of embedding a secret data to
protect a signal. With video watermarking it works with the principle that they should
ensure digital ownership, have the ability to track the source of the digital video, have
quick and easy way to detect video manipulation, have no visible video degradation
and should have easy implementation procedure. The basic characteristics are
imperceptibility, reliability, low complexity of watermarking algorithm and security
of the hiding place.
In general scenario, watermarks are embedded into video signal either during or
after compression ([29], 28]). This has the advantage that watermarking process is not
influenced by the compression algorithm. However, the watermarks are generally not
robust to a transcoding of the video. Thus, embedding the watermark to the
uncompressed video data will have the advantage of being robust to the transcoding
of the video. The challenge here is that since both watermarking technique and
compression algorithm both seek to use irrelevant data, the embedded watermark
might be removed by the compression algorithm. It would be better if the watermark
is embedded in the relevant data so that it is not removed by compression algorithm.
A method based on geometric warping was suggested by [27]. This method was
efficient in terms robustness to lossy compression and improved the quality by using a
block linking approach. This method had the advantage of not requiring the original
video while extracting the watermark. However, this technique reduced the watermark
capacity. This method is referred as base system in this paper. Aggarwal et al. [1]
proposed another efficient watermarking algorithm to improve the payload and
robustness without affecting the image quality. They used a concept of nested
watermarking where one watermark is embedding into another watermark and the
resultant image is embedded into the cover watermark. The concept of nested
watermark improved payload and was able to maintain the quality to a great extent.
Motivated by these two works, in the present work, a single copyright message is
divided into two shares using Visual Cryptography (VC). The second share is then
embedded into first share, thus creating a new watermark image. This image is then
embedded into an uncompressed video signal using geometric warping method as
suggested by [27].
The rest of the paper is organized as follows. Section 2 presents a brief study on the
previous works related to the topic. In Section 3, the basic concepts of visual
cryptography (VC) and geometric warping are described. Section 4 explains the
proposed watermarking algorithm. Section 5 presents the experimental results and
Section 6 concludes the work with future research directions.

2 Previous Studies
Solutions to video watermarking have been studied by various researchers who work
with both raw uncompressed video data [8] and compressed video data [2]. Video
watermarking techniques can be mainly grouped as spatial domain approaches and
frequency domain approaches.
132 T. Jayamalar and V. Radha

The spatial domain watermarking techniques embed the watermark by modifying

the pixel values of the host image directly. Least Significant bit (LSB) technique is
the most frequently used method [16]. Both [30] and [13] embedded the watermark in
saturation on the HIS (Hue, Saturation, Intensity) color space. A variable block size
based adaptive watermarking in spatial domain was proposed by [15]. In a later
period, [32] proposed a probability block based watermarking method for color image
with fixed block size. Recently, an algorithm based on chaotic maps was proposed by
[34] to determine the pixel bit for embedding. The main advantages of pixel based
methods are that they are conceptually simple and have very low computational
complexities. However, they also exhibit some major limitations. The need for
absolute spatial synchronization leads to high susceptibility to de-synchronization
attacks; lack of consideration of the temporal axis results in vulnerability to video
processing and multiple frame collusion; and watermark optimization is difficult
using only spatial analysis techniques.
Frequency domain approaches use transformation techniques to embed the
watermark into a video sequence. Discrete Cosine Transformation (DCT), Discrete
Fourier Transformation (DFT) and Discrete Wavelet Transformation (DWT) are the
three main methods of data transformation. A subsample based watermarking
technique was proposed by [20], where the DCT coefficients of the subimages were
utilized to store the watermark. The authors of [6], proposed an algorithm which
was based on embedding the watermark image three times at three different
frequency bands, namely, low, medium and high and the results proved that the
watermark cannot be totally destroyed by either low pass, medium or high pass
filter. In [29], two complementary watermarks were embedded into the host image
in order to make it difficult for attackers to destroy both of them. The main benefit
obtained from frequency domain techniques is that they can take advantage of
properties of alternate domains to address the limitations of pixel-based methods.
Besides, analysis of the host signal in a frequency domain is a prerequisite for
applying more advanced masking properties of the HVS to enhance watermark
robustness and imperceptibility. Generally, the main drawback of transform domain
methods is the high computational requirement.
Another field where scientists focus is on embedding into a compressed domain
which can be done during or after the encoding process ([27],[28]). Alternatively,
techniques to work on raw format have also received attention. In general, invisible
watermark is embedded into irrelevant information of video data which can not be
perceived by human eyes. This results in problems because compression algorithms,
which also try to remove irrelevant information, might remove watermark
accidentally. In [27], this problem was solved by embedding watermark in the
relevant information of videos in an imperceptible manner. This approach is based on
geometric warping of blocks. This approach though very efficient produced visible
artifacts and video quality after extraction was low. Method to prevent these artifacts
and improve the video quality was proposed by [27]. They used geometric warping
method to embed watermarks, which did not require the original video during
watermark extraction. This method improved [23] by removing artifacts and video
 Improved Video Watermarking Using Nested Watermark, Wavelets 133

quality, but was slow for applications where time was significant. This paper proposes
a solution to solve this problem.

3 Methods Used
The concepts behind the various methods used in the proposed watermarking scheme
are explained in this section.

3.1 Visual Cryptography

Visual Cryptography (VC), a paradigm introduced by [22], was initially used to encrypt
material like written text, printed text, pictures, in a secure manner. It is a Visual Secret
Sharing Scheme (VSSS) which uses the Human Visual System (HVS) to decrypt a
secret message without expensive and complicated decoding process [31]. The basic VC
system starts with the encoding phase, where a secret image is divided into a collection
of ‘m’ black and ‘n’ white pixels. Each collection of m x n pixels is referred to as a
share, which will resemble a noisy image when separately. During decoding phase,
these shares or subset of shares are stacked together which will allow the visual
recovery of the secret message. It has been applied to many applications, including but
not restricted to E-voting system [24], financial documents [11], information hiding [4],
general access structures [3], visual authentication and identification [23]. More detailed
information about visual cryptography can be found in [35].
Out of the many VC schemes proposed, the present work considers the (2,2) VSS
scheme proposed by [22]. In this algorithm, each pixel of the copyright image is
expanded into 2 * 2 pixels. To share a white pixel of the secret image, one row from
the first 6 rows of table1 is chosen randomly. Similarly, the two shares of a black
pixel are determined by a random selection from the 6 last rows of Fig. 1. As a
result, an M*N pixels secret image is expanded into two 2M*2N pixels share-
images.
Considering security of the method, presence of only one share image reveals
nothing about the corresponding secret image, i.e., each 2*2 pixels block of one
share-image may correspond to either a white pixel or a black pixel of the secret
image. Stacking the shares of a black secret pixel results in 4 black subpixels, whereas
only 2 black subpixels is gained by stacking shares of a white secret pixel [12]. So,
secret image is revealed to human eyes by stacking the shares without performing any
cryptographical computations.
The scheme explained here is used in the proposed system and has the following
advantages.
► Simple to implement
► Encryption do not required any heavy computations
► Decryption algorithm not required (uses only Human Visual System).
► Even infinite computation power can not predict the secret message.
134 T. Jayamalar and V. Radha

3.2 Watermarking Using Geometric Warping

The geometric warping method used in this paper follows the methodology proposed
by [27] and the readers are referred to this paper for a detailed explanation. A
consolidated explanation is presented in this section.
The geometric warping works by embedding the watermark in the relevant and
visible part of the video in an imperceptible manner. This approach uses the Peak
Signal to Noise Ratio (PSNR) to perform geometric warping and work on the
principal that the visible-imperceptible approach can be realized if the watermark is
visible according to PSNR definition but imperceptible if someone sees the video.
This method changes the border position of the objects as they are untouched by the
compression algorithms. To decide for each block, which of the object borders to use
during embedding, a Normed Centre of Gravity (NCG) value is used. The NCG value
is calculated as follows.
1. Compute the mean values of the rows and columns of the block and store it
in Mx and My.
2. Using Mx and My, compute x- and y-coordinate of the NCV. (This arranges
the two vectors in two circles).
3. Compute the 2-D vector vk (k=x or y) using Equation (1)

 n  
  m k (i). cos π +  (i − 1). 2π    
 i =1   n    
vk =  n 

(1)

  m k (i).sin π +  (i − 1). 2π    
n
 i =1 n  n    
  
4. For each vector, compute the vector angles (Equation 2) and the vector
lengths Lx, Ly.
nθ k
k= for (k = x or y) (2)
2π
The main advantage of NCG is that it is robust again any compression algorithms.
The NCG of blocks with distinct object borders are robust to compression algorithms
while NCG blocks without distinct object borders are not robust. Distinct object
borders result in high vector lengths, using which the robustness of the NCG x,y-
coordinates can be predicted. The NCG x,y- coordinates distortion decreases with
increase in vector length. Through experiments, it was proved that the distortion for
vector lengths greater than 430 is lower than 0.25 pixels for compression rates upto
QP less than equal to 40. Hence, a watermark can be embedded by quantizing the
values of the x,y- coordinates of blocks with L > 430 b using a quantization step
size of one. This watermark will be robust to compression rates upto QP less than
equal to 40.
 Improved Video Watermarking Using Nested Watermark, Wavelets 135

During embedding of watermark, the first step is to find the vector length Lmin
which ensures the required robustness. After which, the blocks with L > Lmin are
chosen. To embed on watermark bit, the values of the x,y- coordinates of the NCG are
quantized. The quantization is done by geometric warping of the block. The direction
and strength of warping depends on the original and the quantized NCG-coordinates.
To prevent block artifacts to neighbor blocks, the warping process uses different
weights for the strength of warping the single pixels. The weighting is computed by
using a quadratic function.
This geometric warping process is used to change the NCG x,y- coordinates.

4 Proposed Watermarking Scheme

The proposed watermarking algorithm consists of two stages. The first stage is the
embedding stage and the second is extraction of the copyright. The process is shown
in Fig. 1 and Fig. 2.

Copyright Signal Cover Signal

Visual Cryptography Geometric Warping

Share 1 Share 2 Perform Block Selection

DWT Perform Block Linking

HH LH HL LL Embedding Process

IDWT

Nested Watermark Signal

Fig. 1. Embedding Process

136 T. Jayamalar and V. Radha

Watermarked Signal

Extract Embedded bits

Perform DWT

HH LH HL LL

IDWT

Extract Share 1 Extract Share 2

Overlap

Copyright

Fig. 2. Watermark Extraction Process

4.1 Embedding Process

The first step is to create nested watermark and the procedure is given in Fig. 3.

1. Perform Visual Cryptography on the original copyright image and obtain

Share 1 and Share 2.
2. Read Share 1
3. Perform 2D DWT on Share 1 to obtain LL, LH, HL and HH subbands.
4. Read Share 2
5. Add Share 2 in LL subband
6. Perform IDWT to get nested watermark.
7. Calculate Peak Signal to Noise Ratio of nested watermark with original
copyright image.

Fig. 3. Creating Nested Watermark

 Improved Video Watermarking Using Nested Watermark, Wavelets 137

The second step is the watermarking process using geometric warping. To avoid
blocking artifacts the geometric warping method uses a block selection procedure. In
block selection technique, the selected blocks should maintain minimum distance d to
each other in order to prevent artifacts and flickering effects. The minimum distance
can be calculated using Equation (3)
[di = log(d.(isize – 1) + 1) + 0,.5 (3)
where d = distance and i = dimensions (x, y or t).
The outline of the block selection algorithm is given as follows:
(i) The block group is chosen which is of arbitrary forms.
(ii) The minimum distance dx, dy and dt is computed for each dimension x, y
and t of the block group.
(iii) The numbers of blocks inside an ellipsoid are counted for each block. The
semi-axes of the ellipsoids are defined by dx, dy and dt. The current block
defines the center of the current ellipsoid.
(iv) The blocks with the least neighbors are chosen for watermarking. If more
than one chosen block is inside an ellipsoid, only one of them (the first) is
used for watermarking. This process is unambiguous.
(v) The chosen blocks and all blocks within the ellipsoids around the chosen
blocks will not be considered in the next steps.
(vi) The steps 3-4 are repeated until all blocks are chosen for watermarking or
have a distance lower than the minimum to a chosen block.
The blocks chosen using the above algorithm are used for watermark embedding. The
watermarking information is embedded by moving the x, y-coordinates on the
quantization lattice to minima/maxima. This is done by geometric warping.

4.2 Extraction Process

To extract the watermark bits, the blocks with L > Lmin are chosen. From the NCG
coordinates of the chosen blocks the embedded bit values can be computed. Thus, to
extract the watermark, the original video is not needed. The NCG is used to detect the
watermarked blocks and compute the embedded watermark bit. After successful
extraction of the watermarked bits, the next step is to extract the nested watermark.
The procedure used is given in Fig. 4.

1. Identify blocking containing watermark bits and extract them to get the
nested watermark
2. Apply 2D-DWT to obtain LL, LH, HL and HH subbands
3. Extract Share1 and Share 2 parts
4. Overlap to obtain the copyright image.

Fig. 4. Extraction Process

138 T. Jayamalar and V. Radha

5 Experimental Results
Several experiments were conducted to evaluate the performance of the proposed
system. During experimentation several samples were used which included standard
test videos and commercially-available movies. The results observed had similar
pattern. The results projected in this paper are based on a video clip having 1526
frames (352 x 289 frame size). The experiments were conducted on a Pentium IV
with 512 MB memory. The attacks taken into consideration are frame dropping, frame
averaging and compression. After extracting and refining the watermarks, a similarity
measure, called Normalized Correlation (NC) is used (Equation 4) which gives the
cross-correlation, normalized by the reference watermark energy giving unity as the
peak correlation.

 i  j Wij * RWij
NC = (4)
i  j W 2
i j
where Wij is the original watermark and RWij is the extracted watermark.
Number of frames dropped is a common attack encountered by video
watermarking. The reason behind this is the large amount of redundancies that exists
between the frames. The effect of this attack is projected in Fig. 5 when tested with
different percentage of frames dropped. From the results, it could be seen that the
proposed watermarking scheme is better with this type of attack.
Frame averaging and statistical analysis is another common attack to the video
watermark. When attackers collect a number of watermarked frames, they can estimate
the watermark by statistical averaging and remove it from the watermarked video. Fig.
6 shows the results obtained while taking the frame averaging attack into
consideration. The results show that fact that the proposed scheme can resist statistical
averaging quite well than the base model.

1
0.9
Normalized Correlation

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80
% of Frames Dropped
Base System Proposed System

Fig. 5. Number of Frames Dropped

 Improved Video Watermarking Using Nested Watermark, Wavelets 139

1 1
0.9 0.9

Normalized Correlation
Normalized Correlation

0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0

1000
100

200

300

400

500

600

700

800

900
0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Number of frames colluded Quality Factor
Base System Proposed System Base System Proposed System

Fig. 6. Frame Averaging Fig. 7. Compression Attacks

The NC values of the extracted watermarks with different quality factors are
shown in Fig. 7. From the experiment, it is noted that the proposed scheme improves
the robustness for watermark protection. It can further be noted, with high quality
factor, the algorithm is better in resistance.
From the various results obtained, it can be concluded that the proposed
watermarking scheme is superior to the existing system.

6 Conclusion
This paper proposed a technique for watermarking a video signal with a secret
copyright image. The proposed technique is an amalgamation of various techniques
like Visual Cryptography, Wavelets and Geometric Warping. Visual cryptography is a
concept that is normally used to for hiding multiple watermarks. In this paper, a novel
idea of splitting the copyright image into two and creating a nested watermark by
storing one split into another is proposed. This method increases the capacity of the
watermark. Experimental results proved that the proposed technique is efficient in
terms of capacity and robustness against attacks. The present work uses only black
and white copyright images, which can be changed to color images and a color visual
cryptography method can be combined with this. Moreover, applicability of wavelet
packets can be considered for more efficiency.

References
1. Aggarwal, D., Kaur, S., Anantdeep, E.: An Efficient Watermarking Algorithm to Improve
Payload and Robustness without Affecting Image Preceptual Quality. Journal of
Computing 2(4) (2010)
2. Arena, S., Caramma, M.: Digital watermarking applied to MPEG2 coded video sequence
exploiting space and frequency masking. In: Proceedings International Conference on
Image Processing (ICIP 2000), Vancouver, Canada, vol. 3, pp. 438–441 (2000)
140 T. Jayamalar and V. Radha

3. Ateniese, G., Blundo, C., De Santis, A., Stinson, D.R.: Visual cryptography for general
access structures. Information Computation 129, 86–106 (1996)
4. Bonnis, A., Santis, A.: Randomness in secret sharing and visual cryptography schemes.
Theor. Comput. Sci. 314, 351–374 (2004)
5. Checcacci, N., Barni, M., Bartolini, F., Basagni, S.: Robust video watermarking for
wireless multimedia communications. In: Proceedings IEEEE Wireless Communications
and Networking Confernce, WCNC 2000, vol. 3, pp. 1530–1535 (2000)
6. Chun-Shien, L., Shih-Kun, H., Chwen-Jye, S., Mark, L.H.: Cocktail watermarking for
digital image protection. IEEE Transactions on Multimedia 2, 209–224 (2004)
7. Craver, S., Wu, M., Liu, B.: What can we reasonably expect from watermarks? In:
Proceedings IEEE Workshop on the Applications of Signal Processing to Audio and
Acoustics, pp. 223–226 (2001)
8. Ejima, M., Miyazaki, A.: A wavelet-based watermarking for digital images and video. In:
Proceedings International Conference on Image Processing (ICIP 2000), Vancouver,
Canada, vol. 3, pp. 678–681 (2001)
9. Foo, S., Yeo, T., Huang, D.: An adaptive audio watermarking system. In: Proceedings
IEEE Region 10 International Conference on Electrical and Electronic Technology, vol. 2,
pp. 509–513 (2001)
10. Hartung, F., Girod, B.: Watermarking of Uncompressed and Compressed Video. IEEE
Transaction Signal Processing 66(3), 283–301 (1998) (Special issue on Watermarking)
11. Hawkes, L., Yasinsac, A., Cline, C.: An Application of Visual Cryptography to Financial
Documents; technical report TR001001, Florida State University (2000)
12. Houmansadr, A., Ghaemmaghami, S.: A Digital Image Watermarking Scheme Based on
Visual Cryptography. In: International Symposium on Telecommunications, pp. 1–5
(2005)
13. Huang, P.S., Chiang, C.S., Chang, C.P., Tu, T.M.: Robust spatial watermarking technique
for colour images via direct saturation adjustment. IEEE Proceedings Vision, Image and
Signal Processing 152, 561–574 (2005)
14. Kaabneh, K., Youssef, A.: Muteness-based audio watermarking technique. In: Proceedings
International Conference on Distributed Computing Systems Workshop, pp. 379–383
(2001)
15. Kimpan, S., Lasakul, A., Chitwong, S.: Variable block size based adaptive watermarking
in spatial domain. In: IEEE International Symposium on Communications and Information
Technology, ISCIT 2004, vol. 1, pp. 374–377 (2004)
16. Lee, Y.K., Chen, L.H.: High capacity image steganographic model. IEEE Proceedings
Vision, Image and Signal Processing 147, 288–294 (2000)
17. Lin, E., Podilchuk, C., Kalker, T., Delp, E.: Streaming Video and Rate Scalable
Compression: What Are the Challenges for Watermarking? In: Proceedings the SPIE
International Conference on Security and Watermarking of Multimedia Contents III, San
Jose, CA, vol. 4314, pp. 22–25 (2001)
18. Lu, C., Huang, S., Sze, C., Liao, H.Y.M.: Cocktail watermarking for digital image
protection. IEEE Transactions on Multimedia 2, 209–224 (2000a)
19. Lu, C., Liao, M., Chen, L.: Multipurpose audio watermarking. In: Proceedings 15th
International Conference on Pattern Recognition 2000, vol. 3, pp. 282–285 (2000b)
20. Lu, W., Lu, H., Chung, F.L.: Robust digital image watermarking based on subsampling.
Applied Mathematics and Computation 181, 886–893 (2006)
21. Mohammed, A.A., Hussein, J.A.: Efficient Video Watermarking Using Motion Estimation
Approach. In: 2009 Eight IEEE/ACIS International Conference on Computer and
Information Science (ICIS 2009), pp. 593–598 (2009)
 Improved Video Watermarking Using Nested Watermark, Wavelets 141

22. Naor, M., Shamir, A.: Visual cryptography. In: De Santis, A. (ed.) EUROCRYPT 1994.
LNCS, vol. 950, pp. 1–12. Springer, Heidelberg (1995)
23. Naor, M., Pinkas, B.: Visual authentication and identification. In: Kaliski Jr., B.S. (ed.)
CRYPTO 1997. LNCS, vol. 1294, pp. 322–336. Springer, Heidelberg (1997)
24. Paul, N., Evans, D., Rubin, A., Wallach, D.: Authentication for remote voting, workshop
on human-computer interaction and security systems, Fort Lauderdale, Florida (April
2003)
25. Piva, A., Bartolini, F., Barni, M.: Managing copyright in open networks. IEEE
Transactions on Internet Computing 6(3), 18–26 (2002)
26. Praun, E., Hoppe, H., Finkelstein, A.: Robust Mesh Watermarking. ACM SIGGRAPH
(1999)
27. Pröfrock, D., Schlauweg, M., Müller, E.: Video Watermarking by Using Geometric
Warping Without Visible Artifacts. In: Proc. of Information Hiding (8thIH 2006),
Alexandria, USA (2006)
28. Pröfrock, D., Richter, H., Schlauweg, M., Müller, E.: H.264/AVC video authentication
using skipped macroblocks for an erasable watermark. In: Proc. of Visual Communications
and Image Processing - VCIP, vol. 5960 (2005)
29. Qui, G., Marziliano, P., Ho, A.T.S., He, D., Sun, Q.: A hybrid watermarking scheme for
H.264/AVC video. Proc. of the ICPR 4, 865–868 (2004)
30. Ren-Junn, H., Chuan-Ho, K., Rong-Chi, C.: Watermark in color image. In: Proceedings of
the first International Symposium on Cyber Worlds, pp. 225–229 (2002)
31. Tai, G.C., Chang, L.W.: Visual Cryptography for Digital Watermarking in Still Images. In:
Aizawa, K., Nakamura, Y., Satoh, S. (eds.) PCM 2004. LNCS, vol. 3332, pp. 50–57.
Springer, Heidelberg (2004)
32. Verma, B., Jain, S., Agarwal, D.P., Phadikar, A.: A New color image watermarking
scheme, Infocomp. Journal of Computer Science 5(2), 37–42 (2006)
33. Wang, C., Nie, X., Wan, X., Wan, W.B., Chao, F.: A Blind Video Watermarking Scheme
Based on DWT,” iih-msp. In: 2009 Fifth International Conference on Intelligent
Information Hiding and Multimedia Signal Processing, pp. 434–437 (2009)
34. Wu, X., Guan, Z.H.: A novel digital watermark algorithm based on chaotic maps. Physics
Letters A 365, 403–406 (2007)
35. Yang, C.: A note on Efficient Color Visual Encryption, vol. 18, pp. 367–372 (2002)
 Robust Fault Tolerant AOMDV Routing Mechanism in
MANET

Jayalakhsmi Vaithiyanathan and B. SenthilrajaManokar

Computer Applications Department, Sudharsan Engineering College,

Sathyamangalam,Tamilnadu, India
[email protected]

Abstract. Wireless mobile ad-hoc networks are characterized as networks

without any physical connections. Multipath routing is a new extension to many
traditional routing protocols in MANETs. Using multipath routing in MANETs
can save energy, reduce frequent routing update, enhance data transmission
rates, and increase wireless network bandwidth. There are numerous Multi-
paths on-demand routing algorithms exist to discover several paths instead of
one, once the routing is performed. In this paper a new Robust Fault Tolerant
AOMDV (RFTA) algorithm is proposed to improve the fault tolerance and
efficient routing in mobile adhoc network (MANET). The algorithm is
compared with various multi path routing protocols like SMR, SMS and
MDSR. The efficiency of the protocol has been evaluated on different scenarios
using performance metrics such as packet delivery ratio and end-to-end delay.

Keywords: AODMV , end-to-end delay, fault tolerance, packet delivery ratio,

SMR, SMS.

1 Introduction
MANET is the network completely self-organizing and self-configuring, requiring no
existing network infrastructure or administration. Some routing policies [1] [2] [3]
have been proposed in order to establish and maintain the routes in MANET, and
DSR is a generally utilized one. There are many common routing algorithms used in
Adhoc networks but AODV [4, 5] and DSR [6], which both are on-demand
algorithms. In some multi-path algorithms once different paths are discovered, they
are all stored but only one of them is used for transferring the data. The other stored
paths will become useful once the current one is broken. There are also other multi-
path algorithms that transfer data over all discovered paths concurrently which
reduces end to end delay and increases end to end bandwidth.
Some key challenges Associated with an ad hoc environment, and some specific
characteristics.
Dynamic topology: The routing between the different nodes Consists of multi hop,
therefore routes may change Irregular. Due to the nodes Movement, the routing
information will need to be updated more often than its wired counterpart. This
results in more routing overhead information, which in turn increases the use of the

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 142–150, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Robust Fault Tolerant AOMDV Routing Mechanism in MANET 143

radio medium resources and puts great emphasize in choosing routing protocol,
when all the different protocols uses different strategies to establish and maintain
routes [7].
Multi-hop: The data packets have to be forwarded between multiple intermediate
nodes to reach the destination. In a MANET the topology may change constantly, due
to possibility of node movement [8] Bandwidth-constrained.
Variable capacity links: As a contrast to hardwired networks, communication over
wireless medium will continue to have less bandwidth capacity. In addition signal
interference, noise and fading, will most resulting in even less useful bandwidth.
Congestion is very common, will likely approach or exceeds the network capacity, as
the mobile networks often are simply an extension of a fixed infrastructure [9].
The typical problems encountered in designing single path protocols for a mobile
wireless environment also arise in designing multipath protocols. Mobility makes it
difficult or impossible to maintain a global view of the network. Mobility also has
implications in terms of caching policies. If cached information at intermediate nodes
is frequently out of date, caching can degrade routing performance because detecting
inaccurate routing. Information is not instantaneous. In addition to the dynamic
topology, unreliable and range limited wireless transmission makes resiliency a
requirement rather than an enhancement in a routing solution. Since mobile
transmitters are likely to be battery powered, routing protocols need to minimize the
communication for coordinating network nodes. At the protocol level, the design of
multipath routing needs to consider failure models, characteristics of redundant
routes, coordinating nodes to construct routes, mechanisms for locating mobile
destination and intermediate forwarding nodes, and failure recovery.
The paper is organized in 5 sections. Section 2 includes some definitions that gives
a useful background to this paper. Section 3 has a Review of Multipath Routing
Protocols. Section 4 gives the proposed RFTAOMDV algorithm. Section 5 illustrates
the performance evaluation with the other protocols. Section 6 presents conclusions at
the end.

2 Related Work
Most of multi-path routing algorithms such as AOMDV [10] have their basis in a
single-path algorithm due to their efficiency. One drawback with multi-path
algorithms is that they send more routing request packets, comparing to single path
algorithms and this makes them more complicated.

2.1 AOMDV

It is a simple, yet high efficient multi-path algorithm which tries to find Link-Disjoint
paths. In this algorithm the source broadcasts routing request packet to all its
neighbors. Once neighbors of the source node receive the RREQ packet, they add
their own address to the packet; announcing themselves as the founder of a path, and
broadcast it. Every intermediate node by receiving a request packet will check to see
if it has already received a same request and if not, then it reconstructs and broadcasts
144 V. Jayalakhsmi and B. SenthilrajaManokar

the request packet. If the same request was processed previously the following
conditions will be investigated and if met, the packet would be accepted and will be
inserted in the route table; otherwise the packet will be ignored.
These conditions are:
1. Whether the request packet is received from a new neighbor node.
2. Whether the request packet is received from a new path founder node.
3. Whether the number of hops in the request packet is less than the existing ones.
The destination node will send as many responses as number of request packets it has
received and hereby informs the source node of the paths.

2.2 Dynamic Source Routing (DSR)

This routing protocol consists of two mechanisms: the Route Discovery and the Route
Maintenance that permit to completely maintain and automatically determine routes:
Route Discovery is in charge to find routes between the nodes into the network by
broadcasting RREQ. DSR is a reactive protocol, a source node S starts searching
routes only if it needs to send a packet to a destination D. Any intermediate node that
receives a non-duplicate RREQ appends its address to the source route list in the
RREQ packet and rebroadcast it and the destination node sends a RREP back to the
source. If an intermediate node has some requested information, it may send an RREP
back to the source. Finally, the source node obtains several routes to reach the
destination.

3 Review of Multipath Routing

3.1 Split Multipath Routing (SMR)

SMR[11] protocol provides a way of determining maximally disjoints paths. Paths

are maximally disjoint when they are node disjoint, but when there are no node-
disjoint paths available, the protocol minimizes the number of common nodes.
Multiple routes are discovered on demand, one of which is the path with the shortest
delay. The routes established by the protocol are not necessarily equal in length.
When a source needs a route to a destination but no information is available, it floods
a RREQ message on the entire network. Because of this flooding, several duplicates
that traversed through the network over different routes reach the destination. The
destination then selects multiple disjoint paths, and sends RREP packets back via the
chosen routes. Since the destination needs to select disjoint paths, source routing is
used. The complete route information is in the RREQ packets. Furthermore,
intermediate nodes are not allowed to send RREPs, even when they have route
information to the destination. If nodes reply from their cache as in DSR, it is very
difficult to determine maximally disjoint multiple routes since the destination does not
receive enough RREQs and will not know the information of routes formed from
intermediate nodes cache. In SMR a new packet-forwarding approach is introduced.
Instead of dropping all duplicate RREQs, intermediate nodes only forward the
packets, which used a different incoming link than the first received RREQ and
whose hop count is not larger than that of the first received RREQ The SMR
 Robust Fault Tolerant AOMDV Routing Mechanism in MANET 145

algorithm is optimized when the destination selects the two routes that are
maximally). One of the two routes is the route with the shortest delay; the path taken
by the first RREQ the destination receives. This path is used to minimize route
acquisition latency required by on-demand protocols. When the first RREQ is
received, the destination sends a RREP to the source via this path. The RREP includes
the entire path so the intermediate nodes can determine where to forward the packet.
After having sent the first RREP, the destination waits a certain amount of time to
receive more RREQs and determine all possible routes. Because the destination then
knows route information from all possible routes it can determine the maximally
disjoint route to the already replied route.

3.2 Multi-Path DSR (MP-DSR)

M-DSR first extends DSR by labeling the initial route request received at a destination
node the primary source route. additional route requests received are from node disjoint
routes, given that route requests are forwarded at most once by any given node involved
in any given route discovery. These additional routes are generally longer, as these route
requests arrive after the primary source route's request arrives. Route replies for each
route are sent to the source. The M-DSR route usage policy dictates secondary routes are
only used after the primary route has failed. The use of previously discovered secondary
routes reduces route discovery overhead. If only one RREQ is received by the
destination, that means there is only one path from source to destination, M-DSR acts
essentially the same as the single path DSR protocol. M-DSR offers two mutually
exclusive multi-path maintenance heuristics. In M-DSR have been developed two
versions of protocols. It prevents intermediate node route caching for alternate routes,
which implies that only the source is capable of fixing a broken path.

3.3 Shortest Multipath Source (SMS)

SMS routing protocol is proposed based on DSR. It builds multiple partial-disjoint

paths from source S to destination D in order to avoid the overhead of additional route
discoveries and to quick recovery in case of route breaks. Improved performance in
terms of Fraction of packets delivered, end-to-end delay and routing overheads is
obtained. SMS earns multiple partial-disjoint paths that will bypass at least one

(a) (b)

Fig. 1. (a) Partial Disjoint Multiple Paths (b) Select the alternate path
146 V. Jayalakhsmi and B. SenthilrajaManokar

intermediate node on the primary path that is shown in Fig. 1-a.Consider the case of
traffic flowing between nodes Sand D using link A-B as a segment of the path. In case
of a link failure between A and B, the source node will search for an alternate route
that does not contain node B that is shown in Fig. 1-b. An alternative route between
source S and destination D is (S, A, F, C, D).

4 Proposed RFTA Algorithm

In the proposed RFTA( Robust Fault Tolerant AOMDV) algorithm , there is no need
for the intermediate nodes to have Route Cache tables and as a result the destination
will receive all the path-request messages from different paths. Also in the proposed
algorithm, each node should save the RREQ messages it receives from other nodes in
RREQ_Seen table, so that it can respond to the queries it receives from its neighbors.
Also, there is an additional field in RREQ_Seen table of each node, called
Neighbor_Count (N_C) which will be used to count the number of active neighbors
identified after sending the RREQ message. In order to let the subsequent nodes to
know the total number of active neighbors of the traversed nodes along a path, a new
field called Neighbor Count is added to the headers of both RREQ and RREP
messages. Also two new messages, RREQ_Query and RREQ_Query_Reply, are
added to the route discovery process.

4.1 Pseudo Code of RFTA

The steps taken by source node, destination node and intermediate nodes are listed in
the following tables.

Table 1. Algorithm for the Source node in RFTA

Steps

1. If there is data to be sent to a certain destination and there is no

valid path for that destination, broadcast the RREQ packet
2. Wait for first RREP packet to arrive.
3. After receiving the first RREP packet, wait for a certain
amount of time, then from received paths choose those ones
that have the least active neighbors and starts load balancing
data on these paths.

Table 2. Algorithm for the Destination node in RFTA

Step
1. Send back a RREP packet to all the nodes from which a RREQ
packet is received
 Robust Fault Tolerant AOMDV Routing Mechanism in MANET 147

Table 3. Algorithm for the Intermediate node in RFTA

Steps
1. Once a RREQ message is received and is acceptable(based on three
aforesaid conditions in the Related Works section) perform the following
steps:

i. Save this message in the RREQ_Seen table

ii. Construct the RREQ_Query packet

iii. Send the RREQ_Query packet to the neighbours asking them if

they have already seen the same RREQ message before

iv. Wait for a certain amount of time for RREQ_Query_Reply

messages from neighbours

v. Increase the ActiveNeighborCount field based on the positive

responses that are received

vi. Broadcast the RREQ message with the new

ActiveNeighborCount value

2. Once a RREQ_Query message is received one of the following steps are

performed:

i. If based on the RREQ _Seen table this is a new RREQ

message. Store the RREQ in the RREQ Seen table.

ii. If based on RREQ_ Seen table the same RREQ has been
received from the same node, ignore the query.

iii. If the RREQ is an existing one in the RREQ _Seen table

and it has been received from a new neighbor but the same
RREQ had not been broadcasted before ,then only respond
with sending back the positive RREQ_ Query_ Reply.

iv. If the RREQ is an existing one in the RREQ_ Seen table

and it has been received from a new neighbor and the same
RREQ has already been broadcasted then respond with
sending back the positive RREQ_ Query_ Reply and also
increment the After A_N_C field of corresponding RREQ
in the RREQ_ Seen table.

3. Once the RREP message is received add the content of the corresponding
After N_C to the Active NeighborCount field of RREP and send the
RREP.
148 V. Jayalakhsmi and B. SenthilrajaManokar

5 Performance Evaluation Results

5.1 Simulation Environments

The simulation environment is ns2[12]. The network includes 50 mobile host nodes
which are placed in 1000*500 meter flat area in random and the number of source
nodes was 30 of 50 hosts. Each node has a radio propagation range of 250 m and
channel capacity was 2Mbps [5]. The IEEE 802.11 Distribution Coordination
Function (DCF) is used as Medium Access Control (MAC).A traffic generator was
random constant bit rate (cbr) and Tep was established between nodes. The size of
data payload was 512 bytes. The random waypoint model was used as a mobility
model. Sending rate of packets was set to 3 packets per second and the simulation
was run in 30 seconds as a pause time. The maximum and minimum speed was varied
between 0 and 10 m/s [6].
Three important metrics are evaluated:
• Throughput: measured as the ratio of the number of data packets delivered to
the number of data packets sent by the source.
• End-to-end delay: measured as the average end-to-end latency of data
packets.
• Routing loads: measured in terms of the total number of routing packets
transmitted per data packet delivered at the destination. Each broadcast
transmissions was count as one single transmission.

5.2 Simulation Results

Throughput of the four protocols is shown in Fig. 2. From the resu1ts SMS
outperform SMR especially when the mobility increases. In DSR only one route is
used for each session and when that route is invalidated the source uses the cached
route that is learned from overhearing packets. Intermediate nodes that have cached

Fig. 2. Fraction of Packets Delivery vs. Speed

 Robust Fault Tolerant AOMDV Routing Mechanism in MANET 149

Fig. 3. End-to-End Delay vs. Speed

routes to the destination provide those routes to the source by sending RREPs. SMS
outperforms DSR, SMR and MP-DSR due to lower overheads in the route
discovery process since the destination node sends multiple RREP. In moderately
loaded networks, the loss of RREP packets results in the failure of the route
discovery process. In on demand protocols the source uses an exponential back-off
algorithm to limit the rate at which new route discoveries are initiated. Since the
back-off time is large compared to send buffer capacity, application packets within
the back-off time may be dropped due to buffer overflow. But with this proposed
RFTA algorithm which use multiple route replies show higher Throughput than
the other protocols.
The end-to-end delay of the four protocols is shown in Fig 3. SMR has the longest
delay in mobile scenarios because it delivers data packets on routes longer than those
of SMR. In addition, MP-DSR yields longer delays in reconstructing routes and the
period of time the data packets are buffered at the source node during route recovery
results in larger end-to-end delays.
SMR uses the remaining valid route when one of the multiple routes is
disconnected, and hence no route acquisition latency is required. The availability of
shorter alternate routing paths in the former eliminates route discovery latency that
contributes to the delay when an active route fails. When a congestion state occurs in
a path, the source node can re-distribute incoming data packets to alternative routing
paths, thereby reducing the queue waiting time of data packets.
The proposed algorithm routing outperforms the other protocols at lower mobility
and node densities. SMS finds the shortest alternative routes in a single route
discovery process, reducing delay and routing overheads incurred when recovering
from route breaks. SMR overloads the network with control messages and data
packets are dropped at full buffer of intermediate nodes. Routing load savings for M-
DSR at higher nobilities and node densities came from a large saving on route
requests.
150 V. Jayalakhsmi and B. SenthilrajaManokar

6 Conclusion
The proposed Robust Fault Tolerant AOMDV algorithm explores alternate disjoint
routes that can be useful in case the primary route breaks. Two variations are
explored. In the first, only the source gets multiple alternate routes. In the second,
each intermediate node on the primary route gets an alternate route. The key
advantage is the reduction in the frequency of route discovery flood, which is
recognized as major overhead in on demand protocols. In any case, any form of
multipath technique always performs substantially better than single path routing. The
simulation study also demonstrates that even though the multipath routing brings
down routing load, it also increases end-to-end Delay, as alternate routes tend to be
longer than primary routes. This presents subtle tradeoffs between delay and routing
load. Simulation results indicate that the proposed RFTA outperforms SMS, SMR and
MP-DSR on consideration of a number of metrics and is a candidate that, potentially,
offers improved routing performance. And there has been a noticeable improvement
in the packet delivery ratio and also in the reduction of end-to-end delay comparing to
other protocols.

References
[1] Johnson, D.B., Maltz, D.A., Hu, Y.C.: The Dynamic source routing protocol for mobile
ad hoc networks (DSR). IETF draft-ietf-manet-dsr- 10.txt (July 2004)
[2] Perkins, C.E., Belding-Royer, E., Das, S.: Ad hoc on demand distance vector (AODV)
routing. IETF RFC3561 (2003)
[3] Perkins, C.E., Watson, T.J.: Highly dynamic destination sequenced distance vector routing
(DSDV) for mobile computers. In: Proc. of ACM SIGCOMM, London, UK (1994)
[4] Perkins, C.E., Belding-Royer, E.M., Das, S.: Ad hoc On-Demand Distance Vector
(AODV) Routing. RFC 3561 (July 2003)
[5] Perkins, C.E., Royer, E.M.: The Ad hoc On-Demand Distance Vector Protocol. In:
Perkins, C.E. (ed.) Adhoc Networking, pp. 173–219. Addison-Wesley, Reading (2000)
[6] Johnson, B., Maltz, D.A.: Dynamic Source Routing in Ad-Hoc Wireless Networks.
Mobile Computing 353, 153–181 (1996)
[7] Li, X., Cuthbert, L.: A Reliable Node-Disjoint Multipath Routing with Low Overhead in
Wireless Adhoc Networks. In: Proceedings of the Seventh ACM/IEEE International
Symposium on Modeling, Analysis and Simulation of Wireless and Mobile Systems,
Venezia, Italy, pp. 230–233 (2004)
[8] Mueller, S., Ghosal, D.: Multipath Routing in Mobile Ad-hoc Networks: Issues and
Challenges. Invited paper in Lecture Notes in Computer Science (2004)
[9] Nasipuri, A., Castaneda, R., Das, S.: Performance of Multipath Routing for On-demand
Protocols in Mobile Ad-hoc Networks. ACMlKluwer Mobile Networks and Applications
Journal 6, 339–349 (2001)
[10] Marina, M.K., Das, S.R.: On-demand Multipath Distance Vector Routing for Ad Hoc
Networks. In: Proc. of 9 th IEEE Int. Conf. on Network Protocols, pp. 14–23 (2001)
[11] Lee, S.j., Gerla, M.: Split Multipath Routing with Maximally Disjoint Paths in Ad-hoc
Networks. In: Proceedings of IEEE International Conference on Communication (ICC),
Elsinki, Finland, pp. 3201–3205 (2001)
[12] Network Simulator-2, http://www.isi.edu/nsnam/ns
 ZKIP and Formal System

M. Thiyagarajan1 and S. Samundeeswari2

1
Professor,
School of Computing, SASTRA
[email protected]
2
Assistant Professor,
School of Computing, SASTRA
[email protected]

Abstract. Every crypto system is an interactive system, every interactive

system is a formal system, the results of the formal system can be proved
through the Godel’s incompleteness Theorem. Proving techniques in the
solution of problems designed on formal systems employ logical connectivity.
This logical programming is possible from the proof of Godel’s incompleteness
theorem in the formal systems. Any intractable problems can be solved by
heuristic approach. One such approach is the Godel’s attempt, we realize the
automated steps translating the proof of Godel’s theorem as our zero knowledge
interactive steps. The implementation of this approach is being illustrated.

1 Introduction
The implementation validation of protocols as ZKIP is our contribution in this area of
cryptography. On classification of complexity classes which are driven from the study
of formal language and automated theory. Such hypothesis helps us to decide on the
particular complexity class. Some of the problems studied as NP complete problems
adapt themselves to a reduction to polynomial time problem. This type of conversion
is perceived as a function which converts Non Interactive Zero-Knowledge Protocol
[2] to Zero-knowledge Interactive Protocol. Godel’s Incompleteness theorem and the
proof by the earlier researchers in the foundation of mathematics pave the way in the
conversion of Non Interactive Protocol to Interactive protocol.
In the most recent techniques of proving general results on formal systems is the
Zero knowledge Interactive Proof. Here the Godel’s theorem and proof are
represented in the formal system set up between corporate concepts such as
consistency and ω _consistency[6]. The system is developed through Godel number,
Godel sequence and primitive formulas. The connectivity of these concepts with
propositional connectives and quantifiers is correct because they are contained in
ordinary logical connectives. The matching of expression for the Godel formula
through consistency and ω _consistency are used in developing the interaction
between the prover and verifier. The input is a goal which may contain a constraint
such as a time limit and a partial proof, which is a list of proof steps. A proof step
contains a proposition or an assumption or an explanation which gives as arguments,
the proceeding proof steps are used to derive the propositions. A look up table is

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 151–161, 2011.
© Springer-Verlag Berlin Heidelberg 2011
152 M. Thiyagarajan and S. Samundeeswari

supplied by the prover at the request of the verifier, in proving the two main
implications of the Godel’s incompleteness theorem.
ZKIP[3] contains steps which are related to reasoning process of mathematics and
guide the search for the proof reductio ad absurdum is used to prove the negation of a
proposition. We generate formulas and sets which form the central ideas of significant
proof. The architecture and model of operation of ZKIP are illustrated in detail by our
proof of Godel’s incompleteness theorem, which says, every formal number theory
contains an un decidable formula, i.e., neither the formula nor its negation are
provable in the theory.
Here construction in the proof of Godel’s namely Godel number, Godel
sequence and Godel formula together with formal system developed with these put
two different main streams consistency and ω -consistency. This leads to a
learning procedure of a verifier to get the steps from the prover who gives
directions and paves the way for subsequent calculations and decisions leading to
the ultimate proof. Godel’s theorem and the proof are represented in the formal
languages which incorporate concepts such as concepts of substitutions with
special relational procedure which models the ordinary representation of proofs in
soft computing.

2 Preliminaries

2.1 Definition 1

A formal theory is consistent if for no formula F both F and –F are provable. A

formalization is : [ 5]
allT ( fnt(T ) → (consistent(T ) ↔ allF ( provable( F , T ) → not( provable(¬F , T )))))

2.2 Definition 2

A formal theory is ω − consistenet if every formula F (x) has the property: If ├

F (n), i.e., F (n) is provable, for all natural numbers n , then ¬∀xF (x) is not
provable. A formulization is : [5]
allT ( fnt (T ) →
(ω − consistenet (T ) ↔
allF , x(al ln(number ( n) → (1)
provable( sub( F , x, num(n)), T ) →
not ( provable(¬∀xF , T )))))
There is a closed formula F of formal number theory such that
• If the theory is consistent, F is not provable, and
• If the theory is ω − consistent , ¬F is not provable.
 ZKIP and Formal System 153

A formalization is:
allT ( fnt (T ) →
(existsF (closed − formula( F , T ) &
(2)
(consistent (T ) → not ( provable( F , T ))) &
(ω − consistenet (T ) → not ( provable(¬F , T )))))

2.3 Definition 3: Formal System [6]

A formal system is ω -inconsistent if it contains a formula ϕ(x) such that both

(∃x) ϕ( x) ( ∃ x) ϕ (x) (3)

and
ι ϕ(0), ι ϕ(1), ι ϕ( 2),... (4)

are provable in the system. If no such formula exists, then the system is called ω -
consistent. Evidently ω -consistency implies ordinary consistency (since inconsistency
implies all formulas provable), but the converse does not necessarily hold.

2.4 Proposition 1 [6]

If the formal system S is ω -consistent, then the formula (∃y ) Γ(i, y ) is not
provable in S.
2.5 Proof 1

Suppose that it is provable. Then B( j, k ) cannot hold any k, since otherwise

(3.10) is provable in S and S k does B ( j , k ) hold.
is inconsistent. Thus, for no
That is, B ( f (i, i ), k ) holds for no k , implying that G (i , k ) holds for no k .
Thus, for each natural number k . ~ Γ(i, k ) is provable in S . In particular then
all formulas

~ Γ(i ,0), ~ Γ(i ,1 ), ~ Γ(i , 2), (5)

are provable in S. But, since (∃y ) Γ(i, y ) is also provable, this implies that S
is ω -consistent. We must conclude, then, that (∃y ) Γ(i, y ) is not provable in S.

2.6 Theorem 1 [6]

If the system S is ω -consistent, then it cannot be completeness.

154 M. Thiyagarajan and S. Samundeeswari

The main problem in the proof of Godel is a construction of an un decidable

problem. In order to construct such a formula a mathematician may apply elementary
rule for the formulation of formula to a known formula and resulting attempt to prove
one of the resulting formulas is un decidable.

2.7 Proposition 2 [6]

If a formal system, S , such as that described above, is consistent, the formula (3.10)
is not provable in S.
2.8 Proof 2

Suppose that it is provable in S , and let k be the sequence number of such a proof.
Then B (i, k ) holds and, by virtue of j = f (i, i ) , G (i, k ) holds. But

then Γ (i, k ) , is that provable in S , for it may be shown that, whenever particular
numbers are in some recursive relationship, the formula of the system which
(∃y ) Γ(i, y ) is
expresses this fact is formally provable. It then follows easily that
provable. Since our starting assumption was that (∃y ) Γ( z , y ). is provable in S ,
it follows that S is not consistent S . However, S was assumed consistent, so we
conclude that (∃y ) Γ( z , y ). is not provable in S .
To show that the formula

(∃y ) Γ(i, y ) (6)

Is also not provable, Godel used the assumption of “ ω - consistency”;

2.9 Properties of Zero Knowledge Proof

Zero-knowledge proofs (ZKPs) are interactive proofs (protocols) in which the prover
proves the knowledge (or the possession) of a secret without revealing any
information about the secret itself. ZKPs can be used whenever there is critical data to
exchange while only proving the possession of such data is needed, without a real
need for exchanging the critical data. [7]

2.9.1 Completeness Property

For any c > 0 and sufficiently long x ∉ L, Probability ( V accepts x )
−c
> 1− x . In other words, an interactive proof (protocol) is complete if, given an
honest prover and an honest verifier, the protocol succeeds with overwhelming
probability.[1]
 ZKIP and Formal System 155

2.9.2 Soundness Property −c

For any c > 0 and sufficiently long x ∉ L, Probability ( V accepts x ) < x , i.e.
negligible, even if the prover deviates from the prescribed protocol. In other words, if
a dishonest prover P ' can successfully execute the protocol with non negligible
probability, then P ' has knowledge essentially equivalent to the actual secret [1].

2.9.3 Zero Knowledge Proof

An interactive proof < P, V > is called zero-knowledge if for every probabilistic
polynomial-time V * , there exists a probabilistic expected polynomial-time simulator
(algorithm) M v* that on inputs x ∈ L produces probability distributions M v* ( x )
polynomially indistinguishable from the distributions < P, V > ( x ) . [1]
*

2.9.4 Interactive Proof (IP)

An interactive proof < P, V > for language L is a two-party protocol in which a
computationally unrestricted prover, P , interacts with a probabilistic polynomial-
time verifier, V by exchanging message. Both parties share a common input x . At
the end, V either accepts or rejects and both completeness and soundness properties
hold. [1]

3 Objective
In this section, we first describe the model of our ZKIP protocol for Godel. Then, we
present the pseudo code representation of the Godel incompleteness theorem.

3.1 Our Model

Following sequence of steps leading to a goal

1. The proof attempts to have a proposition. If the proposition is not
contained in the proof, we go to universal proposition as implication, an
existential proposition or a negation
2. Existential proposition ∃xp (n)
3. Component – domain and ranges of elementary functions which form
the elementary building stars of formulas and prediction.
4. Basic of negation reduction , contradiction and formulas
5. Data exchanged --basic parameters can reaches through Godel building
blocks.
6. If the theory is consistency F is not provable.
7. If theory is ω -consistent negation F is not provable.
The problem is un decidable if there is no algorithm that takes input
and incidences of the problem and determines whether that answer to
that incidence is yes or no. The theory of un decidability is concerned
with the existence and non existence of algorithm for solving problems
with an infinity of incidence.
156 M. Thiyagarajan and S. Samundeeswari

The main problem in the proof of Godel’s theorem is the construction of an

undecidable formula, for which elementary rules are applied for the formation of
formulas to a known formula and attempt to prove that one of the resulting formulas
is undecidable. An example of a rule for the formation is that ¬F is a formula if
F is a formula. Another rule is that ∀xF is a formula if x is a variable and F a
formula. The application of the two rules to the variable x and the formula
A(x) yields the formulas ¬A(x) and ∀xA(x) .

3.2 Pseudo Code

Step 1: Generation of Godel number and Godel sequence

Step 2: The statement of the formal system and the composition taken in the form
of Godel number and Godel sequence.
Step3: This formalization is translated in to a formal system in which either a
statement A or its negation is true.
Step 4: The process consists of considering statements on the premises as logical
statements using implication and contradiction.
Step 5: These steps are developed in the existence of an undecidable formula. This
formula if it is true for all possible combination of earlier implication, contradiction
and truthful nature makes the statement universe. That is true of all combinations of
truth value for the component statements.
Careful connectivity of following statements yields the proof of the incompleteness
theorem.

Composition

closed − formula (∀ y ¬ A ( k , y ), T ) (7)

Contradiction

allF ( provable( F , T ) → not ( provable(¬F , T )))

assume( provable(∀y¬A(k , y ), T ))
consistenet (T )
(8)
not ( provable(∀y¬A(k , y ), T ))
assume( A(k , n))
not ( A( k , n))
 ZKIP and Formal System 157

Forward

existsn ( gnp ( n , ∀ y ¬ A ( k , y ), T ))
assumen ( gnp ( n , ∀ y ¬ A ( k , y ), T ))
A(k , n)
provable ( A ( k , n ), T (9)
provable ( ∃ yA ( k , y ).T )
provable ( ¬∀ y ¬ A ( k , y ), T )

not ( provable(¬∀yA(k , y ), T )
gnp (n, ∀y¬A(k , y ), T )
(10)
existsn( gnp (n, ∀y¬A(k , y ), T ))
provable(∀y¬A(k , y ), T )

4 Data and Analysis

4.1 Derivation of Godel Number Formula

We introduce the formal system in the foundation of mathematics through the

derivation of Godel number and description of Godel sequence. The crucial concept
w -consistency is used to state and prove the Godel’s incompleteness theorem.
Consider a formula system whose primitive symbols F , G ,..., and further
symbols S ,0, = ι, ⊃ , (formulas,). In the interpreting formulas constructed from
these symbols we shall regard the individual variables as ranging over the natural
numbers, while the functional symbols stand for properties and relations of numbers,
S will denote the successor function (so that S x is the number x + 1), while 0, =,
and the logical connectives have their usual meaning. Thus the formula

( x) F ( x). ⊃ .(∃y )(G ( y ) = SSO)) (3.7)

Thus the formula would be read “If ever number x has the property F , then there
exists a number y such that either y has the property G, or y is 2”. In the formal
system, incidentally, we shall denote SSO by 2 , SSSO by 3 , etc.
We precede to “arithmetize” the formal system as follows. We first assign to each
primitive symbol a coordinate in the following system as follows. We first assign to
each primitive symbol a coordinate in the following manner (in the groupings below,
symbols in the first line are within the system, and the numbers in the second line are
the coordinates of the symbols directly above):
158 M. Thiyagarajan and S. Samundeeswari

For the basic logical symbols we make the correspondence:

0 S  ⊃ ( ) =
2 3 4 5 6 7 8

For the individual variables (which can be denumerable infinite in number):

x y z ...
11 112 113 ...

For functions of the first order (functions of individual variables):

F1 G1 H1
13 132 133
and so on, using powers of the ( n + 5) prime number for function symbols of the
th

nth order.
To a formula in the system will correspond, then, a sequence of numbers, namely,
those corresponding to the successive symbols of the formula.
Thus, to
(x)F1(x) (11)
Will correspond to the sequence
6 11 7 13 6 11 7
We can make a unique number correspond to the formula by taking the product of the
successive prime numbers (in their natural order) with powers equal to the numbers of
the symbols (in the order in which they occur). Thus to formula above will correspond
the number
26 . 311 . 57 . 713 . 116. 1311. 177 (12)
We shall call this the Godel number of the formula.

4.2 Description of Godel Sequence [41]

For any sequence of formulas we may make a correspond unique number

n(k )
2 n ( i ).3n ( 2 )... p k
(13)

Where n(1), n( 2),...n( k ) are the successive Godel numbers of the formulas of
sequence. We call (3.8) the sequence number of the sequence of formulas. In
particular every formal proof (that is proof with in the system) corresponds a
sequence number, since a formal proof of a formula F consists of a finite sequence
of formulas each of which is an axiom, or derived from axioms and preceding
 ZKIP and Formal System 159

formulas by the allowable rules of inference of the system, the last formula of the
sequence being F itself.
Let B ( x, y ) denote that y is the sequence number of a formal proof in which
the last formula has Godel number x. And let f ( x, y ) be the Godel number of the
formula obtained by replacing, in the formula having Godel number x, each
occurrence of the particular individual variable z by y . Godel shows that both B
and f are recursive.
Now form the relation B ( f ( x, x), y ). This is also recursive: denote it by
G ( x, y ). Being recursively definable, it has a counterpart in the system which
denote by Γ ( x, y ). Then by virtue of the rules for constructing formulas, the system
contains the formula.
ι (∃y ) Γ( z , y ). (14)
Let Godel number of this formula be i.
n(k )
A new formula is obtained by replacing z in n (i )
2 .3 n( 2)
... p k
by i
j = f (i, i ) (15)

4.3 Validation

From the above discussion we can decide on the completeness of a formal system. As
many NP complete problem can be reduced to polynomial time problem, the ω -
consistency helps the reduction of formal system to one of polynomial formal system.
Here the Proof of the Godel’s Incompleteness theorem is used as a Zero
Knowledge Interactive Proof. The formal system’s complexity is an undecidable
problem, We do not have any proof to validate a proposition in the formal system or
its negation. The system is converted to an interactive proof system where there is an
efficient proof verification algorithm called a verifier following the conditions of
completeness and soundness. This is an attribute of the Godel’s assertion that every
consistent formal system is not provable and ω _consistency formal system is not
also provable in number theory. This is whole analysis of ZKI protocol developed by
Golwasser, Micali, and Rockoff[]

5 Conclusion
Formal system addresses the problem based on logic programming in almost all
branches of science, engineering and technology. This system is based on semantic
and other aspects of programming techniques and their certainty alone with due
consideration on acceptable valid conclusions from the major premises. On the other
hand ZKIP (zero knowledge interactive protocol) can be applied to situations under
deterministic interactive and non interactive type of proving techniques as well as
160 M. Thiyagarajan and S. Samundeeswari

uncertainties dealt with probabilistic approach. The step involved in the automated
proof of Godel’s incompleteness theorem regarded as interactive steps. This analytical
approach leads to understand crypto system protocol design as ZKIP. This work being
general in nature can be adapted to any type of protocols designing a crypto system as
a formal system. As the logical programming can be implemented in standard
packages, this approach translates the code for implementing in an easy way. We feel
strongly ZKIP attains better status in the proving techniques than formal systems.

References
[1] Almuhammadi, S.: One-Round Zero Knowledge Proofs And Their Application In
Cryptographic Systems. University of California, Berkeley (2005)
[2] Blum, M., Feldman, P., Micali, S.: Non interactive zero-knowledge and its application. In:
Proc. Twentieth Annual ACM Symposium on Theory of Computing, Chicago, pp. 103–
112 (1988)
[3] Diffie, W., Hellman, M.E.: New directions in cryptography. IEEE Trans. Inform.
Theory IT-22(6), 644–654 (1976)
[4] Goldwasser, S., Milcli, S., Rackoff, C.: The knowledge complexity of interactive proof-
systems. SIAM Journal on Computing 18, 186–208 (1989)
[5] Amman, K.: An Automatic proof of Godel’s Incompleteness theorem. Arttificial
Intelligence, Vol 61, 291–306 (1993)
[6] Ray, M., Wilder, L.: Introduction to the Foundation of Mathematics. John Wiley and
Sons, Inc., New York (1965)
[7] Rhee, M.Y.: Cryptography and Secure Communications. McGraw-Hill, Korea (1994)
 ZKIP and Formal System 161

Appendix

Problem for Structured proof of Godel’s Incompleteness theorem and its

analysis
#include<iostream.h>
#include<conio.h>
#include<math.h>
void main()
{
clrscr();
int seq_no,i,seq[20],p[20];
double ans=1;
cout<<"GODEL SERIES IMPLIMENTATION\n";
cout<<"enter the number of sequence to used\n";
cin>>seq_no;
cout<<"\nenter the sequence\n";
for(i=1;i<=seq_no;i++)
cin>>seq[i];//entering the sequence...
//prime number sequence...
p[1]=2;p[2]=3;p[3]=5;p[4]=7;p[5]=11;p[6]=13;p[7]=17;
p[8]=19;p[9]=23;p[10]=29;p[11]=31;p[12]=37;p[13]=41;p[14]=43;
cout<<"\nsequence is\n";
for(i=1;i<=seq_no;i++)
cout<<" "<<seq[i];
cout<<"\n";
for(i=1;i<=seq_no;i++)
cout<<p[i]<<".";
for(i=1;i<=seq_no;i++)
ans*=pow(p[i],seq[i]);
cout<<"\nthe number is:"<<ans;
getch();
}

The input is as follows:

GODEL SERIES IMPLIMENTATION
enter the number of sequence to used
3
enter the sequence
6
11
6
The output is given by:
sequence is
26 311 56
the number is:1.77147e+11
 Personalized Web Based Collaborative Learning in
Web 3.0: A Technical Analysis

S. Padma1 and Ananthi Seshasaayee2

1
Research Scholar, Bharathiar University, Coimbatore
[email protected]
2
Quaid-e-Millath Govt. College for women, Chennai
[email protected]

Abstract. Web 3.0 is an evolving extension of the current web environment.

Information in web 3.0 can be collaborated and communicated when queried.
Web 3.0 architecture provides an excellent learning experience to the students.
Web 3.0 is 3 D, media centric and semantic. Web based learning has been on
high in recent days. Web 3.0 has intelligent agents as tutors to collect and
disseminate the answers to the queries by the students. Completely Interactive
learner's query determine the customization of the intelligent tutor. This paper
analyses the Web 3.0 learning environment attributes .

Keywords: E learning, Web 3.0, Semantic ,Collaborative learning.

1 Introduction
Web 3.0 is an evolving extension of www, in which the information can be shared and
interpreted by other software agent to find and integrate applications to different domains.
Web 3.0 provides integrated real time application environment to the user. The
applications are majorly involved in searching using semantic web, 3D web and are
media centric. Web 3.0 supports pervasive components. Each component and its
relations are represented below.
In web 3.0, web is transformed into database or Data Web wherein the data which
are published in the web is reusable and can be queried. This enables a new level of
data integration and application interoperability between platforms. It also makes the
data openly accessible from anywhere and linkable as web pages do with hyperlinks.
Data web phase is to make available structured data using RDF. The scope of both
structured and unstructured content would be covered in the full semantic web stage.
Attempts will be to make it widely available in RDF and OWL semantic formats.
The driving force for web 3.0 will be artificial intelligence. Web 3.0 will be
intelligent systems or will depend on emergence of intelligence in a more organic
fashion and how people will cope with it. It will make applications perform logical
reasoning operations through using sets of rules expressing logical relationships
between concepts and data on the web. With the realization of the semantic web and
its concepts web 3.0 will move into Service Oriented Architecture.
The evolution of 3D technology is also being connected to web 3.0 as web 3.0 may
be used on massive scale due to its characteristics.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 162–170, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Personalized Web Based Collaborative Learning in Web 3.0 163

Web 3.0 is media centric where users can locate the searched media in similar
graphics and sound of other media formats.
The pervasive nature of web 3.0 makes the users of web in wide range of area be
reached not only in computers and cell phones but also through clothing, appliances,
and automobiles.
Learning process in Web 3.0 is a highly sophisticated for the learners. The learners
have access the unbelievable knowledge source. The tutors are intelligent agents who
are customized for the learners.

1.1 Review of Related Work

Claudio Baccigalupo and Enric Plaza discussed in the paper poolcasting : a social web
radio architecture for Group Customization about Pool casting a social web radio
architecture in which groups of listeners influence in real time the music played on
each channel. Pool casting users contribute to the radio with songs they own, create
radio channels and evaluate the proposed music, while an automatic intelligent
technique schedules each channel with a group customized sequence of musically
associated songs [13].
M.T.Carrasco Benitez discussed in the paper Open architecture for multilingual
social networking about an open architecture for all the multilingual aspects of social
networking. This architecture should be comprehensive and address well-trodden
fields such as localization, and more advanced multilingual techniques to facilitate the
communication among users [14].
Autona Gerber, Alta van der Merwe, and Andries Barnard discussed in the paper A
functional Semantic web architecture about the CFL architecture which depicts a
simplification of the original architecture versions proposed by Bernes-Lee as a result
of the abstraction of required functionality of language layers. Gerber argues that an
abstracted layered architecture for the semantic web with well defined functionalities
will assist with the resolution of several of the current semantic web research debates
such as the layering of language technologies [15].

2 Web Based Learning -- The Current Scenario

Web based learning is the electronically supported learning and teaching which are
procedural in character and aim to effect the construction of knowledge with
reference to individual experience, practice and knowledge of the learner. Information
and communication systems, whether networked or not, serve as specific media to
implement the learning process[1].
The current web environment is on the transition state between web 2.0 and web
3.0. The Earlier Web based learning process was instructional. Using the internet
technologies the students acquired knowledge. In Web 2.0 the learner is powered by
the social softwares like blogs, wiki's , podcasts and virtual worlds. The value
addition of social software has made the learner to be participatory in the learning
process. Social networking software's are participatory knowledge acquisition
phenomenon which shares the knowledge of the society.
A typical environment of learning in Web 2.0 is read and write .The primary
participants are learner and a tutor. The learner can be an active participant in the
164 S. Padma and A. Seshasaayee

social networking sites and acquire knowledge. When the learner opts for undergoing
classes in a particular topic, he learns from the static knowledge of the tutor which are
programs. It includes static contents and the learner can ask questions and answers if
supported by the programs will be displayed. The tutors look and appearance is 2D.
The main disadvantage of Personalized learning in Web 2.0 is although the
environment is participatory, it is static and predefined.
It does not includes the collective knowledge of current advancements in the topic.
The interface is simple to use. The Learner can have only limited queries. The
dissemination of knowledge by the tutor is determined by the developer.

3 Technical Background of Web 3.0

Web 3.0 architecture makes the learner sophisticated. Resource Definition Frame
work (RDF), SPARQL, WebDAV, site Specific API's ,FOAF and SSL are the
components of Web 3.0.
Resource Definition Frame work is a meta data data model . The resources are
transformed into statements of the form subject-predicate-object triple expressions.
The subject denotes the resource, predicate denotes traits or aspects of the resource
and express a relationship between the subject and the object. For example to
represent the notion "The sky has the color blue" in RDF is as the triple: a subject
denoting "the sky", a predicate denoting "has the color", and an object denoting
"blue". A collection of RDF statements intrinsically represents a labeled, directed
multi-graph. As such, an RDF-based data model is more naturally suited to certain
kinds of knowledge representation'2'. In web 3.0 Personalized learning is
accomplished by associating the queries by learners and the knowledge of the tutor
together with the available web resources in RDF format so that the query can be
associated with the knowledge.
SPARQL protocol for RDF query language process query which consist of triple
patterns, conjunctions , disjunctions and optional patterns. SPARQL facilitates
personalized web based learning in web 3.0 by accepting patterns, multiple queries
and complicated queries by the learner [3]. It also empowers the intelligent agent tutor
to collaborate multiple web resources into meaningful and understandable text and
visuals of the learner.
Web-based Distributed Authoring and Versioning (WebDAV) is a set of methods
based on the Hypertext Transfer Protocol (HTTP) for collaboration between users in
editing and managing documents and files stored on World Wide Web servers. The
most important features of the WebDAV protocol include:locking (overwrite
prevention ),properties (creation, removal, and querying of information about author,
modified date),name space management ( ability to copy and move Web pages within a
server's namespace )and collections (creation, removal, and listing of resources)[4].In
Personalized collaborative environment webDAV assists the intelligent agent tutor to
interact with the web servers to collect and produce cognizant knowledge to to the learner.
Site specific API's interacts with the RDF and produces required output of the
learner. Site specific API's provides the functionality to the learner by making the query
understandable to the tutor in precise form. It supports the intelligent agent tutor to
interact with the web data resources in world wide web servers to collaborate the
data[5]. Site Specific API's are useful while modifying the existing personalized
learning web sites of web 3.0 so that the functionality can be reused and remixed again.
 Personalized Web Based Collaborative Learning in Web 3.0 165

FOAF+SSL is a decentralized secure authentication protocol using the FOAF profile

information and SSL security layer. It makes ease the role of the intelligent agent tutor
in web 3.0 by maintaining FOAF cloud of the persons , activities their relations to other
people and objects[5]. The intelligent agent tutor of web 3.0 need not maintain a
centralized machine readable ontology . The security issues are maintained by SSL.
The aggregation of these components is utilized in the personalized collaborative
learning of web 3.0.
Technology plays a vital role in all aspects of personalized learning. The
technologies which are utilized in personalized collaborative learning includes
Artificial intelligent, Automated reasoning, Cognitive architecture, Composite
applications, Distributed computing, Knowledge representation, Ontology, cloud
computing, Grid computing, Scalable vector graphics, Semantic web, Semantic WiKi
and Software agents.

4 Personalized Web Based Learning in Web3.0

In web 3.0 the two major components are the learner and the 3D tutor. The learner is
the human who intends to acquire knowledge about a specific subject. The tutor is an
intelligent agent which delivers the collective knowledge to the learner.
The learner initially starts to take up the tutorial. The learner specify the tutor's
look. He can specify the gender of the tutor. The look or avatar of tutor can also be
specified by the learner. During the learning session, the learner must be completely
interactive. His interactive queries, interpretations and examples defines the
intelligence of the tutor. The learner can request more examples, working 3D models ,
clarifications, justifications, applications at any point of time. The output can be
audio, video, 3D, or text as opted by the learner. It can be a combination of the output
formats also depending on the preferences of the learner. The learner can demand
aggregations and mash up of websites based on the topic. He can have choice of
collaboration of best web resources and technologies available till date.
The tutor's look is specified by the user. The tutor intuitively asses the intelligence
of the learner. Based on the intelligence and personal preferences of learner the tutor
must deliver the knowledge. The tutor must collaborate knowledge from various web
resources , filter the irrelevant knowledge and share it. The tutor must give outputs
according to the personal preferences of the learner and also prefer the alternative
output form if one type is not sufficient. It must understand the language of the
learner and interact accordingly. The agent must aggregate the web resources required
by the learner which is updated till date. The technologies and tools used must be the
suitable by analyzing all the till date available technologies and tools.

5 Attribute Analysis of Personalized Web Based Learning in

Web 3.0
Attributes in web 3.0 are the characteristics which determine the behaviour of the
applications in web 3.0 architecture. Attribute analysis determines the key factors in
developing web 3.0 applications. The attributes considered are 3D,Mash ups,
Preferences regarding e-tutor and Speech recognition.
166 S. Padma and A. Seshasaayee

6 Data Collection
Data collection is a process of preparing and collecting data[10]. The method of Data
collection adopted is Questionnaire. A Questionnaire is designed with a closed type
question. Responses of the questionnaire is got from 100 Post Graduate Compute
stream Students at a private university in Chennai.
The questionnaire sample is given below.
Questionnaire:

s.no Question Response type

1. Do you require a 3D tutor ? Yes/no

2. Can the Demos for your learning be in 3D? Yes/no

3. Do you wish to ask a knowledge based random Yes/no

query to the etutor ?

4. Will you prefer a collective text output ? Yes/no

5. Do you want a collective knowledge Yes/no

of the web sites of your choice ?

6. Do you want the tutor to teach according to your Yes/no

knowledge level ?

7. Will you like to opt the gender of your tutor ? Yes/no

8. Should your tutor know about your preferences ? Yes/no

9. Should the tutor have a collective knowledge of Yes/no

the best web sites available ?

10. Do you wish your speech to be recognized by Yes/no

your tutor ?
 Personalized Web Based Collaborative Learning in Web 3.0 167

Fig. 1. Sample Data

Fig. 2. Comparison of requirements

The Top requirements are Random question, Custom Mash up, Etutor teaches according
to the level Choose look of etutor, concise_knowledge.

7 Finding the Frequent Item Sets in the Attributes Considered

Using Apriori Algorithm
One of the most popular data mining approaches is to find frequent itemsets from a
transaction dataset and derive association rules. Finding frequent itemsets (itemsets
with frequency larger than or equal to a user specified minimum support) is not trivial
because of its combinatorial explosion. Once frequent itemsets are obtained, it is
straightforward to generate association rules with confidence larger than or equal to a
user specified minimum confidence.
Apriori is a seminal algorithm for finding frequent itemsets using candidate
generation [12]. It is characterized as a level-wise complete search algorithm using
anti-monotonicity of itemsets, "if an itemset is not frequent, any of its superset is
never frequent". By convention, Apriori assumes that items within a transaction or
itemset are sorted in lexicographic order. Let the set of frequent itemsets of size k be
Fk and their candidates be Ck . Apriori first scans the database and searches for
frequent itemsets of size 1 by accumulating the count for each item and collecting
168 S. Padma and A. Seshasaayee

those that satisfy the minimum support requirement. It then iterates on the following
three steps and extracts all the frequent itemsets[11].
Apriori algorithm was adopted to the data set. Let R1 represent Random question,
R2 represent Custom Mash up, R3 represent Etutor teaches according to the level ,R4
represent Choose look of etutor, R5 represent concise_knowledge.

8 Results

Requirement matrix:

A requirement matrix is constructed based on the analysis done using apriori

algorithm.

    "

     

     

     

     

"     

From the matrix representation it is evident that,

{R1 => R4}
{R2 => R3 ,R5}
{R3=>R2,R5}
{R4=>R1}
{R5=>R2,R3}

R2,R3,R5 are strongly inter related requirements

R1 ,R4 are strongly inter related requirements
 Personalized Web Based Collaborative Learning in Web 3.0 169

Fig. 3. Defining groups

9 Conclusion
From the analysis ,it is concluded that Custom Mash up, Etutor teaching according to
the level of the learner , and concise_knowledge of the tutor are the interrelated
requirements by the students for designing a tutorial web3.0 site. Learner asking
Random question and learner choosing the look of etutor are the interrelated
requirements by the students for designing a tutorial web 3.0 site. If any one of the
interrelated requirement are required by the specific clients , then the developer can
include all the inter related requirements in the a tutorial web 3.0 product for better
customer satisfaction.

10 Future Work
By identifying the other attributes of web 3.0, the interdependencies of attributes can
be identified to ease the design and development of web developer fraternity.

References
[1] Tavangarian, D., Leypold, M., Nölting, K., Röser, M.: Is e-learning the Solution for
Individual Learning? Journal of e-learning (2004)
[2] http://en.wikipedia.org/wiki/Resource_Description_Framework
[3] http://www.w3.org/TR/rdf-sparql-protocol/
[4] http://en.wikipedia.org/wiki/WebDAV
[5] http://richard.cyganiak.de/2008/papers/rwwod-tr-
[6] http://en.wikipedia.org/wiki/FOAF_(software)
[7] Bahrami, A.: Object Oriented Systems Development –using the unified modeling
language. McGRaw-Hill International, New York
[8] http://www.sequencediagrameditor.com/diagrams/AutomatedRegis
trationSystem.htm
170 S. Padma and A. Seshasaayee

[9] Carrasco Benitez, M.T.: Open architecture for multilingual social networking. Position
paper for the W3CWorkshop on the Future of Social Networking (FSN), January 15-16
(2009)
[10] http://en.wikipedia.org/wiki/Data_collection
[11] http://www.hkws.org/seminar/sem433-2006-2007-no69.pdf
[12] Agrawal, R., Srikant, R.: Fast algorithms for mining association rules. In: Proceedings of
the 20th VLDB Conference, pp. 487–499 (1994)
[13] Baccigalupo, Plaza, C., Poolcasting, E.: A Social Web Radio Architecture for Group
customization Automated Production of Cross Media Content for Multi-Channel
Distribution. In: AXMEDIS apos; 2007 Third International Conference, November 28-
30, pp. 115–122 (2007)
[14] Carrasco Benitez, M.T.: Open architecture for multilingual social networking. Position
paper for the W3C workshop on the Future of Social Networking (FSN), January 15-16
(2009)
[15] Gerber, A., van der Merwe, A., Barnard, A.: A functional semantic architecture,
http://www.eswc2008.org/final-pdfs-for-web-site/fisr-1.pdf
 Rule Acquisition in Data Mining Using a Self Adaptive
Genetic Algorithm

K. Indira1, S. Kanmani2, D. Gaurav Sethia2, S. Kumaran2, and J. Prabhakar2

1
Department of Computer Science,
2
Department of Information Technology,
Pondicherry Engineering College,
Puducherry, India
{induharini,kanmani.n,gaurav.sethia7,kumarane.90}@gmail.com,
[email protected]

Abstract. Rule acquisition is a technique of data mining that is used to deduce

inferences from large databases. These inferences cannot be noticed easily
without data mining. Genetic algorithms (GAs) are considered as a global
search approach for optimization problems. Through the proper evaluation
strategy, the best “chromosome” can be found from the numerous genetic
combinations. In the self-adaptive genetic algorithm, its main thought is to let
control parameter (crossover rate, mutation rate) adjusted adaptively within the
proper range, thus achieve a more optimum solution. It is proved that the self-
adaptive genetic algorithm is with excellent convergence and higher precision
than the traditional genetic algorithm.

Keywords: Association rule mining, Genetic algorithm, Crossover, Mutation,

Fitness, Support, Confidence.

1 Introduction
Mining is used to refer to the process of searching through a large volume of data,
stored into a database, to discover useful and interesting information previously
unknown. Association rule mining is a type of data mining. It is the method of finding
the relations between entities in databases. Association rule mining is mainly used in
market analysis, transaction data analysis or in the medical field. For example, in a
medical database, diagnosis is possible provided the symptoms or in case of
supermarket, the relation between the purchase of different commodities can be
obtained. Such inferences are drawn using association rule mining and can be used for
making decisions.
There are some well known techniques for association rule mining. Some of the
well known algorithms are Apriori, constraint based mining, Frequency Pattern
Growth Approach, genetic algorithm. There have been several attempts for mining
association rules using Genetic Algorithm.
The main reason for choosing a genetic algorithm for data mining is that a GA
performs global search and copes better with attribute interaction when compared
with the traditional greedy methods, based on induction.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 171–178, 2011.
© Springer-Verlag Berlin Heidelberg 2011
172 K. Indira et al.

Genetic algorithm is evolved from Charles Darwin’s ‘Survival of the fittest’ theory.
It is based on individual’s fitness and genetic similarity between the individuals.
Breeding occurs in every generation and eventually it leads to better and optimal
group in the later generations. [1] Analyses the mining of Association Rules by
applying Genetic Algorithms.
[2] introduces CBGA approach that hybridizes constraint-based reasoning within a
genetic algorithm for rule induction. The CBGA approach uses Apriori algorithm to
improve its efficiency.
[3], [4] discuss some variations of the traditional Genetic algorithms in the field of
data mining. [3] is based on a evolutionary strategy and [4] adopts a self adaptive
approach. The self adaptive modification on a GA has never been attempted on
association rule mining before, but as this promises to be very promising in improving
the efficiency, it has been taken up.
The main modules in data mining process are
i. Data cleaning: also so known as data cleansing, is a phase in which noise
data and irrelevant data are removed from the collection.
ii. Data selection: at this step, the data relevant for the analysis is decided on
and retrieved from the large data collection.
iii. Data mining: it is the crucial step in which clever techniques are applied to
extract patterns potentially useful.
A brief introduction about Association Rule Mining and GA is given in Section 2,
followed by the proposed system in section 3. In section 4 the parameters used in
association rule mining using SAGA are defined. Section 5 presents the experimental
results followed by conclusion in the last section.

2 Association Rules and Genetic Algorithm

2.1 Association Rules

An important type of knowledge acquired by many data mining systems takes the
form of if-then rules [5]. Such rules state that the presence of one or more items
implies or predicts the presence of other items. A typical rule has the form
If A, B, C…n then Y
The two parameters with respect to if-then rules are described below.
The confidence [6] for a given rule is a measure of how often the consequent
is true, given that the antecedent is true. If the consequent is false while the antecedent
is true, then the rule is also false. If the antecedent is not matched by a given data
item, then this item does not contribute to the determination of the confidence of
the rule.
The support indicates how often the rule holds in a set of data. This is a relative
measure determined by dividing the number of data that the rule covers, i.e., that
support the rule, by the total number of data in the set.
 Rule Acquisition in Data Mining Using a Self Adaptive Genetic Algorithm 173

2.2 Genetic Algorithm

Genetic algorithm [7] is simulated in the natural environment of biological evolution

and genetics and the formation of an adaptive search algorithm for global
optimization probability. Genetic algorithm is suitable for solving problems
characterized by large space, multi-peak, non-linear, global optimization problem of
high complexity display. The parameters of the problem to be resolved into a binary
code or the decimal code (also into other hexadecimal code) that the gene, a number
of genes form a chromosome (individual), a number of chromosomes is similar to
natural selection, crossover and mutation matching algorithms, after repeated iteration
(that is, hereditary from generation to generation) until the final results of the
optimization. The use of genetic algorithms to solve the problem involved the
following seven key factors: encoding, fitness function, selection operator, the
crossover operator, mutation operator, control parameters.
In the traditional genetic algorithm, the crossover rate and mutation rate are fixed
values which are selected based on experience. Generally we believe that when the
crossover rate is too low, the evolutionary process can easily fall into local optimum
to result in groups of premature convergence due to population size and the lack of
diversity. When the crossover rate is too high, the process is optimized to the vicinity
of optimal point and the individual is difficult to reach optimal point which can slow
the speed of convergence significantly though groups can ensure the diversity.

Fig. 1. Flow Chart Traditional GA

174 K. Indira et al.

3 Proposed System

To overcome the drawbacks of traditional genetic algorithm, SAGA is proposed.

SAGA involves changing the crossover and mutation rates adaptively [8]. The main
purpose of setting mutation operator is to maintain the diversity of population and
avoid stagnation of evolution. In the traditional genetic algorithm the mutation rate is
fixed and after several iterations, the group’s quality will gradually to converge and to
form inbreeding. The organized adaptive genetic algorithm has a higher robustness,
global optimality and efficiency.

Procedure SAGA

Begin
Initialize population p(k);
Define the crossover and mutation rate;
Do
{
Do
{

Calculate support of all k rules;

Calculate confidence of all k rules;
Obtain fitness;
Select individuals for crossover / mutation;
Calculate the average fitness of the n and (n-1) the generation;
Calculate the maximum fitness of the n and (n-1) the generation;
Based on the fitness of the selected item, calculate the new crossover and
mutation rate;
Choose the operation to be performed;

} k times;
}

4 Parameters in Genetic Algorithm

4.1 Selection of Individuals

Chromosomes are selected from the population for breeding. According to Darwin's
evolution theory the best ones should survive and create new offspring. According to
Roulette Wheel Selection, Parents are selected according to their fitness. The better
the chromosomes are, the more chances to be selected they have, because the
selection depends on fitness.
 Rule Acquisition in Data Mining Using a Self Adaptive Genetic Algorithm 175

Fig. 2. Roulette Wheel Selection

4.2 Fitness Function

Given a particular chromosome, the fitness function returns a single numerical

"fitness," or "figure of merit". This value is proportional to the "utility" or "ability" of
the individual which that chromosome represents.
This paper adopts minimum support and minimum confidence for filtering rules.
Then correlative degree is confirmed in rules which satisfy minimum support-degree
and minimum confidence degree. After support-degree and confidence-degree are
synthetically taken into account, fit degree function is defined as follows.

(1)

In the above formula, Rs + Rc = 1 (Rs ≥ 0_Rc ≥ 0) and Suppmin, Confmin are respective
values of minimum support and minimum confidence. By all appearances_ if the
Suppmin and Confmin are set to higher values, then the value of fitness function is also
found to be high.

4.3 Crossover Operator

Crossover selects genes from parent chromosomes and creates a new offspring. The
most common form of crossover is single point crossover in which a crossover point
on both parents is selected and child 1 is head of chromosome of parent 1 with tail of
chromosome of parent 2 and child 2 is head of 2 with tail of 1.

4.4 Mutation Operator

Mutation changes randomly the new offspring. For binary encoding we can switch a
few randomly chosen bits from 1 to 0 or from 0 to 1. Mutation provides a small
amount of random search, and helps ensure that no point in the search has a zero
probability of being examined.
176 K. Indira et al.

4.5 Number of Generations

The generational process of mining association rules by Genetic algorithm is repeated

until a termination condition has been reached. Common terminating conditions are:
• A solution is found that satisfies minimum criteria.
• Fixed number of generations reached.
• The highest ranking solution's fitness is reaching or has reached a
plateau such that successive iterations no longer produce better results.
• Manual inspection.
• Combinations of the above.

4.6 Self Adaptive

The use of a fixed mutation probability Pm, when Pm value is small, does not have an
impact on the mutation operator. When Pm value is great, it could undermine the
group's excellent genes, the algorithm does not even slow down the convergence.
Here, a method of adaptive mutation rate is used as follows:

(2)

pm n is the nth generation mutation rate, pm(n+1) is the (n+1)th generation mutation rate.
The first generation mutation rate is pm0. fi(m) is the fitness of the nth individual
stocks i. fmax(n+1) is the highest fitness of the (n+1)th individual stocks. fi(n) is the fitness of
the nth individual i. m is the number of individual stocks. λ is the adjustment factor.

5 Experimental Studies
The objective of this study is to compare the accuracy achieved on different datasets
using a traditional GA and a SAGA. The encoding of chromosome is binary encoding
with fixed length. The fitness function adopted is as given.
Three datasets namely Lenses, Haberman and Car evaluation from UCI Machine
Learning Repository have been taken up for experimentation. Lenses dataset has 4
attributes with 24 instances. The second dataset is Haberman which has 4 attributes
and 306 instances. The final one is the car evaluation dataset, which has 6 attributes
and 1728 instances. The Algorithm is implemented using Java.
The accuracy and the convergence rate are recorded in the table below. Accuracy is
the count of rules matching between the original dataset and resulting population
divided by the number of instances in dataset. The convergence rate is the generation
at which the fitness value becomes fixed.
The parameters for the fitness function and the other parameters are chosen for best
performance of a traditional GA. The parameters are set such that the convergence is
fastest and the number of matches is maximum and SAGA is run with the same
parameters and the results obtained are tabulated below. Accuracy is in terms of
percentage with respect to number of matches.
 Rule Acquisition in Data Mining Using a Self Adaptive Genetic Algorithm 177

Table 1. Default GA parameters

Parameter Value
Varies as per the
Population Size
dataset
Initial Crossover rate 0.9
Initial Mutation rate 0.1
Roulette wheel
Selection Method
selection
Minimum Support 0.2
Minimum Confidence 0.8

Table 2. Accuracy comparison between GA and SAGA when parameters are ideal for
traditional GA

Dataset Traditional GA Self Adaptive GA

Accuracy No. of Accuracy No. of
Generations Generations
Lenses 75 38 87.5 35
Haberman 52 36 68 28

Car Evaluation 85 29 96 21

Table 3. Accuracy comparison between GA and saga when parameters are according to
termination of SAGA

Dataset Traditional GA Self Adaptive GA

Accuracy No. of Accuracy No. of

Generations Generations
Lenses 50 35 87.5 35
Haberman 36 38 68 28
Car 74 36 96 21
Evaluation

From the above table we can conclude that the Self adaptive GA performs better than
a traditional GA in both aspects, i.e. the convergence rate and the accuracy. The
accuracy in case of the Haberman dataset is low because one of the parameter in the
data is age. As age can take a wide range of values and only perfect matches are
considered, the accuracy comes down.
178 K. Indira et al.

When the algorithm comes to an end, the parameter values for mutation rate and
crossover rate are changed because of self adaptivity. If the new values are set as the
original values for a GA, then the performance of the GA is as below.
The table shows that the accuracy of the traditional GA goes down if the
parameters are set in accordance with the termination Condition mutation rate of
SAGA, this is because, when the SAGA ends, the mutation rate might take a high
value, which when applied to a GA, will bring down the accuracy. The fitness
threshold plays a major role in deciding the efficiency of the rules mined and
convergence of the system.

6 Conclusion
Genetic Algorithms have been used to solve difficult optimization problems in a
number of fields and have proved to produce optimum results in mining Association
rules. When Genetic algorithm is used for mining association rules the GA parameters
decides the efficiency of the system. Once the optimum values are fixed for individual
parameters, then making the algorithm self adaptive increases the efficiency because
it changes the mutation and crossover rate adaptively thus making the algorithm more
intelligent. When the mutation rate is varied with respect to the result from the
previous generation the accuracy increases. The efficiency of the methodology could
be further explored on more datasets with varying attribute sizes.

References
1. Collard, M., Francisi, D.: Evolutionary Data Mining: An Overview of Genetic-Based
Algorithms. In: 8th IEEE International Conference on Emerging Technologies and Factory
Automation, vol. 1, pp. 3–9 (2001)
2. Chiu, C., Hsu, P.-l.: A Constraint Based Genetic algorithm approach for Mining
Classification Rules. IEEE Transactions on Systems, Man and Cybernetics 35, 305–320
(2005)
3. Saggar, M., Agarwal, A.K., Lad, A.: Optimization of Association Rule Mining using
Improved Genetic Algorithms. IEEE, Transaction on System, Man and Cybernetics 4,
3725–3729 (2004)
4. Zhu, X., Yu, Y., Guo, X.: Genetic Algorithm Based on Evolution Strategy and the
Application in Data Mining. In: First International Workshop on Education Technology and
Computer Science, ETCS 2009, vol. 1, pp. 848–852 (2009)
5. Cattral, R., Oppacher, F., Dwego, D.: Rule Acquisition with Genetic Algorithm. In:
Congress on Evolutionary Computation, CEC 1999, vol. 1 (1999)
6. Dai, S., Gao, L., Zhu, Q., Zhu, C.: A Novel Genetic Algorithm Based on Image Databases
for Mining Association Rules. In: IEEE Conference on Computer and Information Science,
pp. 977–980 (2007)
7. Wu, Y.-T., An, Y.J., Geller, J., Wu, Y.T.: A Data Mining Based Genetic Algorithm. In:
IEEE Workshop on Software Technologies for Future Embedded and Ubiquitous Systems
(2006)
8. Li, J., Feng, H.R.: A Self-Adaptive Genetic Algorithm Based on Real Code, pp. 1–4.
Capital Normal University, CNU (2010)
 Dynamic Multi Dimensional Matchmaking Model for
Resource Allocation in Grid Environment

Japhynth Jacob1, Elijah Blessing Rajsingh2, and Isaac Balasingh Jesudasan3

1
Assistant Professor, St. Mother Theresa Engineering College
[email protected]
2
Director, KSCST, Karunya University
[email protected]
3
Director, St. Mother Theresa Engineering College
[email protected]

Abstract. Resource allocation and management in the Grid environment is a

complex undertaking as resources are distributed and heterogeneous in nature.
Optimally assigning jobs to resources is one of the key issues in Grid
environment. This work introduces a new Multi Dimensional Matchmaking
Framework for resource allocation on computational grids. Making the
environment intelligent and monitoring the dynamic behavior of resources gives
accurate prediction of resource allocation. The dynamic dimensional based
resource allocation enhances the overall performance of services.

Keywords: Matchmaking, Intelligence, Guaranteed time.

1 Introduction
Grids are distributed computing models that enable decentralized collaboration by
integrating computers into networks. In the Grid environment, shared resources and
users typically span different organizations. Before resources can be allocated to run
an application, a user or agent must select resources appropriate to the requirements
of the application. This process of selecting resources based on application
requirements is called “resource matching”. Nowadays many matchmaking based
technologies addresses the issue of Grid resource management. The grid
matchmaking process involves three types of agents: consumers (requesters),
producers (providers), and a matchmaking service. A matchmaking service uses a
matching algorithm to evaluate a matching function which computes the matching
degree. This paper introduces a new multi dimensional matchmaking framework. The
multi dimensional matchmaking process considers static, dynamic and behavioral
parameters. The correlation between the request and the resource is ranked and
optimal resource is allocated. Intelligence is added in this model to speed up the
allocation process. User satisfaction level is retrieved for every job.

2 Related Work
In condor G matchmaking mechanism [1] users describe their requirements in class
ads. Classads describes attributes of request (i.e. storage space, library files required

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 179–185, 2011.
© Springer-Verlag Berlin Heidelberg 2011
180 J. Jacob, E.B. Rajsingh, and I.B. Jesudasan

etc…). Resources also describes class ads describing their attributes. Condor
matchmaker matches the request with the resources. This system lacks support for
parallel jobs, lacks integration with GIS, and is not normalized. A semantic matchmaker
service on grid [10] uses separate ontology for request and service (ontology as set-of-
concept-definitions or specification of a conceptualization).Ontology translator
translates various formats into one. Matchmaking is done by using terms defined in
ontologies. It is based on static information. Semantic based matchmaking uses
ontology based matchmaking and uses grid service technology to access it. But in this
system users should express their preferences manually by calculating the ranking
function and there are no application level descriptions in request ontology. In matching
service for an intelligent grid [2], each request is characterized by set of properties. Each
property is a tuple (name, value).Matching service uses matching algorithm to evaluate
a matching function which computes matching degree. It considers only static
information. Online resource matching for heterogeneous grid environments [4] takes
into account the dynamically changing conditions on resources and request. Resource
State Repository (RSR) stores the real time status of all resources. Each job requires
resources from various grid sites. It uses integer programming to solve the problem of
heterogeneity. Each job has one dependency on each resource type. This system lacks
fairness and preemption. Semantic matchmaking of grid services using parameter
matchmaking algorithm matches request with resources based upon three parameters;
input, output and functionalities. Matching degrees are computed by applying different
filters. The matchmaking algorithm semantically compares IOF of request, service and
computes various degrees of matches. Here the least rank is taken as the final rank.
However other higher ranks have lost its significance. There may be a chance for that
parameter to be a significant parameter to compare. In Interoperating grids through
delegated matchmaking, delegation chain has been formed. But it has worst resource
selection. In Grid resource discovery approach based on matchmaking engine overlay
[3] centralized matchmaker becomes a bottleneck when requests and resources increase.
It also leads to single failure problem.

3 Problem Statement

Condor G matchmaking mechanism lacks support for parallel jobs, lacks integration
with GIS and ranks are not normalized. In a semantic matchmaker service on grid,
users should express their preferences manually by calculating the ranking function
and there are no application level descriptions in request ontology. Resource
matchmaking service for intelligent grid considers only static information. Online
resource matching for heterogeneous grid environments has no fairness in allocation
and has no preemption. Delegated matchmaking has worst resource selection.
Problems in matchmaking are summarized as: Ranks of the matched resources are not
normalized, even after allocating best resource, job’s successful execution till the end
is not assured, Needs dynamic updation about the status of jobs, Speed of matching
process needs to be improved.
 Dynamic Multi Dimensional Matchmaking Model for Resource Allocation 181

4 Multi Dimensional Matchmaking Framework

Multi dimensional matchmaking framework assigns weight value for each parameter
to normalize the ranks. It includes behavioral parameters to avoid unpredictable
failures, dynamic parameters to dynamically update the status. The intelligent part
will speed up the matchmaking process. First filtering stage considers Static
parameters like resource type, resource name, IP address, node memory size, node
CPU speed, node hard disk capacity, no of nodes, total memory size, total CPU rates,
total disk capacity, total CPU speed. The next stage of filtering considers dynamic
parameters like Process deadline or guaranteed time, economic efficiency, security or
dedication rate. The final filtering stage considers behavioral parameters like hit ratio,
abandon rate, occupation rate .Parameters in each stage will be given weight value
according to their importance in each stage. This weight value is multiplied with
matching degree. The highest value is taken for each stage. The average of 3 stages is
the final matching value. The process repeats until it searches all possible resources.
The resource with highest value is chosen as the best resource.

Fig. 1. Matchmaking architecture

182 J. Jacob, E.B. Rajsingh, and I.B. Jesudasan

Intelligent scheduler gets feedback from users after they have successfully
completed their jobs. It stores their level of satisfaction. If their level of satisfaction is
nearly equal to the threshold level, it stores the location of resources they have used in
temporary hashing. Next time if same job arrives, the temporary hashing gives the
location of resources that have successfully completed that job previously. Every
similar type of resources in grid site will have a separate capacity value ‘C’. If the
newly arrived resource capacity value is greater than the resource the old resources’
reference will be replaced by the newly joined resource location. Matchmaking
process is analyzed by considering four cases

Case 1: Resource ingress Low and customer ingress Low

Case 2: Resource ingress Low and customer ingress High
Case 3: Resource ingress High and customer ingress Low
Case 4: Resource ingress High and customer ingress High.

4.1 Matchmaking Algorithm

User submitting the request Ri to resource manager RM

RM checks whether Ri = {any one of the old request}
If Ri = {any one of the old request}
{Ri is directed to GM
//GM retrieves all the resource that is already used by the earlier old request (since
it is similar to the newly arrived request now)
//GM then directs the request along with the list of used resources to HMM
Ri {ri i=1-n}HMM
//HMM calculates the transition probability of the retrieved request
//Transition probability of resource is high if
*any new resource which is of the same type but with advanced features/capacity
*any resources left the grid site or any resource newly added into the grid site.
//The state transition probability is low if
*no new resources are being added up with advanced features into the grid site
*no resources left the grid site or no resources newly added into the grid site
If {State transition probability is high
// recalculate the rank with the newly joined resources.
Ri Pi i=1-n
//store the result in rank house
//use the old rank calculated for the request
RiX}}
If Ri = new request
{ //Direct the request to the matchmaking engine.
//Calculate new rank for the resources retrieved.
//Allocate highly ranked resource to the user and stores the rank in rank
housing.
}
 Dynamic Multi Dimensional Matchmaking Model for Resource Allocation 183

5 Performance Evaluation

5.1 Experiment Setup

The objective of this research is to assign optimal resource to job, to speed up the
allocation process, to make the system more reliable. It is assumed that all service
providers are available at all time provided they are free with their resources. The
number of request served per second increases as the system receives more number of
jobs. The optimal resource is selected by considering assignment model, correlation
model, testing of samples and measures of dispersion.

Let represents job, then = { 1 , 2, ….., }

Where is number of parameters.
Let 1, 2, 3… are resources, where is number of resources.
Let 1 = {Y11, Y12, Y13}
2 = (Y21, Y22, Y23}
.
.
= {Yz1, Yz2, Yz3}

The correlation between the request and the resource is calculated using
.∑ ∑ .
(1)
.∑ ∑ .∑ ∑

The behavior of service is calculated by testing of samples

(2)
√

represents old data output

represents expected output
represents deviation from previous behavior
represents number of previous samples
represents behavior of service.

6 Conclusion

In this paper we have proposed a multi dimensional matchmaking model to select

optimal resource to job. The intelligent environment prevents the starvation of low
priority jobs in the remote site. By analyzing the job and resource in multi
dimensions, best resource can be allocated to jobs. The intelligent scheduler reduces
the search time irrespective of number of resources. Job entry and exit dynamicity is
considered while allocating resources for jobs. Measuring user satisfaction level and
threshold value gives accurate prediction in allocation. Our proposed system ensures
184 J. Jacob, E.B. Rajsingh, and I.B. Jesudasan

that higher priority jobs are always treated better than low priority jobs leading to an
overall betterment of the QoS.

References
1. Imamagic, E., Radic, B., Dobrenic, D.: An Approach to Grid Scheduling By Using Condor-
G Matchmaking Mechanism. Journal of Computing and Information Technology -
CIT 14(4), 329–336 (2006), doi:10.2498/cit.2006.04.09.
2. Bai, X., Yu, H., Ji, Y., Marinescu, D.C.: Resource Matching and a Matchmaking Service
for an Intelligent Grid. World Academy of Science, Engineering and Technology 1 (2005)
3. Liu, Y., He, H.: Grid Resource Discovery Approach Based on Matchmaking Engine
Overlay. In: Third International Conference on Semantics, Knowledge and Grid
4. Vijay, C., Naik, L.K., Liu, C., Wagner, J.: On-line ResourceMatching for Heterogeneous
Grid Environments. In: IEEE International Symposium on Cluster Computing and the Grid
(2005)
5. Li, F., Qi, D.: Research on Grid Resource Allocation Algorithm Based on Fuzzy Clustering.
In: Second International Conference on Future Generation Communication and Networking
(2008)
6. Clematis, A., Corana, A., D’Agostino, D., Galizia, A., Quarati, A.: Job_resource
matchmaking on Grid through two-level benchmarking. Future Generation Computer
Systems (June 2010)
7. Shu, G., Rana, O.F., Avis, N.J., Chen, D.: Ontology-based semantic matchmaking
approach. Advances in Engineering Software 38, 59–67 (2007)
8. Han_, W., Shi, X., Chen, R.: Process-context aware matchmaking or web service
composition. Journal of Network and Computer Applications 31, 559–576 (2008)
9. Wanga, C.-M., Chenb, H.-M., Hsuc, C.C., Lee, J.: Dynamic resource selection heuristics
for a non-reserved bidding-based Grid Environment. Future Generation Computer
Systems 26, 183–197 (2010)
10. Grigori, D., Corrales, J.C., Bouzeghoub, M.: Behavioral matchmaking for service retrieval:
Application to conversation protocols. Information Systems 33, 681–698 (2008)
11. Asaduzzaman, S., Maheswara, M.: Utilizing unreliable public resources for higher profit
and better SLA Compliance in computing utilities. J. Parallel Distrib. Computing 66, 796–
806 (2006)
12. de Assunção, M.D., Buyya, R.: Performance analysis of allocation policies for intergrid
resource provisioning. Information and Software Technology 51, 42–55 (2009)
13. Yu, H., Bai, X., Marinescu, D.C.: Workflow management and resource Discovery for an
intelligent grid. Parallel Computing 31, 797–811 (2005)
14. Foster, I., Kesselman, C.: The GRID 2: Blueprint For a New Computer Infrastructure.
Elsevier, Amsterdam (2005)
15. Liu, S.: User-Centric Resource Allocation Hierarchy in Grid Computing. In: The Sixth
International Conference on Grid and Cooperative Computing, GCC 2007 (2007)
16. Altunay, M., Brown, D., Byrd, G., Dean, R.A.: Evaluation of Mutual Trust during
Matchmaking. In: Proceedings of the Fifth IEEE International Conference on Peer-to-Peer
Computing, P2P 2005 (2005)
17. Li, H., Buyya, R.: Model-based simulation and performance evaluation of grid scheduling
strategies. Future Generation Computer Systems 25, 460–465 (2009)
 Dynamic Multi Dimensional Matchmaking Model for Resource Allocation 185

18. Comito, C., Gounaris, A., Sakellariou, R., Talia, D.: A service-oriented system for
distributed data querying and integration on Grids. Future Generation Computer Systems,
doi:10.1016/j.future.2008.11.009.
19. Kosar, T., Balman, M.: A new paradigm: Data-aware scheduling in grid computing. Future
Generation Computer Systems 25, 406–413 (2009)
20. Korkhova, V.V., Moscicki, J.T., Krzhizhanovskaya, V.V.: Dynamic workload balancing of
parallel applications with user-level scheduling on the Grid. Future Generation Computer
Systems 25, 28–34 (2009)
A Comparative Study of CSO and PSO Trained Artificial
Neural Network for Stock Market Prediction

Suresh Chittineni1, Vabbilisetty Mounica1, Kaligotla Abhilash1,

Suresh Chandra Satapathy1, and P.V.G.D. Prasad Reddy2
1
Anil Neerukonda Institute of Technology and Sciences
Sangivalasa-531162, Visakhapatnam, Andhra Pradesh, India
{sureshchittineni,mounica.vabbilisetty,abhilash.kaligotla,
sureshsatapathy}@gmail.com
2
Department of CS&SE, Andhra University Engineering College
Visakhapatnam, Andhra Pradesh, India
[email protected]

Abstract. Stock market prediction is the act of trying to determine the future
value of a company stock of other financial Instrument traded on a financial
exchange. This paper presents a comparison between, PSO and CSO trained
Neural Network to predict the stock rates by preparing data which acts as input.
The data is prepared in such a way that the external factors like traditional
issues can be mitigated. Earlier Neural Network was trained using Back
Propagation algorithm but it converges to local optima and cannot be applied to
discrete functions. Sow we have chosen PSO and CSO optimization algorithm
to train the Neural Network. The results show that training neural network with
such data gives a better performance.

Keywords: PSO Neural Network, and CSO Based Neural Network, Pre
Processed Data.

1 Introduction

The importance of wealth in life is necessary for everything and it brings comfort
and luxury. Wealth doesn’t mean a recurring earning/salary but it does refer to a
fixed asset, an investment or savings so it is a challenging and daunting task to find
out which is more effective and accurate method for stock rate prediction. The
successful prediction of a stock’s future price could yield significant profit. [2]
Prediction of financial markets has long been an attraction in the minds of Equity
investors. Technical Analysis [11] provides a framework for studying investor
behavior, and generally focuses only on price and volume data. Technical Analysis
using this approach has short-term investment horizons, and access to only price
and exchange data. With the advent of powerful computers much attention has
been focused on this field. Equity market prices depend on many influences. Key
factors that influence future equity prices can be broadly divided into quantitative
and qualitative types.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 186–195, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 A Comparative Study of CSO and PSO Trained Artificial Neural Network 187

Primary quantitative factors include open rate, high rate, low rate, close rate and
volume for individual equities. Qualitative factors include socio-economic, political,
international, regional and performance factors to name but a few. The aim of this
paper is to compare the pso and cso based neural network so as to find a technique
that can predict stock market more accurately.
Preliminary research performed on Indian National Stock Exchange market has
suggested that the inputs to the system may be taken as: previous day’s closing rate
and volume of last trading day for PSO[3] and CSO based neural network. After the
inputs have been determined, the data have been gathered for NTPC stock for the
period of Jan-2007 to DEC-2009 for training PSO and CSO based neural network.
For testing purpose we have used testing data of NTPC Stock for the period of 1 year
i.e. JAN-2010 to DEC-2011. Training and testing is performed using two network
architectures.
1) One Hidden Layer PSO-Neural Network.
2) One Hidden Layer CSO-Neural Network.
The results are compared between PSO-NN (Particle Swarm Optimization Based
Neural Network) and CSO-NN (Cat Swarm Optimization Based Neural Network).

2 PSO Based Neural Network

A. Overview

Particle swarm optimization (PSO) is a population based stochastic optimization

technique developed by Dr. Eberhart and Dr. Kennedy in 1995, inspired by social
behavior of bird flocking or fish schooling.
PSO shares many similarities with evolutionary computation techniques such as
Genetic Algorithms (GA). The system is initialized with a population of random
solutions and searches for optima by updating generations. However, unlike GA,
PSO has no evolution operators such as crossover and mutation. In PSO, the potential
solutions, called particles, fly through the problem space by following the current
optimum particles.

Fig. 1. Conceptual diagram explaining PSO

188 Ch. Suresh et al.

Sk: current searching point

Sk+1: modified searching point
Vk: current velocity
Vk+1: modified velocity
VPbest: velocity based on pbest.
Vgbest: velocity based on gbest.

Each particle keeps track of its coordinates in the problem space that is associated
with the best solution (fitness) it has achieved so far. (The fitness value is also
stored.) This value is called pbest. When a particle takes all the population as its
topological neighbors, the best value is a global best and is called gbest.
The particle swarm optimization concept consists of, at each time step, changing
the velocity of (accelerating) each particle toward its pbest and gbest locations.
Acceleration is weighted by a random term, with separate random numbers being
generated for acceleration toward pbest and gbest locations.

B. Training Neural Network with PSO

The algorithm for PSO training is as follows:

Step 1: randomly generate the initial positions and velocities of particles of, say; p
strings each of length d, which depends on the number of weights required.

Step 2: compute the fitness of each particle.

Step 3: get the values of gbest and pbest, where gbest is the global best among all the
particles and pbest is the personal best till this iteration.

Step 4: update the velocities of the particles using the equation (1) below:
Newpvel=C1*pvel + C2* rand * (ppos-gbest)+C3*rand*(ppos-pbest) (1)
Where,
Pvel: current velocity of the particle.
C1, C2, C3: random values.
Ppos: current position of the particle.
Newpvel: updated velocity of the particle.

Step 5: Update the position of the particle using the equation (2) below:
Newppos=ppos+newpvel (2)
Where,
Newppos: updated position of the particle.
Ppos: current position of the position of the particle.

Step 6: continue the above-mentioned steps till some stopping criterion is satisfied.
Step 7: Now with the best particle the stock rates are predicted.
 A Comparative Study of CSO and PSO Trained Artificial Neural Network 189

C. PSO – NN Organization

Hybrid approach offer strong advantage over either rule based or unaided neural
network approaches [8] PSO based neural network offer good generalization [5]
abilities although it may be difficult to determine the optimal network configuration
and network parameters [4][6]. The architecture of PSO based neural network used is
as follows.
1) Input layer with 2 nodes
2) One hidden layer with 2 nodes [7]
3) Output layer with one node.

3 CSO-Based Neural Network

A. Overview

CSO is generated by observing the behavior of cats, and composed of two sub-models
by simulating the behavior of cats[9]. Two of the major behavioral traits of cats are
modeled for the proposed Cat Swarm Optimization (CSO). These are termed
“seeking mode” and “tracking mode”. Combination of these two modes allows CSO
a better performance. Details of the two modes are given below.
Rest and alert – seeking mode: This sub mode is used to model the cat during a
period of resting but being alert – looking around its environment for its next move.
Seeking mode has four essential factors, which are designed as follows: seeking
memory pool (SMP), seeking range of the selected dimension (SRD), counts of
dimension to change (CDC) and self position consideration (SPC). SMP is used to
define the size of seeking memory of each cat, indication any points sort by the cat.
The cat would pick a point from the memory pool, according to rules described later.
SRD declares the mutative ration for the selected dimensions. While in seeking
mode, if a dimension is selected for mutution, the difference between the old and new
values may not be out of range, the range defines by the SRD. CDC tells how many
of the dimensions will be varied. All these factors play important roles in seeking
mode. SPC is a Boolean valued variable, and indicates whether the point at which the
cat is already standing will be one of the candidate points to move to. SPC cannot
influence the value of SMP.
Movement – tracing mode: Tracing mode is the sub-model for modeling the case of
the cat in tracing targets.
Once a cat goes into tracing mode, it moves according to its’ own velocities for
each dimention.

B. Training Neural Network with CSO

As already discussed, CSO has two sub modes, namely seeking mode and tracing
mode. We combine these two modes into the algorithm.
Seeking mode is described below:
190 Ch. Suresh et al.

Step 1: Make j copies of the present position of catk, where j=SMP. If the value of
SPC is true, let j=(SMP=1), then retain the present position as one of the candidates.

Step 2: For each copy, according to CDC, randomly plus or minus SRD percents the
present values and replace the old ones.

Step 3: Calculate the fitness values (FS) of all candidate points.

Step 4: Randomly pick the point to move to from the candidate points, and replace the
position of catk.

The action of tracing mode can be described as follows:

Step 1: Update the velocities for every dimension (vk, d) according to equation (1).

Step 2: Check if the velocities are in the range of maximum velocity. In case the new
velocity is over range, it is set equal to the limit.

Step 3: Update the position of catk according to equation (2).

Vk,d = Vk,d + r1.c1.(xbest, d –xk,d),
d-1, 2, 3, ….M (1)
where xbest,d is the position of the cat, who has the best fitness value; xk,d is the
position of catk, c1 is a constant and r1 is a random value in the range of [0, 1].
Xk,d = Xk,d + Vk,d (2)

C. CSO-NN Organizations

Hybrid approach offer strong advantage over either rule based or unaided neural
network approaches [8]. CSO based neural network offer good generalization
abilities although it may be difficult to determine the optimal network configuration
and network parameters [4][6]. The architecture of CSO based neural network used is
as follows:
1) Input layer with 2 nodes
2) One hidden layer with 2 nodes[7]
3) Output layer with one node.

4 Experiment Results
The sytem has been developed and tested on Windows XP operating system. We
have used MATLAB for programming. Normalization is a key part of the input for
neural networks and should enable more accurate predictable rates. Normalized data
is used for training PSO based neural network and CSO based neural network. We
normalize inputs so that input values lied between 0 and 1. Simulation data was
sourced from Indian National Stock Exchange (NSE).
 A Comparative Study of CSO and PSO Trained Artificial Neural Network 191

A. Preparation of the Data Sets

Stock market depends upon many Qualitative factors like socio-economic, political,
international, regional and performance factors, so the system which we tune should
be in such a way that it is able to predict the stock rates accurately at least to some
extent and quantify some of the external factors. Traders invest their money in
shares, whose minds are closely tied up with sentiments. They sometimes choose a
particularly day and start investing money on that particular day only.
Now the data set required to train such a Neural Network System is prepared by
taking a particular day’s stock values from the entire data of 3 years. This method is
used to aim at a better result than normal data. [1]. Thus in this paper the traditional
issues are taken into account to quantify the external factors and produce good results.
Input attributes should be carefully selected to keep the dimensionality of input
vectors relatively small [10]. As we know close rate and volume are primary
quantitative factors for individual equities and from quantitative factors the key
qualitative factor of the market sentiment can be derived. So we used close rate and
volume of stocks as our input in PSO and CSO based neural network and next stock
rate as our target for training networks. PSO-NN and CSO-NN is trained on data set
of NTPC for the years of Jan 2007 to Dec 2009 after training, testing is done on data
set of 1 year from JAN 2010 to DEC 2010 of NTPC stock. PSO-NN and CSO-NN
performance is compared on stock rates of NTPC for the specified period.

Fig. 2. Block Diagram explaining the working with the prepared data

In Fig3 and Fig4, we have taken the close rate and volume of Friday and predicted
the open rate of Monday, thus 45 days of prediction is done in this NN. The CSO-NN
system demonstrated better performance in comparison to PSO-NN on stock rates of
NTPC for the specified period. On comparing it has been found that, CSO-CNN
system is able to predict stock price movement of NTPC correctly with MSE 0.0012
as shown in Fig 3, while performance of PSO Based NN system with MSE 0.0017
shown in Fig 4.
In Fig5 and Fig6, we have taken the close rate and volume of Wednesday and
predicted the open rate of Thursday, thus 49 days of prediction is done in this NN.
The PSO-NN system demonstrated better performance in comparison to CSO-NN on
stock rates of NTPC for the specified period. On comparing it has been found that,
PSO-NN system is able to predict stock price movement of NTPC correctly with
MSE 0.0013 as shown in Fig 5, while performance of CSO Based NN system with
MSE 0.0017 shown in Fig 6.
192 Ch. Suresh et al.

The following graphs show the prediction of stock market for 45 week days
(Monday). The red lines in the graph indicates the expected trend and the blue lines
indicate the calculated trend obtained by training the neural network system for over a
period 3 years.
The graph in Fig 3 given below represents the output of CSO trained neural
network system that is expected to predict stock rates for 45 days. In this graph the 0
to 10 days period the trend is followed. From 1 0 to 15 days the graph is almost the
same. From 15 to 40 days period the trend remained same. From 40 to 45 days again
the trend of the graph remained same with small deviations. On the whole the trend is
depicted every time.

Fig. 3. Output of CSO trained neural network system for Mondays

This graph is Fig 4 below represents the output of PSO trained neural network
system that is expected to predict stock rates for 45 days. Over here the trend of the
calculated values remained the same all through with the expected values with a few
deviations occurring.

Fig. 4. Output of PSO trained neural network system for Mondays

 A Comparative Study of CSO and PSO Trained Artificial Neural Network 193

The following graphs show the prediction of stock market for 49-week days
(Thursday). The red line in the graph indicates the expected trend and the blue lines
indicate the calculated trend obtained by training the neural network system for over a
period 3 years.
The graph in Fig 5 given below represents the output of PSO trained neural
network system that is expected to predict stock rates for 49 days. Over here the trend
of the calculated values remained the same all through with the expected values with
a few deviations occurring.
This graph in Fig 4 below represents the output of CSO trained neural network
system that is expected to predict stock rates for 49 days. In this graph the 0 to 10
days period the trend is followed. From 1 0 to 17 days the graph is almost the same.
From 18 to 46 days period the trend remained same. From 46 to 49 days, small
deviations have occurred. On the whole the trend is depicted every time.

Fig. 5. Output of PSO trained neural network system for Thursdays

Fig. 6. Output of CSO trained neural network system for Thursdays

194 Ch. Suresh et al.

The deviations in both the above shown graph highlights the fact that the system
that we used is a trained neural network system but in reality due to some external
aspects like social, political and economical factors influence in day-to-day stock
market prediction. This makes the prediction of the stock market movement very
difficult. This variation in the graph can be mitigated if the external factors are
quantified. Quantifying the external factors can be a further research.

Table 1. Comparative details between the two algorithms

S.No Training Data No of No of MSE

algorithm set days days value
trained predicted
1 PSO- NTPC 3 YRS 45 0.0017
Monday
2 CSO- NTPC 3 YRS 45 0.0012
Monday
3 PSO- NTPC 3 YRS 49 0.0013
Thursday
4 CSO- NTPC 3 YRS 49 0.0017
Thursday

5 Conclusion
This paper has compared the forecasting accuracies of PSO based neural network and
CSO based neural network. Results showed that under certain circumstances that
prevailed on Monday, CSO based neural network has outperformed in predicting this
stock movement in comparison to PSO based neural network. While under
circumstances that prevailed on Thursday. PSO based neural network has
outperformed in predicting this stock movement in comparison to CSO based neural
network. By this approach of preparing data we are able to predict stock market rates
better than ordinary data, this result is also not very accurate because stock rates not
only depend on traditional sentiment factors but also on other factors like political
issues, and all other issues related to the company which are difficult to be quantized.
A further research can be done in quantifying the external factors.

References
[1] Suresh, Ch., Abhilash, K., Mounica, V., Satapathy, S.C.: Cat Swarm Optimization Based
Neural Network and Particle swarm optimization Based Neural Network in Stock Rates
Prediction
[2] Khan, A.U., et al.: Genetic Algorithm Based Back propagation Neural Network Performs
better than Back propagation Neural Network in Stock Rates Prediction, Glow Gift,
Bhopal (2005)
 A Comparative Study of CSO and PSO Trained Artificial Neural Network 195

[3] Zhang, X., Chen, Y.: Stock Index Forecasting Using PSO Based Selective Neural
Network Ensemble
[4] Hamid, S.A.: Primer on using neural networks for forecasting market variables. In:
Proceedings of the a Conference at School of Business, Southern New Hampshire
University (2004)
[5] Majhi, R., Panda, G.: Robust Prediction of Stock Indices using PSO based Adaptive
Linear Combiner (2009)
[6] Adya, M., Collopy, F.: How Effective are Neural Networks at Forecasting and Prediction?
A Review and Evaluation. Journal of Forecasting (1998)
[7] Khan, A.U., et al.: Stock Rate Prediction Using Back Propagation Algorithm: Analyzing
the prediction accuracy with different number of hidden layers, Glow Gift, Bhopal (2005)
[8] Bergerson, K., Wunsch, D.: A commodity trading model based on a neural network-
expert system hybrid. In: IJCNN 1991, Seattle International Joint Conference, July 8-14,
vol. 1, pp. 289–293 (1991)
[9] Chu, S.-C.: Computational Intelligence Based on the Behaviour of Cats-And it’s
Applications, Taiwan (2006)
[10] White, H.: Economic prediction using neural networks: The case of IBM daily stock
returs. In: Neural Networks in Finance and Investing, ch. 18, pp. 315–328. Probus
Publishing Company (1993)
[11] Mizuno, H., Kosaka, M., Yajima, H., Komoda, N.: Application of Neural Network to
Technical Analysis of Stock Market Prediction. Studies in Informatic and Control 7(3),
111–120 (1998)
Scaled Conjugate Gradient Algorithm in Neural Network
Based Approach for Handwritten Text Recognition

Haradhan Chel1, Aurpan Majumder2, and Debashis Nandi3

1
Bharat Sanchar Nigam Limited, India
2
NIT Durgapur, Dept. of Electronics and Communication, India
3
NIT Durgapur, Dept. of Information Technology, India

Abstract. Handwritten text and character recognition is a challenging task

compared to recognition of handwritten numeral and computer printed text due
to its large variety in nature. Neural Network based approach provides most
reliable performance in handwritten character and text recognition but
recognition performance depends upon some important factors like no of
training samples, reliable features and no of features per character, training
time, variety of handwriting etc. Important features from different types of
handwriting are collected and are fed to the neural network for training. More
no of features increases testing efficiency but it take longer time to converge the
error curve. To reduce this training time effectively proper algorithm should be
chosen so that the system provides best train and test efficiency in least possible
time that is to provide the system fastest intelligence. In this paper we have used
Scaled Conjugate Gradient Algorithm, a second order training algorithm for
training of neural network. It provides faster training with excellent test
efficiency. A scanned handwritten text is taken as input and character level
segmentation is done. Some important and reliable features from each character
are extracted and used as input to a neural network for training. When the error
level reaches into a satisfactory level (10-12) weights are accepted for testing a
test script. Finally a lexicon matching algorithm solves the minor
misclassification problems.

Keywords: Transition Feature, Sliding Window Amplitude Feature, Contour

feature, scaled conjugate gradient.

1 Introduction
The endeavor of researchers for recognition of handwritten text through computer
learning started in the sixtieth decade. Along with the up gradation of computational
techniques it was always tried to implement the human perception into mathematical
logic which is used to train the computer system. Different researchers took different
approaches [7],[ 9], [20] to recognize handwritten text. Almost all the researchers
have chosen some common logical steps for recognition of handwritten text such as
Segmentation, Feature Extraction and pattern classification technique. Most of the
approaches were on Neural Classifier [6], [14], [23], Hidden Markov Model [9],
Fuzzy Classifier, or hybridized technique like Neuro-fuzzy, Neuro-GA, or some other

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 196–210, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Scaled Conjugate Gradient Algorithm in Neural Network 197

statistical methods [25]. Neural Classifier has high discriminative power [1],[11],[12]
for different patterns. Handwritten text contains wide varieties of styles in nature. It is
really difficult to get real character level segmentation. A lot of works has been done
on segmentation of hand written text in last two decades. Improper segmentation
results inaccurate feature extraction and poor recognition performance. Similarly
reliable and only discriminative features give better recognition performance. If
number of features is increased recognition performance increases but it increases
computational complexities and leads to much longer time of training of the neural
network. The most essential thing is to choose a proper training algorithm which train
faster the network with large no of features and provide best recognition performance.
Many of these algorithms are based on Gradient Descent Algorithm such as Back
Propagation Algorithm [15],[18]. But all the algorithms are practically inapplicable in
large scale systems because of its slow nature and it’s performance also depends on
some user dependent parameters like learning rate and momentum constant. Some
second order training algorithm such as Fleture Reeves [5], Polak Riebier[16],[19] or
Powell-Beale Restarts algorithm[24] may be applied in appropriate applications.
However Scaled Conjugate Gradient algorithm [22] results in much better
recognition performance. By virtue of Scaled Conjugate Algorithm fastest training is
obtained with almost 98 percent recognition efficiency. This paper not only shows the
comparison of different second order training algorithm but also the reliable feature
extraction and efficient lexicon matching technique. A small relevant description of
Scaled Conjugate Algorithm is also given in later section. Different feature extraction
schemes are also described in brief and finally the result of all experiments.

2 Image Preparation and Segmentation

A handwritten text is scanned through optical scanner and stored in bitmap image
format. The image is converted in to a binary image. One important point may be
mentioned regarding image preparation is choosing proper threshold so that image
does not get any discontinuity in any continuous part. Image is segmented [4],[8],[ 3]
text level to word level and word level to character level segmentation all relevant
information such as no of word, no of characters, no of lines are stored. The algorithm
for segmentation is used in image separation may be described below [20].
Algorithm 1:
Step1: The base lines such as upper, lower and middle base lines are calculated.
Step 2: Word is over segmented.
Step 3: The word segmentation points are modified using various rules.
Step 4: Undesired and incorrect segmentation points are removed.
Step 5: The correct segmentation points are used for final segmentation of words.
Step 6: Position of characters and words in the script are stored.
198 H. Chel, A. Majumder, and D. Nandi

3 Feature Extraction
As it was mentioned earlier in introduction that reliable and meaningful feature
increases the recognition performance. Reliable feature means those features which
produce almost same value irrespective of slight variations in different sample images
of same character. In this experiment three kinds of features extracted and a total 88
nos. of features were collected per character. Before feature extraction all the
segmented character is resized [17] in a fixed size (60 x40). One runtime separated
unbounded character image, bounded image and resized image is shown in figure 1. A
short description of all four kinds of features may be mentioned below.

Fig. 1. Unbounded image, Bounded image and Resized Image

3.1 Transition Feature

Suppose the character image has a fixed dimension of X rows and Y columns. An M
x N grid is superimposed over the image as shown in figure 2. The grid is prepared in
such a way that all start and end values of each row and column got an integer value.
The start values of row m and column n may be written as under

1 1

1
2,3 … 2
1

2,3 . . . 3

refers to nearest integer value to x. Assume the intensity of coordinate(x,y) of

the X x Y image is A(x,y). The original image is scanned along every row and
column and the gradient information along each row and columns are collected.

∆ , A , 1 , 4

1,2 … … . 1
And
∆ x, y 1, , 5

1,2 … … . 1
 Scaled Conjugate Gradient Algorithm in Neural Network 199

As the image is a binary image both ∆ , ∆ x, y return one of the three values
0, 1 or -1. 0 indicate no transition 1 means white to black transition and -1 indicate
transition from black to white. A 5x5 grid over character image is shown on figure 2.a
and transitions are also shown in figure 2.b. In this case we would consider only -1
value and in each row and columns of the M x N grid total number of transitions in
every row and columns are important features of the image. As it was mentioned
earlier that the approach of the experiment is neural network based and the immediate
operation after feature extraction is training, it is found experimentally that neural
network works reliably and faster if the input feature values are within 0 to 1 limit. As
in this case all values are more than 1, they are normalized between 0 and 1 by a
nonlinear transfer function as mentioned below.

1
……………0 5 6
1

k is a constant and values may be chosen between .3 to 1 for better result. The upper
limit of x is taken as 5 because maximum possible white to black transition in
handwritten characters found 5. In this way we found (M+N) nos. feature for each
handwritten character. At the time of scanning the original image through all M rows
and N columns using M x N grid and the coordinates of first and last transitions in
each rows and columns are stored and a new kind of transition features were
collected. Suppose in row first and last transition occurs at , and
, then two features can be collected from that row as mentioned below

(7)

(8)

Similarly if in column of the M x N grid first and last transition occurs at

, and , then two features can be collected from that column as
mentioned below

(9)

(10)

Fig.2. a. 5 x 5 grid over original image Fig. 2. b. 5 x 5 grid over grid image
200 H. Chel, A. Majumder, and D. Nandi

3.2 Sliding Window Amplitude Feature

In this feature extraction scheme the image is subdivided into some overlapping
windows such that half of the window breadth and width are overlapped with
previous and next window along row and column wise except the windows at
boundaries and corners. The windows at left boundary have no windows left at left
side that it will overlap. Similarly top, bottom and right boundary side windows do
not overlap with any windows beyond their boundary. The corner block has only two
blocks to overlap for example the top left corner has only right and down windows to
overlap. Black pixel density in each block or window is calculated. The original
image of handwritten character has a fixed size of X rows and Y columns. As shown
in figure (2.a) the image is superimposed with M x N grid. The four corners of the
rectangular window R(m,n) are [(xm,yn) (xm, yn+2) (xm+2, yn) (xm+2, yn+2)] .Values of xm
and yn is defined in eqn. 1-3. Each block produces one feature and the feature is
normalized between 0 and 1 and a total (M-2) (N-2) features are collected with
normalized feature values [13]. In our experiment we have taken both M and N as 8
and a total 36 features were collected per character.

3.3 Contour Feature

The most important feature extraction scheme is contour feature [2][ 3]. Image is
scanned and filtered. The filtered image is superimposed with an 11 X 1 grid as
shown in figure (3.a) to find the external boundary intersection points of the character
and total 22 such points are collected. All the 22 points are numbered in a circular
fashion that from the top to bottom as 1 to 11 and from bottom to top as 12 to 22. All
points are connected with the next point to draw an exterior boundary contour as
shown in figure (3.b). For example the nth line is bound between two coordinate
(xn,yn) and (xn+1, xn+1).. The alignment of all the boundary lines is unique property of
that image and it differs from character to character. The two unique parameters of a
straight line that is the gradient and a constant value which solely depend on the
coordinates of the two points are taken as feature normalizing by a nonlinear
hyperbolic function. Equation of a straight line and its representation by two points
may be mentioned by equation 12-14. Suppose a straight line in equation is bounded
between two points (x1, y1) and (x2, y2). The m is the gradient and c is a constant
parameter of the straight line. Both m and c are two unique parameters which contain
the boundary patterns of the characters.

12

13

14
 Scaled Conjugate Gradient Algorithm in Neural Network 201

Fig. 3.a. 11 X 1 grid superimposed over Fig. 3.b. Contour of image represented with
boundary image straight line

Both the m and c may have any positive and negative values between negative
infinity and positive infinity. As the immediate process after feature extraction is
training of neural network with
1
feature matrix all the feature
0.9 values should be normalized
0.8 between 0 and 1 as it was shown
0.7 by experiment that neural network
0.6
shows better response if the inputs
are in range within 0 to 1 limit.
f(x)

0.5
For normalizing all the values of
0.4
gradient (m) and constant (c) a
0.3 hyperbolic nonlinear function is
0.2 used as transfer function which is
0.1
shown in equation 15 and a
graphical representation is also
0
-100 -80 -60 -40 -20 0 20 40 60 80 100 shown in figure (4).In equation 15
Nonlinear transfer function x
any value of x maps the output
Fig. 4. Nonlinear Hyperbolic Transfer Function between 0 and 1.k is a constant
and value of it depends upon the
no of features extracted. For 22 nos. of contour feature a value range from k=.3 to .5
is selected.
1
15
1
If no of feature is increased its value should be decreased and similarly if no of
features are decreased its value should be increased. In figure 4 plots for x versus f(x)
is shown for an x range from -100 to 100. In this way 22 nos. features are collected
and finally put in to the feature matrix. In our experiment we have taken three kinds
of features extraction schemes and we have taken 30 transition features, 36 sliding
window amplitude feature and 22 contour features and a total 88 nos. features are
collected. For training ten sets of capital letters (A to Z) and ten sets of Small letters
(a to z) written over two A4 size pages are taken and processed as described in the
202 H. Chel, A. Majumder, and D. Nandi

previous sections. Finally the feature matrix having dimension 520x 88 is fed to the
neural network for training. The description of neural network and algorithm used for
training are described in the next section.

4 Training of Neural Network

The immediate operation after feature extraction is training of a neural network. In
this experiment we have used both single and multilayer feed forward network. It is
not surprising that increase in no of feature improves recognition efficiency. But in
opposite side it takes a longer time to train the neural network. In our experiment we
have taken 88 nos. different kinds of features and 520 training character set as a result
total 45760 nos. of features were collected. But theses large nos. of features takes very
longer time to train the network if a proper training algorithm is not chosen. We have
used different conjugate gradient methods for training and the speed of training
improved significantly. All the algorithms are used for experiment are second order
conjugate direction method such as Fleture-Reeves [15], [21], Powell Beale [24],
Polak Ribiere [16],[ 19] and Scaled Conjugate Gradient algorithm by Moller [22] .A
comparative study also been made among the above algorithms and a comparison of
converging performance of all the second order conjugate gradient methods is
presented in result section. It has also been found experimentally that among the
entire training algorithms Powell Beale method and Scaled conjugate methods have
shown better performance in respect of quick convergence of the error curve. Using
Scaled Conjugate Gradient algorithm [22] the Hessian matrix of error equation is
always positive over all iterations. But in all other algorithms mentioned above this
thing is not guaranteed. This property of SCG algorithm increases learning speed
reliably in successive iteration. Let us we define an error function Taylor Series
expansion.

∆w E w E w T ∆w ∆w T E w ∆w 16

Suppose the notations are used as and ’’ . Weight vector in nth

iteration may be mentioned as n. H is the Hessian matrix and A is the local gradient
vector.
We will use a second order approximated error equation for further calculations
and the same equation may be as follows

∆w E w E w T ∆w ∆w T E w ∆w 17

The solution of the quadratic difference equation may be found as follows [15].
∆w H A 18
The above solution can be achieved subject to condition that H is positive definite
matrix. The above equation states that how much amount of shifts is required for all
the weights so that the error curve converges significantly. The above equation (18) is
the essence of Newtown’s Method [15] and from the equation it is found that the
equation can be converged in one step if the inverse of the Hessian matrix is
 Scaled Conjugate Gradient Algorithm in Neural Network 203

calculated. But in practical situation this phenomenon does not occur because of the
some constraints as mentioned below.
a) When number of Weights is more calculation of inverse of the Hessian
Matrix is computationally expansive and highly time consuming matter.
b) There is no guarantee that the inverse of the Hessian Matrix will be positive
definite.
c) The system converges in one step if and only if the error equation is
perfectly quadratic in nature. But generally all error equation has some
higher order terms.
To avoid the above limitations the only way was found to achieve the solution
through an iterative process. Suppose a set of nonzero vectors p1, p2, p3………. pN are
basis vectors in RN and are H conjugate. If H is a positive definite matrix then the
following conditioned must be satisfied [10].

0
0 19)

Let . When k= i= n then xn is nonzero quantity.

20
̃

Here ̃ ” . The idea to estimate the term sn with a non symmetric

approximation [22] it may be written as

̃ Where 0 1 21

The above equation is the essence of second order error estimation used by Hestene
[10] in conjugate gradient method. Scaled Conjugate Gradient method is a slightly
different concept applied over it. In eqn.( 20) the sign of xn may be negative or
positive but by definition if H is a positive definite matrix then xn must be a greater
than zero. Hestene [10] combines the concept of introducing a dumping scalar value
and modifies the equation as in eqn. 22. The other steps are same to the conjugate
direction method. The function of the scalar parameter λ is to compensate

Where 0 1 22

the indefiniteness of value of E”(wk) when it is negative. In every iteration sign of xn

is checked i.e. whether xn >0 or xn <0. When xn is less than zero the value of λn is
increased and similarly when xn is greater than zero the value of λn is decreased. All
other steps are similar to the conjugate direction method.
The whole process is governed by basically two up gradation equation. Firstly the
weight up gradation equation (eqn. 23) and secondly the basis vector up gradation
equation (eqn. 24). The target is to find such a solution set of weight vector that H
become positive definite. As the iterations go forward the first order error gradient
E’(wn ) decreases and at last it reaches very near to zero. The equations are as follows
204 H. Chel, A. Majumder, and D. Nandi

23
̃ 24
Where ’ , αn and βn may be calculated in each iteration by the following
equations.
̃ ̃
25

̃ , ,
̃ , ̃
And 26

The most important matter is to discuss that how the value of λ in eqn. 22 is chosen.
As described earlier that, the function of λn is to check the sign of xn in each iteration
and to set a proper value of λn so that get a positive value.
Let it is found that xn≤ 0 and λn is raised by λ so that sn get a new value

s λ 27

28

Putting the value of from eqn. no 27 in eqn. 28

λ | | 0 29

λ 30
| |

From eqn. 29 some guide line is found that the new should be greater than by -
| |
but no such particular value can be obtained that so that the optimal solution
can be obtained. However we have used

λ 3 | |
31

The above assumption is put into eqn. 29 we get

2 3 | | 32
Combining eqn. 32 and 25 we found the modified value of
̃
| |
33

the above equation clears that with the increase of the value of λn decrease the height
of step size and decrease in λn increases the height of step size which agrees to all the
assumptions made earlier.
In the above steps it was realized that how the Hessian Matrix remain positive in
all iterations using a scalar parameter λ. But the second order approximation of error
equation which was used during entire calculation cannot assure for the best
performance though we get positive definite Hessian matrix over all iterations. The
 Scaled Conjugate Gradient Algorithm in Neural Network 205

above calculation steps assure that the error curve will be converging towards minima
but in some situation choosing a proper value of λ is required otherwise the rate of
convergence become slow. So a proper mechanism for scaling λ is adopted. A
parameter ρ is defined which measure in what extent the second approximation of the
error curve matches the original error curve. The following equation defines ρ as

34

Here measures how finely the second order approximated error equation
matches to the . The above equation tells than the value
of λ nearer to 1 means better the approximation. For faster convergence λ is scaled by
the following equations.
1
.75
3
1
.25 35
| |
In our experiments the initial values was chosen as λ1=10-3 and 10 was taken
and in later steps was assumed. Initially weights are chosen as random
| |
nos. between 0 and 1. At starting of training in first iteration is
assumed. The algorithm may be summarized as below.
a) Initialization of parameters like σ,p,r in first iteration.
b) Calculation of second order parameters like s,x and σ
c) Check whether Hessian matrix H is positive definite or not
d) If false adjust the value of s by increasing λ and recalculate s again.
e) Calculate and readjust the value of λ
f) Calculate step size
g) If error > minimum error limit go to next iteration.
h) Accept the weight vector for test.

During the training opera-

tion the epoch versus error
data is collected and the
nature of the convergence is
also noted. A run time error
curve is plotted in figure 5
which shows that the nature
of convergence is very fast
and it takes just one or two
min for completion of
training and error limit
reaches in 10-12 range within
only 700 epoch.

Fig. 5. epoch versus error curve

206 H. Chel, A. Majumder, and D. Nandi

5 Result and Discussion

The experiment for recognizing handwritten text was conducted in MATLAB
environment where two types of images are used as input i.e. train image and test
image. The purpose of the experiment was to recognize individual’s handwriting by a
neural network which is trained to identify the patterns of handwriting of the same
person. Here both the train and test script was written by same person. For this reason
this experiment was not conducted and compared by any standard data base like
CEDER or IAM data base of handwritten text. Two major features are highlighted in
this experiment which differs it from all other researches [6],[13],[23] done earlier
may be mentioned that is i) the superfast training speed ii) high recognition
performance. Ten sets of handwritten capital and small letters of English alphabets are
taken as training script and various handwritten texts written by same person are used
as test script. A sample train and test scripts are shown in figure 6 and figure 7. All
the characters are written in natural way over A4 sheet. In both training and testing
stage common character level segmentation and feature extraction is done over the
train and test script. Training characters are English alphabets so there is no concept
of forming words but test script is a text or may be called as group of words without
any special symbol like ;, : and ?. In case of test script the segmentation is done from
text level to word level and later from word level to character level. The start and end
locations of each word in the text are stored. After recognition of all the characters in
test script the computer printed words are regrouped by the NN output characters
using the start and end location information of each words. Character level and Word
level efficiency may be defined as under.
Character Level Efficiency = (No of characters recognized correctly)/(Total no of characters in
the script)

Word Level Efficiency = (No of words recognized correctly)/(Total no of words in the script)
One character misclassification in a word decreases the word level recognition
performance. But it gets easily be improved if the same word is checked through a

Fig. 6. Train script Fig. 7. Test script

 Scaled Conjugate Gradient Algorithm in Neural Network 207

Table 1. Comparison between different training algorithms dictionary for

validity of the
No of NN Training No of Train Error word. If the
Features/ structure algorithm epochs efficiency limit word exists in
character
the dictionary
88 88,120, SCG 700 100% 10-12 it returns the
52 same word
88 88,120, Powell- 1480 100% 10-12 otherwise nea-
52 Beale rest word of
-12
same length is
88 88,120, Fleture 5658 100% 10 returned. For
52 Reeves implementing
this one dictionary is formed in MATLAB as an m-file which contains more than two
thousand common words taken from a standard oxford dictionary. In this paper
various algorithms were used for training such as a) Scaled Conjugate direction
Method b) Powell Beale method and c) Fleture-Reeves method among the three
algorithm Scaled conjugate algorithm shown the best training speed. A comparative
statement is described in table.1.the above table shows that total 88 nos. of features
were taken and NN structure indicates 52 nos. of output neurons for 52 characters (a-z
and A-Z) and the result was obtained for 120 nos. of hidden neurons. Both the three
algorithms provide good convergence of the error curve but Scaled Conjugate
Gradient algorithm provides fastest training speed and the training completes within
700 epoch which takes less than one minute for reaching 100 percent train efficiency.
The neural network trained with ten sets of handwritten alphabets that is total 520 nos.
of characters when asked to make response to test feature matrix its response is really
satisfactory only a few misclassification arises for very similar type of characters and
for very bad handwritings. A character level recognition performance is shown in
table 2 which shows that using double hidden layer with 150 and 100 hidden neurons
we get maximum 83.60 percent case sensitive character level recognition and a

Table 2. Character level classification Table 3. Performance before and after

performance lexicon matching

NN structure MSE Train Test efficiency NN Without Lexicon After Lexicon Matching
obtained efficiency (Character Level) Structure Matching
% % Character Word Character Word
Case Case Level (%) Level Level (%) Level
sensitive insens (%) (%)
itive 88,160,52 89.17 62.87 98.43 96.74
88,160,52 9.63x10-13 100 81.43 89.17
88,200,52 89.42 48.65 99.02 97.83
-13
88,200,52 9.75x10 100 82.64 89.42
88,250,52 88.70 47.56 99.02 97.83
88,250,52 9.32x10-13 100 79.25 88.70
88,120,80,52 87.34 38.96 98.43 95.65
88,120,80,52 9.44x10-13 100 74.17 87.24
88,150,100,52 92.32 66.13 98.43 95.65
-13
88,150,100,52 9.69x10 100 83.60 92.32
88,200,100,52 89.72 47.56 99.02 98.91
88,200,100,52 9.98x10-13 100 81.67 87.72
208 H. Chel, A. Majumder, and D. Nandi

maximum 92.32 percent case insensitive character level recognition performance is

achieved. Though the degree of recognition is good but it may further be improved by
lexicon matching technique which was discussed earlier. The word label performance
after lexicon matching is shown in table 3. This shows that the case insensitive
performance shown in table 3 improves by maximum 99.02 percent and because some
of the word which contain lesser misclassified characters get corrected by the lexicon
matching technique which increases word level recognition and when word get
corrected the misclassified characters are also changes to its correct version. As a
whole both the word and character level performance increases. In this experiment
maximum word level and character level recognition performance achieved as 98.91
and 99.02 percent respectively. Though the result of the experiment was very well and
most of the words and characters were recognized correctly except a few
misclassifications that were found during experiment which may be shown in table 4.

Table 4. General misclassifications

Original Recognized Original Recognized

H M ,n n m
r v, Y Y r,y
e c,l,C y r,Y
a Q, u, l ,G c C
f t C c
s a,b, O o
l x o O,b,D
I j j i
b o t f,F,q

6 Conclusion
An efficient approach using a second order training algorithm is presented in this
paper for recognition of handwritten alphabet based text. When no. of features are
higher and no. of output classes are 52 all first order training algorithms basically
fails or generates very poor result in training. Basic reason is the slower convergence
of error curve and proper second order training algorithm become suitable
replacement of those algorithms. Using this algorithm Hessian matrix of error
equation always remain positive definite and in every iteration the error curve
converges in a faster way. In this paper comparison between several second order
training algorithms has been shown and it was found experimentally that Scaled
Conjugate Gradient algorithms works with fastest speed and recognition performance
is also excellent. The experiment was done using a Core 2 Duo (1.73 GHz) processor,
1 GB RAM, windows XP platform and MATLAB software. Training part was very
fast and most of the time it completed within one or two minutes. But regarding the
complexity of the test script, scripts characters are simple, easy to understand by
human eye. Test script contains some inter-line horizontal space and some inter-word
and inter-character space also. This pattern is not always available in natural
handwriting. So it needs more experimental effort for faithful conversion from
 Scaled Conjugate Gradient Algorithm in Neural Network 209

difficult handwriting to text conversion. However, this approach may be a true

guideline for future research for giving computer an intelligence which a human being
applies everyday and at every moment.

References
1. Cho, S.-B.: Neural-Network Classifiers for Recognizing Totally Unconstrained
Handwritten Numerals. IEEE Trans. on Neural Networks 8, 43–53 (1997)
2. Verma, B.: A Contour Code Feature Based Segmentation For Handwriting Recognition.
In: 7th IAPR International conference on Document Analysis and Recognition, ICDAR
2003, pp. 1203–1207 (2003)
3. Strathy, N.W., Suen, C.Y., Krzyzak, A.: Segmentation of Handwritten Digits using
Contour Features. In: ICDAR 1993, pp. 577–580 (2003)
4. Casey, R.G., Lecolinet, E.: A Survey of Methods and Strategies in Character
Segmentation. IEEE Trans. Pattern Analysis and Machine Intelligence 18, 690–706 (1996)
5. Fletcher, R., Reeves, C.M.: Function minimization by conjugate gradients. The Computer
Journal 7, 149–153 (1964)
6. Gorgevik, D., Cakmakov, D.: An Efficient Three-Stage Classifier for Handwritten Digit
Recognition. In: ICPR, vol. 4, pp. 507–510 (2004)
7. Fink, G.A., Plotz, T.: On Appearance-Based feature Extraction Methods for Writer-
Independent Handwritten Text Recognition. In: Proceedings of the 2005 Eight
International Conference on Document Analysis and Recognition (ICDAR 2005), pp.
1520–5263 (2005)
8. Dunn, C.E., Wang, P.S.P.: Character Segmentation Techniques for Handwritten Text A
Survey. In: Proceedings of the IInd International Conference on Pattern Recognition, The
Hague, The Netherlands, pp. 577–580 (1992)
9. Zimmermann, M., Bunke, H.: Optimizing the Integration of a Statistical Language Model
in HMM based Offline Handwritten Text Recognition. In: Proceedings of the 17th
International Conference on Pattern Recognition (ICPR 2004), pp. 1051–4651 (2004)
10. Hestenes, M.: Conjugate Direction Methods In Optimization. Springer, New York (1980)
11. Lee, S.-W.: Off-Line Recognition of Totally Unconstrained Handwritten Numerals Using
Multilayer Cluster Neural Network. IEEE Trans. on Pattern Analysis and Machine
Intelligence 18, 648–652 (1996)
12. Gader, P.D., Mohamed, M., Chiang, J.-H.: Handwritten Word Recognition with Character
and Inter-Character Neural Networks. IEEE Transactions on Systems, Man, and
Cybernetics-Part B:Cybernetics 27, 158–164 (1997)
13. Blumenstein, Verma, B.: Neural–based solutions for the segmentation and recognition of
difficult handwritten words from a benchmark database. In: Proc. 5th International
Conference on Document Analysis and Recognition, Bangalore, India, pp. 281–284 (1999)
14. Chiang, J.-H.: A Hybrid Neural Model in Handwritten Word Recognition. Neural
Networks 1I, 337–346 (1998)
15. Haykin, S.: Neural Networks A comprehensive Foundation, 2nd edn.
16. Duda, R.O., Hart, P.E., Stock, D.G.: Pattern classification, 2nd edn.
17. Pratt, W.K.: Digital Image Processing, 3rd edn.
18. Rajeshkaran, S., Vijayalakshmi Pai, G.A.: Neural Networks, Fuzzy Logic, and Genetic
Algorithms Synthesis and Applications. Eastern Economy
19. Khoda, K.M., Liu, Y., Storey, C.: Generalized Polak –Ribiere Algorithm. Journal of
Optimization Theory and Application 75(2) (November 1992)
210 H. Chel, A. Majumder, and D. Nandi

20. Verma, B., Lee, H.: A Segmentation based Adaptive Approach for Cursive Hand written
Text Recognition Neural Networks. In: International Joint Conference on IJCNN 2007,
August 12-17, pp. 2212–2216 (2007), doi:10.1109/IJCNN.2007.4371301
21. Fletcher, R.: Practical Methods of Optimization. John Wiley & Sons, New York (1975)
22. Moller, M.F.: A Scaled Conjugate Gradient Algorithm For Fast Supervised Learning.
Neural Networks 6, 525–533 (1993)
23. Blumenstein, M., Verma, B.: Neural-based Solutions for the Segmentation and
Recognition of Difficult Handwritten Words from a Benchmark Database. In: Proc. 5th
International Conference on Document Analysis and Recognition, Bangalore, India, pp.
281–284 (1999)
24. Dai, Y.H., Yuan, Y.: Convergence properties of the Beale-Powell restart algorithm. Sci.
China Ser. A 41, 1142–1150 (1998)
25. Pal, U., Sharma, N., Wakabayashi, T., Kimura, F.: Off-line handwritten character
recognition of Devanagari script. In: Proceedings of 9th International Conference on
Document Analysis and Recognition, vol. 1, pp. 496–500 (2007)
 A Linear Framework for Identification of Ship and Its
Velocity Estimation Using SAR Images

P. Subashini and M. Krishnaveni

Department of Computer Science,

Avinashilingam Deemed University for Women, Coimbatore
TamilNadu, India
{mail.p.subashini,krishnaveni.rd}@gmail.com

Abstract. Satellite-based Synthetic Aperture Radar (SAR) provides a powerful

vessel surveillance capability in front of time consuming traditional
reconnaissance methods. Nevertheless, due to the presence of speckle and to the
reduced dimensions of the targets compared to the sensor resolution, the
automatic interpretation of SAR images is often complicated even though
vessels undetected are sometimes visible by eye. Therefore, the main
difficulties of automatic ship detection will be reviewed and the main
drawbacks will be identified here. As a result of this preliminary analysis, the
application of transformations is found to be a useful mathematical object for
this purpose. Hence based on the capabilities of the image analysis, the
objective is achieved in such a way that it provides better understanding and
further exploitation of satellite high resolution images. More specifically, a
novel method for ship detection based on multiscale tools is proposed, in which
SAR images are over here.

Keywords: Ship detection, velocity, transformation, segmentation, SAR images.

1 Introduction
In navy radar applications the presentation of the radar image traditionally has been
the way for the radar operator to interpret the radar information manually. The large
increase in calculation capacity of the image processing in modern radar systems had
the effect that the detection and extraction of targets has been automated[9]. With
powerful image processing techniques and algorithms, modern radar systems has the
possibility to extract targets and their velocity from the surrounding background. A
condition for this automatic detection is that the radar image should be relatively free
from undesired signals. Such undesired signals can be rain clutter, sea clutter,
measuring noise, landmasses, birds etc. Doppler filtering is often used to remove
these undesired signals and extract the interesting part of the radar signal. By
literature given by I-I Lin and Victor Khoo [2], revelation is that with modern image
processing techniques be able to improve the radar image and investigate an
automatic classification and presentation of the objects. The analysis of ship wakes in
SAR imagery with specialized algorithms can provide significant information about a
wake’s associated vessel, including the approximate size and heading of the ship. The

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 211–224, 2011.
© Springer-Verlag Berlin Heidelberg 2011
212 P. Subashini and M. Krishnaveni

velocity of the vessel can be estimated by measuring the disarticulation of the ship
relative to the height of the wake. Image processing algorithms such as the Fourier
Transform and the Radon Transform allow the user to manipulate SAR images in a
way that dramatically increases the chance of detecting ship wakes[14][11]. The
paper is organized as follows: Section 2 deals with the structure of the framework for
target detection. Section 3 deals with the preprocessing the image before
identification. Section 4 converses the proposed algorithm for detecting the ship with
high probability of detection. Section 5 and 6 converse the determination of target
centre and wake representation through transformation Section 7 explains the
experimental results and findings of the target detection and estimation of vessel
velocity. The paper concludes with the possible improvements and the future work of
the system.

2 Structure of the Framework for Target Detection

A digital image is built up with a number of elements, called pixels, usually in matrix
form. The matrix size corresponds to the image size. Other forms of image
representation are chains and graphs. In radar terminology, the pixels in a radar image
are named radar bins or just bins. The pixel value corresponds to its brightness or
color in the picture. The brightness can be used to interpret different pictures in
different ways, depending of what the brightness mirrors.
In surface radar images, the brightness reflects the received signal strength. Images
can be of different kinds. One way to sort images is to divide them into intensity and
binary images. The binary image has only two levels, 0 and 1. Multispectral images,
or intensity images, have more levels. The proposed work is given in figure 1. This
proposed framework combines the existing and recent techniques of image processing
techniques for detecting target. The output of the system delivers a Matlab-based
application development platform intended for detection system. The design let the
user to examine, devise, and evaluate algorithms and applications using SAR images.
It offers standardization for the mentioned purpose that the user does not require
comprehensive knowledge of the target hardware.

3 Preprocessing SAR Image

This is the first and lowest level operation to be done on images. The input and the
output are both intensity images[1]. The main idea with the preprocessing is to
suppress information in the image that is not relevant for its purpose or the following
analysis of the image. The pre-processing techniques use the fact that neighboring
pixels have essentially the same brightness. There are many different pre-processing
methods developed and for different purposes. Interesting areas of pre-processing for
this work is image conversion, noise suppression and edge detection. Image
smoothing is a local pre-processing technique, which mainly is used for noise
suppression. The method calculates new pixel values in a local neighborhood. In
image enhancement, restoration is handled over the SAR image which otherwise
called restoration.
A Linear Framework for Identification of Ship and Its Velocity Estimation 213

5DZUDGDULPDJH

,PSXOVH,QWHUIHUHQFHILOWHU

1RLVHUHGXFWLRQ

(GJH GHWHFWLRQ

&RPSXWHV7,'PHWKRGRSWLPDOWKUHVKROG

&OHDULPDJHERUGHU

,GHQWLW\LQGLYLGXDOWDUJHW

&RQVLGHUHDFKWDUJHW

'HWHUPLQDWLRQRIWDUJHW
&HQWUHDQG2ULHQWDWLRQ

3HUIRUP7UDQVIRUPDWLRQVIRU:DNH'HWHFWLRQ

'HWHUPLQLQJSDUDPHWULFUHSUHVHQWDWLRQRIOLQHV:DNHV

(VWLPDWH6KLS3DUDPHWHUV
9HORFLW\

Fig. 1. A proposed Linear framework for ship and wake detection

214 P. Subashini and M. Krishnaveni

3.1 Noise Removal

According to the characteristics of SAR images it contains speckle noise [7]which

henceforth reduced using three filters namely (i) Inverse filter (ii) Wiener filter (iii)
Lucy Richardson filter .On the experimental verification, the result obtained
articulates that wiener filter gives better outcome for noise suppression.
Inverse filter -Inverse filter is a form of high pass filer. Inverse filtering responds
very badly to any noise that is present in the image because noise tends to be high
frequency. Works better for a noiseless case[5].
Wiener filter - The Wiener filtering is a linear estimation of the original image.It is
also called regularized inverse filter.It removes the additive noise and inverts the
blurring simultaneously.
Lucy Richardson filter - Lucy-Richardson denoising accounts on photon counting
noise with a Poisson distribution. The Lucy-Richardson algorithm generates a
restored image through an iterative method. Lucy-Richardson maximum likelihood
algorithm addresses complex image restoration tasks.

Fig. 2. Subjective evaluation for Noise filters

3.2 Edge Detection

Roberts operator
The Roberts operator performs a simple, quick to compute, 2-D spatial
gradient measurement on an image[7]. The gradient magnitude is
given by

G = G x 2
+ G y 2 (1)

And the angle of orientation of the edge giving rise to the spatial gradient (relative to
the pixel grid orientation) is given by:
 A Linear Framework for Identification of Ship and Its Velocity Estimation 215

θ = arctan(Gy / Gx ) − 3π / 4 (2)

Prewitt operator
The result of using this edge operator is given by.
G = m ax (G i : i = 1 to n ) (3)

For each pixel the local edge gradient magnitude is estimated with the maximum
response of all kernels at this pixel location.

Laplacian of Gaussian (LoG)

The Laplacian L(x, y) of an image with pixel intensity values I(x, y) is given by:
∂ 2I ∂ 2I (4)
L ( x.y ) = +
∂x 2
∂y 2

Since the input image is represented as a set of discrete pixels, a discrete convolution
kernel can approximate the second derivatives in the definition of the Laplacian.

Table 1. Advantages and disadvantages of Edge Detection Techniques

Operator Advantages Disadvantages

Classical  Simplicity,  Sensitivity to

(Prewitt,  Detection of edges and their noise,
Roberts..) orientations  Inaccurate
 Using probability for finding
error rate  Complex
 Localization and response Computations
Canny  Improving signal to Noise  False Zero crossing
ratio  Time Consuming
 Better detection specially in
noise conditions
 Higher weights are assigned
to the pixels close to the  Coupling between
Sobel candidate pixels. image pixels on a
 It masks for horizontal and single scale
vertical edge detection

Canny operator
The Canny method finds edges by looking for local maxima of the gradient of image.
The method uses two thresholds, to detect strong and weak edges, and includes the
weak edges in the output only if they are connected to strong edges. The edge
gradient and direction can be determined using
216 P. Subashini and M. Krishnaveni

G = Gx + G y
2 2

Gy
θ = arctan( ) (5)
Gx

Fig. 3. Subjective evaluation for edge operators

4 Identifying Individual Target

Segmentation of target in SAR images is an important step in any ship detection

system and many different approaches to the problem have been proposed so far[13].
The approaches of the paper are Adaptive thresholding (manual and automatic) and
Distribution thresholding[3]. Two methods for thresholding are taken. The hard part
of the thresholding is (1) Selecting suitable threshold for any input image (2)
Deciding the proper class hierarchy for modular implementation (3) Implementing
path traversal and nonmaximal suppression algorithms. In simple implementations,
the segmentation is determined by a parameter T known as a threshold which is a
result of analysis of the attribute Y(x,y). In the simplest algorithm implementation,
each pixel of the image is compared with the threshold T. If the pixel's value of the
attribute Y(x,y) is lower than the threshold, the pixel is qualified to the background,
otherwise the pixel is marked as object’s pixel.
Manual –A binary output image one containing only 0’s (background) and 1’s
(object),may be created by applying a threshold:

0
Vout (i, j ) =  if otherwise vin (i, j ) < threshold (6)
1
This can be extended to as many different classes as necessary by defining thresholds
t1,t2 such that :
 A Linear Framework for Identification of Ship and Its Velocity Estimation 217

0 if 0 ≤ ν in (i , j )< t1
1 ift 1 ≤ ν in (i, j) < t2 (7)
2 ift 2 ≤ ν in (i, j ) < t3

The most common and reliable method of finding the thresholds is manual selection.

4.1 Iterative Threshold Selection Method

The methodology of the threshold value selection starts from segmenting the
grayscale of the neighborhood into two parts using an initial threshold T0 which is set
to the neighborhood average lightness value LN. Then the means of the gray values
associated with the object’s pixels Lo, and the background pixels LB, are computed.
The new threshold value T1 is now set to the average of these two means as shown in
the following equation:

LB + LO
T1 = (8)
2
The process is repeated based upon a new threshold, until a threshold averages to a
constant value. This situation corresponds to the following equation.

T i = T i + 1
(9)

The size of the neighborhood is determined by a 3x3 size mask.

4.2 Threshold Intensity Distribution Method

The segmentation of the reference image is designed to group pixels of similar gray-
levels[4]. Since the sea surface and the ship targets are usually poorly textured, the
gray-level segmentation is likely to group pixels belonging to objects or background.
If the objects and background occupy different ranges of gray levels, then it can mark
the object pixels by a process called thresholding[6].
In intensity distribution B (i,j) is a binary image (pixel are either 0 or 1) created by
thresholding F(i,j)

B(i,j) = 1 if F (i,j)<t

B (i,j) =0 if F(I,j)>=t (10)

It is assumed that the 1’s are the object pixels and the 0’s are the background pixels.
The Histogram (h) - gray level frequency distribution of the gray level image F.
hF (g) = number of pixels in F whose gray level is g
HF (g) =number of pixels in F whose gray level is <=g
218 P. Subashini and M. Krishnaveni

Fig. 4. Intensity distribution

Fig. 5. Subjective evaluation of Segmentation methods (a) Original image (b) iterative method
(c) intensity distribution method

This method is a probabilistic method that makes parametric assumptions about

object and background intensity distributions and then derives “optimal” thresholds.

5 Determine Target Centre and Finding Velocity Using Wake

Representation
(i) Two features [10] of the target is obtained using property method
Centroid'-- 1-by-n dimensions (L) vector; the center of mass of the region. The first
element of Centroid is the horizontal coordinate (or x-coordinate) of the center of
mass, and the second element is the vertical coordinate (or y-coordinate). All other
elements of Centroid are in order of dimension.

Orientation' -- Scalar; the angle (in degrees) between the x-axis and the major axis of
the ellipse that has the same second-moments as the region. This property is
supported only for 2-D input label matrices.
 A Linear Framework for Identification of Ship and Its Velocity Estimation 219

(ii) Determine parametric representation of wake

The angle f is calculated as , and a = f +90° . The Radon transform is
performed to all ranges of a and r . However, in the case of wake detection, it is not
necessary to apply the Radon transform to all possible a since turbulent wake is parallel
to the ship[2][12]. Nevertheless, to account for possible errors, a range of a is used. This
range is chosen as ± 20° from a with interval every 5° , i.e. the range of a used is í a -20°
, a -15° , a -10° , a -5° , a , a +5° , a +10° ,a +15° ,a +20° ý . Similarly, it is not necessary
to use the full range of r since ship can only travel at some maximum speed and thus
both the azimuth shift D and r are limited. After the transformation, the maximum and
minimum of all the g(a ,r ) are calculated. The maximum will correspond to a bright
turbulent wake while the minimum will be associated with a dark turbulent wake and
the wake position in terms of (a ,r ) can be identified.

Fig. 6. Wake representation using radon transform

6 Performance Evaluation
The measurement of the image quality is divided into two main categories, subjective
and objective. The subjective category is a measurement of how a human observe and
interpret the image. The objective category is a technical measurement and the image
is often compared with a known reference image[8]. The image used for image
processing is in the format of .PNG and the size of the image is 256 x 256. It is
necessary to find out the best method that is suitable of estimation. The performance
evaluation of the adopted operation is quantified by following assessment.
220 P. Subashini and M. Krishnaveni

6.1 Noise Removal

Peak to signal noise ratio (PSNR)
 255 2 
PSNR = 10 log 10   (11)
 MSE 
Larger the PSNR value [40] the performance is good and smaller the MSE value the
error rate is low.

Mean Square Error (MSE)

MSE is calculated using formula (1):
M N 2
1
M SE =
MN
  [g (i,
i =1 i =1
j ) − f (i , j ) ] (12)

where, M and N are the total number of pixels in the horizontal and the vertical
dimensions of image. g denotes the Noise image and f denotes the filtered image.

Fig. 7. Comparison based on PSNR

6.2 Edge Detection

The overall image sharpness is measured by computing the average difference

between pixel pairs at all locations within the input image. The pixel subtraction
operator takes two images as input and produces as output a third image whose pixel
values are simply those of the first image minus the corresponding pixel values from
the second image. If the operator computes absolute differences between the two
input images then the formula is given in the Eq:
Q = P1 ( i , j ) − P2 ( i , j ) (13)
where Q is the absolute difference and the P1 and P2 are the pixel of original image
and filtered image. The absolute difference is calculated between the image in
memory and the current image.
 A Linear Framework for Identification of Ship and Its Velocity Estimation 221

Fig. 8. Comparison based on MSE

Fig. 9. Comparison based on AAD

6.3 Segmentation

Structural content. A structural content is used to find mask structural defects and
noise, but are prone to periodic artifacts. This examines to classify image regions
hierarchically based on the level of structural content.

M N M N
SC =  x j , k 2 2 /  x′j , k 2 2 / (14)
j =1 k =1 j =1 k =1
222 P. Subashini and M. Krishnaveni

Structural content of a segmented image is calculated by summation of the original

image by the division of the segmented image.

Fig. 10. Comparison based on PSNR


6KLSSL[HOV
'HFLVLRQLQIHDWXUHVSDFH

Fig. 11. performance of target detection

6.4 Estimating the Velocity of the Target

The velocity is estimated with the target centre and the wakes angle values and the
calculation in given in figure 12. In figure 12, the SAR image taken indicates the
speed of the vessel as 12.5 m/s. As per the image analysis the estimated value from
the experiment is equal to the insitu value of the SAR image.
 A Linear Framework for Identification of Ship and Its Velocity Estimation 223

Image for Analysis

Ship displaced from wake indicate speed (left

with a speed of 6 m/s, and the right with 12.5
m/s ERS-2 03 APR-1996 03:29:28 Hr
descending mode,incident angle 23 lat:01.64
long 102.69

Calculated velocity

Centroid: [12.1389 10.6667]

Orientation: 21.0129

−1 ρ
υ= ( )
44 cos φ sin α

velocity = 12.31527 m/s

Fig. 12. Estimation of ship velocity

7 Conclusion
In this paper, individual object detection problem is analyzed for different SAR
images. At the first glance, the problem is separated into two classes: identifying the
target and determining the wakes for finding velocity using SAR images. The
solutions proposed in preprocessing are: enhancement, noise removal, edge detection.
The noise elimination method gives promising results in the controlled environments.
The reliability of the edge detection process turns out to be the most important step in
this method. The best results can be obtained, when “canny method” is used for
background extraction. Thresholding and morphological operations steps are also
very important for binarization and localization of the detection results. Automatic
threshold selection is also used for segmentation. The experiments show that adaptive
selection of the threshold for each image pair gives more promising results than a
global threshold selection. Wake detection is handled with all the conventional
transform method and the radon transform gives better solution for finding the
velocity of the ship. The results of this detection scheme are reliable and satisfactory
224 P. Subashini and M. Krishnaveni

while the computation time is quite high compared to background elimination

method. In the validation sequences, the detection problem is much more integrated
with new algorithms to bring out optimal output. However, the segmentation process
is very critical in this approach and it may require manual interaction for
optimization. Therefore, more consistent unification approaches, which include
temporal analysis, is required and is left as a future work.

References
[1] Tekalp, A.M., Parbolic, G.: Digital Restoration of Images scanned from photographic
paper. Journal of Electronic Imaging 2(c), 19–27 (1993)
[2] Lin, I.-I., Khoo, V.: Computer-based algorithm for ship detection from ERS SAR
imagery. ESA.Earthonline.net
[3] Kittler, J., Illingworth, J.: Threshold selection based on a simple image statistic.
CVGIP (30), 125–147 (1985)
[4] Otsu, N.: A thresholding selection method from graylevel histogram. IEEE Transactions
on Systems, Man, and Cybernetics (9), 62–66 (1979)
[5] Hu, J., Yan, H., Hu, X.: Removal of impulse noise in highly corrupted digital images
using a relaxation algorithm. Optical Engineering 36(03), 849–856 (1997); Proc SPIE
[6] Tsai, D.M., Chen, Y.H.: A fast histogram-clustering approach for multilevel thresholding.
Pattern Recognition Letters 13(4), 245–252 (1992)
[7] Robinson, G.: Edge Detection by Compass Gradient Masks. Computer Graphic Image
Processing 6, 492–501 (1997)
[8] Wang, Z., Bovik, A.C.: A universal image quality index. IEEE Signal Processing
Letters 9(3), 81–84 (2002)
[9] Lopes, A., Nezry, E., Touzi, R.: Structure detection and statistical adaptive speckle
filtering in SAR images. International Journal of Remote Sensing 14, 1735–1758 (1993)
[10] Mukherjee, A., Parui, S.K., Chaudhuri, B.B., Krishnan, R., Rao, K.K.: Detection of linear
features in satellite imagery using robust estimation. In: Proceedings of the 12th
International Conference on Pattern Recognition, Jerusalem, pp. 514–516 (1994)
[11] Krogagar, E., Czyz, Z.H.: Properties of the sphere, diplane, helix (target scattering matrix
decomposition). In: Saillard, J., et al. (eds.) Proc. JIPR-3, Nantes, France, March 21-23,
pp. 106–114 (1995)
[12] Cameron, W.L., Youssef, N.N., Leung, L.K.: Simulated polarimetric signatures of
primitive geometrical shapes. IEEE Trans. on GRS 34(3), 793–803 (1996)
[13] Touzi, R., Charbonneau, F.: Characterization of target symmetric scattering using polari-
metric SARs. IEEE Trans. on GRS 40(11), 2507–2516 (2002)
[14] Freeman, A., Durden, S.L.: A three-component scattering model for polarimetric
SARdata. IEEE Trans. on GRS 36(3), 963–973 (1996)
 Creational Patterns to Create Decorator Pattern
Objects in Web Application

Vijay K. Kerji

Department of Computer Science, PDA College of Engineering, Gulbarga, India

[email protected]

Abstract. Creational Design Patterns play an important role in making the

software applications most reusable and adaptable to future changes in software
requirements. Creational patterns abstract (encapsulate) how objects are created
in the software system. They enable client code independent of object creation,
object composition and how the objects are represented. Abstract Factory
creational design pattern is one of the object creational patterns which help
software system independent of how objects are created. In this paper, we
proposed and applied creational patterns (Abstract Factory and Singleton) to
achieve the above mentioned objectives. Creation of Decorator Pattern objects
in DMS application is accomplished by Abstract Factory and Singleton
patterns. This change enabled the application more adaptable to the changes
(e.g. any addition of new user) without changing the client code and enhanced
the reusability of the system with increased ease of system maintenance.
Proposed changes are applied to the DMS application using .NET framework,
C# and ASP.NET.

Keywords: Design Patterns, Abstract Factory, Singleton, Reusability, Web

Application.

1 Introduction
Design Patterns refer to reusable or repeatable solutions that aim to solve similar
design problems during development process. Various design patterns are available to
support development process. Each pattern provides the main solution for particular
problem. From developer’s perspectives, design patterns produce a more maintainable
design. While from user’s perspectives, the solutions provided by design patterns will
enhance the usability of web applications [1] [2].
Normally, developers use their own design for a given software requirement,
which may be new each time they design for a similar requirement. Such method is
time consuming and makes software maintenance more difficult. Adapting the
software application to future requirements changes will be difficult if adequate care
is not taken during design [3].
Adopting Design Pattern while designing web applications can promote reusability
and consistency of the web application. Without the adoption of the design patterns or
wrong selection of them will make the development of web application more complex
and hard to be maintained. Thus knowledge in design patterns is essential in selecting

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 225–232, 2011.
© Springer-Verlag Berlin Heidelberg 2011
226 V.K. Kerji

the most appropriate patterns in the web application. Unfortunately, the ability of the
developers in applying the design pattern is undetermined. The good practice of using
design patterns is encouraged and efforts have to be taken to enhance the knowledge
on design patterns [1].
Creational design patterns abstract the instantiation process. They help make a
system independent of how its objects are created, composed, and represented. A
class creational pattern uses inheritance to vary the class that’s instantiated, where as
an object creational pattern will delegate instantiation to other objects. Creational
patterns become important as systems evolve to depend more on object composition
than class inheritance. As that happens, emphasis shifts away from hard coding a
fixed set of behaviors toward defining a smaller set of fundamental behaviors that
can be composed into any number of more complex ones. Thus creating objects with
particular behaviors requires more than simply instantiating class [4].
In case of web applications, it is sometimes necessary to add additional
responsibilities to the web page dynamically and to cater to the changing requirements.
Such additions should be transparent without affecting other objects. There can be
possibility where in these responsibilities can be withdrawn. In such scenario, Decorator
Pattern can be used to add the additional responsibilities to the default web page (whose
some of the contents are always constant). For Example if the application has different
users who can log into it (such as Admin, Program Manager and Stakeholder).Admin
will have different user interface when compared to other users but some of the user
interface elements may be similar across all users. In such scenario, we can use
Decorator Pattern to decorate the changing user interface elements depending on the
user type, keeping the remaining elements constant across all type of users [3].
In such applications, Decorator Pattern objects can be created using Abstract
Factory creational pattern. In this work, concrete decorator objects are created with
such Abstract Factory pattern without specifying their concrete classes. This
encapsulates the object instantiation process from the client thus making the system
more flexible to future changes in the requirements. Additional new users can be
added to the system without applying changes to the client side code. Client side code
is hidden from the type of current user instantiated by the Abstract Factory pattern
thus making the system more reusable and easier to maintain. Since the application
typically needs only one instance of a concrete factory in abstract factory pattern, we
used Singleton design pattern to implement the concrete factory object.
We applied the required changes to the DMS application [3] to achieve the above
mentioned objectives. The resulting application showed considerable improvement in
performance when multiple users are added. It abstracts the client code from changes
in concrete decorator object instantiation.
The rest of the paper is organized as follows. Section 2 describes Abstract Factory
and Singleton Patterns and their applicability. Section 3 briefs about DMS web based
application. Implementation and sample code is listed in section 4. Finally conclusion
is presented in section 5.

2 Abstract Factory and Singleton Patterns

Abstract factory provide an interface for creating families of related or dependent
objects without specifying their concrete classes. It helps to control the classes of
 Creational Patterns to Create Decorator Pattern Objects in Web Application 227

objects that an application creates. Because a factory encapsulates the responsibility

and the process of creating product objects, it isolates clients from implementation
classes. Clients manipulate instances though their abstract interfaces. Product class
names are isolated in the implementation of the concrete factory; they do not appear
in client code [4].
Abstract Factory declares an interface for operations that create abstract product
objects. Concrete Factory implements the operations to create concrete product
objects. Abstract Product declares an interface for a type of product object. Concrete
Product defines a product object to be created by the corresponding concrete factory.
It implements the Abstract Product interface. Client uses only interfaces declared by
Abstract factory and Abstract Product classes.
The factory completely abstracts the creation and initialization of the product from the
client. This indirection enables the client to focus on its discrete role in the application
without concerning itself with the details of how the product is created. Thus, as the
product implementation changes over time, the client remains unchanged. The most
important aspect of this pattern is the fact that the client is abstracted from both the type
of product and the type of factory used to create the product. Furthermore, the entire
factory along with the associated products it creates can be replaced in a wholesale
fashion. Modifications can occur without any changes to the client[5].

Fig. 1. Structure of Abstract Factory Pattern

228 V.K. Kerji

In our work, Abstract Product is that of Decorator object of the DMS application.
Concrete products are different types of users (Admin, Program Manager and Stake
Holder). Essentially, Decorator Pattern objects are made as that of Abstract Product
and its Concrete Product objects. Individual Concrete Factory objects are created for
each of the user types (Concrete Factory1 for user type Admin, Concrete Factory2 for
user type Program Manager and Concrete Factory3 for user type Stake Holder). In
this design, Client is unaware of which concrete product and concrete factory classes
are created since it access the required objects embedded in abstract class pointer.
Hence the client code is detached from underlying concrete object creation and
making the system more adaptable to future requirement changes and easier to
maintain it.
Singleton Pattern ensures a class has only one instance and provides global point of
access to it. Since there is only one instance of the Concrete Factory class, it is
implemented using Singleton Pattern so that client can have access to it from well
known access point.

3 DMS Application
Development Management System is a web based application which allows different
types of user (Admin, Stake Holder and Program Manager) to login and operate the
DMS application. User page has certain functionality similar across all type of users.
Changed functionality is implemented using corresponding Decorator objects, using
XML input data queried through XPath [3].
With reference to Decorator Pattern, DefaultUserView class is of the type
ConcreteComponent which is sub classed from Component base class. Decorator
class is sub classed from Component class which is a base class for Concrete
Decorators. Concrete decorators can be Admin, Stakeholder or Program Manager
decorator classes. Dynamic user page is decorated with any of the concrete decorators
depending upon the type of user logged into the system. Also, the contents of the
decorator are read from XML using X-path query [3].
In our work, changes are incorporated in Decorator Pattern object instantiation
process. It is achieved through Abstract Factory pattern which abstract clients from
how the objects are represented, composed and instantiated. Admin, Stake Holder and
Program Manager Objects of Decorator design pattern are now become part of the
Abstract Factory pattern. Since there can be only one instance of the Concrete Factory
object for a particular logged in user, Singleton design pattern is used so that client
can access this one and only instance for further usage.

4 Implementation and Sample Code

In this work, we applied changes to a web based application called Development
Management System. This application allows different types of user (Admin, Stake
Holder and Program Manager) to login and operate the DMS application. User page
has certain functionality similar across all type of users. Changed functionality is
implemented using corresponding Decorator objects, using XML input data queried
through XPath [3].
 Creational Patterns to Create Decorator Pattern Objects in Web Application 229

AbstractFactory class(Listing1) is a Singleton Design Pattern class which creates the

required Concrete Factory object depending the type of user logged into the system.
Instance is a static function which returns the single instance of the concrete factory
object. getInstance is a helper function which returns the appropriate concrete factory
object depending the user type. User type can be read from either configuration file or
registry. Abstract factory also has an important abstract function, CreateUser which
will be implemented by the concrete factory objects to create the respective user type
object.
_________________________________________________________
abstract class AbstractFactory
{
//one and only singleton instance
private static AbstractFactory instance = null;

//function to access the single instance object

public static AbstractFactory Instance()
{
if(instance == null)
{
instance = getInstance();
}
return instance;
}
//helper function to create proper factory object
Private static AbstractFactory getInstance()
{
string userType; //get the user type from
//.ini file or registry
//Create factory object depending on user
//type
Switch(userType)
{
Case “Admin”:
return new ConcreteFactory1();
Case “StakeHolder”:
return new ConcreteFactory2();
Case “ProgramManager”:
return new ConcreteFactory3();
}

}
//abstract class & function to create decorator
//object
Public abstract Decorator CreateUser();
}

Listing 1. Abstract Factory Singleton class

230 V.K. Kerji

ConcreteFactory1 class (Listing 2 ) is a sub class of AbstractFactory class which is

responsible for creating AdminDecorator object corresponding to Admin user type.
AdminDecorator takes DefaultUserView as an input argument to its constructor and
AdminDecorator object is returned to the client which has called CreateUser function.
_________________________________________________________
class ConcreteFactory1 : AbstractFactory
{
Public override Decorator CreateUser()
{
//Default user view as an input to decorator
DefaultUserView duv = new DefaultUserView();
//create&return Admin Decorator object
return new AdminDecorator(duv);
}
}
_____________________________________________________________________
Listing 2. ConcreteFactory1 class for Admin Decorator

ConcreteFactory2 class (Listing 3) is responsible for creating stake holder

decorator object. ConcreteFactory is a subclass of AbstractFactory class and
implements CreateUser function declared in AbstractFactory base class. Again the
StakeHolderDecorator object is passed to the client as Decorator object pointer,
without revealing the underlying object in it.
_________________________________________________________
class ConcreteFactory2 : AbstractFactory
{
Public override Decorator CreateUser()
{
DefaultUserView duv = new DefaultUserView();
//create&return Stake Holder Decorator object
return new StakeHolderDecorator(duv);
}
}
_________________________________________________________
Listing 3. ConcreteFactory2 class for Stake Holder Decorator

ConcreteFactory3 class (Listing 4) is responsible for creating ProgramManager

Decorator object. ConcreteFactory is a subclass of AbstractFactory class and
implements CreateUser function declared in AbstractFactory base class. Again the
ProgramManagerDecorator object is passed to the client as Decorator object pointer,
without revealing the underlying object in it.
AdminDecorator is a ConcreteDecorator (Listing 5) which adds the responsibility
of rendering the Admin related XHTML to the web browser. AdminDecorator will
copy the visual component object to its base class visual component member variable.
It implements the Render function which in turn reads the XML file to read its
required contents and develops the XHTML as per the XML data [3].
 Creational Patterns to Create Decorator Pattern Objects in Web Application 231

_________________________________________________________
class ConcreteFactory3 : AbstractFactory
{
Public override Decorator CreateUser()
{
DefaultUserView duv = new DefaultUserView();
//create&return Program Mgr Decorator object
return new ProgramManagerDecorator(duv);
}
}
_________________________________________________________
Listing 4. ConcreteFactory3 class for Program Manager Decorator

________________________________________________________
public class AdminDecorator : Decorator
{
//constructor for Admin Decorator
Public AdminDecorator(VisualComponent vc)
{
base.vc = vc;
}
…
}
_________________________________________________________

Listing 5. Admin Decorator class derived from Decorator

__________________________________________________________________

AbstractFactory af = AbstractFactory.Instance();
Decorator dc = af.CreateUser();

//Client unaware of the type of user

string totalXHTML = dc.Render();
//write the XHTML as string to the browser
Response.Write(totalXHTML);
__________________________________________________________________
Listing 6. Client code to render the XHTML

Client Code (Listing 6) shows how the final XHTML contents are retrieved from
Decorator object. Client code has been simplified and is now unaware of the user type
logged into the system. It gets single instance of concrete AbstractFactory object and
in turn created the proper Decorator Object which will render the XHTML code to the
browser.
Above code snippet reveals us that the client code is hidden from the instantiation
process of concrete decorator objects. AbstractFactory object creation is implemented
with the Singleton design pattern to ensure that only one instance of the concrete
232 V.K. Kerji

factory object is created in the application to provide the global point of access. Hence
the above method of implementation makes the maintenance of software easier and
any changes in the user requirements can be handled without changing the client part
of the code. It also improved the performance of the application when more number
of users is added to the system.

5 Conclusion
Software Design Patterns are proven design methods which when used carefully in
the implementation of software development will lead to several advantages.
Reusability, Encapsulation and Ease of system maintenance are some of the benefits
one can incorporate into software systems if these proven design patterns are used in
the system. Creational patterns belong to that category of design pattern which will
help in abstraction of object instantiation process. They help make a system
independent of how objects are created, composed, and represented. Abstract Factory
and Singleton patterns belong to creational pattern category, which are used in this
work to encapsulate the Decorator Pattern object instantiation process. Concrete
Decorator objects are created by respective Concrete Factory objects depending on
the type of user logged into the system. This resulted in improved application
performance when additional user types are added to the system and easier to
maintain the application code for future requirement changes.

References
1. Thung, P.L., Ng, C.J., Thung, S.J., Sulaiman, S.: Improving a Web Application Using
Design Patterns: A Case Study
2. Pawan, V.: Web Application Design Patterns. Morgan Kaufmann, CA (2009)
3. Kerji, V.K.: Decorator Pattern with XML in web based application. In: ICNCS 2011,
Kanyakumari, India (2011)
4. Gamma, E., Helm, R., Johnson, R., Vlissides, J.: Design Patterns Elements of Reusable
Object Oriented Software. Addison Wesley, Reading
5. Matic, D., Butorac, D., Kegalj, H.: Data Access Architecture in object oriented application
using Design Patterns
 Scalable Geographic Based Multicasting
(SGBM) Protocol

H. Venkateswaran and A. John Prakash

Department of Computer Science, Anna University,

Chennai-600 025
[email protected]

Abstract. The Scalable Geographic Based Multicast (SGBM) protocol is

designed for the Mobile Adhoc Networks (MANET), which overcomes the
difficulties in group membership management, multicast packet delivery and
maintenance of multicast structure over the dynamic network. Furthermore a
zone is divided based on the threshold value, so that the packet delivery is
maintained easily and the time spent on the data forwarding will be less. The
performance of the protocol was studied by extensive simulations. The result
depicts that the protocol we designed is more reliable, efficient and scalable to a
large group size or a network size.

Keywords: Multicast routing, geographic multicast, mobile ad hoc network,

scalable.

1 Introduction

Mobile ad hoc network (MANET) consists of many mobile hosts connected by

wireless links. Each node operates not only as an end-system, but also as a router to
forward packets. In MANET, the network topology may change frequently due to the
nodes’ movements [1]. A routing protocol is always needed to forward packets
quickly to the destination and adapt to the topology changes. Multicasting allows one
entity to communicate efficiently with multiple entities residing in a subset of the
nodes in a network. Although multicasting can be simulated using multiple point-to-
point messages, mechanisms that enable multidestination delivery using a single
group address can provide transparency, greater efficiency, and concurrency [2].
Many multicasting protocols have been designed, but due the dynamic network
topology changes the group membership management is not properly maintained, so
that there is a delay in the multicast packet delivery.
To have an efficient multicast packet delivery a tree structure is designed. However
it is difficult to maintain a tree structure in mobile adhoc network, and the tree
connection is easy to break. So then came the mesh based protocol which is an
alternative for tree based, where we have redundant path. The challenge in the mesh
based protocol is to support reliable and scalable multicast in a MANET with these
based schemes, as it is difficult to manage group membership, find and maintain
multiple paths in a network topology changes [3].

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 233–243, 2011.
© Springer-Verlag Berlin Heidelberg 2011
234 H. Venkateswaran and A. John Prakash

To overcome these problems, we have designed a protocol called SGBM (Scalable

Geographic Based Multicasting) protocol which can scale to a large network size.
Here we introduce a virtual tree structure which is different from other tree structure.
The virtual tree structure can be formed during the multicast packet forwarding with
the help of node positioning. It also introduces a zone based scheme to efficiently
handle the group membership management. The zone is also formed virtually and the
zone where the node is located is calculated based on the node position.
We discuss rest of the paper as follows: In section 2 we see the related work on
MANET multicast protocol. Section 3 follows as the motivation, where the
limitations in the related work are discussed. In section 4, the features of SGBM are
discussed. In section 5, the SGBM architecture and group membership management is
explained. In section 7, the preference analysis and finally in section 8, we conclude
the paper.

2 Related Works
In earlier days the Adhoc multicast protocol was a tree based protocol (e.g.
AMRoute). It creates a bidirectional shared multicast tree using unicast tunnels to
provide connections between multicast group members. Each group has at least one
logical core that is responsible for member and tree maintenance [4]. The mesh-based
protocols (e.g., FGMP [5], Core-Assisted Mesh protocol [6], ODMRP [7]) are
proposed to enhance the robustness by providing redundant paths between the source
and destination pairs at the cost of higher forwarding overhead. The conventional tree
based protocols consist of three components [8]:
1. Group membership management. The group membership changes frequently
as each node may join or leave a multicast group randomly, and the
management becomes harder for a large group.
2. Creation and maintenance of a tree- or mesh-based multicast structure. In
these protocols, the structures are based on some non-geographic
mechanisms, which make the tree-based structure not so robust, while the
mesh-based ones achieve the robustness at the cost of inefficiently utilizing
network resource. Also the nongeographic routing mechanisms prohibit
these protocols from scaling to a large network size.
3. Multicast packet forwarding. The multicast packets are forwarded along the
pre-built tree or mesh structure, but the pre-built paths are vulnerable to be
broken over the dynamic topology, especially in a large network with
potentially longer paths.
Beside these components there is a need for the geographic based multicast protocol
to scale for the large network size and to maintain the group membership
management. The geographic multicast protocols presented in [9], [1] and [10] need
to put the information of all the group members into the packet header, which creates
a lot of overhead for a large group, so they are only applicable for the small group
case [8].
 Scalable Geographic Based Multicasting (SGBM) Protocol 235

3 Motivation
In this we point out the limitations of the related works.
1. Difficult to maintain the tree structure.
2. Difficult to manage group membership.
3. Maintenance of multicast structure over the dynamic network topology for a
large group size or network size.
4. Problem in periodic flooding.
In particular, AMRoute performed well under no mobility, but it suffered from loops
and inefficient trees even for low mobility [4]. AMRIS was effective in a light traffic
environment with no mobility. But its performance was susceptible to traffic load and
mobility [11]. CAMP showed better performance when compared to tree protocols
but with mobility, excessive control overhead caused congestion and collisions that
resulted in performance degradation [6]. ODMRP was very effective and efficient in
most of our simulation scenarios. The protocol showed a trend of rapidly increasing
overhead as the number of senders increased [7]. These limitations motivated us to
develop the protocol called the SGBM (Scalable Geographic Based Multicast)
protocol.

4 Contribution

SGBM features are discussed in this section, The protocol can scale to a large
network size, where the packet transmission takes place in a dynamic mobile
environment. For the packet transmission, it uses several virtual architecture, where
there is no need to maintain the state information. This protocol is different from other
tree based protocol, where there is no need to maintain the tree structure explicitly.
Here it creates the virtual tree structure and transmits the data’s more efficiently. This
protocol also handles the group membership management by a zone based scheme.
The zone structure is also formed virtually. It also avoids the need of network periodic
flooding of source information.

5 SGBM Architecture
The above figure gives the architecture of the SGBM, which is also similar to the
RSGM. SGBM is made up of two-tier architecture. In the lower tier, the zone is built
based on position information and a leader is elected on demand in the zone of the
group members. The leader manages the group membership. At the upper tier, the
leaders of the member zone report the source directly via a virtual reverse tree based
path [3]. SGBM differs from RSGM only in the multicast packet delivery. It is
described in detailed in section 7.
236 H. Venkateswaran and A. John Prakash

Fig. 1. SGBM Architecture

6 Group Management
The group membership management is managed at two tiers. The zone is built on the
lower tier and a leader is elected for each zone. At the upper tier the leader, of the
zone report the source directly via a virtual reverse tree path. A node joining or
leaving the zone should be reported to the zone leader for the periodic membership
change. When a node joins a zone it should be updated to the corresponding zone
leader, if there is no leader in the zone, the node announces it as the leader and
distributes the data.

7 Multicast Packet Delivery

The source takes care of the multicast packet delivery to each zone. Using
optimization technique a threshold value is assigned to each zone, and number of
nodes in the zone is based on the threshold value. If the nodes exceed the threshold
value, then we split the zone into two halves, and checking for the condition that
exceeds the threshold value. If so, again split the zone. Here the time taken to forward
the data packets will be less. The zone leader distributes the data’s within the zone.
Since we split the zone, the packet delivery takes place quickly and more efficiently.
Then the next hop is based on the geographic broadcast forwarding through a virtual
tree based path where it reduces the time delay and forwarding bandwidth.

8 Performance Analysis
In this section, we study the performance of SGBM by simulations [3]; SGBM is
implemented in the Glomosim [15]. The performance of the protocol is impacted on
Packet Delivery Ratio, Control overhead, Average End-to-End delay, joining delay.
Packet Delivery Ratio (PDR) – The ratio of the number of packets received at the
destination and number of packets sent by the source.
 Scalable Geographic Based Multicasting (SGBM) Protocol 237

Control overhead – The ratio of packet delivery time and number of packets
delivered. Average
End-to-End Delay – The ratio of end-to-end delay and number of packets received,
measured in milliseconds. It includes the processing time and queuing time of packet
in each node in the route.
Joining delay – The average time interval between a member joining a group and its
first receiving of the data packet from that group. To obtain the joining delay, the
simulations were rerun with the same settings except that all the members joined
groups after the sources began sending data packets.
The performance of the protocol is impacted on zone size, multicast packet delivery
ratio, time delay, forwarding bandwidth. The following graphs give the end to end
path length delay, control overhead and packet delivery ratio in SGBM (i.e;RSGM
IM).

Table 1. The parameters used in network simulation are listed below

&KDQQHO7\SH &KDQQHO:LUHOHVV&KDQQHO

5DGLR3URSDJDWLRQ0RGHO 3URSRJDWLRQ7ZR5D\*URXQG

1HWZRUN,QWHUIDFH7\SH 3K\:LUHOHVV3K\

0$&3URWRFRO7\SH 0$&

,QWHUIDFH4XHXH7\SH 4XHXH'URS7DLO3UL4XHXH

/LQN/D\HU7\SH //

5DGLR/LQN&RQWURO 5/&

5RXWLQJ3URWRFRO 6*%0

1RRI1RGHV 

6LPXODWLRQ7LPH 6HFRQGV

7HUUDLQ5DQJH 0[0

3DFNHW6L]H E\WHV
238 H. Venkateswaran and A. John Prakash

Fig. 2. SGBM End-to-End Path Length Delay

Fig. 3. SGBM Control Overhead

 Scalable Geographic Based Multicasting (SGBM) Protocol 239

Fig. 4. SGBM Packet Delivery Ratio

Fig. 5. SGBM Group Size vs. End to End delay

240 H. Venkateswaran and A. John Prakash

Fig. 6. SGBM Group Size vs. Control Overhead

Fig. 7. SGBM Group Size vs. Packet Delivery Ratio

 Scalable Geographic Based Multicasting (SGBM) Protocol 241

Fig. 8. SGBM Node Density vs. Average End to End delay

Fig. 9. SGBM Node Density vs. Packet Delivery Ratio

242 H. Venkateswaran and A. John Prakash

9 Conclusion

In this paper, we designed the Scalable Geographic Based Multicasting Protocol

(SGBM) for Mobile Adhoc Network. In SGBM, the data’s are forwarded from source
to destination in a virtual tree based path without the need of explicitly creating and
maintaining a tree structure. Scalable membership management is achieved through a
zone structure.A home zone is defined to provide location and address service for the
sources to avoid the periodic network-range flooding of source information, and the
location service for group members is combined with the membership management to
avoid the overhead of searching for addresses and positions of all group members
through an outside location server. The position information is used in SGBM to
guide the zone structure building, membership management and packet forwarding,
which reduces the maintenance overhead and leads to more robust multicast
forwarding upon the topology changes. We also handle the empty zone problem
which is challenging for the zone-based protocols. Our studies show that our protocol
not only would outperform the existing geographic multicast protocol and
conventional multicast protocol but would also scale to a large group size and large
network size. Specifically, our protocol would be more robust to network
dynamics.Finally we conclude the paper as the protocol SGBM is more better than
ODMRP and RSGM protocol.Hence it is given as RSGMIM(RSGM improved)in the
graph.

References
1. Chen, K., Nahrstedt, K.: Effective Location-Guided Tree Construction Algorithms for
Small Group Multicast in MANET. In: Proc. IEEE INFOCOM, pp. 1180–1189 (2002)
2. Auerbach, J., Gopal, M., Kaplan, M., Kutten, S.: Multicast Group Membership
Management. IEEE/ACM Transactions on Networking 11(1), 166–175 (2003)
3. Xiang, X., Wang, X., Yang, Y.: Stateless Multicasting in Mobile Adhoc Networks. IEEE
Transactions on Computers 59(8), 1076–1090 (2010)
4. Lee, S., Su, W., Hsu, J., Gerla, M., Bagrodia, R.: A Performance Comparison Study of Ad
Hoc Wireless Multicast Protocols. In: Proc.IEEE INFOCOM (2000)
5. Chiang, C.-C., Gerla, M., Zhang, L.: Forwarding Group Multicast Protocol (FGMP) for
Multihop. Mobile Wireless Networks, Cluster Computing 1(2), 187–196 (1998); Special
Issue on Mobile Computing
6. Garcia-Luna-Aceves, J.J., Madruga, E.: The Core-Assisted Mesh Protocol. IEEE J. on
Selected Areas in Comm. 17(8), 1380–1394 (1999)
7. Gerla, M., Lee, S.J., Su, W.: On-Demand Multicast Routing Protocol (ODMRP) for Ad
Hoc Networks, Internet Draft, draftietf-manet-odmrp-02.txt (2000)
8. Xiang, X., Wang, X., Zhou, Z.: An Efficient Geographic Multicast Protocol for Mobile Ad
Hoc Networks, pp. 1–12 (1999)
9. Basagni, S., Chlamtac, I., Syrotiuk, V.R.: Location Aware, Dependable Multicast for
Mobile Ad Hoc Networks. Computer Networks 36(5/6), 659–670 (2001)
10. Mauve, M., Fubler, H., Widmer, J., Lang, T.: Position-Based Multicast Routing for Mobile
Ad-Hoc Networks. In: Proc. ACM MOBIHOC, Poster Section (June 2003)
 Scalable Geographic Based Multicasting (SGBM) Protocol 243

11. Wu, C., Tay, Y., Toh, C.K.: Ad Hoc Multicast Routing Protocol Utilizing Increasing Id-
Numbers (AMRIS) Functional Specification, Internet Draft (November 1998)
12. Paliwoda, K.: Transactions involving Multicast. Computer Communication Rev. 11(6),
313–318 (1988)
13. Deering, S.E.: Multicast Routing in internetworks and Extended LANs. Computer
Communication Rev. 18(4), 55–64 (1988)
14. Deering, S.E.: Host extensions for IP Multicasting, Networking Group, RFC 988 (1986)
15. UCLA Parallel Computing Laboratory, GloMoSim (2010),
http://pcl.cs.ucla.edu/projects/glomosim/
 Median Adjusted Constrained PDF Based Histogram
Equalization for Image Contrast Enhancement

P. Shanmugavadivu1, K. Balasubramanian2, and K. Somasundaram3

1
Associate Professor, Department of Computer Science and Applications, Gandhigram Rural
Institute – Deemed University, Gandhigram, India
[email protected]
2
Associate Professor, Department of Computer Applications, PSNA College of Engineering
and Technology, Dindigul, India
[email protected]
3
Professor, Department of Computer Science and Applications, Gandhigram Rural Institute
– Deemed University, Gandhigram, India
[email protected]

Abstract. A novel Median adjusted Constrained PDF based Histogram

Equalization (MCPHE) technique for contrast enhancement is proposed in this
paper. In this method, the probability density function of an image is modified
by introducing constraints prior to the process of histogram equalization (HE).
This technique of contrast enhancement takes control over the effect of HE so
that it enhances the image without causing any loss to its details. A median
adjustment factor is then added to the result, which normalizes the change in the
luminance level after enhancement. This factor suppresses the effect of
luminance change due to the presence of outlier pixels. The outlier pixels of
highly deviated intensities have greater impact in changing the contrast of an
image. Experimental results show that the proposed method gives better results
in terms of PSNR and SSIM values when compared to the existing histogram
based equalization methods.

Keywords: Contrast Enhancement, Histogram, Histogram Equalization (HE),

Probability Density Function (PDF), Cumulative Density Function (CDF).

1 Introduction
Contrast enhancement plays a vital role in image processing for both human and
computer vision. It is used as a preprocessing step in medical image processing,
texture synthesis, speech recognition and many other image/video processing
applications [3], [7], [9]. Different techniques have already been developed for this
purpose. Some of these methods make use of simple linear or nonlinear gray level
transformation functions [1] while the rest use complex analysis of different image
features such as edge, connected component information and so on.
Histogram is a statistical probability distribution of each gray level in a digital
image. Histogram Equalization (HE) is a very popular technique for contrast
enhancement of images[1]. It is the most commonly used method due to its simplicity

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 244–253, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Median Adjusted Constrained PDF Based Histogram Equalization 245

and comparatively gives better performance on almost all types of images. HE

transforms the gray levels of the image based on the probability distribution of the
input gray levels.
The histogram-based equalization techniques are classified into two principal
categories as global and local histogram equalization. Global Histogram Equalization
(GHE) uses the histogram information of the entire input image in its transformation
function. Though this global approach is suitable for overall enhancement, it fails to
preserve the local brightness features of the input image. Normally in an image, the
high frequency gray levels dominate the low frequency gray levels. In this situation,
GHE remaps the gray levels in such a way that the contrast stretching is restricted to
some dominating gray levels having larger image histogram components and causes
significant contrast loss for other smaller ones.
Local histogram equalization (LHE) [1] tries to eliminate such problems. It uses a
small window that slides through every pixel of the image sequentially. The block of
pixels that are masked by the window are considered for HE. Then, the gray level
mapping for enhancement is done only for the center pixel of that window. Thus, it
makes use of the local information remarkably. However, LHE requires high
computational cost and sometimes causes over-enhancement in some portions of the
image. Moreover, this technique has the problem of enhancing the noises in the input
image along with the image features. The high computational cost of LHE can be
minimized using non-overlapping block based HE. Nonetheless, these methods
produce an undesirable checkerboard effects on enhanced images.
Histogram Specification (HS) [1] is another method in which the expected output
of image histogram can be controlled by specifying the desired output histogram.
However, specifying the output histogram pattern is not a simple task as it varies from
image to image. A method called Dynamic Histogram Specification (DHS) [10]
generates the specified histogram dynamically from the input image. Though this
method preserves the original input image histogram characteristics, the degree of
enhancement is not significant.
Brightness preserving Bi-Histogram Equalization (BBHE) [2], Dualistic Sub-
Image Histogram Equalization (DSIHE) [4] and Minimum Mean Brightness Error Bi-
Histogram Equalization (MMBEBHE) [6] are the variants of HE based contrast
enhancement. BBHE divides the input image histogram into two parts based on the
mean brightness of the image and then each part is equalized independently. This
method tries to overcome the problem of brightness preservation. DSIHE method uses
entropy value for histogram separation. MMBEBHE is an extension of BBHE method
that provides maximal brightness preservation. Though these methods can perform
good contrast enhancement, they also cause more annoying side effects depending on
the variation of gray level distribution in the histogram. Recursive Mean-Separate
Histogram Equalization (RMSHE) [5], [8] and Recursive Sub-image Histogram
Equalization (RSIHE) [11] are the improved versions of BBHE. However, they are
also not free from side effects. Recursively Separated and Weighted Histogram
Equalization (RSWHE) is presented in [15]. To achieve sharpness in images, a
method called Sub-Regions Histogram Equalization (SRHE) is recently proposed [16]
in which the image is partitioned based on the smoothed intensity values, obtained by
convolving the input image with a Gaussian filter. In this paper, we propose a
Median adjusted Constrained PDF based Histogram Equalization (MCPHE) method
by extending Weighted Thresholded HE (WTHE) [12] method for contrast
enhancement.
246 P. Shanmugavadivu, K. Balasubramanian, and K. Somasundaram

Section 2 discusses the popular HE techniques. In section 3, the principle of the

proposed method is presented. Section 4 presents the information about two widely
used statistical parameters to assess the image quality. The results and discussions are
given in section 5. In section 6, the conclusion is given.

2 HE Techniques
In this section, the existing HE approaches such as GHE, LHE, various partitioning
approaches and WTHE are reviewed briefly.

2.1 Global Histogram Equalization (GHE)

For an input image F(x, y) composed of discrete gray levels in the dynamic range of
[0, L-1], the transformation function C(rk) is defined as:
k k
ni
S k = C ( rk ) = P ( ri ) =  (1)
i=0 i=0 n
where 0 ≤ Sk ≤ 1 and k = 0, 1, 2, …, L-1, ni represents the number of pixels having
gray level ri , n is the total number of pixels in the input image and P(ri) represents the
Probability Density Function of the input gray level ri. Based on the PDF, the
Cumulative Density Function is defined as C(rk). The mapping given in equation (1)
is called Global Histogram Equalization or Histogram Linearization. Here, Sk is
mapped to the dynamic range of [0, L-1] by multiplying it with (L-1).
Using the CDF values, histogram equalization maps an input level k into an output
level Hk using the level-mapping equation:

H k = ( L − 1) × C ( rk ) (2)

For the traditional GHE described above, the increment in the output level Hk is given
by:

Δ H k = H k − H k −1 = ( L − 1) × P ( rk ) (3)

The increment of level Hk is proportional to the probability of its corresponding level

k in the original image. For images with continuous intensity levels and PDFs, such a
mapping scheme would perfectly equalize the histogram in theory. But, the intensity
levels and PDF of a digital image are discrete in practice. In such a case, the
traditional HE mapping is not ideal and it results in undesirable effects where the
intensity levels with high probabilities often become over-enhanced and the levels
with low probabilities get less enhanced and their frequencies get either reduced or
even eliminated in the resultant image.

2.2 Local Histogram Equalization (LHE)

GHE takes the global information into account and can not adapt to local light
conditions. Local Histogram Equalization (LHE) performs block-overlapped
histogram equalization. LHE defines a sub-block and retrieves its histogram
 Median Adjusted Constrained PDF Based Histogram Equalization 247

information. Then, histogram equalization is applied at the center pixel using the CDF
of that sub-block. Next, the window is moved by one pixel and sub-block histogram
equalization is repeated until the end of the input image is reached. Though LHE
cannot adapt to partial light information, it still over-enhances certain portions
depending on its window size. However, selection of an optimal block size that
enhances all parts of an image is not an easy task to perform.

2.3 Histogram Partitioning Approaches

BBHE tries to preserve the average brightness of the image by separating the input
image histogram into two parts based on the input mean and then equalizing each of
the parts independently. RMSHE, RSIHE and RSWHE partition the histogram
recursively. Here, some portions of histogram among the partitions cannot be
expanded much, while the outside region expands significantly that creates unwanted
artifacts. This is a common drawback of most of the existing histogram partitioning
techniques since they keep the partitioning point fixed throughout the entire process
of equalization. In all the recursive partitioning techniques, it is not an easy job of
fixing the optimal recursion level. Moreover, as the recursion level increases,
recursive histogram equalization techniques produce the results same as GHE and it
leads to computational complexity.

2.4 Weighted Thresholded HE (WTHE)

WTHE is a fast and efficient method for image contrast enhancement [12]. This
technique provides a novel mechanism to control the enhancement process, while
being adaptive to various types of images. WTHE method provides a good trade-off
between the two features, adaptivity to different images and ease of control, which are
difficult to achieve in the GHE-based enhancement methods. In this method, the
probability density function of an image is modified by weighting and thresholding
prior to HE. A mean adjustment factor is then added with the expectation to normalize
the luminance changes. But, while taking the mean of the input and reconstructed
images, the highly deviated intensity valued pixels known as outlier pixels are also
taken into account. This will not effectively control the luminance change in the
output image, which is a major drawback of this method. The outlier pixels are the
pixels which are usually less in number and are having distant intensity values than
the other pixels.

3 Median adjusted Constrained PDF Based Histogram

Equalization (MCPHE)
The proposed method, MCPHE is an extension of WTHE which performs histogram
equalization based on a modified histogram. Each original probability density value
P(rk) is replaced by a Constrained PDF value Pc(rk) yielding:

Δ H k = ( L − 1) × Pc ( rk ) (4)
248 P. Shanmugavadivu, K. Balasubramanian, and K. Somasundaram

The proposed level mapping algorithm is given below.

i. Input the image, F(i, j) with a total number of ‘n’ pixels in the gray level range
[0, L-1]
ii. Compute the Probability Density Function (PDF), P(rk) for the gray levels
iii. Find the median value of the PDFs, M1
iv. Compute the upper constraint ‘Pu’ :
Pu = v * max(PDF) where 0.1 < v < 1.0
v. Set power factor ‘r’ where 0.1 < r < 1.0
vi. Set lower constraint factor Pl be as low as 0.0001
vii. Compute the constrained PDF, Pc (rk )

Pc (rk ) = Τ( P (rk ))
 Pu if P (rk ) > Pu 
 r 
 P (rk ) − Pl   (5)
=   * Pu if Pl < P (rk ) ≤ Pu 
 Pu − Pl  
0 if P (rk ) ≤ Pl 
 
viii. Find the median value of the constrained PDFs, M2
ix. Compute the median adjustment factor as:
Mf = M2 – M1
x. Add Mf to the constrained PDFs
xi. Compute cumulative probability density function, Cc(F(i,j))
xii. Apply the HE procedure (level mapping) as:

F ' ( i , j ) = ( L − 1) × C c ( F ( i , j )) (6)

where F′(i, j) is the enhanced image.

The transformation function T(.) in equation (5) transforms all the original PDF
values between the upper constraint Pu and the lower constraint Pl using a normalized
power law function with index r > 0.
In this level-mapping algorithm, the increment for each intensity level is decided
by the transformed histogram given in equation (5). The increment is controlled by
adjusting the index r of the power law transformation function. For example, when
r<1, the power law function gives a higher weightage to the low probabilities.
Therefore, the lower probability levels are preserved and the possibility of over-
enhancement is less. The effect of the proposed method approaches that of the GHE,
when r→1. When r > 1, more weight is shifted to the high-probability levels and
MCPHE would yield even stronger effect than the traditional HE.
The upper constraint Pu is used to avoid the dominance of the levels with high
probabilities when allocating the output dynamic range. The lower constraint Pl is used
to find the levels whose probabilities are too low. The Pl value is set to be as low as
0.0001. Any pixel having its probability less than the lower constraint Pl is having very
low impact in the process of contrast enhancement. It can be observed from equation (5)
that when r = 1, Pu=1 and Pl=0, the proposed MCPHE reduces to the GHE.
 Median Adjusted Constrained PDF Based Histogram Equalization 249

After the constrained PDF is obtained from equation (5), the median adjustment
factor (Mf) is calculated by finding the difference between the median value of the
constrained PDFs and the median value of the original PDFs. Then, the Mf is added to
the constrained PDFs which will effectively control the luminance change, since the
outlier pixels are ignored while computing the medians.
The cumulative density function (CDF) is obtained as:
k
C c ( k ) =  Pc ( m), for k = 0, 1, ..., L − 1 (7)
m =0

Then, the HE procedure is applied as given in equation (6). This will not cause serious
level saturation (clipping), to the resulting contrast enhanced image.

4 Measurement Tools to Assess Image Quality

There are two well known statistical parameters namely Peak Signal to Noise Ratio
(PSNR) and Structural Similarity Index Matrix (SSIM) which are usually employed to
assess the after-enhancement image quality [11].

4.1 Peak Signal-to-Noise Ratio

The peak signal-to-noise ratio (PSNR) is computed using the source image X(i, j) of
size M by N pixels and the enhanced image Y(i, j). Errors are computed only on the
luminance signal so that the pixel values of X(i, j) ranges between black (0) and white
(255). First the mean square error (MSE) of the reconstructed image is computed as
follows:
M N

 [X (i, j ) − Y (i, j )]
i =1 j =1
2
(8)
MSE =
M ×N
The root mean squared error (RMSE) is the square root of MSE. The PSNR in
decibels (dB) is computed as:

 max(Y (i, j )) 
PSNR = 20 log10   (9)
 RMSE 

4.2 Structural Similarity Index Matrix (SSIM)

The Structural Similarity Index Matrix (SSIM) is defined as:

(2 μ X μY + C1 )(2σ XY + C2 )
SSIM ( X , Y ) = (10)
( μ X2 + μY2 + C1 )(σ X2 + σ Y2 + C2 )
250 P. Shanmugavadivu, K. Balasubramanian, and K. Somasundaram

where X and Y are the reference and the output images respectively;
μX is the mean

of X,
μY is the mean of Y, σX is the standard deviation of X, σY is the standard
deviation of Y, σ XY is the square root of covariance of X and Y, whereas C1 and C2
are constants. The SSIM value between two images X and Y is generally in the range
0 to 1. If X=Y, then the SSIM is equal to 1 which implies that when the SSIM value is
nearing 1, the degree of structural similarity between the two images are more.

5 Results and Discussion

The performance of the newly developed method, MCPHE is tested on various
standard images such as Einstein, Village, Bottle, House and Peppers out of which
Einstein is given in Fig. 1(a). To compare the performance of MCPHE, the image is
enhanced with the contemporary enhancement techniques like GHE, LHE, BBHE,
DSIHE, HS, RMSHE and WTHE. For all these methods, the performance is
measured qualitatively in terms of human visual perception and quantitatively by
computing the Peak Signal-to-Noise Ratio (PSNR) values and Structural Similarity
Index Matrix (SSIM) values which are given in Table 1 and Table 2 respectively.
The original Einstein image is given in Fig. 1(a). The enhanced images of the same
by GHE, LHE, BBHE, DSIHE, HS, RMSHE, WTHE and are shown in Fig. 1(b) - 1(h)
respectively. Fig. 1(i) is the result of MCPHE which clearly shows the improvement in
image quality than the other HE techniques. Similarly, MCPHE is found to produce
better results for the other images such as village, Bottle, House, Peppers etc in terms
of visual perception, compared to other methods. The PSNR values obtained for
various images including the above said images are furnished in Table 1. It can be
noted that for all the images, the PSNR values of MCPHE is higher than all other
methods. The SSIM values are calculated for various methods including MCPHE and
are presented in Table 2. MCPHE is found to generate higher SSIM value than the
contemporary methods. Hence, the results of the study clearly reveal that MCPHE
produces better image enhancement compared to the contemporary methods in terms
of PSNR and SSIM.

Table 1. Comparison of PSNR values of various methods and MCPHE

Method
RMSHE
GHE LHE BBHE DSIHE HS WTHE MCPHE
(r=2)
Image

Einstein 14.9772 6.3961 16.1678 16.5590 16.1721 24.3070 22.2114 26.6039

Village 17.4181 5.7367 22.2929 21.4134 19.0391 26.8144 34.5230 39.5615
Bottle 12.6761 8.7862 18.4965 17.6775 12.9664 28.4096 30.6490 34.3248
House 10.7989 4.3828 14.4538 12.9997 11.0700 24.1902 24.8067 33.0999
Peppers 22.6708 4.9869 26.6785 26.6894 24.8050 27.1802 32.5050 37.0061
 Median Adjusted Constrained PDF Based Histogram Equalization 251

Table 2. Comparison of SSIM values of various methods and MCPHE

Method
RMSHE
GHE LHE BBHE DSIHE HS WTHE MCPHE
(r=2)
Image
Einstein 0.6672 0.0046 0.7099 0.6950 0.7430 0.8033 0.8770 0.9699
Village 0.7814 0.0068 0.8924 0.8843 0.8596 0.8868 0.9950 0.9982
Bottle 0.7517 0.0067 0.8847 0.8700 0.7685 0.9433 0.9796 0.9919
House 0.6582 0.0042 0.7475 0.7098 0.6727 0.8936 0.9475 0.9942
Peppers 0.9286 0.0052 0.9542 0.9553 0.9577 0.9365 0.9918 0.9974

(a) Original (b) GHE (c) LHE

(d) BBHE (e) DSIHE (f) HS

(g) RMSHE (r=2) (h) WTHE (i) MCPHE

Fig. 1. Einstein image (a) Original; results of (b) GHE (c) LHE with window size = 5 (d)
BBHE (e) DSIHE (f) HS (g) RMSHE (r=2) (h) WTHE (i) MCPHE
252 P. Shanmugavadivu, K. Balasubramanian, and K. Somasundaram

6 Conclusions
The proposed MCPHE contrast enhancement method is proved to be a better
approach for low contrast images. Experimental results on standard images show that
the degree of enhancement measured in terms of PSNR and SSIM values for the
proposed method is higher than those of the existing histogram based equalization
techniques. Moreover, this method preserves the details of the input images during
enhancement. This method can effectively be employed in video processing, medical
image processing etc.

References
1. Gonzalez, R.C., Woods, R.E.: Digital Image Processing, 2nd edn. Prentice Hall,
Englewood Cliffs (2002)
2. Kim, Y.: Contrast enhancement using brightness preserving bihistogram equalization.
IEEE Transactions on Consumer Electronics 43(1), 1–8 (1997)
3. Wahab, A., Chin, S.H., Tan, E.C.: Novel approach to automated fingerprint recognition.
IEEE Proceedings on Vision, Image and Signal Processing 145(3), 160–166 (1998)
4. Wan, Y., Chen, Q., Zhang, B.: Image enhancement based on equal area dualistic sub-
image histogram equalization method. IEEE Transactions on Consumer Electronics 45(1),
68–75 (1999)
5. Chen, S., Ramli, A.R.: Contrast Enhancement using Recursive Mean-Separate Histogram
Equalization for Scalable Brightness Preservation. IEEE Transactions on Consumer
Electronics 49(4), 1301–1309 (2003)
6. Chen, S., Ramli, A.R.: Minimum mean brightness error bi-histogram equalization in
contrast enhancement. IEEE Transactions on Consumer Electronics 49(4), 1310–1319
(2003)
7. Pei, S.C., Zeng, Y.C., Chang, C.H.: Virtual restoration of ancient Chinese paintings using
color contrast enhancement and lacuna texture synthesis. IEEE Transactions on Image
Processing 13(3), 416–429 (2004)
8. Chen, S., Ramli, A.R.: Preserving brightness in histogram equalization based contrast
enhancement techniques. Digital Signal Processing 14, 413–428 (2004)
9. Torre, A., Peinado, A.M., Segura, J.C., Perez-Cordoba, J.L., Benitez, M.C., Rubio, A.J.:
Histogram equalization of speech representation for robust speech recognition. IEEE
Transactions on Speech Audio Processing 13(3), 355–366 (2005)
10. Sun, C.C., Ruan, S.J., Shie, M.C., Pai, T.W.: Dynamic contrast enhancement based on
histogram specification. IEEE Transactions on Consumer Electronics 51(4), 1300–1305
(2005)
11. Sim, K.S., Tso, C.P., Tan, Y.Y.: Recursive sub-image histogram equalization applied to
gray-scale images. Pattern Recognition Letters 28, 1209–1221 (2007)
12. Wang, Q., Ward, R.K.: Fast image/video contrast enhancement based on weighted
thresholded histogram equalization. IEEE Transactions on Consumer Electronics 53(2),
757–764 (2007)
13. Menotti, D., Najman, L., Facon, J., Araujo, A.A.: Multi-histogram equalization methods
for contrast enhancement and brightness preserving. IEEE Transactions on Consumer
Electronics 53(3), 1186–1194 (2007)
 Median Adjusted Constrained PDF Based Histogram Equalization 253

14. Sengee, N., Choi, H.K.: Brightness Preserving Weight Clustering Histogram Equalization.
IEEE Transactions on Consumer Electronics 54(3), 1329–1337 (2008)
15. Kim, M., Chung, M.G.: Recursively Separated and Weighted Histogram Equalization for
Brightness Preservation and Contrast Enhancement. IEEE Transactions on Consumer
Electronics 54(3), 1389–1397 (2008)
16. Ibrahim, H., Kong, N.S.P.: Image Sharpening Using Sub-Regions Histogram Equalization.
IEEE Transactions on Consumer Electronics 55(2), 891–895 (2009)
 Isolated Word Recognition System Using Back
Propagation Network for Tamil Spoken Language

V. Radha, C. Vimala, and M. Krishnaveni

Department of Computer Science,

Avinashilingam Deemed University for Women,
Coimbatore, Tamilnadu, India
[email protected]
{vimalac.au,krishnaveni.rd}@gmail.com

Abstract. Recently with the wide development of computers, various forms of

information exchange between human and computer are discovered. At present,
interacting with the computer using speech is one of the active scientific
research fields especially for people with disabilities who face variety of
barriers to computer use. Such research in Automatic Speech Recognition
(ASR) is investigated for different languages because each language has its
specific features. This paper presents a speech recognition system for
individually spoken word in Tamil language using multilayer feed forward
network which falls under the category, "Networks for Classification and
Prediction" and has widespread interesting applications and functions related to
speech processing. To implement the above system, initially preprocessing is
done with the input signal and the speech features being the main part of speech
recognition system, are analyzed and extracted via Mel Frequency Cepstral
Coefficients (MFCC). These feature vectors are given as the input to the Feed-
Forward Neural Network for classifying and recognizing Tamil spoken word.
Experiments are done with sample Tamil speech signals and its performance are
measured based on Mean square error (MSE). The results indicate that the
adopted network with specified parameters have produced the best MSE.

Keywords: Elliptic Filter, Mel Frequency Cepstral Coefficients, Feed-forward

neural networks, Tamil Speech Recognition, Mean square error, and Peak
signal to Noise Ratio.

1 Introduction
Speech is one of the oldest and most natural means of information exchange between
human beings. As the new generation of computing technology, it comes as the next
major innovation in man-machine interaction. Obviously such interface would yield
great benefits which can be accomplished with the help of Automatic Speech
Recognition (ASR) system. It is a process by which a machine identifies speech. It
takes a human utterance as an input and returns a string of words or phrases as output.
Recently, Neural Network (NN) [1] is considered as one of the most successful
information processing tool that has been widely used in speech recognition [2][3].

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 254–264, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Isolated Word Recognition System Using Back Propagation Network 255

The conventional neural networks Multi-Layer Perceptron (MLP) have been

increasingly used for word recognition and also for other speech processing
applications. The main objective of this paper is to implement classification and
recognition system for isolated Tamil [6][7]spoken words. To carry out this task
preprocessing [4][5] is done before feature extraction[2][10] which includes filtering,
[4] and windowing. The best filtered speech signal is chosen and fed as an input.
These filtered outcomes are evaluated based on MSE and PSNR values. For
extracting specific features from the speech the popular feature extraction [2][10]
technique MFCC is used. Using this 12 feature vectors are extracted and they are
given as the input to the network. Finally, the speech recognition [2][3] is developed
using feed-forward neural networks [1]and their performances are evaluated based on
its default measure MSE.
The paper is organized as follows. Section 2 explains the system overview. Section
3 explains the preprocessing [5] steps involved in this system. Section 4 gives details
about feature extraction technique based on MFCC. Section 5 deals with Feed
forward neural network techniques and its performance in Tamil [6][7] speech signal.
Section 6 explores the performance evaluation of the adopted method. Finally, the
conclusion is summarized in section 7 with future work.

2 System Overview
The performance of the speech recognition [2][3] system depends on the classification
techniques such as Neural Networks, Hidden Markov Models, Bayesian networks and
Dynamic Time Warping etc. Among these approaches Neural networks (NNs) [1]

Raw
Speech

Speech Noise
reduction
Pre Processing

Speech Feature Cepstral

frames extraction coefficients

Pattern Training
Recognition Classification Matching using vectors
NN

Fig. 1. System overview of speech recognition based on NN

256 V. Radha, C. Vimala, and M. Krishnaveni

and pattern recognition. Rather than being used in general-purpose speech recognition
applications it can handle low quality, noisy data and speaker independence
applications. Such systems can achieve greater accuracy than HMM based systems, as
long as there is training data and the vocabulary is limited.
One of the most commonly used networks based on supervised learning algorithm
is multilayer feed forward network is implemented in this paper for classifying and
recognizing Tamil speech [6][7]. In Tamil language, the pronunciation of independent
letters and group of letters forming words are not different. Tamil speech recognizing
system [7][8] does not require the support of a dictionary. Thus the recognizing
process in Tamil speech [9] is fairly simple compared to English.
To implement the system, initially the speech data is preprocessed using filtering,
framing [4] and windowing techniques. Subsequent to that, the enhanced signal is
given as the input to the MFCC algorithm to extract features. These feature vectors
also called cepstral coeffients given as the input to the network. After that the network
is trained with these input vectors and the target vectors. Finally classification and
recognition is done based on pattern matching. The figure 1 demonstrates the overall
structure of the system.

3 Preprocessing
To enhance the accuracy and efficiency of the speech signals, they are generally pre-
processed before further analysis. Since the selected dataset usually needs to be pre-
processed prior to being fed into a NN this is done here. Speech signal pre-processing
includes digital filtering, [4] and windowing.

No of Original Signal Output signal

samples
1

Fig. 2. Original signals and the resulted signals after applying filters
 Isolated Word Recognition System Using Back Propagation Network 257

3.1 Filtering

One of the foremost processing steps in speech processing is to apply filters to the
signal. It is essentially used for speech enhancement by reducing the background
noise and to improve the quality aspects of speech. To reduce noise in the signal three
types of filters are used namely pre emphasis, median and elliptic. All these filters are
implemented and based on the experimental results, it was found that the elliptic filter
performs well among the other filters. Elliptic filter provides the lowest order
implementation of the classical filters for the same frequency and rejection
requirements. Elliptic filters are equiripple in the pass band and the stop band. The
finite zeros of transmission, which allow the filter to have a narrower transition band,
determine the ripple response in the stop band. The above taken filter results are
evaluated both subjectively and objectively. This resulted signal is used for further
processing of a system. The figure 2 shows the subjective evaluation of the above
three filters.

3.2 Framing and Windowing

The further preprocessing [4] step includes framing [4] and windowing. In this step
the continuous speech signal is blocked into frames of N samples, with adjacent
frames being separated by M (M < N). The first frame consists of the first N samples.
The second frame begins M samples after the first frame, and overlaps it by N - M
samples and so on. This process continues until all the speech is accounted for within
one or more frames. Typical values for N and M are 256.
The next step in the preprocessing [5] is to applying window for each individual
frame so as to minimize the signal discontinuities at the beginning and end of each
frame. It is used for minimizing the spectral distortion by using the window to taper
the signal to zero at the beginning and the end of each frame. It is defined
as w( n), 0 ≤ n ≤ N − 1 , where N is the number of samples in each frame and then
the result of windowing is the desired signal. Windowing is done by using the
equation(1).

yl (n) = xl (n) w(n), 0 ≤ n ≤ N − 1 (1)

Window choice is an important task in any signal processing applications. Among the
different types of windows like triangular, Blackman etc, the hamming window best
suites for speech signal processing. Typically the Hamming window is defined as in
the equation (2).

 2πn 
w( n) = 0.54 − 0.46 cos , 0 ≤ n ≤ N − 1 (2)
 N −1

4 Feature Extraction Using MFCC

The motivation of feature extraction [2] [10] is to convert speech waveform to a
parametric representation at a lower information rate for further analysis. Speech
258 V. Radha, C. Vimala, and M. Krishnaveni

signals have non-stationary characteristics. As a result, the speech waveforms are

commonly small frames (typically 5 ms to 40 ms) in which the signal characteristics
are considered quasi-stationary to allow for short-term spectral analysis and feature
extraction. A wide range of feature extraction techniques are available such as Linear
Predictive Coefficients (LPC), Linear Predictive Cepstral Coefficients (LPCC) and
Mel Frequency Cepstral Coefficients (MFCC). In this research work the most
frequently used MFCC parameters are considered to determine the best feature set for
the Tamil[6][7] speech database. The steps involved in MFCC can be seen in the
following figure 3 which illustrates the functioning process of deriving the Mel
Frequency Cepstral Coefficients.

Input speech
Pre Fast Fourier Absolute
Processed
Transform Value
input using
Elliptic filter

Discrete Mel-scaled
Cosine
Log
filter bank
Feature Transform
Vectors

Fig. 3. Steps of MFCC

Fig. 4. MFCC coefficients

Initially, the preprocessed speech sample is taken as an input and FFT is applied
which provides the absolute value. After that, the bank filter is used for warping the
Mel frequency. According to this, the width of the triangular filters varies and so the
log total energy in a critical band around the center frequency is included. After
warping the numbers of coefficients are obtained which is done by the equation(3).
 Isolated Word Recognition System Using Back Propagation Network 259

The extracted 12 feature vectors are as follows:

Table 1. Sample Feature vectors of MFCC

No of 12 coefficients values of MFCC

Samples

1 3.262109483760267 2.183415580698431 -3.211822134287218

7.959173232010254 -1.489845448833876 0.787602786178414

5.759973017258244 -2.252154484623268 3.439043146191077

1.25409424132334 -0.129311243184615 0.300936550868309

2 3.798478622210788 1.74797281436433 -2.489657282715821

8.522089004065837 -3.47052936681966 0.302290925428428

6.570803874923982 -2.833757855234683 4.19327339637328

1.515215876790111 -0.349448852656072 0.614526124621857

3 3.262109483760267 2.183415580698431 -3.211822134287218

7.959173232010254 -1.489845448833876 0.787602786178414

5.759973017258244 -2.252154484623268 3.439043146191077

1.25409424132334 -0.129311243184615 0.300936550868309

4 8.418557880841314 5.409326742576696 0.103873482360843

10.423558371669165 -1.562692756940896 1.840185046791586

3.560342622423488 -6.684312284440906 1.217165610703996

-0.449879780800645 -0.702780799553466 0.272368699987399

5 3.556871227862671 1.880082703099456 -2.677798789621644

9.584605229643586 -1.913588118828719 1.378513202846451

5.8273459608221 -6.567998823813931 2.083529875349709

0.019901182114678 -1.264406283818704 0.614171471695149

260 V. Radha, C. Vimala, and M. Krishnaveni

N /2
Y (k ) =  S ( j )H ( j )
k (3)
j =1

Finally the Inverse Discrete Fourier Transformer is used for the cepstral coefficients
calculation. It transforms the log of the quefrench domain coefficients to the
frequency domain where N is the length of the DFT which is defined in equation (4).

N ' −1 k 2π

 Y (k )e
1 j
N'
n
c(n)= (4)
N' k =0

For this work, the MFCC is implemented with 24 filters and 12 dimensions. It uses
256 size of NFFT. The following figure 4 and table 1 shows the 12 feature vectors
derived from MFCC for the sample Tamil [7][8]speech signals. These coeffients are
taken as the input vector for the network.

5 Feed Forward Neural Network for Tamil Spoken Word

Recognition
A feed forward neural network [10] is a biologically inspired classification algorithm. It
consists of a (possibly large) number of simple neuron-like processing units, organized
in layers. Every unit in a layer is connected with all the units in the previous layer.
These connections are not all equal and may have a different strength or weight. The
weights on these connections encode the knowledge of a network. They usually consist
of three to four layers in which the neurons are logically arranged. The first and last
layers are the input and output layers respectively and there are usually one or more
hidden layers in between the other layers. The term feed-forward means that the
information is only allowed to "travel" in one direction. This means that the output of
one layer becomes the input of the next layer, and so forward. Feed-forward networks
are advantageous as they have the fastest models to execute.
In the Network, the input layer does have a part in calculation, rather than just
receiving the inputs. The raw data is computed, and activation functions are applied in
all layers. This process occurs until the data reaches the output neurons, where it is
classified into a certain category. The operation of this network can be divided into
two phases:
1. The learning phase
2. The classification phase
During the learning phase the weights in the FFNet will be modified. All weights are
modified in such a way that when a pattern is presented, the output unit with the
correct category, hopefully, will have the largest output value. In the classification
phase the weights of the network are fixed. A pattern, presented as the input will be
transformed from layer to layer until it reaches the output layer. Now classification
can occur by selecting the category associated with the output unit that has the largest
output value. In contrast to the learning phase classification is very fast. The
experimental results are presented in the following section.
 Isolated Word Recognition System Using Back Propagation Network 261

6 Performance Evaluation
Performance evaluation is done for both filtering techniques and the NN separately.
The best filtering signal for feature extraction is evaluated based on Mean Square
Error Rate (MSE) and Peak Signal to Noise Ratio (PSNR) values. The MSE value is
calculated by the equation(5)

 ∧

  yi − y i 
MSE =  
(5)
n− p
and PSNR value is calculated by the equation (6)

 R2 
PSN R = 10 log 10   (6)
 M SE 
The average values of 10 speech samples were taken for the performance evaluation.
The figure 5 (a) and (b) presents the performance evaluation of three filters based on
MSE and PSNR values. Among these filters Elliptic filter offer least MSE value and
high peak signal value.

Comparision of Filters based on Mean Square

Error Rate
0.7
0.6
0.5
MSE Ratio

0.4
0.3
0.2
0.1
0
Preemphasis Elliptic Filter Median Filter
Filter
Types of Filters

Fig. 5(a). Performance evaluation of filters based on MSE

262 V. Radha, C. Vimala, and M. Krishnaveni

Comparision of Filters based on Peak Signal to

Noise Ratio
900

850
PSNR Ratio

800
750

700
650
Preemphais Filter Elliptic Filter Median Filter
Types of Filters

Fig. 5(b). Performance evaluation of filters based on PSNR

In this research work, the input speech signal is preprocessed with elliptic filter. This
filtered signal is divided into short frames and hamming window is applied to reduce
the discontinuities between these frames. For further process of creating and training a
network, the values given for the input vector are 12 coefficients of MFCC. There are
12 neurons in the first layer and 2 neurons in the second (output) layer. The transfer
function used in the first layer is tan-sigmoid, and the output layer transfer function is
linear. With standard back-propagation, the learning rate is held constant throughout
training. The performance of the algorithm is very sensitive to the proper setting of

Fig. 6(a). Actual and predicted value

 Isolated Word Recognition System Using Back Propagation Network 263

Fig. 6(b). Performance of the Network

the learning rate. If the learning rate is set too high, the algorithm may oscillate and
become unstable. If the learning rate is too small, the algorithm will take too long to
converge. However this sensitivity can be improved if it allows the learning rate to
change during the training process. For this research work the learning rate of 0.03 is
used and epoch count was limited to 300. The following figure 6 (a) and (b) shows the
experimental results of actual and predicted value of a sample Tamil speech[7][8]
signal. The performance goal is achieved with the parameters specified for the network.
The mean square error rate is the default performance evaluation measure for the
adopted network. It is computed by "summing the squared differences between the
predicted values vs. actual value, and then divide the sum by the number of
components. A threshold of 0.01 mean square error is good. Our research work
achieved minimum MSE for the Tamil speech [9] signal i.e 0.00571769/0.01.

7 Conclusion
The latest developments in the research of neural networks are providing new and
improved methods of tackling complex problems and tedious tasks such as speech
recognition. For human beings, speech is a natural and simple task. However, it is an
extremely complex and difficult job to make a computer respond to spoken
commands. In this paper such an important task is carried out for recognizing Tamil
spoken words. The important feature vectors are extracted from the enhanced speech
signal using MFCC and they are given as the input to the feed forward neural
network. The adopted network is trained with these input and target vectors. In the
successful implementation the results were found to be satisfactory considering less
number of training data. In future more number of isolated and continuous words to
be trained and tested with this network which may provide functional alternatives to
the standard input operations using keyboard and mouse. This preliminary experiment
264 V. Radha, C. Vimala, and M. Krishnaveni

will helps to develop Automatic Speech Recognition system for Tamil language using
different approaches like Hidden Markov Models or with other hybrid techniques.

References
[1] Ahmad, A.M., Ismail, S., Samaon, D.F.: Recurrent Neural Network with Backpropagation
through Time for Speech Recognition. In: Intemational Symposium on Communications
and Information Technologies (ISClT 2004), Sapporo, Japan, October 26-29 (2004)
[2] Saheli, A.A., Abdali, G.A., Suratgar, A.A.: Speech Recognition from PSD using Neural
Network. In: Proceedings of the International MultiConference of Engineers and
Computer Scientists, IMECS 2009, Hong Kong, March 18-20, vol. I (2009)
[3] Hasnain, S.K., Beg, A.: A Speech Recognition System for Urdu Language. In:
International Multi-Topic Conference (IMTIC 2008), Jamshoro, Pakistan, pp. 74–78
(2008)
[4] Hasnain, S.K.: Recognizing Spoken Urdu Numbers Using Fourier Descriptor and Neural
Networks with Matlab. In: Second International Conference on Electrical Engineering,
March 25-26, University of Engineering and Technology, Lahore (2008)
[5] Melin, P., Urias, J., Solano, D., Soto, M., Lopez, M., Castillo, O.: Voice Recognition with
Neural Networks, Type-2 Fuzzy Logic and Genetic Algorithms. Engineering Letters,
13(2), EL_13_2_9 (2006); Advance online publication: 4 August
[6] Murugavel, M.: Speech Recognition Model for Tamil Stops. In: Proceedings of the World
Congress on Engineering, WCE 2007, London, U.K, July 2-4, vol. I (2007)
[7] Saraswathi, S., Geetha, T.V.: Design of language models at various phases of Tamil
speech recognition system. International Journal of Engineering, Science and
Technology 2(5), 244–257 (2010)
[8] Saraswathi, S., Geetha, T.V.: Implementation of Tamil Speech Recognition System Using
Neural Networks. In: Manandhar, S., Austin, J., Desai, U., Oyanagi, Y., Talukder, A.K.
(eds.) AACC 2004. LNCS, vol. 3285, pp. 169–176. Springer, Heidelberg (2004),
doi:10.1007/978-3-540-30176-9_22(2004).
[9] Thangarajan, R.: Word and Triphone Based Approaches in Continuous Speech
Recognition for Tamil Language. WSEAS Transactions on Signal Processing 4(3) (March
2008)
[10] Uma Maheswari, N., Kabilan, A.P., Venkatesh, R.: A Hybrid model of Neural Network
Approach for Speaker independent Word Recognition. International Journal of Computer
Theory and Engineering 2(6), 1793–8201 (2010)
 Enhanced Hybrid Compression
Models for Compound Images

D. Maheswari and V. Radha

Department of Computer Science,

Avinashilingam Deemed University for Women,
Coimbatore, India
[email protected]
[email protected]

Abstract. This paper presents an efficient compound image compression

method based on block and layer–based segmentation techniques. Two hybrid
models are proposed for segmenting compound images. The first model
combines layer-based and block-based techniques for segmentation and
compression of compound images. Several experiments were conducted to
evaluate both models in terms of compression ratio, PSNR and time.

Keywords: Compound Image, Layer–based, Block-based, Segmentation,

Compression.

1 Introduction

Compound images are different kind of images that contain both palletized regions
[1], which has text or graphics, and continuous tone regions. The example of typical
compound images are hybrid documents, contents captured from screen, slides, and
web-pages. In this fast transmission world, compound images need to be compressed
and transmitted in a secure way. Compressing compound images with a single
algorithm that simultaneously meets the requirements for text, image and graphics has
been elusive and thus requires new algorithms that can competently reduce the file
size without degrading the quality [2]. The compound image compression
performance basically depends on the segmentation result. A segmentation process is
used where regions of similar data types are grouped together. After successful
segmentation existing techniques that best suits each data type can be used to achieve
best compression results[3].Most of the segmentation algorithms proposed in the
literature belong to three categories, namely object-based, block-based and layer-
based segmentation [4]. Many researchers follow either layer or block-based methods
[5][6][7] some commercial compressors like DjVu[8.] ,Digipaper [9] , JPEG 2000[10]
also exists. This paper proposes yet another hybrid method combining both block and
layer-based segmentation techniques for compressing compound image compression.
The main goal is to develop a hybrid model to handle the overlapping regions of
compound images. For this purpose two hybrid models were proposed. Model 1

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 265–274, 2011.
© Springer-Verlag Berlin Heidelberg 2011
266 D. Maheswari and V.Radha

combines block-based segmentation with wavelet packets for analyzing compression

of overlapping regions. Model 2 combines both the layer -based and block -based
techniques for efficient handling of overlapping images.
The paper is organized as follows: Section 1 gave a brief introduction to the topic
under discussion. Section 2 discusses the compound image segmentation. Section 3
explains proposed hybrid models. Section 4 presents the results and evaluates the
proposed system. Section 5 concludes the paper.

2 Compound Image Segmentation Techniques

The compound image segmentation has three categories. Normally researchers follow
two types of compound image segmentation techniques, They are

1. Layer- based segmentation and

2. Block-based segmentation.

This paper combines both of these techniques. In layer-based segmentation, a page is

divided into rectangular layers. In block-based segmentation page is divided into
rectangular blocks where each block follows exact object boundaries.

2.1 Layer-Based Segmentation

The steps followed during layer–based segmentation and compression, are,

• Segmentation of image into layers

• Production of a binary mask image
• Data filling of sparse foreground image
• Data filling of sparse background image
• Compression of each layer with an appropriate encoder
• Package of the whole representation into one meaningful stream for
transmission or storage.
Most layered coding algorithms use the standard three layers Mixed Raster Content
(MRC) representation.

2.2 Block-Based Segmentation

In this technique the compound image is divided in blocks of a certain size (e.g. 8x8)
and a classification method is applied to the block to segment the image into
blocks, which can later be compressed using different techniques. The classification
information is also coded before coding the block pixels and usually it depends only on
the values of the pixels inside a block and possibly the decision on the neighboring
blocks.
 Enhanced Hybrid Compression Models for Compound Images 267

3 Hybrid Models
The block and layer-based techniques separately gives better performance.
Combining the two segmentation methods gives much better results than the
individual. For this purpose hybrid model 1 and model 2 are proposed.

3.1 Model 1

Model 1 uses block classification method to identify various blocks in the compound
image. As the first step, the image pixels are grouped into three classes, namely Low-
Gradient, Mid-Gradient and High-Gradient pixels. In this model the compound image
is first divided into 16x16 blocks. Using the gradient values and a combination of
decision rules, the block is classified into five different regions, namely, background
block, text block, picture block, graphics block and overlapping blocks. Out of all the
five blocks, the main challenge was faced with the overlapping block, which had
mixed contents. For this purpose a wavelet packet-based compression technique is
used. The overall process is given in Figure 1.

Compound
images

Block
Segmentation (16 x 16 blocks) Type Map

Back Pure Overlap Graphics Pure

ground Text Picture

Fig. 1. The Proposed Model 1 Segmentation Technique

During segmentation, two features, namely, histogram and gradient of the block is
used. Using the gradient values, the histogram distribution for each pixel group is
computed. For this purpose the following decision rules are used, that are shown in
Figure 2.
268 D. Maheswari and V.Radha

Rule 1 : If HG + LG < T1 → Picture block

Rule 2 : If HG < T2 and LG > T3 and Gray level Range > T4 → Background block

Else → Picture block

Rule 3 : If Major colors Number ≤ 2 and HGP>T6 → Text block

Else if colors ≥ 2 → Graphics block

Else → Picture bock

Rule 4 : If HG ≥ T7 → Overlap block

Else → Picture block

Fig. 2. Decision Rules of Model 1

After segmenting a compound image using wavelet packet tree the image is
divided into background, pure text, overlap, graphics and pure picture regions. The
background block is compressed using arithmetic coder, the pure text block, graphics,
pure picture blocks are compressed using XML coder [3], Palette based coder and
modified JPEG coder respectively. Then the overlap block is compressed using the
proposed wavelet packet tree coder. The overlapping block, which cannot be
compactly represented both in spatial and frequency domain contain strong high
frequency signals due to text edges. In order to fully utilise these during compression
a wavelet packet tree based compression is used.
The tree construction procedure is as follows: Information content of decomposed
component (approximation and details) may be greater than the information content
of components, which has been decomposed. This method reduces the time
complexity of wavelet packets decomposition and selects the sub-bands that have
vital information. The decision of further decomposition and cost calculation is based

1. Let level counter = 1

2. Let the current node = input image.
3. Decompose current node using wavelet packet tree.
4. Find the entropy of current node.
5. Find the entropy of decomposed components.
6. Compare the entropy of parent node with the sum of the entropy of child node. If
the sum of the entropy of child nodes is less than that of parent node, then child
node will be considered as a leaf node of a tree and repeat the steps 3-6 for each
child nodes considering it as current node. Otherwise parent node acts as a leaf
node of a tree.
7. Stop.

Fig. 3. Algorithm for Overlapping Block

 Enhanced Hybrid Compression Models for Compound Images 269

on a strategy, which decides during run time whether to retain or prune the
decomposed components. After the best basis has been selected based on cost
function, the image is represented by a set of wavelet packets coefficients, which is
further compressed using run length encoder to give better overall compression. The
algorithm is given in Figure 3.

3.2 Model 2

In model 2, the two techniques layer-based and block-based are combined to form a
hybrid compound image compression technique. The method proposed uses model 1
for segmentation, but differs in the handling of the overlapping block. The
overlapping block is dealt with separately and uses the layer-based approach to
segment the block into three layers text, picture and mask. Then the various
compression schemes suiting the need of block or layer are applied. The text layer is
compressed using the XML coder , The mask layer is compressed using JBIG coder
and the picture layer is compressed using the modified JPEG coder. This model
combines the advantages of both layer-based and block-based approaches and hence
can be considered as an improvement of both the approaches.
The segmentation technique for overlapped block in model 2 is given in Figure 4.

Overlapped Block

Layer-Based Segmentation

Text Mask Picture

Fig. 4. Overlapped block Model 2

4 Experimental Results
The main objective of the tests was to find which of the proposed algorithms is best
suited for compound color image compression.

4.1 Test Images

In order to verify the performance of the proposed schemes, experimental evaluation

is performed using a series of different benchmark images. The experimentation is
done in three types of compound images, which are scanned document images,
computer generated images and photographic images with compound nature. The
following Figure 5 is the test images.
270 D. Maheswari and V.Radha

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Fig. 5. Test Images

For experimentation, 9 images named (a to i) were selected and the selection was
done in such a way that they were combination of all classes of compound images.

4.2 Performance Metrics

The performance of the proposed models was evaluated using different parameters
like compression ratio, compression and decompression time. One of the objectives of
the proposed compression models is to maintain the visual quality of the
decompressed image. The quality of the decompressed image was ascertained by
using the quality metric Peak Signal to Noise Ratio (PSNR).
Compression Ratio. The compression ratio of the test images were presented in
Table 1. The proposed models are compared with traditional compressors JPEG and
DjVu.
 Enhanced Hybrid Compression Models for Compound Images 271

Table 1. Compression Ratio(%)

Images Model 1 Model 2 JPEG DjVu

a 45.02 47.91 37.43 39.89
b 39.56 43.02 35.00 36.51
c 39.12 42.51 34.78 36.24
d 42.77 45.91 36.39 37.67
e 39.12 40.32 36.47 37.12
f 39.25 41.67 35.92 36.23
g 42.26 44.32 35.64 37.62
h 43.38 45.66 36.60 38.84
i 40.67 43.95 36.67 38.74

From the above Table 1, it can be seen that both model 1 and model 2 are enhanced
versions of compression schemes. However model 2 produces better compression
ratio for all the test images than model 1.
Compression Time and Decompression Time. The following Table 2 shows the
result of compression and decompression time for all the models. The compression
and decompression time are the time taken for encoding and decoding an image.

Table 2. Compression and Decompression Time (Seconds)

Images Model 1 Model 2 JPEG DjVu

CT DT CT DT CT DT CT DT

a 0.77 0.65 0.75 0.62 1.02 0.79 0.89 0.73

b 0.72 0.66 0.72 0.64 1.02 0.82 0.89 0.75

c 0.71 0.65 0.71 0.63 0.98 0.83 0.84 0.77

d 0.73 0.67 0.72 0.66 0.87 0.83 0.83 0.72

e 0.82 0.74 0.81 0.72 0.89 0.83 0.85 0.75

f 0.84 0.75 0.83 0.73 0.96 0.89 0.93 0.75

g 0.72 0.65 0.71 0.62 0.92 0.84 0.82 0.76

h 0.73 0.64 0.71 0. 63 0. 93 0.84 0.84 0.76

i 0.73 0.63 0.72 0.62 0.94 0.75 0.89 0.71

272 D. Maheswari and V.Radha

From above Table 2, again model 1 and model 2 achieves compression in a

minimum period of time compared with other traditional compressors.

Peak Signal to Noise Ratio. The PSNR is used to judge the quality of decompressed
image with that of the original image. The Figure 6 compares the PSNR values of
proposed algorithms.

db

Fig. 6. PSNR Value of Four Models

The high PSNR value is obtained by model 2 when compared with other models.
To show the improvement of the proposed algorithms, a Visual comparison of the
image (h), before and after compression is shown in Figure 7.

Original Image Model 1

Fig. 7. Visual Comparison of Test Image

 Enhanced Hybrid Compression Models for Compound Images 273

Model 2 JPEG

DjVu

Fig. 7. (Continued)

From the above results, Model 1 and Model 2 achieves better results compared
with the traditional models.

5 Conclusion
This paper proposed two hybrid compound image compression models. The first
model divides the image into five blocks and applies block-based segmentation. The
second model uses block-based segmentation which is used in the first model and for
overlapping block it again applies layer-based segmentation. The proposed models are
compared with traditional JPEG and DjVu. During experimentation, it was found that
the proposed Model 2 produces good results compared with other models.

References
1. Liu, D., Ding, W., He, Y., Wu, F.: Quality-biased Rate Allocation for Compound Image
Coding with Block Classification, pp. 4947–49500 (2006), 0- 7803-9390-2/06/$20.00
©2006 IEEE
2. Maheswari, D., Radha, V.: Enhanced Layer Based Compound Image Compression. In:
Proceedings of the First Amrita ACM-W Celebration of Women in Computing, pp. 209–
216 (2010) ISBN:978-1-4503-0194-7
274 D. Maheswari and V.Radha

3. Maheswari, D., Radha, V.: Secure Layer Based Compound Image Compression using
XML Compression. In: 2010 IEEE International Conference on Computational
Intelligence and Computing Research, pp. 494–498 (2010) ISBN:978-1-4244-5966-7
4. Ding, W., Lu, Y., Wu, F.: Enable efficient compound image compression in H.264/AVC
intra coding. In: IEEE International Conference on Image Processing, vol. 2, pp. II-337–II-
340 (2007)
5. Radha, V., Aparna, R., Maheswari, D.: Performance evaluation of H.264/AVC intra
compound image compression system. International Journal of Computer Applications,
Foundation of Computer Science 1(10), 48–54 (2010)
6. Xi, Q., Wu, X., Zhang, S.: Compound Image Compression with Multi-step Text Extraction
Method. In: Fifth International Conference on Intelligent Information Hiding and
Multimedia Signal Processing, pp. 1270–1273 (2009)
7. Zaghetto, A., de Queiroz, R.L.: Iterative pre-and postprocessing for MRC layers of
scanned documents. In: 15th IEEE International Conference on Image Processing, ICIP
2008, pp. 1009–1012 (2008)
8. Bottou, L., Haffner, P., Howard, P., Simard, P., Bengio, Y., LeCun, Y.: High quality
document image compression using DjVu. Journal of Electronic Imaging 7(3), 410–425
(1998)
9. Huttenlocher, D., Felzenszwalb, P., Rucklidge, W.: DigiPaper: A versatile color document
image representation. In: Proc. ICIP, vol. I, pp. 219–223 (1999)
10. Sharpe, L.H., Buckley, R.: JPEG 2000 .jpm file format: a layered imaging architecture for
document imaging and basic animation on the web. In: Proceedings of SPIE, vol. 4115, pp.
464–475 (2000)
 Index Based Ordered Broadcast with Status (IOBS)
Algorithm for Consistent Data Broadcast

K. Chitra1, K. Senthamarai Kannan2, and S. Abirami3

1
Asst.Prof.,Computer Science,
Govt. Arts & Science college for women,
Pudukottai, Tamilnadu
2
Professor & Head of Statistics,
Manonmaniam Sundaranar University,
Tirunelveli-12, Tamilnadu
3
Lecturer in BCA , Dept. of Mathematics,
Thiagarajar college (Autonomous),
Madurai-9, Tamilnadu

Abstract. Mobile computing system is more popular as people need data on

move. Many emerging mobile database applications demand broadcast based
data dissemination, ie, the data broadcasted to more number of clients without
any request. In this paper we address the problem of preserving data
consistency and currency among mobile transactions, while the broadcast and
update transactions executed concurrently on server without any control. We
propose an efficient serializability based algorithm (IOBS) for ensuring the
consistency of entire database. This algorithm has the following properties.
• The order of update transactions seen by the different mobile clients
will be same.
• Consistency and currency are ensured without contacting the server.
• The size of control information for conflict checking is smaller.

Keywords: Data Dissemination, Consistency , Serializability, Currency.

1 Introduction
Data broadcasting method is very effective technique for disseminating information to
massive number of clients when the data base size is small. There are two types[13]
[16] of information delivery for mobile data applications .First is push based data
delivery(Data broadcast method) broadcast the data items in continuous cycles to the
mobile clients, who were all tuned in to the channel. Second is on demand data
delivery (pull based delivery) broadcast the data to specific clients who issue request
to the server. For example, in E-commerce application and road traffic information
systems the database server broadcast data continuously to mobile clients. Mobile
clients can get the broadcasted data without the communicating the server.
In data broadcast[3] [15]applications, the asymmetric data flow from servers to
clients is more than data flow from clients to servers. Data items stored on database
server are highly sensitive. The database server in mobile computing system has to

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 275–286, 2011.
© Springer-Verlag Berlin Heidelberg 2011
276 K. Chitra, K. Senthamarai Kannan, and S. Abirami

broadcast the fresh data items while the update processes executed concurrently with
broadcast processes. If there is no control over broadcast transaction and update
transaction then the mobile client may read outdated data items (inconsistent
data).Concurrency control [14] is one of the important components of transaction
management. Traditional concurrency control algorithm (2PL) is not suitable for this
kind of system due to the overhead in setting locks. Since with the limited bandwidth
available for clients, broadcasting consistent data is challenging research issue. In this
paper we guarantee the broadcast of consistent data items (correct data items) and
current data item (fresh data items) to the mobile clients. We propose a new
algorithm(IOBS)for consistency and currency of data items using serialization
criterion. Serializability is a commonly accepted criteria for correctness of set of
transactions which are executed concurrently[1].The transactions has serializable
property ,if the effect of concurrent transactions execution should be as though all the
transactions executed in some serial order. This algorithm ensures data consistency
for read only mobile transactions[2].
The following summarizes the contributions of the paper.
1. This algorithm is suitable for large size of updates.
2. The size of control information is very smaller. Since it is suitable for low
bandwidth network.
3. This algorithm broadcast the same ordered set of updated data items to all
mobile clients.
4. IOBS algorithm broadcast, rebroadcast updated data items with New status for
consistency and currency of data items.
5. IOBS ensures consistency over multiple broadcast cycles.

The organization of remaining part of the paper is as follows, Section 2 reviews

related works on consistent delivery, Section 3 describes our system model, Section 4
discuss the data inconsistency problem with examples. Section 5 states our new
algorithm. finally conclusion and future work of the paper is in Section 6.

2 Related Work
Mobile computing system becomes a very active research area. Concurrency control
is the most important component of broadcast based data delivery. However
traditional concurrency control protocols (2PL) are not suitable for mobile computing
system. Due to their over head in detecting conflicts and overhead for locking
conflicts [4].
In [5], two approaches were proposed. One is based on broadcasting multiple
versions of data items and another one is based on conflict serialization graph. Multi
version broadcast approach broadcast the list of different versions of updated data. In
invalidation based consistency scheme, server broadcast the control information,
invalidation report and serialization graph for each cycle. Upon the receipt of graph,
client update its local copy of graph using invalidation report and control information.
Each client maintains its local copy of the graph to ensure the serializability. But this
approach suffers from the overhead in broadcasting serilization graph and also each
client has to listen the channel to maintain its updates.
 Index Based Ordered Broadcast with Status (IOBS) Algorithm 277

In [6] F-Matrix and R-Matrix algorithms were proposed for data concurrency. This
algorithm broadcast the matrix (m x n) for each cycle. ie. C(i,j) represents the latest
cycle number in which some transactions affects the value of Obj and writes to Obi.
Then Mobile client performs consistency checking using matrix before reading the
broadcast data item. Even though this algorithm broadcast fresh data items but it
needs high bandwidth to broadcast the matrix control information and the server has
to do expensive computation for control matrix.
In [7], [9], [10], they suggest serializability based mechanisms to maintain the
freshness of broadcast data. These approaches abort the update transactions if any one
of the following condition meet. If a transaction updates the data items of Not
Read Set and Already Read Set.If a transaction updates data items of Not Read Set
and read the data items of NUS(set of data .items written by the update transaction
after the broadcast transaction). It shows simple but it suffers from overhead of
aborting update transactions. Since, these protocol are not suitable for broadcasting
sensitive data items.
In multi version data broadcast approach [11],the server broadcast multiple version
of each data item. To reduce the latency of client transactions, caching mechanism
,broadcast disk organization approach were suggested. But these approaches require
high bandwidth for broadcasting multiple versions.
In [12], they suggested hybrid rebroadcast/ invalidation report mechanism. The
server process is same as that of UFO algorithm[4]. The client maintains the schedule
between update transaction and mobile transaction to be serializable by combining the
rebroadcast and IR method. This algorithm broadcast the ordered read set to mobile
clients, But the client hast to contact with the server even for read only transactions.
it‘s performance suffer from restart process of mobile transactions.
To our knowledge, UFO algorithm in [4] has been shown to be more efficient
algorithm.In this algorithm serializability is used as consistency criterion. During the
execution phase the set of data items of broadcast and update transactions are
compared. If there is any overlap (ie,updated data item is in already broadcast list)
this algorithm rebroadcast them. Otherwise the serialization order is U→BT. This
approach broadcast the fresh data item without any control information and it has
minimal impact on server. But the different mobile transactions are allowed to
perceive different serialization orders of update transactions[8] (ie, the order of read
operation of mobile clients is not taken in to account) It is not suitable if the data
conflict size is larger and there is a overhead due to the division of execution of
update transaction in to two phases.

3 Transaction Organization
A mobile computing system provides data, voice and video services between clients
from anywhere, any time and in any form. Mobile computing system consists of
database servers and a number of clients. In such environment, the server has high
broadcast bandwidth while the band width of client is low. The database server
maintains the dynamic information of external environment. Such information are
highly sensitive. Such as stock quotes information system, location of moving object,
Ecommerce applications. To maintain the freshness, newly updated data items of
database server have to be committed and broadcasted instantly while the status of
278 K. Chitra, K. Senthamarai Kannan, and S. Abirami

external environment has changed. Otherwise the values of updated data items may
become useless. The mobile clients may get outdated data values. For example, in
stock information system if a client tune in to the mobile network to read the current
value of a stock. The updated data item(fresh data item) is very important to his
business decision.
Our system model based on a broadcast environment in which database server
periodically broadcast the data items to clients. The period for which broadcasting
data items is called a broadcast cycle. The length of the broadcast cycle may be fixed
or variable. [4]. No cache is in mobile clients. Mobile clients can fetch the data items
from the broadcast channel without contacting data base server. In further, we
assumed that 2PL (Two phase lock in method) [1] is used for concurrency control
among update transaction at data base server.

3.1 Mobile Client’s Read Only Transactions and Server’s Update Transactions

Mobile clients do not need to continuously listen broadcast. They can tune in to read
specific item. To do so, clients must have the prior information about the structure of
broadcast. The mobile clients issue the request (Read Only), called mobile
transactions(MBT). We have modeled the broadcast process of database server as
broadcast transaction (BRT) and update process generated by external environment as
update transaction (UPT). They can execute concurrently on data base server.

3.2 Assumptions

• The data base server maintains each data item in a row with unique row
index.
• All the update transactions (UPT) are executed only on server.
• All the mobile transactions are read only
• All the mobile clients can get the same ordered set of updated data.
• The number of update transactions on server is lower than the number of
broadcast transactions.
• For each update transaction , Broadcast transaction broadcast latest version of
data item, updated status information.

4 Inconsistency Problem in Data Dissemination

If data base server allows the interleaved execution of update and broadcast
transaction without any control, the mobile transactions may read inconsistent data
item. In this paper, consistency of data item is ensured using serializability
criterion[1]. In the following three examples [4], we discuss the data inconsistency
problem of stock information system . we consider D1 as stock value of Reliance, D3
as stock value of Ranbaxy, D5 as stock value of Tata.
Example 1: The data conflicts between update transaction and mobile transaction.
Suppose the update transaction (UPT) update the stock values of D1 and D3, a mobile
transaction (MBT) needs to read the stock value of D1 and D3. Then the schedule is,
 Index Based Ordered Broadcast with Status (IOBS) Algorithm 279

i).BRT broadcast D3.

ii).MBT reads D3.
iii).UPT updates D1.
iv).UPT updates D3.
v).BRT broadcast D1.
vi).MBT reads D1.
The mobile transactions (MBT) may read inconsistent data items . The serialization
graph is cyclic, such as MBT→UPT→MBT it is non serializable[4]. The reason is
that BRT broadcast D1 before to the execution of update transaction. MBT couldn’t
read updated value of D1.
Example 2: The data conflicts with more update transactions (UPT).
Suppose two update transaction (UPT1, UPT2), such as UPT 1 updates the value
of D3, D1 and UPT 2 updates the values of D1, D5. Then the schedule is,
i).BRT broadcast D1.
ii).MBT reads D1.
iii).UPT1 updates D3.
iv).UPT1 updates D1.
v).UPT2 updates D1.
vi).UPT2 updates D5.
vii).BRT broadcast D5.
viii).MBT reads D5.
The serialization graph is cyclic, such as UPT2→MBT→UPT1→UPT2.
Example 3: Data conflicts with multiple broadcast cycles.
Suppose a mobile transaction may not able to find all his data items in a single
broadcast cycle. In consistency may occur over more than one broadcast transactions.
For example, if the schedule is,
i).BRT1 broadcast D1.
ii).MBT reads D1.
iii).End of BRT1
iv).UPT1 Updates D1
v).UPT1 Updates D3
iv).BRT2 broadcast D3.
vii).MBT reads D3.
The serialization graph is cyclic, such as MBT→UPT1→ MBT.
The data inconsistency problems in the three examples can be resolved using our
IOBS algorithm.

5 Index Based Ordered Broadcast with Status Algorithm (IOBS)

When the broadcast transaction is being processed, there can be two types of update
transactions. One is update transaction executed before the broadcast transaction,
another is update transaction executed after broadcast transaction. Since instead of
280 K. Chitra, K. Senthamarai Kannan, and S. Abirami

aborting update transaction, the IOBS algorithm detects the data conflicts among
update and broadcast transaction to maintain the data consistency and currency.
The basic process of IOBS algorithm is to preserve the data consistency for the
update transactions serialized before broadcast transaction and the update transaction
serialized after broadcast transaction. The implementation of IOBS algorithm will be
discussed in detail in section 5.4.

5.1 Details of IOBS Algorithm

The IOBS algorithm will be executed for each broadcast cycle indefinitely at data
base server. This algorithm has three phases,
• Update phase
• Commit phase
• Broadcast phase.

1. Update phase

During update phase, it is assumed that a well formed concurrency control protocol
2PL [1] is adopted for preserving data consistency among update transactions. In this
phase the broadcast transaction is blocked. Since we have assumed the number update
transactions is smaller than broadcast transaction. This phase maintains ordered
updated data items and the ordered (based on the order of updations) row indices of
updated data items.

2. Commit Phase
After the completion of update phase, commit phase will be executed. During this
phase each row index from ordered list of row indices of updated data items (UPI) is
compared with the index of data items to be broadcasted (BRI). If there is a match,
the server broadcast the corresponding data item from UPD and with the status value
“New”. Then the newly broadcasted data is removed from BRD to avoid duplicated
data broadcast. Since the serialization order is UPT→ BRT. Otherwise serialization
order is BRT→ UPT. (ie. The index of updated data item may be in the index list of
already broadcasted list) This algorithm rebroadcast newly updated data item from
UPD and with the status value “New”. At the end of this phase UPI, UPD are
reinitialized for the next update transaction.

3. Broadcast Phase
During this phase IOBS algorithm broadcast the data items alone from BRD .Since
this algorithm resolved the data inconsistency problem by rebroadcasting newly
updated values with the status of “New”. Then all the mobile transactions can read the
fresh data items in the order of update transactions executed at server. The IOBS
algorithm detect the data conflict across the broadcast transactions also.

5.2 The IOBS Algorithm

In this section, we will state the IOBS algorithm. First we defined the following
quantities.
 Index Based Ordered Broadcast with Status (IOBS) Algorithm 281

UPD=The set of updated data items of an update transaction on current broadcast

cycle.
UPI= Ordered indices set of updated data items of an update transaction on
current broadcast cycle.
BRD=The set of data items to be read by the broadcast transactions.
BRI= The set of indices of data items to be read by the broadcast transaction .
The data base server execute the following algorithm for each cycle indefinitely.
Ser()
{
While (true){
If (UPT is in execution)
{
Block BRT.
UPI+= Ordered Index of updated data item.
UPD+= Ordered Updated data item.
}
else if (UPT is committed)
{
for each ( X in UPI)
{
If (X є BRI)
{//serialization order is UPT →BRT.
Remove data item from BRD where row index=X
Broadcast data item T, status =New from UPD where
row index = X.
}
else
{//serialization order is BRT →UPT.
Rebroadcast data item T, status = New from UPD where
row index = X.
}
}
UPI={ }, UPD={ }
}
else
{//no dependency between UPT and BRT.
Broadcast each data item T from BRD.}
}// End while}

5.3 Examples
Referring to the examples in section 4, we will discuss the schedule under IOBS
algorithm.
Example 1
i).BRT broadcast D3.
ii).MBT read D3.
282 K. Chitra, K. Senthamarai Kannan, and S. Abirami

iii).UPT1 updates D1.

iv).UPT1 updates D3.
v).UPI={index of D1, D3}.
UPD={updated values of D1, D3.}
BRD={D1 }
BRI={index of D1}
Compare the two sets, UPI and BRI.
vi).BRT broadcast D1 with status=New
and remove it from BRD.
vii).MBT reads D1 with the updated New.
viii).BRT rebroadcast D3 with status =New.
ix).MBT reads D3 with status =New.

Example 2
i).BRT broadcast the stock value of D1.
ii).MBT reads D1 .
iii).UPT1 updates D3.
iv).UPT1 updates D1.
v).UPI={index of D3, D1}
UPD={new value of D3, D1}.
BRD={D5}.
BRI={ index of D5}.
Compare the two set of UPI and BRI.
vi).BRT rebroadcast D3 with status=New
vii).MBT reads D3 with status=New
viii).BRT rebroadcast D1 with status=New
ix).MBT reads D1 with status=New
x).UPT2 updates D1.
xi).UPT2 updates D5.
xii)UPI={index of D1, D5}
UPD={new value of D1, D5}
BRD={D5 }
BRI={ index of D5}
Compare UPI and BRI.
xiii).BRT rebroadcast D1 with status=New
xiv).MBT reads D1 with status=New
xv).BRT broadcast D5 with status= New.
and remove it from BRD.
xvi).MBT reads D5 with status=New
Example 3
i).BRT1 broadcast D1
ii).MBT reads the D1
iii).End of BRT1
iv).UPT1 updates D1
v).UPT1 updates D3
vi).UPI={index of D1, D3}
UPD={new values of D1, D3}
 Index Based Ordered Broadcast with Status (IOBS) Algorithm 283

BRD={ D3}
BRI={ index of D3}
Compare UPI and BRI.
vii).BRT2 Rebroadcast D1 with status= New
ix).MBT reads D1 with status= New
xi).BRT2 broadcast D3 with status=New and remove it from BRD
x).BRT2 reads D3.

5.4 Implementation of IOBS Algorithm

5.4.1 Properties
The important properties of IOBS algorithm are
i).The size of the control information used to retain the data consistency is
minimum.
ii).It ensures the data consistency and ordered read set for mobile clients.
iii).The impact of this algorithm on server is minimum.
The implementation of IOBS algorithm is simple and does not require any algorithm
for mobile clients. Mobile transaction can access the broadcast data without
contacting server. All the conflict checks will be executed on server. The overheads
of the algorithm are:
1.comparison between index set of updated data items and index set of
broadcasted data items to maintain the serializability.
2. The broadcast transaction is blocked while the update transaction is in execution.
3.rebroadcast of conflicting data items.
4.broadcast of updated data items with status information.
The waiting time of broadcast transaction is short since number of update transactions
are smaller than broadcast transactions. The IOBS algorithm broadcast data items
along with status only for updated data items .Hence overhead for broadcasting
status should be minimum.
The advantage of rebroadcast is that every time most updated data items are
broadcasted. Index set comparison speed up the conflict checking process and the
usage of index set simplifies the large size of updates.

5.4.2 Correctness
We have seen how IOBS algorithm has low over head and minimal impact on server.
Four tasks of IOBS algorithm summarized in the following fig.1.
Task 1
It corresponds to the first if part in 5.2. In this task broadcast transaction is stopped,
ordered index set of updated data items(UPI) and UPD framed.
Task 2
It corresponds to the second if part in 5.2. In this task, the two sets UPI and BRI
were compared. For each index of UPI, if the index is in BRI then this phase
284 K. Chitra, K. Senthamarai Kannan, and S. Abirami

broadcast the fresh value of updated data item with New status and it removes the
corresponding data item, index from BRD and BRI to avoid the duplicated data
broadcast. ie, this task will be executed if the update transactions executed before
broadcast transactions.

Yes
UPT transaction in
Task1
Execution

No

No
UPT committed Task4

Yes

UPT serialized Yes

before broadcast Task2

No Task3

Fig. 1. Three phases of IOBS Algorithm

Task 3
It corresponds to the third if part in 5.2.In this task , the index of UPI may be in index
set of already broadcasted list then the algorithm rebroadcast data item with New
status. ie, this task would be executed if the update transactions executed after
broadcast transactions.
Task 4
It corresponds to the last if part in 5.2.If there is no update transactions ,this task
broadcast the data items without any status.

6 Conclusion and Future Work

The IOBS algorithm is an efficient algorithm to control the concurrent execution of
update and broadcast transactions. This algorithm broadcast s the same ordered set of
updated data items to all mobile clients. The ordered broadcast is very useful for
dynamic databases. Mobile clients can take instant decisions by using the dynamic
data bases . IOBS algorithm is suitable for large size of updations like video, Image
,etc., since row index is used for data conflict checks.
In addition, IOBS algorithm broadcasts the fresh data along with the New status
information. Broadcasting status information is use full for the mobile clients to get
 Index Based Ordered Broadcast with Status (IOBS) Algorithm 285

current and consistent data items even on network disconnections. The followings are
threads for future work,
1)The mobile client can issue only read-only query to database server.
2) Detection of data conflict is complicated.
3)Overhead due to rebroadcast of updated data items while the update transactions
serialized after broadcast transactions.

References
1. Date, C.J.: An Introduction to Database systems, 7th edn. Pearson Education, London
(2000)
2. Varadarajan, R., Chithra, M.: Scheduling data reads and writes using feedback on a weakly
connected environment. International Journal of Computer Applications in
Technology 34(3) (2009)
3. Varadarajan, R., Chithra, M.: Reducing latency through efficient channel allocation
methods for multiversion data dissemination in mobile computing environment.
International Journal of Internet Technology and Secured Transactions 1(3/4) (2009)
4. Lam, K.Y., Chan, E., Au, M.-W.: Broadcast of Consistent Data to Read-Only Transactions
from Mobile Clients. In: Proceedings of 2nd IEEE Workshop on Mobile Computing
Systems and Applications, New Orleans, Louisiana, USA (February 1999) (3)
5. Pitoura, E.: Supporting Read-Only Transactions in Wireless Broadcastin. In: Quirchmayr,
G., Bench-Capon, T.J.M., Schweighofer, E. (eds.) DEXA 1998. LNCS, vol. 1460. Springer,
Heidelberg (1998)
6. Shanmugasundaram, J., Nithrakashyap, A., Sivasankaran, R., Ramamritham, K.: Efficient
Concurrency Control for Broadcast Environments. In: Proceedings of ACM SIGMOD
International Conference on Management of Data, Philadelphia, June 1-3 (1999) (8)
7. Lee, V.C.S., Lam, K.-W., Wu, S., Chan, E.: Broadcasting Consistent Data in Mobile
Computing Environments. In: Proceedings of the seventh Real Time Technology and
Applications Symposium, RTAS 2001. IEEE Computer Society, Washington, DC (2001)
8. Young, C.-R., Chiu, G.-M.: An Efficient Protocol for Disseminating Consistent Data in
Broadcast Environments. In: International Workshop on Ubiquitous Data Management
(2005)
9. Poon, B.H.C., Lam, K.-W., Lee, V.C.S.: Broadcasting Consistent Data in Mobile
Environments. In: 2004 IEEE International Conference on Mobile Data Management
(MDM 2004), p. 169 (2004)
10. Hin, B., Poon, C., Lam, K.-W., Lee, V.C.S.: Server-side Broadcast Transaction in Mobile
Computing Environment. In: International Conference on Information Technology: Coding
and Computing (ITCC 2004), vol. 1, p. 511 (2004)
11. Pitoura, E., Chrysanthis, P.K.: Multiversion Data Broadcast. IEEE Transactions on
Computers 51(10), 1224–1230 (2002)
12. Lam, K.-Y., Chan, E., Leung, H.-W., Au, M.-W.: Concurrency Control Strategies for
Ordered Data Broadcast in Mobile Computing Systems. Inf. Syst. 29 (2004)
13. Krishna Mohan Rao, S., Venugopal Reddy, A.: Data Dissemination in Mobile Computing
Environment. BVICAM’S International Journal of Information Technology 1(1) (January-
July 2009)
286 K. Chitra, K. Senthamarai Kannan, and S. Abirami

14. Moiz, S.A., Rajamani, L., Pal, S.N.: Commit Protocols in Mobile Environments:Design&
Implementation. International Journal of Database Management Systems (IJDMS) 2(3)
(August 2010)
15. Acharya, S., Alonso, R., Franklin, M., Zdonik, S.: Broadcast Disks: Data Management for
Asymmetric Communication Environments. The Kluwer International Series in
Engineering and Computer Science 353, 331–361 (1996)
16. Acharya, S., Franklin, M., Zdonik, S.: Balancing Push and Pull for Data Broadcast. In:
Proceedings of the 1997 ACM SIGMOD International Conference on Management of Data,
vol. 26(2) (1997)
 Capturing High-Level Semantics of Images in Web
Documents Using Strength Matrix

P. Shanmugavadivu1, P. Sumathy1, and A. Vadivel2

1
Department of Computer Science and Applications
Gandhigram Rural Institute, Dindigul, Tamil Nadu, India
[email protected], [email protected]
2
National Institute of Technology, Trichy, Tamil Nadu, India
[email protected]

Abstract. The multimedia information retrieval from distributed environment

specifically from World Wide Web is considered as one of the challenging
issues to be addressed by the researchers. It has been perceived that the
description of images using low-level features apparently increases the semantic
gap. The information present in a HTML document in the form of textual
keywords can be extracted for capturing semantic information with the view to
narrow the semantic gap. The high-level textual information of images can be
extracted and associated with the textual keywords that narrow down the search
space and improve the precision of retrieval. In this paper, a strength matrix, is
being proposed, which is based on the frequency of occurrence of keywords and
the textual information pertaining to image URLs. The strength of these textual
keywords are estimated and used for associating these keywords with the
images present in the documents. The high-level semantics of the image is
described in the HTML documents in the form of image name, ALT tag,
optional description, etc., is used for estimating the strength. The effectiveness
of information retrieval of the proposed technique is found to be comparatively
better than many of the recently proposed retrieval techniques. The
experimental results of the proposed method endorse the fact that image
retrieval using image information and textual keywords is better than those of
the text based and the content-based approaches.

Keywords: Multimedia Information Retrieval, Web Image Retrieval, High-

level Features, Textual Keywords.

1 Introduction
The revolutionized advent of internet and the ever-growing demand for information
sprawled in the World Wide Web has escalated the need for cost-effective and high-
speed information retrieval tools. Many attempts have been made to use image
contents as a basis for indexing and images retrieval. In early 1990, researchers have
built many image retrieval systems such as QBIC [1], Photobook [2], Virage [3],
Netra [4] and SIMPLIcity [5] etc., which are considered to be different from the
conventional image retrieval systems. These systems use image features such as color,

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 287–296, 2011.
© Springer-Verlag Berlin Heidelberg 2011
288 P. Shanmugavadivu, P. Sumathy, and A. Vadivel

texture, shape of objects and so on whereas the recently devised image retrieval
systems use text as well as image features. In the text based approach, the images are
manually annotated by text descriptors and indexed suitably to perform image
retrieval. However, these types of systems have two major drawbacks in annotating
the keywords. The first drawback is that a considerable level of human intervention is
required for manual annotation. The second one is the inaccuracy annotation due to
the subjectivity of human perception.
To overcome the aforesaid drawbacks in text-based retrieval system, content based
image retrieval (CBIR) has been introduced [6]. However, these systems are suitable
only for domain-specific applications. The low-level features such as color, texture
and shape are efficiently used for performing relevant image retrieval [7] - [8]. Color
histograms such as Human Color Perception Histogram [9] - [10] as well as color-
texture features like Integrated Color and Intensity Co-occurrence Matrix (ICICM)
[11] - [12] show high precision of retrieval in such applications. However, the
inevitable semantic gap that exists between low-level image features and the user
semantics drift the performance of CBIR still far from user’s expectations [13]. In
addition, the representation and storage of voluminous low-level features of images
may be a costly affair in terms of processing time and storage space. This situation
can be effectively handled by using keywords along with the most relevant textual
information of images to narrow down the search space.
For this purpose, initially it is essential to explore the techniques to extract
appropriate textual information from associated HTML pages of images. Many
research methods have been proposed on the content structure of HTML document,
including image title, ALT tag, link structure, anchor text and some form of
surrounding text [20]. The main problem of these approaches is that the precision of
retrieval is lower. This disadvantage has triggered the task of developing adaptive
content representation schemes, which can be applied to a wide range of image
classification and retrieval tasks. Further, the design techniques are also required to
combine the evidences extracted from text and visual contents with appropriate
mechanisms to handle large number of images on the web. Many recent techniques
classify the images into one or more number of categories by employing learning
based approach to associate the visual contents extracted from images with the
semantic concept provided by the associated text. The principal challenge is to devise
an effective technique to retrieve web-based images that combine semantic
information from visual content and their associated HTML text.
This paper proposes a faster image retrieval mechanism, which is tested on a large
number of HTML documents. For this purpose, the HTML documents are fetched,
using a web crawler. The content of the HTML documents is segregated into text and
images and HTML tags. From the text, keywords are extracted and these keywords
are considered to be the relevant keywords to represent the high level semantics of the
images contained in the same HTML document.
In the following sections of the paper, the related works are presented. In section 3,
the proposed method is elaborated along with a suitable example. The experimental
result is presented in section 4. The conclusion is given in the last section of the
paper.
Capturing High-Level Semantics of Images in Web Documents Using Strength Matrix 289

2 Related Works
The retrieval of images from the Web has received a lot of attention recently. Most of
the early systems have employed text based approach, which exploits how images are
structured in web documents. Sanderson and Dunlop [16] were among the first to
model image contents using a combination of texts from associated HTML pages. The
content is modeled as a bag of words without any structure and this approach is found
to be ineffective for indexing. Shen et al [14] have built a chain related terms and
used more information from the Web documents. The proposed scheme unifies the
keywords with the low-level visual features. The assumption made in this method is
that some of the images in the database have been already annotated in terms of short
phrases or keywords. These annotations are assigned either using surrounding texts of
the images in HTML pages or by speech recognition or manual annotations. During
retrieval, user’s feedback is obtained for semantically grouping keywords with
images. Color moments, color histograms, Tamura’s features, Co-occurrence matrix
features and wavelet moments are extracted for representing low-level features.
Keywords in the document title of HTML page and image features have been
combined for improving the retrieved documents of news category [17]. In this
technique, from a collection of 20 documents chosen from one of the news site has
been used and 43 keywords along with HSV based color histogram are constructed.
While constructing histogram, saturation and hue axes are quantized into 10 levels to
obtain H×S histogram with 100 bins. However, this technique is found to perform
well for a small number of web pages and images only. In general, image search
results returned by an image search engine contain multiple topics and organizing the
results into different semantic clusters may help users. Another method has been
proposed for analyzing the retrieval results from a web search engine [20]. This
method has used Vision-based Page Segmentation (VIPS) to extract semantic
structure of a web page based on visual presentation [18]. The semantic structure is
represented as a tree with nodes, where every node represents the degree of coherence
to estimate the visual perception.
Recently, a bootstrapping approach has been proposed by Huamin Feng et al
(2004), to automatically annotate a large number of images from the Web [20]. It is
demonstrated that the co-training approach, combines the information extracted from
image contents and associated HTML text. Microsoft Research Asia [21] is
developing a promising system for Web Image Retrieval. The purpose is to cluster the
search results of conventional Web, so that users can find the desired images quickly.
Initially, an intelligent vision based segmentation algorithm is designed to segment a
web page into blocks. From the block containing image, the textual and link
information images are extracted. Later, image graph is constructed by using block-
level link analysis techniques. For each image, three types of representations are
obtained such as visual feature based representation, textual feature based
representation and graph based representation. For each category, several images are
selected as non representative images, so that the user can quickly understand the
main topics of the search results. However, due to index based retrieval, the time for
processing is found to be on the higher side. Rough set based model has proposed for
decompositions in information retrieval [22]. The model consists of three different
knowledge granules in incomplete information system. However, while WWW
290 P. Shanmugavadivu, P. Sumathy, and A. Vadivel

documents are presented along with images as input, the semantic of images are
exactly captured and thus retrieval performance is found to be lower. Hence, it is
important to narrow down the semantic gap between the images and keywords present
in WWW.
The textual information of WWW documents, which is the high-level semantics,
can be effectively used for defining the semantics of the images without really using
the low-level features. This kind of approach simplifies the semantic representation
for fast retrieval of relevant images from huge voluminous data. This paper proposes a
scheme for extracting semantic information of images present in WWW documents
using only the textual information. The relationship between the text and images
present in WWW documents estimated with frequency of occurrence of keywords and
other important textual information present in image link URLs. Based on the
experimental results, it is observed that the performance of the system is better than
that of Google (www.google.com).

3 Strength Matrix between Keyword and Image High-Level

Features for Capturing Image Semantics

Let H be the number of HTML pages, I be the number of images and K be the set of
keywords. Thus,

H = {h1 , h2 , h3  hn }, I = {i1 , i2 , i3  i m } and K = {k1 , k 2 , k 3  k l }

where n, m and l denote the total number of HTML pages, images and keywords
respectively.
In order to effectively capture the semantics of I present in H, K can be used. The
association between K and H can be written as:

 stg (K ) 
stg (K ↔ H ) =   (1)
 Max(stg (K )) 

This denotes the association between keywords to HTML pages and can be estimated
using the frequency of occurrence of K in H. Since, K is the total number of keywords
in a single HTML page ‘hp’ may contain only ‘kq’of keywords. Now, the relation
between each keyword ‘kj’ where ( j = 1,2,3, , q ) with a single HTML document
can be written as

 stg (k j ) 
stg (k j ↔ h p ) =  
 Max(stg (k )) 
(2)
 j  j =1, 2 ,, q

The above equation denotes the association between each keyword Kj in a single
HTML document ‘hp’. Similarly, the relationship between a page and image can be
derived. From Eq. 2, we get the importance i.e. strength of each keyword ‘kj’ in a
Capturing High-Level Semantics of Images in Web Documents Using Strength Matrix 291

document ‘hp’. The strength is the measure of the frequency of occurrence and a
keyword with a maximum frequency of occurrence is assigned higher strength value,
which is 1. Similarly, all the keywords present in a particular document is assigned a
strength value by which the importance of that particular keyword is estimated. The
example depicted in Table 1 elucidates the estimation of strength using frequency of
occurrence of a keyword in a HTML document.
Let the number of keywords in a HTML document is 5 and maximum frequency of
occurrence (FOC) is 30.

Table 1. Strength Matrix using Frequency of occurrence

HTML Page FOC Strength

K1 10 0

K2 3 1

K3 30 1

K4 8 0

K5 5 1

From the above example, we can observe that not all keywords are equally
important. It is sufficient to consider only a set of keywords kstg such that the strength
of these keywords is greater than a threshold tstg. In our approach, we have fixed this
threshold as 25% of the maximum strength value. Now, it is important to estimate the
strength of the keywords with the images. We have used Image Title, ALT tag, link
structure and anchor text as high-level textual feature and feature vector is
constructed. Thus, the relation

( )
stg k j ↔ I h p = stg (k j ↔ hp ) + m (TAG, k j ) + m (IN , k j ) + m (LS , k j ) + m (AT , k j ) (3)

in above equation j = 1, 2, q and m is a string matching function with either 0 or 1

as the output. The output value of each component of above equation is in the rage of
[0-1]. This relation is effective in capturing the importance of a keyword in a
document and that of images. Both the strength value as well as image related textual
292 P. Shanmugavadivu, P. Sumathy, and A. Vadivel

information are combined to narrow down the semantic gap of image Sample strength
matrix with all features has been depicted in table 2.

Table 2. Strength Matrix using both frequency of occurrence and image high-level features

Keyword stg ⋅(k j ↔ hp ) [m (TAG, k )]

j m (IN, k j ) m (LS, k j ) m ( AT, kj )
FOC

K1 1 0.033 1 0 0 0

K2 30 1 1 1 1 0

K3 20 0.66 1 0 0 1

K4 8 0.26 1 1 1 0

K5 5 0.16 1 0 0 1

In the above table, K i is a keyword extracted from HTML documents. While

extracting the keywords, the stop words are removed and the stemming operation is
carried out. The high-level semantic information of the images can be extracted from
the above table. Say for example, frequency of K 2 is high and also it is matching
with most of the image related textual string and thus K 2 has more association with
the HTML page and captures the semantics of the images.
( ) ( ) ( ) ( )
m TAG , k j , m IN , k j , m LS , k j and m AT , k j are the match functions
to match the similarity between images TAG and keyword, image name and keyword,
link structures and keyword and Anchor Text with keyword respectively.

4 Experimental Results
For the purpose of crawling, the HTML documents along with the images, an internet
crawler was also developed. The various stages of experimental setup is shown in
Fig. 1
Capturing High-Level Semantics of Images in Web Documents Using Strength Matrix 293

Fig. 1. Stages of Experimental Setup

The performance of the proposed technique is measured based on the crawled

HTML pages along with the images from WWW. The textual keywords are extracted
and the frequency of occurrence of all the keywords is calculated. The text
information from URL link of images is also extracted. The strength matrix is
constructed using this information stored in a repository. During querying, the query
keyword is found in the repository and based on the strength value, the result is
ranked. In the experiment, many web documents from various domains such as sports,
news, cinema, etc have been crawled. Approximately, 10,000 HTML documents with
30,000 images have been extracted and preprocessed for retrieval. The web crawler
provides HTML pages along with the image content. The text in each page is
categorized into normal text and image related text. While extracting the keywords,
the stop words are removed and the stemming operation is carried out. Once the
keywords and image related text are extracted, the weighted matrix is constructed and
it is stored in the repository. For each page, this matrix is constructed and stored in the
repository along with a relation between the document and image present. While
making a query based on textual keywords, search will be carried out in the repository
and the final ranking of retrieval is based on the weighted value. The images with
higher weights are ranked first and will tip the retrieval result.
Fig.2 illustrates the obtained sample result for a keyword “A VadiveI” in
www.google.com. The fourth, fifth and nineteenth images are relevant to the query
and however, these images are ranked lower. The outcome of the proposed retriever
for the same keyword as query, is shown below in Fig. 3 It is observed from Fig. 3
that the proposed system has retrieved the relevant image from WWW and ranked the
relevant images higher. This is due to the fact that the strength matrix constructed
294 P. Shanmugavadivu, P. Sumathy, and A. Vadivel

from each page effectively captures the association between images and keywords. In
addition, manually looked into the textual content of each HTML page and estimated
the strength and are given for each image as the percentage of the estimated strength
value was also computed manually. For evaluating the performance of the proposed
approach in our system, the precision of retrieval is used as the measure. Moreover,
the obtained results are compared with some of the recently proposed similar
approach and are presented in Fig. 4. The average precision (P) in percentage for 10,
20, 50 and 100 nearest neighbors is given. We have compared performance with
Hierarchical clustering (HC) [21]; Rough Set (RS) based approach [22] and Bootstrap
Framework (BF) [20]. From the results shown in Fig. 4, it is observed that the
performance of the proposed approach is quite encouraging. The precision of retrieval
using the strength matrix is found to be high compared to others. The reason for this
performance enhancement is due to the effectiveness of strength matrix in capturing
the high level semantics of the images.

Fig. 2. Sample query for retrieving images from Google.com

98% 90% 85% 80% 79%

78% 76% 76% 72% 70%

Fig. 3. Sample retrieval result from the proposed system for query keyword “A Vadivel”
Capturing High-Level Semantics of Images in Web Documents Using Strength Matrix 295

HC
100 RS

Perceived Precision
BF
80 Pr o p o sed

60
40
20
0
10 20 50 100
No. of Nearest Neighbors

Fig. 4. Comparison of Precision of Retrieval using strength matrix

5 Conclusion

The role of textual keywords for capturing high-level semantics of an image in HTML
document is studied. It is felt that the keywords present in HTML documents can be
effectively used for describing the high-level semantics of the images present in the
same document. Additionally, a web crawler was developed to fetch the HTML
document along with the images from WWW. Keywords are extracted from the
HTML documents after removing stop words and performing stemming operation.
The strength of each keyword is estimated and associated with HTML documents for
constructing strength matrix. In addition, textual information presents in image URL
is also extracted and combined with the strength matrix. It is observed from the
experimental result that both textual keywords and keywords from image URL
achieves high average precision of retrieval.

References
1. Faloutsos, C., Barber, R., Flickner, M., Hafner, J., Niblack, W., Petkovic, D., Equitz, W.:
Efficient and Effective Querying by Image Content. J. Intell. Inf. Syst. 3(3-4), 202–231
(1994)
2. Pentland, A., Picard, R.W., Scaroff, S.: Photobook: Content-based manipulation for image
databases. Int. J. Comput. Vision 18(3), 233–254 (1996)
3. Gupta, R.J.: Visual Information Retrieval. Commun. ACM 40(5), 70–79 (1997)
4. Ma, W.Y., Manjunath, B., Netra: A toolbox for navigating large image databases. In:
Proceedings of International Conference on Image Processing, pp. 568–571 (1997)
5. Wang, J.Z., Li, J., Wiederhold, G.: Simplicity: semantics-sensitive integrated matching for
picture libraries. IEEE Trans. Pattern Mach. Intell. 23(9), 947–963 (2001)
6. Liu, Y., Zhang, D.S., Lu, G., Ma, W.-Y.: A survey of content-based image retrieval with
high-level semantics. Pattern Recognition 40(1), 262–282 (2007)
7. Long, F., Zhang, H.J., Feng, D.D.: Fundamentals of content-based image retrieval
Multimedia Information Retrieval and Management. Springer, Berlin (2003)
8. Rui, Y., Huang, T.S., Chang, S.-F.: Image retrieval.: Current techniques, promising
directions, and open issues. J. Visual Commun. Image Representation 10(4), 39–62 (1999)
296 P. Shanmugavadivu, P. Sumathy, and A. Vadivel

9. Vadivel, A., Shamik, S., Majumdar, A.K.: Robust Histogram Generation from the HSV
Color Space based on Visual Perception. International Journal on Signals and Imaging
Systems Engineering 1(3/4), 245–254 (2008)
10. Gevers, T., Stokman, H.M.G.: Robust Histogram Construction from Color Invariants for
Object Recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence 26(1), 113–118 (2004)
11. Vadivel, A., Shamik, S., Majumdar, A.K.: An Integrated Color and Intensity Co-
Occurrence Matrix. Pattern Recognition Letters 28(8), 974–983 (2007)
12. Palm, C.: Color Texture Classification by Integrative Co-Occurrence Matrices. Pattern
Recognition 37(5), 965–976 (2004)
13. Liu, Y., Zhang, D., Lu, G.: Region-based image retrieval with high-level semantics using
decision tree learning. Pattern Recognition 41, 2554–2570 (2008)
14. Shen, H.-T., Ooi, B.-C., Tan, K.-L.: Giving meaning to WWW images. ACM Multimedia,
39–47 (2000)
15. Yanai, K.: Generic image classification using visual knowledge on the web. ACM
Multimedia, 167–176 (2003)
16. Sanderson, H.M., Dunlop, M.D.: Image retrieval by hypertext links. ACM SIGIR, 296–
303 (1997)
17. Zhao, R., Grosky, W.I.: Narrowing the Semantic Gap—Improved Text-Based Web
Document Retrieval using Visual Features. IEEE Transactions on Multimedia 4(2), 189–
200 (2002)
18. Cai, D., He, X., Ma, W.-Y., Wen, J.-R., Zhang, H.: Organizing WWW Images based on
the Analysis of Page Layout and Web Link Structure. In: Proc. Int. Conf. on Multimedia
Expo, pp. 113–116 (2004)
19. Cai, D., Yu, S., Wen, L.R., Ma, W.Y.: VIPs: a vision based page segmentation algorithm.
Microsoft Technical Report, MSR-TR-2003-79 (2003)
20. Feng, H., Shi, R., Chua, T.-S.: A bootstrapping framework for annotating and retrieving
WWW images. In: Proceedings of the ACM International Conference on Multimedia
(2004)
21. Cai, D., He, X., Li, Z., Ma, W.-Y., Wen, J.-R.: Hierarchical clustering of WWW image
search results using visual, textual and link information. In: Proceedings of the ACM
International Conference on Multimedia (2004)
22. Wu, C., Hu, X.: Applications of Rough set decompositions in Information Retrieval.
International Journals of Electrical and Electronics Engineering 4(4) (2010)
Multiple Access Scheme Using MSE-OFDM with Pseudo
Random Cyclic Prefix

Biswajit Datta1, Anupam Karmakar2, and Mrinal K. Naskar3

1
Department of Electronics and Communication Engineering,
College of Engineering and Management,
Kolaghat, West Bengal,
Pin – 721171, India
[email protected]
2
Department of Electronic Science,
University of Calcutta, Kolkata,
West Bengal,
Pin – 700 009, India
[email protected]
3
Department of Electronics and Telecommunication Engineering,
Jadavpur University,
West Bengal,
Pin – 700 032, India
[email protected]

Abstract. A new multiple access scheme using Multi-Symbol Encapsulated

Orthogonal Frequency Division Multiplexing (MSE-OFDM) is proposed. This
is basically a time division multiple access (TDMA) technique utilizing the
MSE-OFDM. The Bit Error Rate (BER) performance of the TDMA system
using conventional OFDM is slightly better than that of the proposed system.
This weakness is compensated as the proposed technique exhibits an
improvement in bandwidth efficiency compared to the TDMA using
conventional OFDM.

1 Introduction

Orthogonal Frequency Division Multiplexing (OFDM) is a digital multicarrier

modulation technique. The technique of multicarrier modulation was first proposed in
1966 by R.W. Chang. But at that time the technique did not become popular because
of complexity in its implementation. The concept of OFDM was first proposed in
mid-1980s [1]. But the use of OFDM for wireless communications became possible
only in early 1990s with the advances in hardware for digital signal processing. The
main advantage of OFDM is its robustness to frequency selective fading and
multipath propagation which has made it a natural choice for high speed data
communication in mobile environment. In the recent years OFDM has found its
applications in Digital Audio Broadcasting (DAB), Digital Video Broadcasting
(DVB) and High Speed Digital Subscriber Line (HDSL). OFDM has been set

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 297–306, 2011.
© Springer-Verlag Berlin Heidelberg 2011
298 B. Datta, A. Karmakar, and M.K. Naskar

as the standard of air interference for Wireless LANs (802.11a and 802.11g).
In spite of its growing applications OFDM suffers from some inherent drawbacks
like lower bandwidth efficiency due to the Cyclic Prefix (CP) overhead, high
Peak-to-Average-Power-Ratio (PAPR) and stringent requirements for time and
frequency synchronisation. The Multiple-Symbol Encapsulated OFDM (MSE-
OFDM) technique [2,3] overcomes some of the above mentioned drawbacks to a
great extent.
In MSE-OFDM technique [2,3] a number of OFDM symbols are grouped
together and are protected by a single CP. CP in this case comprises of last few
samples of the last OFDM symbol. The length of CP is at least equal to the length of
the Channel Impulse Response (CIR). There are two proposed variations of MSE-
OFDM [2], viz., (i) CP reduced MSE-OFDM and (ii) FFT-size reduced MSE-OFDM.
In CP reduced MSE-OFDM, shown in Figure 1, a number of conventional OFDM
symbols are put together into a frame with a single CP. The fame is known as an
MSE-OFDM frame. As there is a single CP for a group of OFDM symbols the CP-
reduced MSE-OFDM exhibits an improvement of bandwidth efficiency compared to
the conventional OFDM. In case of FFT-size reduced OFDM, shown in Figure 2, the
size of the MSE-OFDM frame is made equal to that of a conventional OFDM symbol.
Thus the size of individual OFDM symbol is much less than that of the conventional
one. This implies a less number of subcarriers and hence a lower PAPR. The smaller
number of subcarriers also implies higher spacing between subcarriers. The frequency
synchronisation error depends on the relative frequency offset, i.e., the ratio of the
actual frequency offset to the subcarrier spacing. Hence more subcarrier spacing
means less frequency synchronisation error. The time synchronisation error and
bandwidth efficiency in this case, however, remain unchanged compared with the
conventional OFDM.
Considering the advantages of MSE-OFDM, in this paper we propose a multiple
access scheme using MSE-OFDM for mobile environment. Here we consider a
typical mobile environment consisting of a single Base Station (BS) or Access Point
(AP) and multiple Mobile Stations (MS) which communicate to an external network
via the BS. In this case we propose a multiple access scheme for data transmission
from the BS to different MS’s. Two multiple access schemes using MSE-OFDM have
been proposed in [4] and [5]. In the first case [4] a complete MSE-OFDM frame is
necessary at the receiver for frequency equalisation. In the second case [5] also all
OFDM symbols of a complete frame are required at the receiver for the removal of
Inter Symbol Interferences (ISI) and Inter Carrier Interferences (ICI). Hence the data
destined for different MS’s cannot be put together into one MSE-OFDM frame as
proposed in [4,5]. Also both the schemes are not suitable for packet transmission
because in both cases continuous transmission from all sources have been proposed.
In that case there will be reduction in battery life of the MS’s.
The rest of the paper is organised as follows. The proposed multiple access
schemes using MSE-OFDM has been described in Section 2. This section discusses
the transmitter, receiver and the equalisation technique for the proposed multiple
access scheme. The system performance has been discussed in Section 3 using the
simulation results. Section 4 concludes the paper.
 Multiple Access Scheme Using MSE-OFDM with Pseudo Random Cyclic Prefix 299

(i-1)th ith (i+1)th

Symbol Symbol Symbol

One OFDM Symbol

(a)
(a)

1st 2nd Mth

Symbol Symbol Symbol 1st 2nd Mth
Symbol Symbol Symbol
One MSE-OFDM Frame
One MSE-OFDM Frame
(b) (b)

Repeated part of Repeated part of

Cyclic Prefix OFDM symbol Cyclic Prefix OFDM symbol

Fig. 1. Structure of CP-reduced MSE OFDM. Fig. 2. Structure of FFT-size reduced MSE
(a) Conventional OFDM symbols. (b) An OFDM. (a) Conventional OFDM symbols.
MSE-OFDM frame. (b) An MSE-OFDM frame.

2 Proposed Multiple Access Scheme

In this proposed system the input data packets for different users are stored into a
buffer at the BS. From the buffer the packets are transmitted to the destination MS’s
in sequence of their arrivals. Each data packet forms a complete MSE-OFDM frame
so that the complete frame is demodulated at the destination MS only. Several MSE-
OFDM frames will comprise a one multi-access frame. The packet for each MS will
get transmission opportunity in its allocated slot in the multi-access frame. In our case
we proposed an MSE-OFDM system using pseudo random cyclic prefix [6]. Here
each MSE-OFDM frame uses one pseudo random sequence as the CP and different
pseudo random sequences will be used for different destination MS’s. At each MS the
received CP will be correlated with its own pseudo random sequence code. The MS
that results the highest value of correlation will be able to receive the rest of the
packet. Also the same PN sequence will be used at the end of each frame so that the
received signal has a cyclic look.

2.1 Transmitter

MSE-OFDM transmitter [6] is shown in Figure 3(a). In this case each OFDM symbol
consists of N complex exponentials or subcarriers which are modulated by the
complex data symbol X. Thus an OFDM symbol may be represented by
300 B. Datta, A. Karmakar, and M.K. Naskar

3VHXGRUDQGRP &3*HQHUDWRU
6HTXHQFH/LEUDU\

  
%LQDU\ 04$0 63 13RLQW 36 %XIIHU
LQIRUPDWLRQ 6\PERO WR1 ,))7 1[0 &3 D/A
VRXUFH 0DSSLQJ LQVHUWLRQ

(a)

&3'HWHFWRU 3VHXGRUDQGRP
6HTXHQFH/LEUDU\

&3    
 5HPRYDO 1[0 5DWH 13RLQW 36 04$0
'$ DQG63 3RLQW &RQYHUVLRQ ))7 6\PERO
WR (TXDOLVHU 1[0WR1 'HFRGHU
1[0

(b)

Fig. 3. Block Diagrams (a) MSE-OFDM Transmitter, (b) MSE-OFDM Receiver

nk
j 2π
1 N −1 N , n = 0,1,2,..., N − 1
x[ n ] =  X ( k )e (1)
N k=0

Each MSE-OFDM frame consists of M such N-Point OFDM symbols so that the lth
frame of the MSE-OFDM signal can be written as

N −1 M −1 N −1
sl =  Cl , p ( k )ψ 1 ( n, k ) +   X l ,i ( k )ψ 2 ( n − iN − P, k ) (2)
k =0 i =0 k = 0

where Cl,,p is the cyclic prefix using a pseudo random sequence. The symbol i
indicates the ith OFDM symbol of the lth frame. ψ1(n,k) and ψ2(n,k) are two
multiplexing window functions corresponding to the cyclic prefix and the M
information carrying OFDM symbols defined as follows

0 ,0 ≤ n ≤ P − 1
ψ 1 ( n, k ) =  (3)
1 , elsewhere
 Multiple Access Scheme Using MSE-OFDM with Pseudo Random Cyclic Prefix 301

and

 j 2πk ( n − P − iN ) , P ≤ n ≤ MN + P − 1
 1 N (4)
ψ 2 ( n − iN − P , k ) =  e
, elsewhere
N

0

At the transmitter side of the base station data packets from various sources will be
stored in the buffer from where the packets will be transmitted either in sequence of
their arrival or according to some pre-defined priorities. Each data packet will
comprise one complete frame of MSE-OFDM signal. The CP will be appended at the
transmitter as per the destination MS address. Different CP’s will be used for different
mobile stations.

2.2 Receiver

In this case the transmission medium is broadcast in nature. So at the receiver of each
MS it will be necessary to perform time domain correlation analysis between the
received CP and a predefined pseudo random code generated within each MS. The
receiver that generates the highest correlation will be able to receive the complete MSE-
OFDM frame. Block diagram of the MSE-OFDM receiver is shown in Figure 3(b).

2.3 Equalisation

The equalisation technique used here is different from that of the conventional OFDM
technique because of the unique frame structure of the MSE-OFDM. In this case an
MN-Point equalisation is employed. Hence in this case an additional MN-Point DFT
and IDFT will be required to be done before and after the frequency domain
equalisation respectively. As a pseudo random sequence has been used as CP, the
same sequence must also be added to the end of each MSE-OFDM frame to give the
signal a cyclic look so that the received signal will be the cyclic convolution between
the transmitted signal and the channel impulse response. Thus the transmitted signal
in (2) can be written as a vector

sl=[Cl,p(0),…,Cl,p(P-1),xl,0(0),…,xl,M-1(0),…,xl,M-1(N-1),Cl,p(0),…Cl,p(P-1)] T (5)

The MSE-OFDM signal has now been modified according to accommodate the
pseudo random sequence postfix. The postfix is stripped off at the receiver along with
the CP.
The received signal vector rl corresponding to the transmitted signal vector (6) can
be expressed as
rl = C .sl + wl (6)

where wl is an additive white Gaussian noise (AWGN) vector with the size of sl and
C is the channel matrix of size [MN+3P-1,MN+2P] as given by
302 B. Datta, A. Karmakar, and M.K. Naskar

h0 0 0
h1 h0

hL −1 h1 h0
C= 0 0 (7)
hL −1 h1 h0
hL −1 h1

0 0 hL −1
A cyclic convolution occurs between the CP removed received signal ~
sl and h so that
the following transform pair holds.

~ ~ ⇔ DFT ( ~ ~
sl ⊗ h + wl sl ).H + Wl = DFT ( ~
rl ) (8)

~ ~ . The size of the DFT here is

where H and Wl are the Fourier Transform of h and w l
MN points. It is assumed that the channel matrix is known from the channel
estimation. The effect of the channel on the received signal can now be removed by a
single tap frequency domain equalisation and the signal is converted back to time
domain for FFT demodulation according to the following relation

DFT ( ~
rl )
= IDFT { } + wl FEQ
~ FEQ
rl (9)
H
where wlFEQ has the same statistical property as that of the AWGN. The signal ~
rl
FEQ

is broken down into M-OFDM symbols in sequence for demodulation using N-Point
FFT.

3 Performance Analysis and Discussions

As already discussed, there are two different implementations of MSE-OFDM. The

CP-reduced MSE-OFDM configuration shows improvement of bandwidth efficiency
as there is reduction in the number of CP insertion. This system has reduced
robustness in time-selective channel and hence it is suitable for static or slowly
varying channels. However the FFT-size reduced MSE-OFDM configuration does not
show any improvement in terms of bandwidth efficiency as the MSE-OFDM frame
duration is kept the same as the symbol duration of the conventional OFDM. It is a
better option in time varying channels. Our multiple access system is proposed for the
time invariant channels; therefore we use the CP-reduced case. We thus get a definite
improvement in terms of bandwidth efficiency of the system.
 Multiple Access Scheme Using MSE-OFDM with Pseudo Random Cyclic Prefix 303

Fig. 4. Bandwidth Efficiency Curve

3.1 Bandwidth Efficiency

We consider that each subcarrier conveys a discrete symbol taken from a two-
dimensional signal space containing 2ν signal points. Then for the N-point
conventional OFDM with a CP of length P the bandwidth efficiency η1 is given by [7]

N
η1 = ν b/s/H (10)
N+P

In the case of MSE-OFDM, the OFDM frame consists of MN signal points, the CP
and a postfix by CP in order to introduce cyclic convolution. Hence the bandwidth
efficiency is given by

MN
η2 = ν b/s/H (11)
( MN + 2 xP )

The improvement in bandwidth efficiency is given by:

NP ( M − 2)
Δη = η 2 − η1 = (12)
( N + P )( MN + 2 xP )

Figure 4 shows both the bandwidth efficiency of MSE-OFDM symbols and also the
trend in improvement of bandwidth efficiency for different values of M. The analysis
has been done taking N=64 and P=16; modulation type being BPSK. It shows that
there will be improvement in bandwidth efficiency of the system with increasing the
value of M. However, it can also be observed that the improvement in bandwidth
efficiency approaches the saturation value of P/(N+P) = 0.2.
304 B. Datta, A. Karmakar, and M.K. Naskar

3.2 BER Performance of Multiple Access MSE-OFDM System

We consider that Δk represents the relative frequency offset (ratio between actual
frequency offset and subcarrier spacing) and Δn is the relative timing offset (ratio
between timing offset and sampling interval). Then the bit error rate for the BPSK
system for the kth subcarrier in one OFDM symbol can be written as [4]

Eb
Pk ,e = erfc (13)
2 2
N o + σ ICI + σ ISI

where the variance of ICI is given by

2  
2
2 N −1 2  sin (πΔk )
σ ICI =  X ( l ) H (l )   (14)
l =0,l ≠ k  N sin (π (l − k + Δk ) / N 
2 2

H(k) is the channel transfer function at the kth subcarrier and the variance of ISI is
given by
2
2 2  ∞ [( jπΔk ( N − 1) + j 2πkΔn + j 2πΔkΔn) / N ] n 
σ ISI = X ( k )    (15)
 n=1 n! 
The average bit error rate can be given as
1 N −1
Pe =  P (16)
N k = 0 k ,e
One aim of this study is to analyse the BER performance of the proposed multiple
access system and also to compare it with the same using conventional OFDM
system. Therefore, we conducted a set of simulations for both the cases of
conventional OFDM and CP-reduced MSE-OFDM. The parameters used in
simulations are N=64, P=16 and the channel impulse response (CIR) for three
different users:

h1 = [1 0 0 0 0 0]
h2 =[0.9285 0 0.3714 0 0 0]
h3 =[-0.6154 0.7692 0 0.1538 0.0769]

One simulation we have conducted for the conventional OFDM system. Here we have
used a set of random bit sequences for three different users and inserted the
appropriate CPs. Above three different frequency selective paths have been selected
for the three users. At the receivers the frequency equalization has been performed
using the appropriate transfer functions. Another set of simulation have been
conducted on the same data for the proposed MSE-OFDM system. Here we have used
the same set of CIRs, but the frequency equalisation has been performed after MN-
point FFT on the received MSE-OFDM frame.
 Multiple Access Scheme Using MSE-OFDM with Pseudo Random Cyclic Prefix 305

Fig. 5. BER performance curve for user 1 Fig. 6. BER performance curve for user 2

The results of simulation are shown in Figures 5, 6 and 7. From the Figures it can
be observed that there are little differences in the BER performance of MSE-OFDM
compared to that of the conventional OFDM. The reason behind this is that in case of
MSE-OFDM the individual OFDM symbols are not separated by guard time whereas
in case OFDM individual symbols are separated by the guard intervals. Thus the
symbols in MSE-OFDM suffer from more ISI.
It can further be said that the better performances by both the OFDM and MSE-
OFDM compared to the theoretical values are due to the channel inversion that takes
place at the receiver equalizer.

Fig. 7. BER performance curve for user 3

4 Conclusions
The proposed multiple access method is basically a time division multiple access
technique. This method is obviously better than the other multi-access techniques
306 B. Datta, A. Karmakar, and M.K. Naskar

[4,5] because in this proposed technique all users receive their individual MSE-
OFDM frames completely for frequency equalization and detection during its
assigned time slot. There can be improvement in the BER performance in case of
multiple access MSE-OFDM if the receiver structure with reduced complexity [6]
using hybrid equalization technique is adopted. But in that case the receiver becomes
complex.

References
1. Cimini Jr., L.J.: Analysis and Simulation of a Digital Mobile Channel Using Orthogonal
Frequency Division Multiplexing. IEEE Transactions on Communications 33, 665–675
(1985)
2. Wang, X., Wu, Y., Chouinard, J.-Y.: On the comparison between conventional OFDM and
MSE-OFDM systems. In: IEEE Global Telecommunications Conference, GLOBECOM
2003, December 1-5, vol. 1, pp. 35–39 (2003)
3. Chouinard, J.-Y., Wang, X., Wu, Y.: MSE-OFDM: A new OFDM transmission technique
with improved system performance. In: Proceedings of IEEE International Conference on
Acoustics, Speech, and Signal Processing (ICASSP 2005), March 18-23, vol. 3, pp. iii/865–
iii/868 (2005)
4. Singh, P., Chakrabarti, S., Raja Kumar, R.V.: Multiple Access Using MSE-OFDM. In: 5th
IEEE Consumer Communications and Networking Conference, CCNC 2008, January 10-
12, pp. 403–406 (2008)
5. Singh, P., Chakrabarti, S., Kumar, R.V.R.: A multiple access scheme using MSE-OFDM
with reduced complexity. In: IEEE Region 10 Conference TENCON 2008, November 19-
21, pp. 1–4 (2008)
6. Wang, X., Wu, Y., Wu, H.C., Gagnon, G.: An MSE-OFDM System with Reduced
implementation Complexity Using Pseudo Random Prefix. In: IEEE Global Telecomm.
Conference, GLOBECOM 2007, vol. 1, pp. 2836–2840 (December 2007)
7. Wang, X., Wu, Y., Chouinard, J.-Y., Wu, H.-C.: On the design and performance analysis of
multisymbol encapsulated OFDM systems. IEEE Transactions on Vehicular
Technology 55, 990–1002 (2006)
 Elimination of Irrelevancy during Semantic Service
Discovery Using Clustering Approach

Chellammal Surianarayanan and Gopinath Ganapathy

Department of Computer Science, Bharathidasan University, Khajamalai Campus,

Tiruchirappalli, 620 023, Tamil Nadu, India
[email protected]
[email protected]

Abstract. Semantic web service discovery process consumes a significant

amount of time even to perform a typical service match of limited functional
capabilities. Modern business integration requires services need to be dynami-
cally discovered and composed quickly. In this work, a similarity based clus-
tering approach is suggested for quick elimination of irrelevant services during
discovery. In this method, all the available published services are clustered by
service description similarity in prior to semantic matching. This yields clusters
of similar services. After services are clustered, when a query is submitted,
firstly, a particular cluster to which the query belongs to is found out. Then se-
mantic matching of capabilities will be performed only to that particular cluster
ignoring all other service clusters as irrelevant to the query. The proposed me-
thod is tested for its performance with a small test collection and results are
presented.

Keywords: semantic service discovery, semantic reasoner, service clustering,

candidate service, semantic matching.

1 Introduction
In automatic service discovery, a programme will autonomously match a given query
with the available published services and discover a list of matched services without
human intervention. Traditional Universal Description Discovery and Integration
(UDDI) based discovery supports only manual keyword based searching and category
based searching facilities that limits the capability and efficiency of queries [1].
Though few of the limitations of UDDI described in [2-3] are met with Web Service
Description Language (WSDL), lack of explicit semantics in WSDL prevents clients
from automatically discovering services without human assistance [4]. Semantic
service description languages such as [5-6] enable service providers to express their
intended semantics and bring maximal automation and dynamism in service discovery
and composition [7].
Generally, semantic service discovery frameworks employ semantic reasoners such
as Pellet and RACER for finding semantic relations between different concepts during
service matching. Despite its accuracy, semantic reasoning is heavy in time [8].
Generally business transactions involve multiple services from different domains need
to be composed dynamically in a complex chain [9].

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 307–315, 2011.
© Springer-Verlag Berlin Heidelberg 2011
308 C. Surianarayanan and G. Ganapathy

Optimization techniques viz., optimization at semantic reasoner and optimization

at matcher [10] are used to achieve the two critical needs, viz., ‘performance’ and
‘scalability’ during service discovery. As the first method depends on the internal
architecture of the reasoner, the second method is used predominantly. The second
method attains improved performance by reducing the number of services that really
need semantic matching of capabilities. Generally, for any given query, there may be
a huge number of obviously irrelevant services available. For example, services that
serve financial functionalities are irrelevant to a query which asks for temperature of a
location. Optimization at matcher removes such irrelevant services quickly and pro-
duces a reduced list of ‘candidate services’ that really need semantic matching.
In this work, a simple agglomerative clustering based approach is suggested for
identifying candidate services for semantic matching. In this approach, all the avail-
able published services are grouped into clusters by similarity in prior to discovery.
When a query is submitted, firstly the specific cluster to which the query belongs to is
found out. Then the semantic matching of capabilities is employed only to that spe-
cific cluster, considering all other clusters as irrelevant to the query.

2 Related Work
Various techniques such as clustering, inverted indexing, caching, WordNet based
similarity metric, etc. are used to achieve optimization at matching component. An
extended Jacquard similarity metric based star clustering is used in [11] for efficient
web service retrieval. In [12], K-Means clustering is used to filter out irrelevant ser-
vices along with Probabilistic Latent Semantic Analysis method to capture hidden
semantics of the query. In [13], Semantic Web Services Clustering method is used as
a layer between UDDI registry and Internet to group services into domains. In [14], a
single linkage based clustering with Jaccard similarity is used to reduce the overhead
significantly and locate the suitable services quickly. A method proposed in [15] uses
WordNet similarity for finding similar services and clusters them into various catego-
ries. An inverted indexing based approach presented in [16] accelerates the removal
of irrelevant services and facilitates automatic service composition.
The present work is a complementary approach in using a simple text based simi-
larity to find the ‘service description similarity’ among services. It finds the service
description similarity between all pair-wise combinations of services. A dissimilarity
matrix is constructed. Then the dissimilarity matrix is given as input for hierarchical
agglomerative clustering. This yields clusters of similar services. The method
quickly identifies the candidate services as it simply uses text similarity approach.

3 Proposed Method
A three-step clustering approach is proposed to find candidate services for semantic
matching. The steps include computation of pair-wise similarity score between all
possible service pairs, construction of dissimilarity matrix and clustering of services
by similarity. The methods of similarity computation, construction of dissimilarity
matrix and clustering of services are explained in the following sub sections. The
overview of the proposed approach is shown in Fig. 1.
 Elimination of Irrelevancy during Semantic Service Discovery 309

Compute pair-wise similarity

Service
Repository of Prior to semantic
published discovery process
services Construct dissimilarity matrix

Cluster similar services using

similarity threshold Only the specific
Service cluster of the
Query query is chosen
for semantic
Perform semantic matching matching

Fig. 1. Proposed clustering approach

3.1 Similarity Computation

Irrespective of various kinds of semantic service description languages which are used
to describe the interface of services, any service description contains details about
service name, service description and a set of operations. This first level service pro-
file is used in computing the service description similarity among services.
A service (S) is expressed as S= (SN, SD, OPi) where SN represents the service
name, SD represents the textual description of the service and OPi represents names of
all operations associated with the service.
The service description similarity between any two services is computed using the
following steps.
Step 1: Extract service name, service description and names of all operations from
service description as individual terms for the pair of services for which the service
description similarity score is to be found out
Step 2: Perform the following pre-processing steps on the contents extracted from
the pair of service descriptions under similarity computation. The pre-processing
steps include removal of common words (stop words) such as ‘this’, ‘the’, ‘returns’,
‘and’, etc., removal of insignificant service terms and conversion of words in different
Parts-Of-Speech (POS) into words of single POS.
For example, let us consider a textual service description as “this finds temperature
of a location”. This description contains the common words, ‘this’, ‘of’ and ‘a’.
From an investigation with service description files, it is found that, most of the
service descriptions contain insignificant terms such as ‘request’, ‘response’, ‘get’,
‘soapIn’, ‘soapOut’, ‘HttpRequest’, ‘’HttpResponse’, ‘portType’, ‘Result’, ‘GetIn’,
‘GetOut’, ‘interface’, etc. Operation names may also contain their service names as
prefixes. For example, the operation name, ‘CurrencyExchange_getRate’ contains its
service name, ‘CurrencyExchange’ as its prefix. During pre-processing common and
insignificant terms are removed. Further the service terms which occur in plural
forms are converted into singular forms.
310 C. Surianarayanan and G. Ganapathy

Step 3: Construct ‘representative strings’ using tokens obtained in Step 2 with dup-
licate terms eliminated.
Step 4: Find the normalized cosine similarity between the representative strings of
pair of services. The above similarity represents the service description similarity
score between the services.
The computation of service description similarity is illustrated as follows. Let us
consider two services ‘PhoneValidator’ and ‘PhoneVerification’ as S1 and S2.
Step 1: Extraction of service name, service description, and operation names for S1
and S2 as individual terms.

Contents extracted for S1. Service Name: PhoneValidator. Service Description:

Validates area codes and prefixes Operation names: PhoneValidatorSoap, GetRe-
mainingHits.

Contents extracted for S2. Service Name: Phone3T Service Description: Validates
a phone number and provides additional information Operation names: PhoneVerify
Step 2: Pre-processing of the contents extracted from S1 and S2
The common words ‘and’ in service description of S1 and ‘a’ and ‘and’ in S2 are
removed. Then, the service insignificant terms, ‘Get’, ‘Soap’ in S1 and ‘3T’ in S2 are
removed. While converting words in different POS into words of single POS, the
words, ‘validator’ and ‘validates’ in S1 are converted into single POS term in verb
form as ‘validate’. The words ‘prefixes’ and ‘hits’ are converted into their singular
forms.
Step 3: Let T1 and T2 represent all the individual terms obtained after pre-
processing the descriptions of S1 and S2 respectively.
T1 = {phone, validate, validates, area, codes, prefix, phone, validate, prefix, re-
maining, hits}
T2 = {phone, validate, phone, number, provides, additional, information, phone,
verify}.
With removal of duplicate words, the representative strings for S1 and S2 are
formed as
RS1 = {phone, validate, area, codes, prefix, remaining, hit}
RS2 = {phone, validate, number, provides, additional, information, verify}
Step 4: Normalized Similarity (NS) between RS1 and RS2 is computed using co-
sine similarity metric.

3.2 Construction of Dissimilarity Matrix

The construction of dissimilarity matrix is illustrated with 4 services as follows. Let

the services are denoted as S1, S2, S3 and S4. The possible combinations of service-
pairs are (S1-S1), (S1-S2), (S1-S3), (S1-S4), (S2-S2), (S2-S3), (S2-S4), (S3-S3), (S3-
S4) and (S4-S4). The normalized dissimilarity score between a pair of services is
given as Dissimilarity, D=(1-NS) where NS is the normalized cosine similarity score
of the pair of services After computing dissimilarity values, a dissimilarity matrix is
constructed. For the above 4 services, the dissimilarity matrix is constructed as
shown in Fig. 2.
 Elimination of Irrelevancy during Semantic Service Discovery 311

Fig. 2. Dissimilarity matrix for four services

3.3 Clustering of Services

After the construction of dissimilarity matrix, the services are clustered using hierar-
chical agglomerative clustering with group average scheme as clustering criterion.
During clustering a similarity threshold of 30% is specified as stopping criterion.

4 Experimental Setup
The following steps are used to test the performance of the proposed approach.
Step: 1 Prepare test data which should contain groups of related/similar services.
For example, in a test data of 10 services, 5 are ‘zip-code’ related services, 2 are ‘sms’
related services and remaining 3 are ‘email’ related services.
Step 2: Compute all pair-wise cosine similarity scores. Derive dissimilarity scores
and construct dissimilarity matrix
Step 3: Employ hierarchical agglomerative clustering method
Step 4: Compare the clusters returned in Step 3 with the original groups of related
services to evaluate the accuracy of the proposed approach.
An experimental setup as shown in Fig 3 is constructed. After the construction of
dissimilarity matrix the services are clustered using Matlab clustering tool.
A test data of 10 test services with 3 service groups is constructed using services
collected from “www.xmethods.com” and “www.remotemethods.com. The related
service groups along with the number of services in each group is given in Table 1

All published Pre-processing Representative

Services steps Strings

Construct Dissimilarity Cosine Similarity computation

matrix using Simmetrics.jar (open source)

Clusters of
Cluster using Matlab similar services

Fig. 3. Experimental setup for the proposed method

312 C. Surianarayanan and G. Ganapathy

Table 1. Related service groups in test data

# Service group Number of similar

services in each group
1 Zip code service group 5
2 Email 3
3 SMS 2

Table 2. URLs of WSDL files of Test Services

Service id URL of WSDL

S1 http://ws.strikeiron.com/zipdistance?wsdl
S2 http://www.info-me-sms.it/ws.php?wsdl
S3 http://www.ripedevelopment.com/webservices/ZipCode.asmx?wsdl
S4 http://ws.xwebservices.com/XWebEmailValidation/V2/XWebEmail
Validation.wsdl
S5 http://www.codebump.com/services/zipcodelookup.asmx?wsdl
S6 http://www.webservicemart.com/uszip.asmx?wsdl
S7 http://ws2.fraudlabs.com/mailboxvalidator.asmx?wsdl
S8 http://www.barnaland.is/dev/sms.asmx?wsdl
S9 http://ws.strikeiron.com/zipinfo3?wsdl
S10 http://ws.strikeiron.com/EmailVerify4?wsdl

Table 3. Representative Strings of Test Services

Service id Representative Strings

S1 {zip, distance, city, remaining, hit}
S2 {send, sms, check}
S3 {zip, code, city, state, area, time}
S4 {validate, email}
S5 {zip, code, three, digit, distance, user, place}
S6 {validate, zip}
S7 {mail, validate}
S8 {send, sms}
S9 {zip, city, state, remaining, hit, status}
S10 {email, verify, status, remaining,, hit}

The test services are given service ids as S1, S2, S3, S4, S5, S6, S7, S8, S9 and S10.
The URLs of the WSDL files and the representative strings of the test services are
given in Table 2 and Table 3 respectively.
The normalized service description similarity score is calculated using the experi-
mental setup. From the similarity score, the dissimilarity score is computed for each
pair. For each service, the dissimilarity score with every other possible service pair is
computed. For illustration, dissimilarity score of all pairs of S1 and S2 are shown in
Table 4.
 Elimination of Irrelevancy during Semantic Service Discovery 313

Table 4. Dissimilarity Score of all Service Pairs for S1

Service Dissimilarity Service Dissimilarity

Pairs for S1 score Pairs for S2 score
S1-S2 1 S2-S1 1
S1-S3 0.635 S2-S3 1
S1-S4 1 S2-S4 1
S1-S5 0.662 S2-S5 1
S1-S6 0.684 S2-S6 1
S1-S7 1 S2-S7 1
S1-S8 1 S2-S8 0.184
S1-S9 0.27 S2-S9 1
S1-S10 0.8 S2-S10 1

Fig. 4. Dendrogram obtained for test data

The test data results in a 10 × 10 dissimilarity matrix with all zeros in the diagonal.
The dissimilarity matrix is symmetric. The matrix is given as input to Matlab cluster-
ing tool in vector format. The dendrogram obtained is shown in Fig. 4.

5 Results and Conclusion

From dendrogram, it is found with the dissimilarity of 70% (i.e. with 30% similarity
threshold) three clusters are formed. Let the clusters be cluster-1, cluster-2 and clus-
ter-3. The clusters obtained are given in Table 5.

Table 5. Clusters obtained using proposed approach

# Cluster Services in the cluster

1 Cluster -1 (sms group) S2 and S8
2 Cluster -2 (Zip group) S1, S9,S3,S10 in and S5
3 Cluster-3 (email group) S4,S6,S7

From the above results, it is noted that out of 10 services, two services namely, S6
and S10 are grouped in different groups. i.e. S6 is expected to be grouped in Cluster-
2. But the approach grouped it in Cluster-3. Similarly, the service S10 is expected to
314 C. Surianarayanan and G. Ganapathy

be grouped in Cluster-3. But the approach grouped it in Cluster-2. When the results
are analyzed, it is found that the cause of wrong grouping is because of the presence
of terms ‘validate’ and ‘remaining’ and ‘hit’. These terms are commonly used both in
email and zip group of services. The proposed method results in a clustering accuracy
of 80% with this simple data set. It is planned to apply the proposed method to huge
data set and compare it with other standard procedures for further evaluation.

References
[1] Sohali, A., Zamanifar, K.: Matching Model for Semantic Web Services Discovery. Jour-
nal of Theoretical and Applied Information Technology 7(2), 139–144 (2009)
[2] Lee, K.-H., Lee, M.-Y., Hwang, Y.-Y., Lee, K.-C.: A Framework for XML Web Services
Retrieval with Ranking. In: IEEE International Conference on Multimedia and Ubiquit-
ous Engineering, pp. 773–778 (2007)
[3] Plebani, P., Pernici, B.: URBE: Web Service Retrieval based on Similarity Evaluation.
IEEE Transactions on Knowledge and Data Engineering 21(11), 1629–1642 (2009)
[4] Guo, R., Le, J., Xia, X.L.: Capability Matching of Web Service Based on OWL-S. In:
Proceedings of the 16th International Workshop on Database and Expert Systems Appli-
cations, pp. 653–657. IEEE Computer Society, Los Alamitos (2005)
[5] Kopecky, J., Vitvar, T., Bournez, C., Farrell, J.: SAWSDL: Semantic Annotations for
WSDL and XML Schema. IEEE Internet Computing 11(6), 60–67 (2007)
[6] Klusch, M., Kaufer, F.: WSMO-MX: A hybrid Semantic Web service matchmaker. Web
Intelligence and Agent Systems 7(1), 23–42 (2009)
[7] Sycara, K., Paolucci, M., Ankolekar, A., Srinivasan, N.: Automated discovery, interac-
tion and composition of Semantic Web Services. Journal of Web Semantics 1(1), 27–46
(2003)
[8] Mokhtar, S.B., Kaul, A., Georgantas, N., Issarny, V.: Towards Efficient Matching of
Semantic Web Service Capabilities. In: International Workshop on Web Services Model-
ing and Testing, pp. 137–152 (2006)
[9] Milanovic, N., Malek, M.: Current Solutions for Web Service Composition. IEEE Inter-
net Computing 8(6), 51–59 (2004)
[10] Lv, Q., Zhou, J., Cao, Q.: Service Matching Mechanisms in Pervasive Computing Envi-
ronments. In: International Workshop on Intelligent Systems and Applications, ISA, pp.
1–4 (2009)
[11] Shou, D., Chi, C.-H.: A Clustering-Based Approach for Assisting Semantic Web Service
Retrieval. In: IEEE International Conference on Web Services, Beijing, pp. 838–839
(2008)
[12] Ma, J., Zhang, Y., He, J.: Efficiently finding web services using a clustering semantic
approach. In: ACM International Conference Proceeding Series, Beijing, vol. 292, Ar-
ticle No.5 (2008)
[13] Nayak, R., Lee, B.: Web Service Discovery with additional Semantics and Clustering. In:
Proceedings of the IEEE/WIC/ACM International Conference on Web Intelligence, pp.
555–558. IEEE Computer Society, USA (2007)
[14] Liu, P., Zhang, J., Yu, X.: Clustering-Based Semantic Web Service Matchmaking with
Automated Knowledge Acquisition. In: Liu, W., Luo, X., Wang, F.L., Lei, J. (eds.)
WISM 2009. LNCS, vol. 5854, pp. 261–270. Springer, Heidelberg (2009)
 Elimination of Irrelevancy during Semantic Service Discovery 315

[15] Qu, X., Sun, H., Li, X., Liu, X., Lin, W.: WSSM: A WordNet-Based Web Services Simi-
larity Mining Mechanism. In: Computation World: Future Computing, Service Computa-
tion, Cognitive, Adaptive, Content, Patterns, pp. 339–345. IEEE Computer Society,
Greece (2009)
[16] Gao, T., Wang, H., Zheng, N., Li, F.: An Improved Way to Facilitate Composition-
Oriented Semantic Service Discovery. In: International Conference on Computer Engi-
neering and Technology, Singapore, pp. 156–160 (2009)
 Leaf and Flower Recognition Using Preferential Image
Segmentation Algorithm

N. Valliammal and S.N. Geethalakshmi

Avinashilingam Deemed University for Women, Coimbatore-641043

[email protected]
[email protected]

Abstract. Automatic plant classification systems are essential for a wide range
of applications including environment protection, plant resource survey, as well
as for education. With the aid of advanced information technology, image
processing and machine learning techniques, automatic plant identification and
classification will enhance such systems with more functionality, such as
automatic labeling and flexible searching. Image segmentation and object
recognition are two aspects of digital image processing which are being
increasingly used in many applications including leaf recognition. In this paper,
the Preferential Image Segmentation (PIS) method is used to segment an object
of interest from the original image. A probabilistic curve evolution method with
particle filters is used to measure the similarity between shapes during matching
process. The experimental results prove that the preferential image
segmentation can be successfully applied in leaf recognition and segmentation
from a plant image.

Keywords: Preferential image segmentation, Tree of shapes, Filters, Segmentation.

1 Introduction
Digital image processing is the use of the algorithms and procedures for operations such
as image enhancement, image compression, image analysis, mapping, geo-referencing,
etc. The influence and impact of digital images on modern society is tremendous and is
considered as a critical component in variety of application areas including pattern
recognition, computer vision, industrial automation and healthcare industries. The topics
in image processing can be roughly categorized as follows: imaging, image denoising,
image restoration, image coding, image segmentation, image transformation, object
representation and recognition, and content-based image retrieval. Typical tasks for
computer vision are scene interpretation, object recognition, optical character
recognition, registration, feature extraction and video tracking.
There are many kinds of plants that exist on the earth. Plants play an important role
in both human life and other lives that exist on the earth. Due to various serious issues
like global warming, lack of awareness of plant knowledge, etc., the categories of
plant is becoming smaller and smaller. Fortunately, botanists are realizing the
importance of protecting plants and are discovering methods to protect them. In a

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 316–325, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Leaf and Flower Recognition Using Preferential Image Segmentation Algorithm 317

similar fashion, there is also a need for recognizing a plant by its category, to help
agriculturists / farmers during reproduction of extinct plants. The main challenge
faced by them is the identification of the category to which a particular plant belongs
to. To help them in this challenging venture, several researches (Man et al., 2008; Lee
and Chen, 2006) are conducted to automatically classify a plant into a category, when
its leaf sample image is given as input.
Plant has plenty use in foodstuff, medicine and industry. And it is also vitally
important for environmental protection. However, it is an important and difficult task
to recognize plant species on earth. Designing a convenient and automatic recognition
system of plants is necessary and useful since it can facilitate fast classifying plants,
and understanding and managing them (Tzionas et al., 2005). The classification of
plant leaves is a vital process in botany and in tea, cotton and other industries.
Moreover, it can also be used for early diagnosis of certain plant diseases.
The paper is organized as follows. Section 2 describes the general architecture used
by plant classifiers. Section 3 explains the preferential segmentation, while section 4
summarizes the results and section 5 shows the conclusion and references.

2 General Architecture
The general process for plant classification through leaf recognition is given in Figure 1.

Fig. 1. General architecture of Automatic Plant Classification system

318 N. Valliammal and S.N. Geethalakshmi

The performance of object recognition depends on the performance of

preprocessing and segmentation. Segmentation is defined as the process of
partitioning a digital image into multiple segments, where the pixels inside each
segment share certain visual characteristics such as color, intensity or texture.
Adjacent regions are significantly different with respect to the same characteristics.
The goal of segmentation is to simplify and/or change the representation of an image
into something that is more meaningful and easier to analyze (Shapiro and Stockman,
2001). Image segmentation locates objects and boundaries (lines, curves, etc.) in
images. After segmentation, the image is represented by a set of segments, which
collectively represents the entire image. The process of plant identification through
leaf recognition is explained below.

2.1 Capturing Leaf Image

The leaf image is acquired by scanners or digital cameras. Since classification is a

complex operation that needs high memory usage [13], normally, lossy compression
format like JPEG / BMP / GIF is used. For convenience sake all the leaf images will
be of the same size and resolution. Plants are basically classified according to shapes,
colors and structures of their leaves and flowers [14]. However, to recognize the plant
based on 2D images, it is difficult to analyze shapes and structures of leaves since
they have complex 3D structures. In addition, the colors of leaves are always green;
shades and the variety of changes in atmosphere and season cause the color feature
having low reliability. Therefore, to recognize various plants using their leaves, the
obtained leaf image in RGB format will be converted to gray scale before
preprocessing. The formula used for converting the RGB pixel value to its grayscale
counterpart is given in Equation 1.
Gray = 0.2989 * R + 0.5870 * G + 0.1140 * B (1)
where R, G, B correspond to the color of the pixel, respectively.

2.2 Preprocessing
The leaf image preprocessing refers to the initial processing of input leaf image to
correct the geometric distortions, calibrate the data radiometrically and eliminate the
noise and clouds that present in the data. These operations are called preprocessing
because they normally carried out before the real analysis and manipulations of the
image data occur in order to extract any specific information. The aim is to correct the
distorted or degraded image data to create a more faithful representation of the real leaf.

Preprocessing Example
 Leaf and Flower Recognition Using Preferential Image Segmentation Algorithm 319

Various preprocessing techniques are then used to enhance the leaf image obtained.
Several techniques like boundary enhancement, smoothening, filtering, noise
removal, etc. can be applied to improve the quality of the leaf image.
Several approaches to segmentation have been proposed which can be categorized
into cluster based techniques (Pauwels and Frederix, 1999), histogram based techniques
(Zwiggelaar, 2010), compression based techniques (Ballester et al., 2007), edge
detection, region growing methods (Chen and Georganas, 2008), Graph partitioning
methods (Sumengen and Manjunath, 2006), watershed transformation (Wenzhong and
Xiaohui, 2010), etc. Another popular method used is the curve evolution methods
(Farzinfar et al., 2010). These methods evolve the initialized curve(s) to the boundaries
of objects in an image. The evolution of the curves may be driven by image gradient
information (Kichenassamy et al., 1995; Caselles et al., 1997), region information
(Chan and Vese, 2001; Jehan-Besson et al., 2003), or their combination (Tsai et al.,
2001). The image segments generated are enclosed by closed contours and are
unsupervised which make it an ideal candidate for image post processing. These
methods are more suitable for simple images and their performance degrades both in
terms of segmentation and complexity, with complicated images (Paragios and Deriche,
2000). In such cases, the utilization of prior information is necessary for curve evolution
methods in complicated image segmentation applications.
Several methods have been proposed that utilize prior information for supervised
image segmentation. These methods usually propose new variational energy functions
which integrate both the prior information and the gradient/region information in the
image to be segmented. The minimizations of these functions can lead to
segmentation results. Shape priors are utilized in (Cremers and Funka-Lea, 2005;
Cremers et al., 2002; Chen, 2002) and (Leventon et al., 2000). Both intensity priors
and shape priors are applied in Leventon et al., (2000). Natural image statistics are
employed for natural image segmentation in Heiler and Schnorr (2005). These
methods usually work better than unsupervised methods. However, shape priors are
primarily incorporated in supervised segmentation methods, and the information
contained in the intensities is not always fully utilized. For example, the methods in
Freedman et al. (2005) and Freedman and Zhang (2004) make use of the intensity
histograms only, and the spatial information is ignored. Furthermore, these methods
usually have initialization problems because the energy functions have multiple local
minimums. The methods are sensitive to the initial locations of the evolving curves.
In some cases, the user may be interested only in finding the location of an object
within the image. In such cases the background and other non-interested objects need
not be clustered. To perform such a task, the process of segmentation is combined
with object recognition. This concept is the key idea behind “Preferential Image
Segmentation” (PIS). PIS method is described as a method to preferentially segment
objects of interests from an image and ignore the remaining portions of the image for
this application. An example is shown in Figure 2.
PIS can be considered as a method similar to object detection, where rough
location of the object of interest is identified (Agarwal et al., 2004). The drawback
behind this method is that it often neglects to provide the exact boundaries which are
very important for segmentation. Similarly, classification methods (Aujol et al.,
2003), image parsing methods (Tu et al., 2005), wavelets (Aijol et al., 2003) and
variational methods (Samson et al., 2000) can be considered similar to PIS. However,
these methods classify the image as a whole instead of separating it into meaningful
objects and additionally they are computationally intense.
320 N. Valliammal and S.N. Geethalakshmi

(a) (b) (c) (d)

Fig. 2. Sample Leaf images

a b c d

Fig. 3. Sample Flower images

This paper analysis the usage of PIS to segment leaves from a plant. Segmenting
leaves from plant / tree is considered important in plant identification system.
Automatic plant identification by computers is a very important task for agriculture,
forestry, pharmacological science and etc. The general plant identification system is
composed of four steps, namely, (i) preprocessing (ii) segmentation (iii) feature
extraction and (iv) matching process. In this paper a method for leaf segmentation as
proposed by Pan et al. (2009) is analyzed. The paper is organized as below. This
section provided a brief introduction to the topic under discussion. Section 2 gives an
overview of the PIS technique used. Section 3 tabulates and discusses the results
obtained during experimentation and Section 4 presents a conclusion.

3 Preferential Image Segmentation (PIS)

The PIS is explained using an example image. Consider a situation where the image
in Figure 2a has to be identified in a clutter image as in Figure 2b.
Tree of shapes provides a natural way to represent the spatial relationships of the
shapes in an image. It represents images based on the techniques of contrast-invariant
mathematical morphologies. This method is based on the theory of image
representation using connected components of set of finite perimeters in the space of
functions with Weakly Bounded Variations (WBV). It shows that an image, if taken
as a function of weakly bounded variation is guaranteed to be decomposed into
connected components with closed boundary this is an extension to classical methods
where an image is taken as a piecewise- smooth function. The representation of an
image using a tree of shapes utilizes the inferior or the superior of a level line to
represent an object, and considers the boundary of the inferior area as the shape of the
 Leaf and Flower Recognition Using Preferential Image Segmentation Algorithm 321

object. Therefore, only closed shapes are generated. This representation also provides
a tree structure to represent the spatial relationship for the objects in an image.
Concept of “shape” is introduced to generate a unique inclusion tree for an image. A
shape is defined as the connected components of a level set and the holes inside them.
The whole image acts as the root of the tree located at the top level. The shapes in the
same level are spatially disjoint in the image. The shapes in the lower level
are spatially included in the shapes in the next higher level. The tree of shapes,
therefore, provides a natural way to represent the spatial relationships between the
shapes in the image.
The tree of shapes is used as a candidate to obtain the intensity information in the
priori image which is used in preferential segmentation. A common method used to
construct tree of shapes to both images (Figures 3a and 3b), whose features are then
compared to identify the ratio of similarity between the images. However, the main
disadvantage of this approach is the amount of shapes (features) generated, which is
very huge. For examples, the example picture produces 3832 shapes (Figure 2c). The
closed curves in Figure 3c corresponds to the boundaries of the shape in the tree of
shapes. These discovered boundaries provide valuable information which can be
utilized during shape similarity measure.
As the number of shapes increases, the computational complexity also increases,
thus necessitating to reduce the number of shapes used during segmentation and shape
comparison. This is performed using the intensity information that is contained inside
the shape. The intensity information by means of the tree of shapes includes the
following features.
1) The number of objects (Nd) contained directly in the shape, which
corresponds to the number of direct children of the shape in the tree, Nd=2
for the indicated shape in Figure 2a.
2) The total number of objects (Nt) contained in the shape, which corresponds
to the total number of children below the shape in the tree Nt=3 for the
indicated shape in Figure 2a.
3) The relative area change between the shape and its direct children. Suppose
the area of the shape is S, and the areas of its direct children are Si, where 1 ≤
i ≤ Nd, the relative area change is then defined as

NdS
A= ∏ i
i =1 S
4) The rough similarities of the boundaries of the shapes, which can be
represented as the ratio of the circumferences (C) squared to the area (S) of
the boundaries of the shapes (Gonzalez and Woods, 2002), i.e., R=C2/S.
These features for the two shapes should be very close to match. Exact matching may
not be achieved because of the differences between shapes, the effects of noise,
lighting changes and cluttered background. Thresholds should be set for coarse
matching to affect a rapid exclusion of most candidate shapes and can be adjusted
according to the application.
322 N. Valliammal and S.N. Geethalakshmi

The number of candidate shapes decreases substantially by means of the intensity

features extracted from the tree of shapes. In the case of Figure 3b, the feature
decreases the number of candidate shapes from 3832 to 977; the feature decreases the
number from 977 to 83; the feature decreases the number from 83 to 1; the feature
retains this candidate. The candidate shape left over matches the prior shape, as
shown in Figure 2d. The process takes 1.70 s.
In most cases, however, the candidate shapes will decrease from thousands to tens.
Curve matching is then performed on the boundaries of the resulting candidate shapes
and the prior shapes. The candidate shape which best matches the prior’s curve is
taken as the preferential segmentation result.
The curve matching steps are outlined below:
1) Extraction of the level lines for each image. The level set representations (1),
(2) are utilized here for the extraction. The level lines are defined as the
boundaries of the connected components.
2) Affine filtering of the extracted level lines at several scales. This step is
applied to smooth the curves using affine curvature deformation to reduce
the effects of noise.
3) Local encoding of pieces of level lines after affine normalization. Both local
encoding and affine normalization are designed for local shape recognition
methods. This step will help to deal with occlusions in real applications.
4) Comparison of the vectors of features of the images. Euclidean distance is
utilized to compare the feature vectors.
The performance of curve matching between two curves is calculated after affine
filtering, curve normalization and local encoding. Suppose C1 and C2 and are two
curves for matching, and S1 and S2 are pieces from and C1 and C2, respectively, then
the performance to take S1 and S2 as the matching between C1 and C2 is calculated
using the formula as in (1).

l1 xl 2
Score = (1)
L1 xL 2

where l1=arclength(S1), l2=arclength(S2), L1= arclength(C1) and L2 = arclength(C2).

The maximum score over all possible matching pieces is taken as the matching
between the curves C1 and C2. The implementation and description is provided in
Lisani et al. (2003).
This paper introduced the usage of PIS segmentation for plant and flower
recognition. The PIS method integrated image segmentation and object recognition
using the tree of shapes, which is an approach to the mathematical description of
morphology. The system was tested with several images and the results are discussed
in the next section.

4 Results
The system was evaluated using a series of benchmark images belonging to two
categories. The first is the leaf images along with plant images. The second is flower
images and their corresponding images. A total of 500 plant and flower images were
 Leaf and Flower Recognition Using Preferential Image Segmentation Algorithm 323

collected (Set I). All these images were RGB color images of 256 x 256 pixels size. A
total of 100 leaf and single flower images were stored separately (Set II). These
pictures also belonged to RGB color space and care was taken to collect different
sized images ranging from roughly 0.8 to 2 times the size of plants in the first set.
Sample images of plants from Set I are shown in Figure 4. In a similar fashion, Figure
5 shows sample images of flowers in Set I. Example leaf and flower images in Set II
are shown in Figure 4 and Figure 5 respectively.

a b c d

Fig. 4.

a b c d

Fig. 5.

Fig. 6. Leaf Recognition

The visual results obtained while using preference image segmentation with plants
and flowers are given in this section. Figure 6 and 7 shows the image after PIS.
324 N. Valliammal and S.N. Geethalakshmi

Fig. 7. Flower Recognition

Fig. 8. Tree of Shapes for leaf and flower

The tree of shapes obtained for Figure 6 and Figure 7 are shown in Figure 8. From
the results, it is evident that the PIS algorithm is a good candidate for leaf and flower
segmentation and object recognition from plants and set of flowers.

5 Conclusion
A probabilistic curve evolution method with particle filters is used to measure the
similarity between shapes during matching process. The experimental results prove
that the preferential image segmentation can be successfully applied in leaf
recognition and segmentation from a plant image. In summary, the proposed method
models an image with a function of bounded variation and utilizes a tree of shapes for
 Leaf and Flower Recognition Using Preferential Image Segmentation Algorithm 325

image representation. This method encodes the prior information for preferential
segmentation as a tree of shapes. The method is invariant to translation, rotation and
scale transformations because both feature extraction and boundary matching are
invariant to these transformations. All these features and the positive results obtained
prove that the Preferential Image Segmentation method can be successfully exploited
in leaf identification for plant recognition systems. Future research directions include
testing the image with more flowers and leaf images and incorporating the PIS
method into the leaf recognition system.

References
[1] Agarwal, S., Awan, A., Roth, D.: Learning to detect objects in images via a sparse, part-
based representation. IEEE Trans. Pattern Anal. Mach. Intell. 26(11), 1475–1490 (2004)
[2] Aujol, J., Aubert, G., Blanc-Féraud, L.: Wavelet-based level set evolution for
classification of textured images. IEEE Trans. Image Process. 12(12), 1634–1641 (2003)
[3] Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. Int. J. Comput.
Vis. 22(1), 61–79 (1997)
[4] Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Trans. Image
Process. 10(2), 266–277 (2001)
[5] Chen, L., Georganas, N.D.: Region-based 3D Mesh Compression Using an Efficient
Neighborhood-based Segmentation. Simulation, Society for Computer Simulation
International 84(5), 185–195 (2008)
[6] Chen, Y.: Using prior shapes in geometric active contours in a variational framework.
Int. J. Comput. Vis. 50(3), 315–328 (2002)
[7] Cremers, D., Funka-Lea, G.: Dynamical statistical shape priors for level set based
sequence segmentation. Variational and Level Set Methods Comput. Vis., 210–221
(2005)
[8] Cremers, D., Tischhauser, F., Weickert, J., Schnorr, C.: Diffusion snakes: Introducing
statistical shape knowledge into the Mumford–Shah functional. Int. J. Comput.
Vis. 50(3), 295–313 (2002)
[9] Farzinfar, M., Xue, Z., Teoh, E.K.: A novel approach for curve evolution in segmentation
of medical images. Comput. Med. Imaging Graph 34(5), 354–361 (2010)
[10] Freedman, D., Zhang, T.: Active contours for tracking distributions. IEEE Trans. Image
Process. 13(4), 518–526 (2004)
[11] Cotton Incorporated USA (2009), The classification of Cotton,
http://www.cottoninc.com/ClassificationofCotton
[12] Kumar, N., Pandey, S., Bhattacharya, A., Ahuja, P.S.: Do leaf surface characteristics
affect agrobacterium infection in tea [camellia sinensis (l.) o kuntze]? J. Biosci. 29(3),
309–317 (2004)
[13] National Institute for Agricultural Botany. Chrysanthemum Leaf Classification,
Cambridge (2005)
[14] Tzionas, P., Papadakis, S., Manolakis, D.: Plant leaves classification based on
morphological features and a fuzzy surface selection technique. In: Fifth International
Conference on Technology and Automation, Thessaloniki, Greece, pp. 365–370 (2005)
[15] Wu, S.G., Bao, F.S., Xu, E.Y., Wang, Y., Chang, Y., Xiang, Q.: A Leaf Recognition
Algorithm for Plant Classification Using Probabilistic Neural Network. In: IEEE
International Symposium on Signal Processing and Information Technology, pp. 11–16
(2007)
Analysis of Next Hop Selection for Geocasting in VANET

Sanjoy Das and D.K. lobiyal

School of Computer and Systems Sciences

Jawaharlal Nehru University,
New Delhi, India
{sdas.jnu,lobiyal}@gmail.com

Abstract. Geocasting is a special variant of multicasting, where data packet or

message is transmitted to a predefined geographical location i.e., known as
geocast region. The applications of geocasting in VANET are to disseminate
information like, collision warning, advertising, alerts message, etc. The
topology of VANET is changes frequently and rapidly due to the fast movement
of vehicles. Therefore, the link time between two neighboring vehicles exist for a
short span of time. This makes delivery of messages to a geocast region most
challenging issue in VANET. In this paper, we have proposed an analytical
model based on connectivity and selection of next node. For the better
network connectivity in multi hop vehicular ad hoc network each cell
should have at least one node. The numerical results show how connectivity
between nodes depends on the transmission range of nodes.

Keywords: VANET, Geocast, neighbour node, probability, Cell.

1 Introduction

VANET is a special class of Mobile Ad hoc Network (MANET).A Mobile Ad hoc

network is a dynamically reconfigurable wireless network with no fixed
infrastructure. Every node in this network behaves like a router to relay a message
from one node to another. In MANET; nodes are laptops, PDAs, palmtops, and other
mobile devices whereas in VANET [1] nodes are vehicles. In addition, the other
characteristics which differentiate VANET from MANET are mobility of nodes;
structure of the geographical areas, delay constraint, privacy, etc.., node movement
depends on the structure of road or structure of city or terrain etc.
While delivering message from source to destination node, if destination node is
not within the transmission range of source node then the source node send the
message to the destination node with the help of intermediate nodes. The ad hoc
network is multi-hop in nature and message delivery in depends on the connectivity
among the intermediate nodes. The message delivery from one location to another
location is done with the help of intermediate nodes. The aim of Geocast routing
protocols is to deliver a message from one location (i.e. sender) to a predefined
location known as Geocast region with optimal number of node and time period. It is

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 326–335, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Analysis of Next Hop Selection for Geocasting in VANET 327

desirable that protocols should maintain the low end to end delay and high accuracy
in delivering message i.e., high delivery success ratio, low overhead and total
numbers of participating hops should be minimum.
The rest of paper is organized as follows. Section 2 presents work related to the
geocast protocols. In section 3 overview of our proposed mathematical model is
presented. In section 4 probabilistic analysis of one hop neighbor and effect of various
network parameters is discussed. Finally, the work presented in this paper is
concluded in section 5.

2 Related Work

Extensive works have been carried out by researchers, academicians and industries
for successfully dissemination of messages from source to destination. There are
several projects [2], [3], [4], [5] on VANET i.e. [CarTalk, FleetNet–Internet on the
Road, NoW (Network on Wheel)] are going on for its deployment in the real world.
The main focus of all these projects is to provide safety, and timely dissemination of
message from one location to another location. One of the message delivery protocols
proposed for VANET tries to deliver a message to a geographic region rather than to a
node called geocast routing.Many protocols have been developed for geocast routing
such as LAR [6], LBM [7], GeoTORA [8] is modified TORA, GRID protocol is
modified to GeoGRID [9], DREAM [10], GRUV [11], etc. A comprehensive survey
of geocasting protocol is presented in [12]. A Voronoi diagram based geocast protocol
has been proposed in [13]. None of these protocols considered the road structure since
they have been primarily designed for MANET. In[22], the authors proposed different
modified LAR [6] algorithms. They have modified the request zone. Through
simulation, the authors have established that their proposed algorithms reduces route
request overhead as compared to original LAR. The performance analysis shows that
their method outperforms original LAR especially, in a dense and highly dynamic ad
hoc network. The mathematical analysis of VANET and ad hoc network
characteristics and network connectivity are proposed in [14], [15], [16], [17],
[18],[19].
The geocasting is the variant of multicasting, in geocasting source node will send
message to a particular geographical area. We divided the geocast method in two
phases. In the Phase-I, the source node will send a message to a node inside the
geocast region and in the phase-II, the receiving node of the geocast region will
deliver the message to the entire member in the geocast region. The node which is
move to the geocast region will automatically become the member of the geocast
region. In Phase-II, the message delivery inside the geocast region is done by simple
flooding techniques. Here, in this paper we are only focused our work on the phase –I.
In phase I, we have described the probabilistic analysis of nodes and effect of various
network parameters i.e., transmission range of node, node density.
328 S. Das and D.K. lobiyal

3 Proposed Model
We have considered the multi-hop environment, because it’s very rare that source and
destination node fall in each other transmission range. As there is no direct
connectivity between source and destination node, to route the message intermediate
nodes plays a vital role. The intermediate nodes are act as relay node. We have
considered highway scenario to deliver message from source to geocast region shown
in fig-1.

… Geocast
Region
Source
Node
| Phase –I | Phase –II

Fig. 1. Simple Scenario of Geocast on Highway

The highway is divided into small size cells [17] shown in fig.2. The cell size is
considered in this paper as a rectangular structure. The cell length is equal to the
transmission range of any nodes i.e. R and width is W. The transmission range for all
nodes is fixed. The highway is divided into n number of equal sized cells. The
advantage of cell is that it is used to identify the location of the vehicles.

Cell1 Cell2 Cell3 … Celln

Fig. 2. Single lane highway scenarios

The minimum number of nodes required to send a message from source to

destination is depend on node’s transmission range and distance between them. To
compute the distance between two nodes we have to know the coordinate of both the
nodes. Let’s take (x1, y1) and (x2, y2) are the coordinates of two nodes. The
Euclidian distance D between two nodes can be calculated by using equation-1

2 1 2 1 . (1)
The two nodes are able to communicate to each other if the Euclidian distance
between them is equal to or less than the transmission range.

4 Probabilistic Analysis of Neighbour Nodes

For our analysis, we assume that all nodes in the network have the same transmission
range. The others parameters used in the analysis are given in the table-1
 Analysis of Next Hop Selection for Geocasting in VANET 329

4.1 Analysis of One Hop Neighbour

We have considered that a highway is divided into number of cell of rectangular

shape. The size we considered here for highway is more realistic. The length of cell is
L and width is w. The area of each cell will be Ci= L*W. The total area of network
will be A= Ci*Nc. The probability of a node having m numbers of neighbours depends
on the size of a cell and transmission range of nodes. The probability of a node having
m numbers of neighbour nodes can be calculated according to theorem-1 proposed in
[20] with the help of the following equation.

N W R W R N
P (Nneigh=m)= 1 ;0 m N 1 . (2)
C N C N

and the average number of neighbours per node can be calculated as follows in
equation -3.
1 . (3)

Table 1. Symbol notation

Symbols Description
N Total number of nodes in network
A Area of network
R radio range
m Average number of one hop neighbours
Nneigh Number of one hop neighbours
Nc Total number of cell.
Ci Area of Cell.

4.1.1 Numerical Results

Fig 3 shows the one hop neighbour probability with varying the number of cells for
300 nodes in the network. We have considered value of radio transmission range 250,
number of nodes fixed to 300 and number of cells 20, 30 and 50 given in Table 2. The
result shows that when the number of cells is 20, the probability of having maximum
number of one hop neighbours is 0.1. When the number of cells increased to 30 and
50 the probability of having maximum number of one hop neighbours is 0.13
and 0.17. The probability of having maximum number of one neighbour node is
high when the no of cell is 20. This probability decreases as the numbers of cell
increases.

Table 2. Simulation parameters

Parameter Values
Nc 20,30,50
330 S. Das and D.K. lobiyal

0.18
Nc=20
0.16 Nc=30
Probability of one hop Neighbours
Nc=50
0.14

0.12

0.1

0.08

0.06

0.04

0.02

0
5 10 15 20 25 30 35 40 45 50
No of one hop Neighbours

Fig. 3. Probability of one hop neighbours with varying number of cells

Table 3. Simulation parameters

Parameter Values
N 100,200,300

0.35
N=100
N=200
0.3 N=300
Probability of one hop Neighbours

0.25

0.2

0.15

0.1

0.05

0
5 10 15 20 25 30 35 40 45 50

No of one hop Neighbours

Fig. 4. Probability of one hop neighbours with varying number of nodes

 Analysis of Next Hop Selection for Geocasting in VANET 331

Fig 4 shows the one hop neighbour probability with varying the number of nodes
for 50cells in the network given in Table 3. For a network size of100 nodes, the
probability of having maximum one hop neighbour is nearly 0.27. For a network size
of 200 and 300 nodes, the probability of having maximum number of one hop
neighbours is 0.2 and 0.17, respectively. The probability of having maximum
numbers of one hop neighbour is high when the number of node is high.

4.2 Probabilistic Analysis of Nodes Present in a Cell

We assume that the highway is a two dimensional structure. We have divided the
entire road structure in to number of cells of size w*L. The total number of nodes N is
uniformly distributed over L0 to Lmax area on the Highway. Here, we have analyzed
the probability of number of nodes, no present in each cell. In case each cell has n0
numbers of node, the network will be fully connected. It is very essential in multi hop
ad hoc environment that connectivity among nodes always exists for uninterrupted
successful data transmission.

Cell0 Cell1 Cell2 Celln

L0 L1 L2 L3 Lmax

Fig. 5. Cells are divided in to subarea

According to [21] we computed the probability of number of nodes, n0 present in a

cell. We assume that n0out of N nodes are located in the interval [Li, Lj], where 0 ≤ Li≤
Lj≤Lmax and i<0 and j<0.The total area of the network is denoted by A and each cell
area is denoted by Ci where 0<i<n. The expected number of nodes present in a unit
area is denoted by ρ =N/A.
A random variable d denotes the number of nodes within the given interval. The
probability that a node is placed within the interval [Li, Lj] is p = (Lj-Li)/Lmax. The
probability that n0 of N nodes are placed in a cell area Ci is

C N C N
A
P (d=n0) = *e A
!

C
= *e C . (4)
!
2
A node can transmit data within Π*R area. But in our proposed model the area is not
circular, as we divided the road into cells of size w*L. Therefore, the probability that
a randomly chosen node has n0neighbours is

ρ W R ρ W R
P (d=n0) = e . (5)
!
For large value of N and A, we can approximate this solution with a Poisson
distribution, i.e.
332 S. Das and D.K. lobiyal

N N
P (d=n0) = e . (6)
!

The node density ρ = N/A is the expected number of nodes present per unit area. For
given density ρ we can calculate the probability that n0 nodes are present in an
interval of length L21 = L2− L1 as

ρL ρL
P (d=n0) = e . (7)
!

4.2.1 Numerical Results

In fig. 6 shows the probabilitics analysis of numbers of nodes to be present in each
cell area.The different parameters considered for the analysis is given in table-4. We

Table 4. Simulation parameters

Parameter Values
N 50,100,150,200,250,300,350…2000

Fig. 6. Probability of having n0 numbers of node present in each cell Vs Node Density
 Analysis of Next Hop Selection for Geocasting in VANET 333

have fixed the number of cells and the transmission range of node. The node density
is vary here. We have shown the probability of 0,1,2,3 and 4 nodes present in each
cell.The initial value of node is 50 and incremented by 50 maximum is 2000. When
the number of node is 50 probability having zero node present in a cell is greater than
0.2. The probability of 3 and 4 nodes present in each cell is near to 0. When the
number of nodes are increases upto 500 the probability of having 2,3 and 4 nodes in
each cell is 1. As the number of nodes exceeds 500, it shows that 3 to 4 nodes are
always present in each cell. As the number of nodes increases, presence of nodes in
each cell is also increases and the network becomes fully connected.
In fig. 7 shows the probabilitics analysis for numbers of nodes to be present in each
cell area. The parameters considered for this analysis is given in Table-5. The node
density and the transmission range vary in this analysis. When the transmission range

Table 5. Simulation parameters

Parameter Values
R 100,110,120,…300

1.4
One Node
Two Node
1.2
Three Node
Four Node
1
Probability of n0 nodes in each

0.8

0.6

0.4
cell

0.2

0
100 150 200 250 300
Transmission Range

Fig. 7. Probability of having n0 numbers of node present in each cell Vs Transmission Range
334 S. Das and D.K. lobiyal

is 100 and the node density is low , the probability of having 0 or 1 node in a cell is
0.3. When the node density and the transmission range is increasing probability of
having nodes in cells is also increasing. The probability that at least one node is
available in any cell becomes constant at 0.9 after transmission range reach to 150
with increasing number of nodes. Further, any cell having atleast 2,3 or 4 node its
probability become constant to 1 after transmission range exceed 200. If we transmit
data with high transmission range, the nodes available in the network will be always
connected and successfully deliver the data to the destination.

5 Conclusion
In this paper we have presented a geocast protocol for Vehicular Ad hoc network. In
this protocol, we have proposed a model for highway scenario in which the highway
is divided in to number of cells. The probabilistic analysis of nodes present in each
cell is shown in our analysis. For the better network connectivity in multi hop
vehicular ad hoc network each cell should have at least one node. We have analyzed
the impact of various network parameters like node density, transmission range and
network size and their impact in network connectivity. The simulation work is
implemented using Matlab. The results shows that when the number of node is 50
probability having zero node in a cell is greater than 0.2. As the number of nodes
exceeds 500, it shows that 3 to 4 nodes are always present in each cell. Further, with
respect to transmission range the probability of at least one node is present in any
cell becomes constant at 0.9 after transmission range reach to 150 with increasing
number of nodes. The probability of having more than 1 node become constant to 1
after transmission range exceed 200.

References
1. Schmitz, R.: Analysis of Path Characteristics and Transport Protocol Design in Vehicular
Ad Hoc Networks. In: IEEE 63rd International Conference Vehicular Technology VTC,
pp. 528–532 (2006)
2. http://vanet.info/projects
3. The NoW: Network on wheels Project,
http://www.network-on-wheels.de/about.html
4. CarTalk, http://www.cartalk2000.net/
5. FleetNet–Internet on the Road, http://www.fleetnet.de/
6. Ko, Y.B., Vaidya, N.: Location-aided routing (LAR) in mobile ad hoc networks. In:
Proceedings of the ACM/IEEE International Conference on Mobile Computing and
Networking (MOBICOM 1998), pp. 66–75 (1998)
7. Ko, Y.B., Vaidya, N.: Geocasting in Mobile Ad-hoc Networks: Location-Based Multicast
Algorithms. In: 2nd IEEE Workshop on Mobile Computing Systems and Applications,
New Orleans, Louisiana (February 1999)
8. Ko, Y.B., Vaidya, N.: GeoTORA: A Protocol for Geocasting in Mobile Ad Hoc Networks.
In: IEEE International Conference on Network Protocols, Osaka, Japan (2000)
 Analysis of Next Hop Selection for Geocasting in VANET 335

9. Liao, W.H., Tseng, Y.C., Sheu, J.P.: GeoGRID: A Geocasting Protocol for Mobile Ad Hoc
Networks Based on GRID. J. Internet Tech. 1(2), 23–32 (2000)
10. Basagni, S., Chlamtac, I., Syrotiuk, V.R., Woodward, B.A.: A distance routing effect
algorithm for mobility (DREAM). In: Proc. MOBICOM, pp. 76–84 (1998)
11. Zhang, G., Chen, W., Xu, Z., Liang, H., Dejun, M., Li, G.: Geocast Routing in Urban
Vehicular Ad Hoc Networks. In: Lee, R., Hu, G., Miao, H. (eds.) Computer and
Information Science 2009, vol. 208, pp. 23–31. Springer, Heidelberg (2009)
12. Maihofer, C.: A Survey of Geocast Routing Protocols. IEEE Communications Surveys &
Tutorials 6(2), 32–42 (2004)
13. Stojmenovic, I.: Voronoi Diagram and Convex Hull Based Geocasting and Routing in
Wireless Networks, University of Ottawa, TR-99-11 (1999)
14. Antunes, N., Jacinto, G., Pacheco, A.: On the minimum hop count and connectivity in one-
dimensional ad hoc wireless networks. Telecommunication Systems 39(2), 137–143
(2008)
15. Seada, K., Helmy, A., Govindan, R.: On the Effect of Localization Errors on Geographic
Face Routing in Sensor Networks. In: IEEE/ ACM 3rd International Symposium on
Information Processing in Sensor Networks IPSN, Berkeley, California, USA, pp. 71–80
(2004)
16. Seada, K., Helmy, A.: Efficient Geocasting with Perfect Delivery in Wireless Networks.
In: IEEE Wireless Communications and Networking Conference WCNC, pp. 2551–2556
(2004)
17. Stojmenovic, I.: Geocasting with Guaranteed Delivery in Sensor Networks. IEEE Wireless
Communications 11(6), 29–37 (2004)
18. Khabazian, M., Ali, M., Mehmet, K.: A Performance Modelling of Connectivity in
Vehicular Ad Hoc Networks. IEEE Transactions on Vehicular Technology 57(4), 2440–
2450 (2008)
19. Harb, S.M., McNair, J.: Analytical Study of the Expected Number of Hops in Wireless Ad
Hoc Network. In: Li, Y., Huynh, D.T., Das, S.K., Du, D.-Z. (eds.) WASA 2008. LNCS,
vol. 5258, pp. 63–71. Springer, Heidelberg (2008)
20. Chen, Q.J., Kanhere, S.S., Hassan, M.: Modeling path length in wireless ad hoc network.
UNSW-CSE-TR-0617 (August 2006),
ftp://ftp.cse.unsw.edu.au/pub/doc/papers/UNSW0617.pdf
21. Bettstetter, C.: On the minimum node degree and connectivity of a wireless multi-hop
network. In: Proceedings of the ACM MobiHoc, pp. 80–91 (2002)
22. DeRango., F., Ha, A., Molinaro, A., Marano, S.: A Modified Location Aided Routing
Protocol for the Reduction of Control Overhead in Ad-hoc Wireless Networks. In: ICT
2003, vol. 2, pp. 1033–1037 (2003)
 Multi Objective Genetic Approach for Solving Vehicle
Routing Problem with Time Window

Padmabati Chand and J.R. Mohanty

School of Computer Engineering,

KIIT University, Bhubaneswar,
Odisha, India
{padma024,jnyana1}@gmail.com

Abstract. Vehicle routing problem with time window (VRPTW) is a NP-

Complete and a multi-objective problem. The problem involves optimizing a
fleet of vehicles that are to serve a number of customers from a central depot.
Each vehicle has limited capacity and each customer has a certain demand.
Genetic Algorithms maintain a population of solutions by means of a crossover
and mutation operators. For crossover and mutation, best cost route crossover
techniques and exchange mutation procedure is used respectively. In this paper,
we focus on three objectives of VRPTW i.e. number of vehicles, total
cost (distance), and time window violation (routing time). The proposed
Multi Objective Genetic Algorithm (MOGA) finds optimum solutions
effectively.

Keywords: Vehicle Routing Problem (VRP), Genetic Algorithm, Multi-

Objective Optimization, Best Cost Route Crossover, Exchange Mutation.

1 Introduction
Routing and scheduling problems are important elements of many logistics systems.
A lot of researches are devoted to finding effective methods to tackle these complex
combinatorial problems [1]. One of the major research topics has been the Vehicle
Routing Problem with Time Windows (VRPTW).The Vehicle Routing is a complex
combinatorial optimization problem. Vehicle routing problem as the efficient use of a
fleet of vehicles must make a number of stops to pick up and deliver passengers or
products. The term customer is used to denote the stops to pick up and deliver the
product. Every customer has to be assigned to exactly one vehicle in a specific order,
which is to be done with respect to the capacity in order to minimize the total cost.
The problem can be considered as a combination of the two well-know optimization
problems i.e. the Bin Packing Problem (BPP) and the Travelling Salesman Problem
(TSP). Relating this to the VRP, customers can be assigned to vehicles by solving
BPP and the order in which they are visited can be found by solving TSP. The rest of
the paper is organized as follows - section 2 gives a back ground study of the VRP,
section 3 gives the multi objective genetic search of VRP, section 4 describes on
experimental results.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 336–343, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Multi Objective Genetic Approach for Solving Vehicle Routing Problem 337

2 Background
The VRPTW can be defined as follows. Let G = (V, E) be a connected digraph
consisting of a set of n + 1 nodes, each of which can be reached only within a
specified time interval or time window, and a set E of arcs with non-negative weights
representing travel distances and associated travel times.Let one of the nodes be
designated as the depot.Each node i, except the depot, requests a service of size qi. In
the VRPTW all customers request the same type of service: pickups or deliveries.
Moreover, each customer must be serviced once, thus prohibiting split service and
multiple visits. The primary objective of the VRPTW is to find the minimum number
of tours; such that each node is serviced within its time window and that the
accumulated service up to any node does not violate the vehicle capacity [2]. The
tours correspond to feasible routes starting and ending at the depot. Often a secondary
objective is imposed to minimize the total distance traveled or to minimize total
schedule time. All problem parameters, such as customer demands, travel times and
time windows, are assumed to be known with certainty. The VRPTW is a basic
distribution management problem that can be used to model many real-world
problems [3]. Some of the most useful applications of the VRPTW include bank
deliveries, postal deliveries, industrial refuse collection, national franchise restaurant
services, school bus routing, security patrol services and JIT (just in time)
manufacturing [4]. The VRPTW has been the subject of intensive research efforts for
both heuristic and exact optimization approaches. The main objectives of VRPTW
are:
• Minimizing the total distances traveled by the vehicles.
• Minimizing the number of vehicles used.
• Minimizing the time window violation (routing time).

3 Multi-Objective Genetic Search for the VRPTW

The ability of genetic algorithm (GA) is to simultaneously search different regions of
a solution space which makes it possible to find a diverse set of solutions for difficult
problems with non-convex, discontinuous and multi-modal solutions spaces. Multi-
objective optimization, also known as multi-criteria optimization, is the process of
simultaneously optimizing two or more conflicting objectives subject to certain
constraints [5]. If a multi-objective problem is well formed, there should not be a
single solution that simultaneously minimizes each objective to its fullest. In each
case we are looking for a solution (for which each objective has been optimized) to
the extent that if we try to optimize it any further, then the other objective(s) will
suffer as a result. Finding such a solution and quantifying how much better this
solution (compared to other such solutions) is the goal of setting up and solving a
multi objective optimization problem [6].
In GA, each chromosome in the population pool is transformed into a cluster of
routes. The chromosomes are then subjected to an iterative evolutionary process until
a minimum possible number of route clusters is attained or the termination condition
is met [7]. The evolutionary part is carried out as in simple GA using selection,
crossover, and mutation operations on chromosomes. Tournament selection is used to
338 P. Chand and J.R. Mohanty

perform fitness-based selection of individuals for reproduction. The crossover

operator of GA may exploit structures of good solutions with respect to different
objectives to create new non-dominated solutions in unexplored parts of the Pareto
front. A crossover operator that ensures solutions generated through genetic
evolution are all feasible is also proposed. Hence, both checking of the constraints and
repair mechanism can be avoided, thus resulting in increased efficiency. In addition,
most Multi-Objective Genetic Algorithm (MOGA) do not require the user to
prioritize, scale, or weight objectives. Therefore, GA has been the most popular
heuristic approach to multi-objective design and optimization problems [8].

3.1 Chromosome Representation and Initial Population Creation

In our approach, a chromosome representing route of length N, where N is the

number of customers in a particular problem instance. A gene in a given chromosome
indicates the original node number assigned to a customer, while the sequence of
genes in the chromosome indicates the order of visitation of customers, ‘-’ indicates a
node representing a group of clustered customers that have already been committed to
a given vehicle [9]. Thus, the chromosome consists of integers, where new customers
are directly represented on a chromosome with their corresponding positive index
number and each committed customer is indirectly represented within one of the
groups (shown by a ‘–‘ mark ) representing a given deployed vehicle [10].

1485- 396 - 72

A route of 4 customers (1,4, 8,5) Route separator

Fig. 1. Chromosome Representation. This figure represents a route of three sub routes (1 4 8 5),
(3 9 6), (7 2). Each sub route is separated by ‘-‘symbol.

3.2 Fitness Evaluation

In this approach, we use dominance-information of the individuals of the population

by calculating for each individual i the number of alternatives Xi from which this
individual is dominated. For a population consisting of np (no. of population)
alternatives we get values of:
0 ≤ Xi ≤ np −1 (1)
Individuals that are not being dominated by others should receive a higher fitness
value than individuals that are being dominated.
We calculated the fitness value for each individual i by a linear normalization.
Individuals with the lowest values of Xi (X i = 0) receive the highest corresponding
value of Fi = Fmax.
Fi=Fmax - Xi (2)
 Multi Objective Genetic Approach for Solving Vehicle Routing Problem 339

3.3 Cross Over

Initial experiments using standard crossover operators such as Partially-Mapped-
Crossover (PMX) and uniform order crossover (UOC) yielded non-competitive
solutions. Hence, we utilized a problem-specific crossover operator that generates
feasible route schedules. An example of the procedure utilized by the proposed
crossover (Best Cost Route Crossover, BCRC) is given in Fig.2. According to Fig.2,
two parents A and B are selected from the population. A route from each parent
chromosome is randomly selected and the customer orders present in each route are
removed from the other parent. Since ‘–‘ marks represent existing vehicles, their
customers are left untouched. This means only integers which represent uncommitted
customers are reinserted into the current chromosome. Then the customers that have
been removed are reinserted at the location which minimizes the overall cost of the
entire tour. This requires computing the cost of inserting each of the remaining
customers at each location in the chromosome without constraint violation. If no
insertion location for a particular customer is found, a new route is created.

A
B

2436-1 9 - 5 7 8 879-316 - 2 5 4

2 43 6 1 9 5 7 8 3 1 6 2 5 4
step1 8 7 9

step2 1 9 2 5 4

87 3 6 2 5 4
step3 36 1 9 7 8

step4 2 5 4 1 9

step5 1
5

step6 87 3 6 2 5 4
36 1 9 7 8

step7 7 8 871 3 6 2 5 4
5 3 6 1 9

9
4
step8

5 36 871 3 6 9 25 4
step9 1 94 7 8

step10 2

step11 5 36 1 942 7 8

Fig. 2. Best Cost Route Cross Over

340 P. Chand and J.R. Mohanty

In the Fig. 2 A and B are two parents. In step 1 for parent A there is three routes
(r1:2 4 3 6 r2:1 9 r3:5 7 8) similarly in B (r1:8 7 9 r2:3 1 6 r3:2 5 4). In step 2 we
selected randomly a route from step 1 of parent A (r2:1 9) and in B (r3:2 5 4). In step 3
the route we selected from step 2 of parent A (1 9) that we removed from parent B of
the given routes in step 1. Similarly the route we selected from step 2 of parent B (2 5
4) that we remove from parent A of the given routes in step 1. In step 4 the route we
are deleted from parent A (2 5 4) and B (1 9) that we have again inserted. From the
route (2 5 4) we have randomly selected a route suppose 5 that we have inserted again
to the route (the route we got in step 3) by satisfying all the constraints i.e. vehicle
capacity and also after inserting there should be optimum solution means distance
should be minimum and minimum number of vehicles. The suitable location is
darkening in figure and also there is an arrow mark in step 6. After inserting, the new
route is (r1:5 3 6 r2:1 9 r3:7 8). Similarly we inserted 4 and 2. If any removed node is
not satisfying constraints then make a new route. Similarly in parent B from 1 and 9 we
have randomly selected a customer and that we have inserted by satisfying all the
constraints. If not satisfying make a new route. So finally the optimum solution we got
from parent A (r1:5 3 6 r2:1 9 4 2 r3:7 8) and B (r1:8 7 1 r2:3 6 9 r3:2 5 4) [11].

3.4 Mutation

Mutation is done by doing exchange mutation. Randomly we selected two genes from
two sub routes and exchange it. But after exchanging the new route should satisfy the
constraints [12]. Mutation is shown in the Fig. 3.

4 3 9 8 -5 2 -1 7 6

After exchanging 4 6 9 8- 5 2- 1 7 3

Fig. 3. Mutation. In this figure we randomly selected two genes from two different sub routes
(gene 3 from sub route (4 3 9 8) and gene 6 from sub route (1 7 6)). Then two genes exchanged.
After exchanging the new sub routes (4 6 9 8) and (1 7 3).

4 Computational Results
This section describes computational experiments carried out to investigate the
performance of the proposed GA. By our given fitness function, it minimizes both
number of vehicles and travel costs without bias. The algorithm was coded in mat lab
and run on an Intel Pentium IV 1.6 MHz PC with 512 MB memory and it gives
optimum result. We have applied MOGA to some instances from Solomon’s
benchmark set, it gives better result. Table 1 represents genetic algorithm (GA) parameters.
This table shows number of parameters used to implement genetic algorithm. Number
of times the program runs is 30. Cross over type is best cost route cross over (BCRC)
and crossover rate 80. Mutation type is exchange mutation and mutation rate is 0.10.
Table 2 represent experimental results of Solomon’s data set. This table represents
experimental results from some instance of Solomon’s data set. Like C201_100 requires
total no. of vehicles 3, total distance 897.24 and routing time 8456.29.
 Multi Objective Genetic Approach for Solving Vehicle Routing Problem 341

Table 1. Represents Genetic Algorithm Parameters

Parameter type Values

Initial Population 100
Number of runs 30
Crossover rate 80
Mutation rate 0.10
Selection type Tournament
Crossover type BCRC
Mutation type Exchange

Table 2. Represents Experimental Results of Solomon’s data set

Solomon’s data set Number of vehicles(NV) Distance(D) Routing Time

C201_100 3 897.24 8456.29
R101_50 10 1324.57 2032.79
R201_50 2 876.34 1234.63
R201_100 8 1289 3068.76

The following figures represent experimental results according to the above table (table
2).Fig 4 represents initial population of 100 solutions. Fig 5 represents pareto optimal front
(all the non-dominated solutions in a curve).Fig 6 and 7 based upon Solomon’s data set.
Both the figures represents route networks, where Fig.6 is the instance of R201_100
(number of vehicle=8, distance=1289, routing time=3068.76), Fig.7 is the instance of
C201_100 (number of vehicle=3, distance=897.24, routing time=12107.76).

INITIAL POPULATION
10

7
no.of routes------------>

0
0 500 1000 1500
distance--------->

Fig. 4. Represents initial population (total no. of solutions). This figure shows x-axis represents
total distance and y-axis represents total no. of routes per solution.
342 P. Chand and J.R. Mohanty

PARETO OPTIMAL FRONT

10

no of routs----> 7

0
0 500 1000 1500
distance---------->

Fig. 5. Pareto Optimal Front. This figure represents number of non-dominated solutions. The
above Figures Fig.4 represents initial population of 100 solutions. Fig.5 represents Pareto
optimal front (represents all non-dominated solutions).Fig.6 and Fig.7, represents route
networks of instance of R201_100 and C201_100 (Solomon’s data set).

Fig. 6. Route network for Instance R201_100

Fig. 7. Route network for instance C201_100

 Multi Objective Genetic Approach for Solving Vehicle Routing Problem 343

5 Conclusion
In this paper we presented a GA based approach for the static VRPTW. The
approach was tested using problem instances reported in the literature, derived
from publicly available Solomon’s benchmark data for VRP. The experimental
results showed that the GA approach was able to find high quality solutions. Our
future goal is to generate larger problem instances, and further evaluate the GA’s
performance on these problems by considering other objectives like number of
time window and speed of the vehicle.

References
1. Desrosier, J., Dumas, Y., Solomon, M.M., Soumis, F.: Time constraint routing and
scheduling. In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (eds.)
Handbooks in Operations Research and Management Science, vol. 8, pp. 35–139. Elsevier
Science Publishers, Amsterdam (1995)
2. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory of NP-
Completeness. W. H. Freeman and Company, New York (1979)
3. Braysy, O., Gendreau, M.: Vehicle routing problem with time windows, Part 1: Route
construction and local search algorithms, SINTEF Applied Mathematics Report,
Department of Optimization, Norway (2001)
4. Cordeau, J.F., Desaulniers, G., Desrosiers, J., Solomon, M.M., Soumis, F.: The VRP with
time windows. In: Toth, P., Vigo, D. (eds.) The Vehicle Routing Problem. SIAM
Monographs on Discrete Mathematics and Applications, ch. 7 (2001) (to appear)
5. Lenstra, J.K., Rinnooy Kan, A.H.G.: Complexity of vehicle routing problem with time
windows. Networks 11, 221–227 (1981)
6. Braysy, O., Gendreau, M.: Vehicle routing problem with time windows, Part II:
Metaheuristics, SINTEF Applied Mathematics Report, Department of Optimization,
Norway (2001)
7. Kohl, N.: Exact methods for time constrained routing and related scheduling problems,
PhD Thesis, Department of Mathematical Modeling, Technical University of Denmark
(1995)
8. Zhu, K.Q.: A diversity-controlling adaptive genetic algorithm for the vehicle routing
problem with time windows. In: Proceedings of the 15th IEEE International Conference on
Tools for Artificial Intelligence (ICTAI 2003), pp. 176–183 (November 2003)
9. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison Wesley, Reading (1989)
10. Liu, B., Zhao, R.: Random programming and fuzzy programming, pp. 15–57. Tsing Hua
University Publishing House, Beijing (1988)
11. Ombuki, B., Ross, J., Brian, H.F.: Multi Objective Genetic Algorithms for Vehicle
Routing Problems with Time Windows. Applied Intelligence 24, 17–30 (2006)
12. Haupt, L., Randy, H., Ellen, S.: Practical Genetic Algorithms, 2nd edn., pp. 43–44. A John
Wiley & Sons, Inc., Pennsylvania (2004)
 Application of Modified NSGA-II Algorithm to
Reactive Power Optimization

S. Ramesh1 , S. Kannan2 , and S. Baskar3

1
Arulmigu Kalasalingam College of Engineering, Krishnankoil, 626 126, India
2
Kalasalingam University, Anand Nagar, Krishnankoil, 626 126, India
3
Thiagarajar College of Engineering, Madurai, 625 015, India

Abstract. The optimal Reactive Power Dispatch (RPD) problem is a

nonlinear constrained optimization problem. This paper presents true
multi-objective solution set for multi-objective RPD problem. The ob-
jective functions are real power loss minimization and control variable
adjustment costs. In this paper, a Modified Non-Dominated Sorting Ge-
netic Algorithm version II (MNSGA-II) is proposed for solving RPD
problem. For maintaining good diversity in the performance of NSGA-
II, the concept of Dynamic Crowding Distance (DCD) is implemented in
NSGA-II algorithm and given name as MNSGA-II. The standard IEEE
30-bus and IEEE 118 bus test systems are used. The results obtained
using MNSGA-II are compared with NSGA-II and validated with refer-
ence Pareto front generated by conventional weighted sum method using
Covariance Matrix Adapted Evolution Strategy (CMA-ES). The perfor-
mance of NSGA-II and MNSGA-II are compared with respect to best,
mean, worst and standard deviation of multi-objective performance mea-
sures.The results show the effectiveness of MNSGA-II and confirm its
potential to solve the multi-objective RPD problem.

Keywords: Covariance Matrix Adapted Evolution Strategy, dynamic

crowding distance, Multi-objective, Non-dominated sorting genetic algo-
rithm, optimization, Reactive Power Dispatch.

1 Introduction
In general, RPD is a complex combinatorial optimization problem for a large
scale power system involving nonlinear functions having multiple local minima
and non-linear and discontinuous constraints. The aim of RPD problem is to
minimize the network real power loss and to reduce the cost of control variable
adjustments by regulating generator bus voltages, switching and changing trans-
former tap-settings [20]. With the increased loading of existing power transmis-
sion systems, the problem of voltage stability and voltage collapse, has become
a major concern in power system planning and operation [22,14,3].
To solve the RPD problem, a number of conventional optimization techniques
have been proposed. They include Gradient-based method, Non-linear program-
ming (NLP), Quadratic programming (QP), Linear programming (LP) and In-
terior point methods [7,18,9,4]. Several evolutionary computation techniques like

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 344–354, 2011.

c Springer-Verlag Berlin Heidelberg 2011
Application of Modified NSGA-II Algorithm to Reactive Power Optimization 345

Genetic algorithm (GA) [12], Evolutionary programming (EP) [23] and Swarm
intelligence [17] have been applied to solve the optimal RPD problems. However,
this approach only results in a single optimal solution where trade-off between
different components of the objective function must be fixed in advance of the
solution. In order to provide a means to assess trade-off between two conflicting
objectives, one may formulate the RPD problem as a multi-objective problem.
Recently, the RPD dispatch problem is formulated as a multi-objective opti-
mization problem. However, the problem is not treated as a true multi-objective
problem [8,11]. It was converted to a single objective problem by linear combi-
nation of different objectives as a weighted sum [8]. Unfortunately, this requires
multiple runs as many times as the number of desired Pareto-optimal solutions.
In contrast, Evolutionary Algorithms (EAs) can find multiple optimal solutions
in single simulation run due to their population approach [10,27]. Recently, some
successful application of EAs to optimal RPD have been reported in [1,27], where
minimizing voltage differences have been considered as an objective in addition
to loss minimization. A similar multi objective RPD problem can be found in
[25], where a fuzzy system is employed to adaptively adjust the parameters of
PSO, such as the inertia weight and learning factors, during the evolutionary
process. The NSGA-II algorithm has been demonstrated to be the most efficient
algorithm for multi-objective optimization on a number of benchmark problems
[6,5,27]. In addition, it has been shown that NSGA-II outperforms two other
contemporary multi-objective evolutionary algorithms: Pareto-archived evolu-
tion strategy (PAES) [16] and Strength-Pareto Evolutionary Algorithm (SPEA)
[26] in terms of finding a diverse set of solutions and in converging near the true
Pareto-optimal set. Recently the NSGA-II algorithm is used in various power
system problems such as economic dispatch [15] and Generation Expansion Plan-
ning [13]. But still NSGA-II algorithm suffers in maintaining diversity among
the solutions in the Pareto front. This paper makes use of diversity maintenance
strategy which is based on Dynamic Crowding Distance (DCD) introduced in
[19].
The objective of this paper is to solve the RPD problem as a multi-objective
optimization problem and solve it by using a modified NSGA-II (MNSGA-
II) algorithm by means of incorporating DCD. To validate the performance of
MNSGA-II and NSGA-II, conventional weighted sum method using Covariance
Matrix Adapted Evolution Strategy (CMA-ES) is used.
The paper is organized as follows: Section 2 describes the multi-objective RPD
problem formulations. Section 3 deals with MNSGA-II implementation to the
RPD problem. Section 4 provides test results, and Section 5 sums up.

2 Problem Formulation

The optimal RPD problem is to optimize the steady state performance of a

power system in terms of one or more objective functions while satisfying several
equality and inequality constraints. Generally the problem can be represented
as follows:
346 S. Ramesh, S. Kannan, and S. Baskar

2.1 Objective Functions

Real Power Loss (PL ). This objective is to minimize the real power loss in
transmission lines that can be expressed through the following equation

nl
f 1 = PL = gk [Vi2 + Vj2 − 2Vi Vj cos(δi − δj )] (1)
k=1

where,
nl is the number of transmission lines;
gk is the conductance of the kth line;
Vi ∠δi and Vj ∠δj are the voltages at end buses i and j of the k th line
respectively.

Cost of Control Variable Adjustment. The proposed objective is to min-

imise the control variable adjustments that can be expressed as
NC
  
 uk − u0k 
f2 = 
wk  max  (2)
u
k − umin 
k
k=1

where NC is the number of control variables; uk and u0k are the current and
initial settings of the k th control variable respectively; and wk is a weighting
factor to reflect the relative cost of the k th control variable. umax
k and umin
k are
th
the maximum and minimum limits of the k control variable.
The objective functions are subjected to following constraints
Equality Constraints:


0 = Pi − Vi Vj (Gij cos θij + Bij sin θij )i ∈ NB−1
j∈N
i (3)
0 = Qi − Vi Vj (Gij sin θij + Bij cos θij )i ∈ NP Q
j∈Ni

where,
Vi voltage magnitude at bus i.
Gij , Bij mutual conductance and susceptance between bus i and j
θij voltage angle difference between bus i and j
NB−1 total number of buses excluding slack bus
NP Q set of P Q buses
Ni number of buses
Inequality Constraints: In the control variables, the generator bus voltages i.e.
Automatic Voltage Regulator operating values (VG ) are taken as continuous
variable, where as the transformer tap settings (Tk ) is taken as discrete values.
The load bus voltages (VP Q ), reactive power generation (QG ) and Transmission
line loading (Sl ) are taken as state variables.
Continuous Control Variables:
VGmin
i
≤ VGi ≤ VGmax
i
; i ∈ NG (4)
Application of Modified NSGA-II Algorithm to Reactive Power Optimization 347

Discrete Control variables:

Tkmin ≤ Tk ≤ Tkmax ; k ∈ NT (5)

State Variables:

VPmin max
Qi ≤ VP Qi ≤ VP Qi ; i ∈ NP Q
min max
QG ≤ QG ≤ Q G ; i ∈ N G (6)
Sl ≤ Slmax

The objective function is calculated by using Newton-Raphson load flow calcu-

lation with the above mentioned equality and inequality constraints.

3 Implementation of MNSGA-II

The NSGA-II algorithm uses non-dominated sorting [6] method for ranking the
individuals. The simulated binary crossover and polynomial mutation generates
new offspring, and tournament selection is then used to select the population
for the next generation. Elitism addresses the problem of losing good solutions
during the optimization process due to random effects. One way to deal with
this problem is to combine the old population and the offspring, i.e., the mating
pool after variation, and to apply a deterministic selection procedure instead of
replacing the old population by the modified mating pool. In this paper, NSGA-
II algorithm with DCD (MNSGA-II) is used to solve the RPD problem.

3.1 Importance of MNSGA-II

It is of great importance for Pareto front with good convergence and diver-
sity characteristics with respect to true Pareto front. A good diversity can give
decision-makers more reasonable and efficient selections. Two aspects must be
considered in the diversity, one is the span of Pareto front, and the other is
the uniformity. The diversity maintenance strategy (DMS) is often realized in
the process of population maintenance, which use truncation operator to wipe
off individuals when the number of non-dominated solutions exceeds population
size. In this paper, a DMS which is based on dynamic crowding distance (DCD)
is discussed in [19].
In MOEAs, the horizontal diversity of Pareto front is very important. The
horizontal diversity is often realized by removing excess individuals in the non-
dominated set (NDS) when the number of non-dominated solutions exceeds pop-
ulation size. NSGA-II uses crowding distance (CD) measure as given in equation
(7) to remove excess individuals. The individuals having lower value of CD are
preferred to individuals with higher value of CD in removal process.
r
1   k k 

CDi = fi+1 − fi−1 (7)
r
k=1
348 S. Ramesh, S. Kannan, and S. Baskar

k
where, r is the number of objectives, fi+1 is the k th objective of the i + 1th
individual and fi−1 is the k objective of the (i−1)th individual after sorting the
k th

population according to the CD. The major drawback of CD is lack of uniform

diversity in the obtained non-dominated solutions. To overcome this problem,
dynamic crowding distance (DCD) method is recently suggested in [19].
In this approach, one individual with the lowest DCD value is removed every
time and recalculates DCD for the remaining individuals. The individuals DCD
are calculated as follows:
CD
DCD =   (8)
log V1i

where CDi is calculated by equation (7) and Vi is based on equation (9),

r
1   k k 
 2
Vi = fi+1 − fi−1 − CDi (9)
r
k=1

Vi is the variance of CDs of individuals which are neighbors of the ith individual.
Vi can give information about the difference variations of CD in different objec-
tives. Suppose the population size is N, the non-dominated set at tth generation
is Q(t) and its size is M. If M > N, DCD based strategy is used to remove
M − N individuals from non-dominated set. The summary of DCD algorithm is
given below:
Step 1: If |Q(t)| ≤ N then go to Step 5, else go to Step 2.
Step 2: Calculate individuals DCD in the Q(t) based on equation (8).
Step 3: Sort the non-dominated set, Q(t) based on DCD.
Step 4: Remove the individual which has the lowest DCD value in the Q(t).
Step 5: If |Q(t)| ≤ N, stop the population, otherwise go to Step 2 and continue.

3.2 MNSGA-II Algorithm

The following steps can be adopted for the implementation of MNSGA-II algo-
rithm.

Step 1: Identify the control variables like generator voltage limits, transformer
tap setting limits for RPD problem
Step 2: Select the parameters like number of population, maximum number of
generation, crossover and mutation probabilities.
Step 3: Generate initial population and evaluate the objective functions
(i.e., f1 , f2 ) for initial population.
Step 4: Set the generation count.
Step 5: Perform Simulated Binary crossover and polynomial mutation [5] for the
set of individuals.
Step 6: Perform non-dominated sorting. (i.e., sorting the population according
to each objective function value in ascending order of magnitude).
Step 7: Calculate DCD between the solutions based on the algorithm discussed
in Section 3.1.
Application of Modified NSGA-II Algorithm to Reactive Power Optimization 349

Step 8: Perform selection based on tournament selection [5] thereby a higher

fitness is assigned to individuals located on a sparsely populated part of
the front.
Step 9: Increment the generation count and repeat the steps from 6 to 9 until
the count reaches the maximum number of generation.

4 Test Results

The implementation was done using MATLAB version 7.4, on a IBM PC with
Intel core 2 quad processor having 1.86GHz speed and 3 GB RAM.

4.1 IEEE 30-Bus System

The representation of the IEEE 30 bus test system and the detailed data are
given in [18]. The system has 6 generators, 4 transformers; therefore, the number
of the optimized variables is 10 in this problem. For generator buses 1, 2, 5, 8,
11, and 13 the lower voltage magnitude limits are 0.95 p.u. and the upper limits
are 1.1 p.u., and for the remaining buses are 1.05 p.u.. The transformer tappings
are varied between 0.9 and 1.1 p.u. with the step size of 0.0001. The following
control parameters are considered in the RPD problem.

1. Automatic Voltage Regulator (AVR) operating values (continuous variable).

2. On-Load Tap Changer (OLTC) (discrete variable).
These control parameters are treated in load flow calculation as follows: AVR op-
erating values are treated as voltage specification values and OLTC tap positions
are treated as tap ratio to each tap position.

Generation of Reference Pareto-front. To compare the performance of

MNSGA-II and NSGA-II multiple run generated reference Pareto front is used.
One of the classical methods for getting Pareto-optimal set is Weighted Sum
method [[5], pp.50]. Using this method only one solution may be found along
the Pareto front in single simulation run. Recently, CMA-ES has been applied to
solve a non-linear multi-modal complex optimization problem successfully. One
of the major advantages of CMA-ES algorithm is its consistency [21]. Hence in
this paper CMA-ES is employed to obtain the Pareto front by using weighted
sum of objectives. The problem was treated as single objective optimization
problem by linear combination of objectives as follows.

M inimize C = w1 × f1 + w2 × f2 (10)

where w1 and w2 are weighting factors and the sum of weighting factors must
be 1. To get 30 non-dominated solutions the algorithm is applied 30 times with
varying weighting factors as a random number varying between 0 and 1. The
weight w1 is calculated by selecting random number between 0 and 1. The weight
w2 is calculated by using the expression w2 = 1−w1 , so that the sum of weighting
350 S. Ramesh, S. Kannan, and S. Baskar

factors is always 1. For each run, 50000 function evaluations are performed for
objective function evaluation.
In CMAES technique, the weights are randomly varied from 0 to 1 for each
trial. A population size of 100 is used. About 30 points are obtained to construct
the Pareto front through CMAES.

Parameter Selection. The results are sensitive to algorithm parameters. Hence,

it is required to perform repeated simulations to find suitable values for the pa-
rameters. Optimal parameter combinations for different methods are experimen-
tally determined by conducting a series of experiments with different parameter
settings before conducting actual runs to collect the results. The population size
is selected as 50 and generation size is selected as 500. The crossover probability
(Pc ) is selected between 0.8 and 0.95, in steps of 0.01 and for each Pc perfor-
mance is analyzed. It is found that Pc = 0.89, produces the best results. Other
parameters such as mutation probability (Pm ) crossover index (ηc ) and mutation
index (ηm ) are selected as 0.01, 5 and 20 respectively.

MNSGA-II Implementation. The NSGA-II and MNSGA-II algorithm are

applied to solve the RPD problem with 10 control variables. For analyzing
the performance of MNSGA-II algorithm, same parameters are used. The best
Pareto front obtained among 10 trial runs is represented in the Figure 1. The
representation of Pareto front has been compared with NSGA-II, MNSGA-II
and multiple run Pareto front in a two dimensional space.

Fig. 1. Comparison of MNSGA-II, NSGA-II and Reference Pareto-front

In this paper, the four metrics namely gamma, spread [5], minimum spacing
[2], and Inverted Generational Distance (IGD) [19] are considered. These metrics
were helpful for evaluating closeness to the reference pareto-optimal front and
also for evaluating diversity among Non-dominated solutions. The metrics are
calculated with 50 numbers of non-dominated solutions. For the IEEE 30 bus
system, 10 trial runs were conducted for NSGA-II and MNSGA-II. The extreme
Application of Modified NSGA-II Algorithm to Reactive Power Optimization 351

solutions of CMA-ES, NSGA-II and MNSGA-II can be seen from Table 1. Table
2 represents the statistical comparison of performance metrics for the test system
and the bold numbers indicate the best values. It can be seen from the Table
1 that almost all the mean values of performance metrics obtained by NSGA-II
with DCD is smaller than NSGA-II which means that NSGA-II with DCD can
improve convergence at the same time of the improvement of diversity.

4.2 IEEE 118-Bus System

The implementation of MNSGA-II for standard IEEE 118-bus system is considered
to study the convergence and diversity characteristics. The system has 54 genera-
tors, 186 branches, nine transformers (all are treated as OLTCs). The voltage set-
tings and tap settings are taken from [24]. Here, the total number of variables to be
optimized is 63. For the initial loading condition the real power loss is observed to
be 132.86 MW and the corresponding cost of control variable adjustment is zero.
Since the number of variables involved in 118-bus system is more than that of 30-
bus system, the numbers of iterations require more to achieve better convergence
characteristics of a Pareto front. Here, number of iterations used is 1800. The best
results obtained for 118 bus-system is shown in Table 3.
Table 1. Best Results for IEEE 30-bus system

Objective Function CMA-ES NSGA-II MNSGA-II

f1 in MW 4.7740 4.7450 4.7560
f2 0.0000 0.0068 0.0056

Table 2. Statistical comparison of performance measures for IEEE 30-bus system

Performance Measures NSGA-II NSGA-II+DCD

Gamma (γ) Mean 0.1077 0.1073
σ 0.0014 0.0004
Delta(Δ) Mean 1.2973 1.3141
σ 0.0011 0.0014
Minimum Mean 0.0780 0.0749
Spacing σ 0.0038 0.0002
Inverted Generational Mean 0.1794 0.1789
Distance(IGD) σ 0.0362 0.0138

Table 3. Best Results for IEEE 118-bus system

Objective Function CMA-ES NSGA-II MNSGA-II

f1 in MW 106.0000 122.5000 120.0000
f2 0.0000 0.0713 0.0669
352 S. Ramesh, S. Kannan, and S. Baskar

MNSGA-II (NSGA-II with DCD) algorithm is performing better for RPD

problem. By observing the Pareto fronts obtained from IEEE 30-bus and IEEE
118-bus system and from performance metrics it is concluded that the DCD has
convergence to the reference Pareto front and also better lateral diversity char-
acteristics among the set of solutions. The diversity characteristic of obtained
Pareto front from NSGA-II with DCD can help the decision maker to have more
choices in the selection of solution in a different operating condition of the sys-
tem. This justifies the importance of multi-objective formulation of the RPD
problem.

5 Conclusion
The Multi-objective RPD problem with conflicting objectives such as real power
loss minimization and cost of control variable adjustments minimization is solved
by MNSGA-II algorithm. For the first time, a true multi-objective solution set
is given in this paper. The concept of DCD was employed in NSGA-II algorithm
to improve convergence and maintain diversity among the set of solutions in
the Pareto front. The NSGA-II, NSGA-II with DCD generated Pareto fronts
are compared with reference Pareto front. With the help of statistical multi-
objective performance measures, it can be concluded that the performance of
NSGA-II with DCD is better than NSGA-II. The diversity characteristic of
obtained Pareto front from NSGA-II with DCD can help the decision maker
to have more choices in the selection of solution. The approach is quite flexible
so that other formulations using different objectives and/or a larger number of
objectives are possible.

Acknowledgement. The authors are thankful to the Department of Science

and Technology, New Delhi for its support through the project SR/S4/MS/350/06
to the second author.

References
1. Abido, M.A., Bakhashwain, J.M.: A Novel Multi-objective Evolutionary Algorithm
for Reactive Power Dispatch Problem. In: ICECS 2003, pp. 1054–1057 (2003)
2. Bandyopadhyay, S., Pal, S.K., Aruna, B.: Multi-objective GAs, quantitative indices
and pattern classification. IEEE Trans. Syst., Man, Cybern.-B 34(5), 2088–2099
(2004)
3. Belhadj, C., Mohamedi, R.: Voltage Stability Modelling and Real-Time Monitoring
Using Expert System for Operation Assistance. IEEE Trans. Power Syst. 11, 1037–
1045 (1996)
4. Da Costa, G.R.M.: Optimal reactive dispatch through primal-Dual method. IEEE
Trans. Power Syst. 12, 669–674 (1997)
5. Deb, K.: Multi-Objective Optimization using Evolutionary Algorithms. John Wiley
& Sons, Ltd., Singapore (2002)
6. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A Fast and Elitist Multi-objective
Genetic Algorithm: NSGA-II. IEEE Trans. Evol. Comp. 6, 182–197 (2002)
Application of Modified NSGA-II Algorithm to Reactive Power Optimization 353

7. Dommel, H.W., Tinny, W.F.: Optimal Power Flow Solutions. IEEE Trans. Power
App. Syst. PAS-87, 1866–1876 (1968)
8. Durairaj, S., Kannan, P.S., Devaraj, D.: Multi-Objective VAR dispatch using Par-
ticle Swarm Optimization. Int. Journ. Emerg. Elect. Power Syst. 4(1) (2005)
9. Granville, S.: Optimal reactive power dispatch through interior point methods.
IEEE Trans. Power Syst. 9(1), 136–146 (1994)
10. Gujarathi, A.M., Babu, B.V.: Multi-Objective Optimization of Industrial Styrene
Reactor: Adiabatic and Pseudo-Isothermal Operation. Chemical Engineering Sci-
ence 65(6), 2009–2026 (2010)
11. Hsaio, Y.T., Chaing, H.D., Liu, C.C., Chen, Y.L.: A Computer Package for Optimal
Multi-objective VAR Planning in Large Scale Power Systems. IEEE Trans. Power
Syst. 9, 668–676 (1994)
12. Iba, K.: Reactive power optimization by genetic algorithms. IEEE Trans. Power
Syst. 9(2), 685–692 (1994)
13. Kannan, S., Baskar, S., McCalley, J.D., Murugan, P.: Application of NSGA-II
Algorithm to Generation Expansion Planning. IEEE Trans. Power Syst. 24, 454–
461 (2009)
14. Kessel, P., Glavitsch, H.: Estimating the voltage stability of power systems. IEEE
Trans. Power Syst. 1(3), 346–354 (1986)
15. King, R.T.F.A., Rughooputh, H.C.S.: Elitist multiobjective evolutionary algorithm
for environmental/economic dispatch. In: The 2003 Cong. Evol. Comp., vol. 2, pp.
1108–1114 (2003)
16. Knowles, J., Corne, D.: The Pareto archived evolution strategy: A new baseline
algorithm for multiobjective optimization. In: Proc. of the 1999 Congress on Evo-
lutionary Computation, pp. 98–105 (1999)
17. Lai, L.L., Nieh, T.Y., Vujatovic, D., Ma, Y.N., Lu, Y.P., Yang, Y.W., Braun, H.:
Swarm Intelligence for Optimal Reactive Power Dispatch. In: IEEE / PES Trans.
Distrib. Conf. and Exhibition, Asia and Pacific, Dalian, China (2005)
18. Lee, K.Y., Park, Y.M., Ortiz, J.L.: A United Approach to Optimal Real and Re-
active Power Dispatch. IEEE Trans. Power App. Syst. PAS-104, 1147–1153 (1985)
19. Luo, B., Zheng, J., Xie, J., Wu, J.: Dynamic Crowding Distance - A New Di-
versity Maintenance Strategy for MOEAs. In: ICNC 2008, 4th Int. Conf. Natural
Computation, vol. 1, pp. 580–585 (2008)
20. Mamandur, K.R.C., Chenoweth, R.D.: Optimal Control of Reactive Power Flow
for improvements in voltage profiles and for real power loss minimization. IEEE
Trans. Power App. Syst. PAS-100, 3185–3194 (1981)
21. Manoharan, P.S., Kannan, P.S., Baskar, S., Willjuice Iruthayarajan, M., Dhanan-
jeyan, V.: Covariance matrix adapted evolution strategy algorithm-based solution
to dynamic economic dispatch problems. Taylor & Francis - Engg. Optim. 41(7),
635–657 (2009)
22. Reis, C., Barbosa, F.P.M.: A Comparison of voltage stability indices. In: IEEE
Mediterranean Electrotechnical Conference, MELECON 2006, pp. 1007–1010
(2006)
23. Wu, Q.H., Ma, J.T.: Power System Optimal Reactive Power Dispatch Using Evo-
lutionary Programming. IEEE Trans. Power Syst. 10, 1243–1248 (1995)
24. Yan, W., Lu, S., Yu, D.C.: A Novel Optimal Reactive Power Dispatch Method
Based on an Improved Hybrid Evolutionary Programming Technique. IEEE Trans.
Power Syst. 19, 913–918 (2004)
354 S. Ramesh, S. Kannan, and S. Baskar

25. Zhang, W., Yutian Liu, Y.: Multi-objective reactive power and voltage control
based on fuzzy optimization strategy and fuzzy adaptive particle swarm. Elect.
Power and Energy Syst. 30, 525–532 (2008)
26. Zitzler, E.: Evolutionary algorithms for Multi-objective optimization: Methods and
applications. Doctoral dissertation ETH 13398, Swiss Federal Institute of Technol-
ogy (ETH), Zurich, Switzerland (1999)
27. Zhihuan, L., Yinhong, L., Xianzhong, D.: Non-dominated sorting genetic
algorithm-II for robust multi-objective optimal reactive power dispatch. IET Gen.
Trans. and Dist. 4(9), 1000–1008 (2010)
 A Revised TESOR Algorithm for Perfect Routing

P.M. Joe Prathap1, G.R. Brindha2, and W. Vinil Dani3

1
IT Dept., RMD Engineering College,
R.S.M. Nagar, Kavaraipettai, Tiruvallur District, India
[email protected]
2
ICT Dept., School of Computing,
SASTRA University, Thanjavur, India
[email protected]
3
EEE Dept., St.Xavier’s Catholic College of Engineering,
Chunkankadai, Nagercoil
[email protected]

Abstract. In the modern world everything is faster. People are expecting

optimality in each and every instance. In this today’s era, shopping is an
inseparable & essential activity. But in reality it creates lot of stress and utilizes
not only our time but also energy. Recently cities are equipped with shopping
malls, but still confusion and unnecessary wandering are unavoidable. Already
we proposed a novel idea called TESOR (Time and Energy Saving Optimal
Routing) algorithm using shortest path algorithm as a base one. Now in this
article we further extended the TESOR based on number of adult customer,
which leads great time and energy saver one. With that Revised TESOR
almost 50% reduction can be achieved in terms of distance travel and spending
time. The simulation analysis shows the effectiveness of the algorithm with
proper analysis and comparative chart. The target domain of this paper is that of
providing 100% end user satisfaction.

Keywords: routing, shortest path, optimization, energy efficient, TESOR.

1 Introduction
After focusing on the development, now it is the time for optimization. So many
developments came to do efficient and faster work [1], [2], [3], but in some areas like
shopping in large malls, the delay is unavoidable one. The shopping mall phenomenon
came to existence by 1980’s. Earlier days we have shop in different places, i.e. for each
category we have to go different areas. We have to wander, to find the shops and then
the products. Now shops are grouped together and available in a single shopping mall.
Even though we have E-shopping we can’t get real satisfying shopping experience. So
real time shopping is an ever existing, inseparable concept in human life.
The malls try to attract customer with so many offers and services. But yet people
are facing some unnecessary delays and confusion which will further burden their
hectic schedule. The customer has to go here and there in order to do their shopping;
for same category of products, we may have several shops inside the malls and there
is a chance of non availability of shop for a particular category or some shops may

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 355–366, 2011.
© Springer-Verlag Berlin Heidelberg 2011
356 P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani

declare holiday for a particular day. Already we proposed an idea in our innovative
TESOR (Time and Energy Saving Optimal Routing) [4] algorithm, in which the
optimized path and shop selection can be done and the customer will get a printout
with the optimal routing graph. With this guidelines customer can shop all the items
with good focus. Especially customer can avoid closed shops as well as he/she can get
the assurance that the product is available in the shop in advance. Now, with revised
TESOR even shopping become easiest one, because here we are checking the number
of adult customers and with that value we are dividing the shopping area and allowing
parallel shopping.
The reason behind this application is the motivation by different researches and
their analysis. Some of them analyzed with dynamic programming algorithm [5] to
solve the problem which is time dependant. Another algorithm determining the fastest
travel time between two nodes based on a given starting time [6], [7], [8].
Subsequently, some authors proved that applying Dijkstra’s naming method to the
time-expanded network is guaranteed to return an optimal solution [9, [10], [11], [12],
[13], [14]. More number of authors have tackled different variation of the least cost
shortest path. These includes, multiple shortest paths in least cost [15] with single
source [16] and turn constraints [17]. And also point to point connection problem,
multicast trees with minimum cost are analyzed with TSP [18], [19], [20], [21], [22].
The rest of this paper is organized as follows. Section II analyzes the requirements
by taking an ideal shopping mall as an example. Section III describes the revised
TESOR algorithm and explains it with examples. Section IV demonstrates the
soundness of the proposed scheme with system architecture and the management
using revised TESOR. Section V depicts the simulation analysis with proper charts.
Conclusions and areas of future research are presented in Section VI.

2 Requirement Analysis

Fig. 1 shows an idle shopping mall with 40 shops. Each shop has its own category of
products. For example, sports shop may sell fitness equipments and sports
accessories. Especially some categories have more than one shop in the mall. For

Fig. 1. Idle shopping mall Fig. 2. Node Representation

 A Revised TESOR Algorithm for Perfect Routing 357

example, for jewellary we have 3 distinct shops in this mall itself. Moreover this idle
mall has two entries E1 and E2. Customers can choose any one of the entry based on
their willingness. To analyze further this idle shopping mall is converted into mesh
circuit as shown in Fig. 2. Here a unique shop id is assigned for each shop range from
S1 to S40. This id is used as an individual node name for our graph formation in
future. While formulating graph care should be taken about similar shops. To do
optimal shopping, end user need to choose only one shop from similar category.

3 Revised TESOR Algorithm

The existing TESOR algorithm has the following five major sections. Preliminary
section explores the initial setup required like entry selection, distance between nodes,
distance with entry, unique id assignment, etc. Second part tells about the customer
interaction like category entry. Third section concentrates on multiple graph
construction. Fourth part covers finding the shortest path in the graph. Fifth part focus
on displaying optimal path graph and distance measures to customer.
The revised TESOR include all the above sections along with some extensions.
Here from users we are getting the adult customer count. For example, a family
contains totally 5 members with 3 adults and also they want to do parallel shopping,
then they can enter the total number of adult as 2. So, one member can take care of
children. Other two can do shopping without any tension. Obviously they can reduce
the shopping time considerable amount.
REVISED TESOR ALGORITHM:
Begin
Initialize Total No. of Adult Customer  AC
If AC= 1
TESOR();
Else
OverallShopArea  CONSTANT
Division of shops based on AC
No. of Dividend area = OverallShopArea / AC;
//Entry pt based division
For i= 1 to No.of Dividend area
TESOR();
Display graph:
No. of display graphs  ∏
If No. of display graphs >1
For i=1 to No. of display graphs
Compare the shop category for duplicity
If duplicity = TRUE
Construct totally ∏ possible graphs
End for
For i= 1 to No. of Dividend area
For i=1 to nodes in G
For k := 1 to n in G
For j := 1 to n in G
path[i][j] = min ( path[i][j],
358 P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani

path[i][k]+path[k][j] )
OD[i]=∑ /parts rep. total dividend area
at an instance.
From OD[i] sort the values (Ascending)
According to the minimum dist. Display the
shortest cycle graph
TESOR()
Begin
PRELIMINARIES
For i = 1 to No.of shops
ASSIGN
Unique_ID= Si;
Distance with Entry_Counters;
Intershop distance;
Shop Holiday details;
End for
If Entry_Counter =E1
Origin_Node = E1;
Else
Origin_Node = E2;
End If.
Continue ;
SHOP SELECTION PROCESS:
GET list of categories from Customers
List of Categories  List_of_Cat.
For i=1 to List_of_Cat
CHECK availability of shops
If Shop_Availability = TRUE and Shop_Holiday = FALSE
Shop_ID  G (Vi,E)
Else
Pass status information to Customers
STATUS FALSE
End for
GRAPH CONSTRUCTION:
Group the nodes which are same category using Si’s and
name them as C1, C2,..
Mark shops with single category as permanent node
P1,P2…
REPEAT
Construct Graph Gi=(Vi,Ei)
Include V’s all P1, P2,…
Include V’s from C1,C2,….. one at a time
Until TotalGraphs  ∏
End Repeat
SHORTEST CYCLE:
Initial node  entry Counter
For i=1 to nodes in G
 A Revised TESOR Algorithm for Perfect Routing 359

For k := 1 to n in G
For j := 1 to n in G
path[i][j] = min ( path[i][j], path[i][k]+path
[k][j] )
Return the values
OPTIMAL PATH DISPLAY:
According to the minimum distance, display the shortest
cycle graph G (V,E)
END TESOR
The algorithm was explained further by taking one real time scenario. Let us assume a
family wants to visit thirteen different shops listed below in Table 1. The family has
two adult members, and they want to do parallel shopping. Assume that preliminary
setup already done as per our algorithm with AC=2. Assume that the entry point is
selected as E1.

Table 1. Shopping item category list

Sl. No Category Customer-1 Customer-2

1 Leather S2,S4 S39
2 Musicals S1,S12 NA
3 Bakery S3 S3
4 Furniture NA S33
5 Sports S7 NA
6 Laptops NA S35
7 Plastics NA S38
8 Cosmetics S17 NA
9 Shoes NA S31, S40
10 Men's Wear NA S25
11 Mobiles S19 NA
12 Glass vessels NA S27
13 Pet Shop NA NA

Since AC value is 2, the area of the shopping mall should be divided into
approximately two half’s. The division is always based on entry point only. (i.e.
From E1 division takes place. This is applicable for all kind AC values.) Now the
divided section renamed as customer 1 and customer 2. The required shops for
shopping also depicted in Table 1. From this table we got a clear picture about
duplication. See for example, we have three distinct shops for leather (S1, S4 & S39).
Our ultimate aim is the customer should avoid duplication and anyone of the customer
can buy a single category. That is we need to avoid intra group duplication as well as
inter group duplication. Note that some category may not be available in the shopping
mall. In our example pet shop is not available. In this case this information should be
passed to the customer. Moreover the holiday details about the shops also taken into
account. So that if a shop is on holiday for that particular day, the node is
automatically removed from the proposed graph formation.
360 P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani

With the revised TESOR algorithm, after preliminary setup, the TESOR function
is called separately for customer 1 and customer 2. The details are shown in Table 2
and Table 3 respectively.

Table 2. Customer 1 possible paths Table 3. Customer 2 possible paths

Gra Possible Path cycle Total Gra Possible Path cycle Total
ph dista ph distan
nce ce
G1 E1ÆS3ÆS2ÆS1ÆS7ÆS1 57 G5 E1ÆS33ÆS24ÆS25ÆS3 70
7ÆS19ÆE1 9ÆS38ÆS40ÆS35ÆE1
G2 E1ÆS12ÆS3ÆS2ÆS7ÆS 53
17ÆS19ÆE1 G6 E1ÆS33ÆS24ÆS31ÆS2 58
5ÆS39ÆS38ÆS35ÆE1
G3 E1ÆS4ÆS3ÆS1ÆS7ÆS1 59
7ÆS19ÆE1 G7 E1ÆS33ÆS24ÆS25ÆS3 58
G4 E1ÆS12ÆS4ÆS3ÆS7ÆS 52 1ÆS38ÆS39ÆS35ÆE1
17ÆS19ÆE1

From Table 2 G4 graph is taken as optimal one which has lowest distance value,
whereas in customer 2 we have G6 and G7 with equal value 58. Arbitrarily anyone of
the graph will be selected if we have equal values. In this example we pick G6.
Now, we have two optimal graphs G4 and G6. But here again we have to check for
duplication among inter graphs. That is, G4 and G6 may have common category. For
example (S4, S39) & (S3,S24) are belongs to same category. Now the next step in the
algorithm is to find all possible combinations by including anyone of the shop at any
one side of the customer without any repetition. After finding the combinations we
have to sort the values by adding the lengths of various instances. These details are
shown in Table 4.

Table 4. Possible instance details with total distance

C1 C2
Ins Le Le
tan ngt ngt Tot
ce Path cycle h Path cycle h al
E1-S33-S24-S31-S25-S39-S38-S35-
A E1-S12-S19-S17-S7-E1 46 E1 63 109
B E1-S12-S4-S7-S17-S19-E1 50 E1-S33-S24-S25-S31-S38-S35-E1 53 103
E1-S33-S24-S31-S25-S38-S35-E1 60 110
C E1-S12-S3-S7-S17-S19-E1 51 E1-S33-S35-S31-S25-S39-S38-E1 68 119
E1-S33-S35-S25-S31-S38-S39-E1 63 114
E1-S12-S4-S3-S7-S17-S19-
D E1 52 E1-S33-S35-S31-S25-S38-E1 59 111
E1-S33-S38-S25-S31-S38-E1 53 105
 A Revised TESOR Algorithm for Perfect Routing 361

From the Table 4, instance B become the best optimal path for both customer 1
and customer 2. Clearly by doing parallel shopping, overall distance travelled by an
individual customer is almost reduced half. And the overall time spent in the
shopping mall too reduced considerable amount of time.
Fig. 3 represents the graphical representation of the two final optimized distinct
graphs. Suppose if we use ordinary TESOR for the same example (i.e. AC=1), we
will get the following solution. These shops are needed to be visited. {[39,S4,S2],
[S1,S12], [S3,S24], [S40,S31], S7, S17, S19, S25, S33, S35, S38} the same category
of shops indicated within brackets. As per our TESOR formula ∏ we will get
(3x2x2x2) = 24 combinations [21]. Further, in some combinations, the distance from
one node to other nodes equal. In that case again we have to check inner combinations
also. With that we got overall 40 distinct combinations. The optimal path cycle is
given below. E1-S12-S4-S3-S7-S17-S19-S25-S31-S38-S33-S35-E1.

Fig. 3. Optimized graph for two customers

4 System Architecture and Management

Fig.4 depicts the overall system architecture of the TESOR system. All the shops in
the malls are connected with the database content server. GUI for end user is

Fig. 4. The system architecture of TESOR Fig. 5. Functional structure of TESOR

362 P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani

available in the entry area. This end user interface has the direct contact with the
content server. So all the shop related updates are easily accessible by the end user.
Information like holidays, non availability of certain product, discount details can be
delivered to the end user easily.
Fig. 5 explores the functional structure of our TESOR software. The revised
TESOR state transition diagram was depicted in Fig.6.

Fig. 6. Revised TESOR state transition diagram

5 Simulation Analysis
A simulation environment was developed with the following setup. The algorithm
was trained up to 50 customers with shopping category range from 5 to 30. For every
trial different values for AC taken ranges from 1 to 4. With each scenario various
graph combinations and shortest length were analyzed. Especially for each case with
TESOR and without TESOR path selection was performed and the performances
were plotted further.

Fig. 7. Revised TESOR GUI Fig. 8. Revised TESOR Simulation

 A Revised TESOR Algorithm for Perfect Routing 363

Fig. 7 depicts the GUI for revised TESOR algorithm. Here the total number of
adult customer will be entered. Based on that, the page is redirected to the simulation
setup.
Fig. 8 shows the GUI for the simulation environment. Based on the values
given by the user the optimal graphs will be generated as Fig. 9 by running revised
TESOR.

Fig. 9. Revised TESOR sample route sheet

Fig. 10. Overall time utilization. Fig. 11. Time utilization for different AC

Fig. 10 shows the comparative analysis of overall time utilized by the customer
through TESOR algorithm, revised TESOR algorithm and with the regular shopping
for the same scenario. From the below graph we can clearly understand the
importance and advantages of our revised TESOR algorithm.
Fig. 11 compare the revised TESOR time utilization for various values of AC. If
AC values increases automatically number of parallel shopping also increased. This
will lead a great time minimization as clearly depicted in Fig.11.
364 P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani

Fig. 12. Performance of Revised TESOR

Simulation work suggests that with our algorithm, the time spent inside the malls
by a customer reduce approximately half way. This method helpful for people who
are in hurry to do all the stuff( that is for fast moving people and modern world).
Fig. 12 shows the enhancment achieved in the performance of the shooping mall
after adopting the proposed revised TESOR algorithm. With that the number of
customer and popularity are increased drastically. Overall profit and customer
satisfaction achiedved thus made a great impact in the society.
It also explores the various future enhancements that can be applied along with our
proposed work.

6 Conclusion and Future Enhancement

In this paper, we have projected a innovative algorithm for finding an optimal
solution based on shortest path problem with a specially-structured, for shop
selection, graph construction and optimal path display function. We have shown that
our previous algorithm TESOR is for AC with the value 1, but in revised TESOR
further it si improved with AC values 2 or 3. We have also compared the difference
and performance of both the algorithm.
Shopping malls also faced lot of competitions nowadays. To sustain its name and
their customer lot of efforts were taken by them. Our proposed work is major break
breakthrough for them if they applied it successfully. Our product not only provides
better solution but also open further enhancements. This will play a small but better
part for developing faster world. The shops can provide the offer pamphlets or
information during the selection of their shop. So that that the customer can select the
shops based on offers. This will create a competition environment between the service
providers, from that the customer is benefited and service provider also increase their
profit. The Graph can also include information about toilets, play area, water car
parking, ATM and first aid area indication.
 A Revised TESOR Algorithm for Perfect Routing 365

References

1. Leng, J.F., Zeng, W.: An Improved Shortest Path Algorithm for Computing One-to-One
Shortest Paths on Road Networks. In: Fac. of Inf. Eng., China Univ. of Geosci., April 26.
IEEE, Wuhan (2010)
2. Kaplan, H., Lewenstein, H., Shafrir, N., Sviridenko, M.: Approximation Algorithms for
Asymmetric TSP by Decomposing Directed Regular Multigraphs. In: Proc. 44th on
Foundations of Comput. Sci., pp. 56–65. IEEE, Los Alamitos (2004)
3. Retvari, G., Biro, J., Cinkler, T.: On Shortest Path Representation. IEEE/ACM
Transactions on Networking 15(6), 1293 (2007)
4. Joe Prathap, P.M., Brindha, G.R.: An Innovative TESOR Algorithm for Perfect Routing.
In: International Conference on Signal Processing, Communication, Computing and
Network Technologies, NCACNSP (2011) (accepted for publication)
5. Cooke, K.L., Halsey, E.: The shortest route through a network with time-dependent
internodal transit times. Journal of Mathematical Analysis and Applications 14(3), 493–
498 (1966)
6. Zhu, S., Huang, G.M.: A new parallel and distributed shortest path algorithm for
hierarchically clustered data networks. IEEE Transactions Parallel and Distributed
Systems 9(8) (September 1998)
7. Huang, G., Zhu, S.: A new distributed shortest path algorithm for hierarchically clustered
data networks. In: Proceedings American Control Conference, 1995, June 21-23, vol. 3.
Dept. of Electr. Eng., Texas A&M Univ., College Station, TX, IEEE (1995)
8. Orda, A., Rom, R.: Shortest-path and minimum-delay algorithms in networks with time-
dependent edge-length. Journal of the Association for Computing Machinery 37(3), 607–
625 (1990)
9. Soltani, A.R., Tawfik, H., Goulermas, J.Y., Fernando, T.: Path planning in construction
sites: performance evaluation of the Dijkstra, Ap, and GA search algorithms. Advanced
Engineering Informatics, 291–303 (May 29, 2003)
10. J.M.: A walk over the shortest path: Dijkstra’s Algorithm viewed as fixed-point
computation. Information Processing Letters 77, 197–200 (2001)
11. Jeffrey Solka, L., James Perry, C., Brian Poellinger, R., George Rogers, W.: Fast
computation of optimal paths using a parallel Dijkstra algorithm with embedded
constraints. Neurocomputing, 195–212 (February 10, 1994)
12. Cantone, D., Faro, S.: Two-Levels-Greedy: a generalization of Dijkstra’s shortest path
algorithm. Electronic Notes in Discrete Mathematics 17, 81–86 (2004)
13. Sedgewick, R., Vitter, J.: Shortest paths in Euclidean graphs. Algorithmica 1(1), 31–48
(1986)
14. Sherali, H., Ozbay, K., Subramanian, S.: The time-dependent shortest pair of disjoint paths
problem: complexity, models, and algorithms. Networks 31, 259–272 (1998)
15. Rees, W.G.: Least-cost paths in mountainous terrain, Scott Polar Research Institute,
University of Cambridge, Lensfield Road, Cambridge CB2 1ER, UK. Computers &
Geosciences 30, 203–209 (2004)
16. Trtiff, J.L.: An experimental comparison of two distributed single-source shortest path
algorithms. Parallel Computing, 1505–1532 (April 1995); Elsevier Science
17. Boroujerdi, A., Uhlmann, J.: An efficient algorithm for computing least cost paths with
turn constraints. Information Processing Letters 67, 317–321 (1998); Elsevier Science
(November 1997)
366 P.M. Joe Prathap, G.R. Brindha, and W. Vinil Dani

18. Natu, M., Fang, S.-C.: The point-to-point connection problem - analysis and Algorithms,
Operations Research und industrial Engineering. Discrete Applied Mathematics 78, 207–
226 (1997); Elsevier Science B.V. (December 1996)
19. Li, S., Melhem, R., Znati, T.: An efficient algorithm for constructing delay bounded
minimum cost multicast trees. J. Parallel Distrib. Comput. 64, 1399–1413 (2004)
20. Kosaraju, S.R., Park, J.K., Stein, C.: Long tours and short superstrings. In: Proc. 35th Ann.
Symp. on Foundations of Comput. Sci., pp. 166–177. IEEE Computer Society, Los
Alamitos (1994)
21. Bellman, R.: Combinatorial Processes and Dynamic Programming. In: Bellman, R., Hall
Jr, M. (eds.) Proceedings of Symposia in Applied Mathematics Combinatorial Analysis,
pp. 217–249. American Mathematical Society, Providence (2004)
22. Kohn, S., Gottlieb, A., Kohn, M.: A Generating Function Approach to the Traveling
Salesman Problem. In: ACM Annual Conference, pp. 294–300. ACM Press, New York
(1997)
 An Accomplishment of New Level Register File
Management Architecture for Montgomery
Algorithm in Elliptic Curve

M. Prabu1 and R. Shanmugalakshmi2

1
Anna University Coimbatore,
Tamil Nadu,
India
2
CSE,
Government College of Technology,
Tamil Nadu, India
{prabu_pdas,shanmuga_lakshmi}@yahoo.co.in

Abstract. In this paper, we present a background on Elliptic Curve

Cryptography, along with the Montgomery operation, which plays major role
on Elliptic Curve Cryptosystem and side channel attacks on EC-Montgomery
operations. We have also provided a brief background on File Register
architecture and proposed a new design level of EC-Montgomery addition
algorithm. These design level architecture are designed in an efficient manner.
Realizing the proposed design level of FPGA technology. Through this FPGA
based technology very advantageous results were found when compared against
other existing designs. Even though our design bears an extra computation time,
it reduces the area space or register gate area space due to reassigning the
design and power analysis attack.

Keywords: Montgomery Algorithm, Elliptic Curve Cryptosystem, Computation

time, FPGA technology.

1 Introduction

In this article, a hardware implementation on register file architecture has been

efficiently compared with two more algorithms namely Montgomery algorithm and
modified Montgomery algorithm. And we have focused on Side Channel Attacks and
its countermeasures on EC-Montgomery algorithm. Side Channel Attacks consists of
a lot of attacks, here we mainly concentrated and worked on Power Analysis and
Fault Analysis attacks[1,2]. Both the Power analysis and Fault Analysis can be
worked based on logical designs. Many proposed implementations of EC- system are
based on Montgomery algorithms.
Here we took a design level architecture as a major part and implemented both
there algorithms (Montgomery and modified Montgomery). With the help of this

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 367–375, 2011.
© Springer-Verlag Berlin Heidelberg 2011
368 M. Prabu and R. Shanmugalakshmi

design we got the quality rich design for the Montgomery algorithm, which is based
on the countermeasures of Power Analysis and Fault Analysis.

1.1 Organization of This Paper

The rest of this paper is structured as follows. In Section 2, we recall the basic of the
Elliptic curve cryptosystems, Montgomery Algorithm and Side Channel Attack. In
Section 3, we introduce a Montgomery algorithm. In Section 4, the model approach
is presented to performance of operation, and then we discuss its security and
implementation properties. In Section 5 & 6, we turn our attention on File Register
Architecture and design level. In Section 7, we make concluding remarks.

2 Background

2.1 Elliptic Curve Cryptography

ECC is a newly emerged public key cryptographic algorithm. It is younger than RSA
and other cryptographic algorithms. Although, as Hankerson, Menezes and Vanstone
(2004) state, Elliptic curve have been focused by researchers and algebraists in the
middle of nineteenth century. In 1985, the first elliptic curve cryptography application
has been discovered by Neal Koblitz and Victor Miller. The main advantage of ECC
over RSA is its key size. For instance, it is widely accepted that 160-bit ECC offers
equivalent security as 1024-bit RSA. This significant difference makes ECC
especially attractive for applications on constrained environments as shorter key sizes
are translated to less power and storage requirements, and reduced computing times.
The level of security provided by ECC increases faster with key size. ECC can also
provide a faster implementation than RSA and DL systems, and utilize power and
reduced bandwidth. In the last few years, confidence in the security of ECC had also
arised, to the point where they had now been included in internationally recognized
standards (especially IEEE std 1363, ANSI X9.62).

2.2 Side Channel Attacks

In recent years, cryptosystems have come under harass from various forms of side
channel attacks. Kocher et al[4] discovered that cryptosystems implementations leak
information, which can help an attacker to access secret data. One such technique for
retrieving secret information is SPA. SPA involves monitoring the power
consumption of cryptographic algorithms in single execution.
When a cryptographic algorithm is actually implemented, its weakness can be
induced with some unanticipated ways. The attackers can track these weaknesses to
circumvent the security of the underlying algorithms. These kind of attacks are
referred and reviewed by Kelsey[9], as so called side channel attacks.
In fact, hardware encryption devices are ideal targets for side channel attacks.
There attacks are divided into passive attacks (power consumption analysis,
electromagnetic leaking analysis) and active attacks.
 An Accomplishment of New Level Register File Management Architecture 369

2.2.1. Power Analysis

The Power analysis can be classified into two parts, that is Simple Power
Analysis(SPA) and Differential Power Analysis(DPA).

Simple Power Analysis

Simple Power Analysis is generally based on looking at the visual representation of
the power consumption unit while an encryption operation is being performed. SPA is
a technique that involves direct interpretation of power consumption measurements
collected during various cryptographic operations in[9,10]. SPA can yield information
about a device’s operation as well as key material.
Differential Power Analysis
The goal of the Differential Power Analysis is to identify the secret key, or part of the
secret key, for ciphering the information. The basic idea is to correlate the encrypted
data and the power consumed by the device.

2.3 EC-Montgomery Algorithm

Elliptic curve is one of the strongest public-key cryptosystem, which is generally used
for authentication The performance of such cryptosystems are primarily determined
by effectively implementing modular multiplication and exponentiation. For
performance as well as for physical security reasons, it is often advantageous to
recognize them by using hardware. Hardware implementations of Elliptic curve
cryptosystems are widely studied.
Montgomery algorithm developed a very efficient technique to compute in the
group associated to an elliptic curve over a non-binary finite field in which
prime field is The Montgomery algorithm proves useful for point compression in
ECC. More precisely, instead of sending a point as part of some cryptographic
protocol, there are variety of ways to perform the Elliptic Curve based Montgomery
operations. Our aim, in this thesis is to give a crystal clear view of its operations and
functions. The operations may be classified as Montgomery addition and
Montgomery multiplication. A number of Montgomery multiplier, have been
suggested [12, 13] and their hardware architectures are designed to deal with a
maximum number of registers. However, when the computer processing power
increases, it provides the same level security. Then the cryptosystem will verify the
level of register management to be equalized.

3 Previous Work
In this section we introduce Yong Ki.L’s and Ingrid.V’s Montgomery compact scalar
multiplication algorithm, which uses a common projective coordinate system [2] .The
algorithm, is shown in Fig 1. A non-super singular elliptic curve over 2 is a
set of coordinate points , satisfying 2 3 2 with the point at
infinity O, where , , , 2 and 0.
370 M. Prabu and R. Shan
nmugalakshmi

Input : A point , and a positive integer 2 2

Output:
If 0 0 thenn output 0,0 and stop
, 1, ,
,
For 2 0
If then

( , ) ( , , , ) (1)

, (2)
return ( , , , )

Fig. 1. Monttgomery Addition with Yong –Ingrid algorithm

The adding formula of is definedd in

Eq.1

(3)

(4)

is the conversion of projective coordinate to afffine

coordinate. Yong-Ingrid adding algorithm is described in Fig 2. Where .
The total number of regiisters in Fig 2. is six , i.e[6], the registers for , , ,
, and . The tootal number of operations of Adding algorithms are 2
multiplications, 1 square, an
nd 2 additions.
Adding Algorithm:

, , , ,

Fig. 2.. Yong Ki L- Ingrid. V Adding Algorithm

Our proposed design of Montgomery Addition algorithm’s implementattion

table shows six level regisster used for making two operations. This 6 level regiister
concept is already done in[4].
i But in our approach same level of registers are
used in calculation part butt there is a reduction in the implementation steps shownn in
Fig 7.
 An Accomplishment of New Level Register File Management Architecture 371

Table 1. Register File Management for Adding Algorithm

Steps Field Operation Reg Reg Reg Reg Reg Reg6

1 2 3 4 5
1 Initial X1 X2 Z
2 Shift X1 X2 Z
3 Copy X1 X1 X2 Z
4 Add T2x1+x2 T2 X1 X2 Z
5 Shift T2 X1 X2 Z
6 Copy T2 T2 X1 X2 Z
7 Shift T2 T2 X1 X2 Z
8 Multiply T2T22 T2 X1 X2 Z
9 Shift*4 X1 X2 Z T2
10 Multiply T1X1xX2 T1 X2 Z T2
11 Switch T1 X2 Z T2
12 Load x X1x X1 T1 X2 Z T2
13 Shift*2 T2 X1 T1 X2 Z
14 Switch T2 X1 T1 X2 Z
15 Multiply X1T2xX1 X1 T2 T1 X2 Z
16 Switch T2 X1 T1 X2 Z
17 Shift*5 X1 T1 X2 Z T2
18 Add X1X1+T1 X1 T1 X2 Z T2

Fig. 3. Initial level of Design

Fig. 4. Second Level of Design

372 M. Prabu and R. Shanmugalakshmi

The following two steps can be explained through the hardware design[3]. The
inputs are X1 & X2 for the both multiplier and adder. After Multiplier the output (T2)
can be extracted by the Square concept.
After completing the level 1 design, the second level is extracted by the level 1
design’s output that is (T2)2 .After getting the result from the level one, it interacted to
the input again X1(x), which is given by the user[7,8]. After that process, the resulting
value will be stored in X1.

Fig. 5. Final Designing Section for Montgomery Algorithm

The final level. The latest X1 value is interacted with the adder with T1.

4 Performance Operation
In this article, we present a , modified version of file management algorithm, which is
suitable for both hardware and software implementations. The shift operation shifts
the register values in circular and the switch operation switches the values of Reg1

Fig. 6. Performance of operations

 An Accomplishment of New Level Register File Management Architecture 373

and Reg2. The copy operation replaces the value of Reg1 with Reg2. Note that the
copy operation is required for field square process which is implemented as the field
multiplication with two operands of the same value.

5 Algorithm for File Register Architecture

5.1 Proposed Algorithm
Initially declare T1 X1.Z2,T2Z1.X2
Step 1: MT1+T2
Step 2:Z1M2
Step 3: NT1.T2
Step 4: Mx.Z2
Step 5:X1M+N

In our algorithm, we have to increase the temporary registers such as T1, T2, M and
N. But our main role is to reduce the number of registers. For this problem,in our
proposed work we proposed a new operation swap. It will help to swap the values from
one register to another register while it will become empty. With the help of swap
operation[5] , we had easily reduce the no registers as well as steps of execution.

5.2 Our File Register Architecture

Table 2. Register File Management for Our Modified Adding Algorithm

Steps Field Operation Reg Reg Reg Reg Reg Reg6

1 2 3 4 5
1 Initial X1 X2 Z
2 Shift X1 X2 Z
3 copy X1 X1 X2 Z
4 Add T2 X1+ T2 X1 X2 Z
X2
5 shift T2 X1 X2 Z
6 copy T2 T2 X1 X2 Z
7 Shift T2 T2 X1 X2 Z
8 T2 T22 T2 X1 X2 Z
9 Shift*4 X1 X2 Z T2
10 T1 X1x T1 X2 Z T2
X2
11 Shift T2 T1 X2 Z
12 swap T2 T1 X2 Z
13 X1x. T2 X2 T2 T1 X2 Z
14 Shift X1 T2 T1 X2 Z
15 swap X1 T2 T1 X2 z
16 X1X+ T2 X1 T1 T2 X2 z
374 M. Prabu and R. Shanmugalakshmi

6 Design Level Architecture

6.1 Proposed Design Level

Fig. 7. Modified Design Level

The above mentioned design level, consist of eight registers in that 4

registers(T1,T2,M and N) were temporary. The modified process can be explained
through the hardware design. The inputs are X1, X2, Z1,Z2 for the multiplier and
adder.

7 Discussion and Conclusion

Table 3. Comparison of Computational workload

Field Operation Yong Ki.L and Ingrid.V Our algorithm

algorithm
Addition 3 2
Multiplication 2 2
Square 1 1

In order to verify the design and the algorithm, we synthesized the proposed
architecture and design using FPGA based Xilinx, Virtex-III platform. Through the
Xilinx we designed our design architecture Fig.7[11]. After verifying the nature of
algorithm, the computational load can be calculated and measured. In comparison
with the known results of the algorithm more suitable for real-time applications.
Reasons are optimistic about the results of our comparative study for all researchers
on this important topic.
 An Accomplishment of New Level Register File Management Architecture 375

References
1. Cilardo, A., Mazzeo, A., Mazzocca, N.: Representation of elements in F2m enabling
unified field arithmetic for elliptic curve cryptography. Electronics Letters 41(4) (July 7,
2005)
2. Montgomery, P.: Speeding the Pollard and elliptic curve methods of factorization.
Mathematics of Computation 48, 243–264 (1987)
3. López, J., Dahab, R.: Fast multiplication on elliptic curves over GF(2m) without
precomputation. In: Koç, Ç.K., Paar, C. (eds.) CHES 1999. LNCS, vol. 1717, pp. 316–327.
Springer, Heidelberg (1999)
4. Meloni, N.: Fast and Secure elliptic Curve Scalar Multiplication Over Prime Fields Using
Special Addition Chains. Cryptology ePrint Archive: listing for 2006 (2006/216) (2006)
5. Paar, C.: Light-Weight Cryptography for Ubiquitous Computing. Invited talk at the
University of California, Los Angeles (UCLA). Institute for Pure and Applied
Mathematics (December 4, 2006)
6. Sakiyama, K., Batina, L., Mentens, N., Preneel, B., Verbauwhede, I.: Smallfootprint ALU
for public-key processors for pervasive security. In: Workshop on RFID Security, 12 pages
(2006)
7. Batina, L., Mentens, N., Sakiyama, K., Preneel, B., Verbauwhede, I.: Low-costElliptic
Curve Cryptography for wireless sensor networks. In: Buttyán, L., Gligor, V.D., Westhoff,
D. (eds.) ESAS 2006. LNCS, vol. 4357, pp. 6–17. Springer, Heidelberg (2006)
8. OztAurk, E.A., Sunar, B., Savas, E.: Low-power elliptic curve cryptography using scaled
modular arithmetic. In: Joye, M., Quisquater, J.-J. (eds.) CHES 2004. LNCS, vol. 3156,
pp. 92–106. Springer, Heidelberg (2004)
9. Satoh, A., Takano, K.: A Scalable Dual-Field Elliptic Curve Cryptographic Processor.
IEEE Transactions on Computers 52(4), 449–460 (2003)
10. Menezes, J., van Oorschot, P.C., Vanstone, S.A.: Handbook of Applied Cryptography.
CRC Press, Boca Raton (1997)
11. Walter, C.: Systolic Modular Multiplication. IEEE Trans. Computers 42(3), 376–378
(1993)
12. Kornerup, P.: A Systolic, Linear-Array Multiplier for a Class of Right-Shift Algorithms.
IEEE Trans. Computers 43(8), 892–898 (1994)
13. Tsai, W.C., Shung, C.B., Wang, S.J.: Two Systolic Architecture for Modular
Multiplication. IEEE Trans. VLSI 8(1), 103–107 (2000)
 Modified Partial Differential Equations Based Adaptive
Two-Stage Median Filter for Images Corrupted with
High Density Fixed-Value Impulse Noise

P. Shanmugavadivu and P.S. Eliahim Jeevaraj

Department of Computer Science and Applications,

Gandhigram Rural Institute, Deemed University,
Gandhigram-624 302, Tamil Nadu, India
[email protected], [email protected]

Abstract. This paper presents a newly devised noise filter viz., Modified Partial
Differential Equation based Adaptive Two-Stage Median Filter (MPATS) to
denoise the images corrupted by fixed-value impulse noise. The performance of
the proposed filter is proved to be better in terms of Peak Signal - to - Noise
Ratio and human visual perception. This filter is effective in denoising the
highly corrupted images with the noise probability of even 90%.

Keywords: Fixed-value impulse noise, salt and pepper noise, standard median
filter, PDE, highly corrupted image.

1 Introduction
Image restoration is a vital component of digital image processing which targets at the
elimination or reduction of noise in corrupted images. The unwanted and irrelevant
information present in any image is referred to as noise that degrades the quality of
image and makes it unsuitable for high level processing [1]. Digital images are
commonly corrupted due to faulty sensors in the capturing devices, atmospheric
turbulence etc., during acquisition and during communication due to the limitations of
the transmission media. If an image is corrupted with the maximum and minimum
intensity of the dynamic ranges, it is termed as fixed-value impulse noise, whereas its
variant namely random – value impulse noise corrupts the images by randomly
assuming any value between the intensity intervals [2].
Review of literature reveals the fact that denoising potentials of non-linear filters
are greater than that of linear filters like mean filters. Further, the non-linear filters are
deemed meritorious over linear filters for their edge and detail preserving potentials.
However, these filters have a major drawback of modifying the intensities of
uncorrupted pixels too, during denoising, as they do not adopt any mechanism to
discriminate between the corrupted and uncorrupted pixels. The next generations of
filters known as adaptive filters invariably employ the strategy of noise detection
prior to noise correction [3]-[4].
The median based filters such as Progressive Switching Median Filter (PSM),
Iterative Median Filter (IMF), Signal Dependent Rank Order Median Filter (SDROM),

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 376–383, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Modified Partial Differential Equations Based Adaptive Two-Stage Median Filter 377

Two-state Recursive Signal Dependent Rank Order Median Filter (SDROMR), Non-
Recursive Adaptive-Center Weighted Median Filter (ACWM), Recursive Adaptive-
Center Weighted Median Filter (ACWMR) , Russo’s Fuzzy Filter (RUSSO), Zhang’s
filter (ZHANG) , Sun and Neuvo’s Switching Based Filter (SUN) , Impulse Rejecting
Filter (IRF) and Tri-State Median Filter (TSM) are devised with two fold benefit of
noise detection and correction[5]-[6].
The proposed filter, “Modified Partial Differential Equation based Adaptive Two-
Stage Median Filter for Images Corrupted with High Density Fixed-Value Impulse
Noise (MPATS)”, performs noise correction on the corrupted images using the
principle of median and Partial Differential Equation (PDE). The Peak Signal-to-
Noise Ratios (PSNR) of the proposed filter are observed to be higher than those of
aforementioned filters. Moreover, the visual perception of the filtered image
substantiates the inferences of the obtained PSNR values.
The noise model of fixed-value impulse noise is given in section 2. The principle
of standard median filter is explained in section 3. The principle of PDE and the
proposed algorithm is detailed in section 4 and 5 respectively. The results and
discussion is presented in section 6 and the conclusions in 7.

2 Noise Model
The probability distribution of fixed value impulse noise in an image is uniform with
respect to the minimum and maximum values of the intensity interval [7].
For a noise probability of p, the distribution of noise is described as:
 I max with probabilit y p1
N ij =  (1)
I min with probabilit y p 2
where Nij are the corrupted pixels and p1=p2=p/2. The noisy image X΄ij is represented
as
 N ij with probability p
X ij' =  (2)
I with probability 1 - p
where, I represent the uncorrupted pixels in the image.

3 Standard Median Filter (SM)

For an input and output image, X () and Y () respectively, the SM filter uniformly
substitutes the central pixel of each window of size n × n by the median of pixels
covered by W. The output of standard median filter is given as:

Yi , j = median{X i − s , j− t | (s, t ) ∈ W } (3)

The coordinates of the pixels in W are given as:

W = {(s, t ) | − k ≤ s ≤ k ,− k ≤ t ≤ k}, (4)

where k = (n-1)/2.
378 P. Shanmugavadivu and P.S. Eliahim Jeevaraj

This process is successively repeated over the entire input image. Though SM is a
standard image restoration technique, it removes the details such as thin lines and
edges while denoising [8].

4 Partial Differential Equations (PDE)

In a networked mesh that is split into small grids, if the central point of the grid is
unknown, then it is found from its neighboring points, using the principle of PDE [9].
The central point is estimated using Laplace equation as given in equation (5).

∂2u ∂2u
+ =0 (5)
∂x 2 ∂y 2
In a 3×3 grid (Fig. 1), if the derivatives are replaced by their difference
approximations, then equation (5) transforms into equation (6) as:

M i, j =
1
4
[
M i −1, j−1 + M i +1, j−1 + M i −1, j + 1 + M i +1, j+1 ] (6)

Mi-1,j-1 Mi-1,j+1 X X X

Mij X Mij X

Mi+1,j-1 Mi+1,j+1
X X X

Fig. 1. Illustration of Diagonal 5-point Formula Fig. 2. Illustration of 3×3 window

It is otherwise known as Diagonal 5 point formula. This second order PDE, finds
application in the edge detection and noise removal [10]-[12].

5 Proposed Technique
The proposed filter Modified Partial Differential Equation based Adaptive Two-Stage
Median Filter for Image Corrupted with High Density Fixed Value Impulse Noise
(MPATS), works in two distinct phases as noise detection and noise correction.
In the noise detection phase, each pixel of the image is processed to identify the
corrupted pixels. If the dynamic grayscale of the image is [Imin , Imax] then, the pixel of
intensity Imin and Imax are graded as corrupted pixels.
 Modified Partial Differential Equations Based Adaptive Two-Stage Median Filter 379

The proposed techniques MPATS, performs noise correction in two levels. The
entire image is divided into a discrete set of subimages in the order of window W of
size n × n. The layout of pixels aggregated by W is portrayed in Fig 2. The central
pixel is the window is labeled as Mij and the other neighbouring pixels are labeled as
X. The algorithm checks whether the central pixel of the window (Mij) is corrupted or
not. If found uncorrupted, the algorithm progresses to next subimage. Otherwise, the
algorithm checks if any of the Diagonal 5 point formula pixels (CDP) of Mij is
corrupted or not. It groups the uncorrupted pixels of W as UP and the corrupted pixels
of Diagonal 5 –point formula as CDP. If all the pixels of Diagonal 5-point formula are
uncorrupted then the median of those pixels is computed as DMED and Mij is replaced
by DMED. If those pixels are corrupted, the algorithm computes the median of Up in W
as UMED and replaces each CDP in the Diagonal 5-point formula pixels by UMED.
According to equation (6), the mean is computed for those pixels. But, this algorithm
computes median of those four pixels as DMED, using which the corrupted central
pixel Mij is replaced. The computation of median has an edge over mean when the
distribution of data is uneven, as evidenced in images.

5.1 Algorithm of MPATS

The algorithmic description of the proposed filter is given below:

Step 1: Read the corrupted image X΄.

Step 2: Divide X΄ into subimages using overlapping sliding window W of sixe n × n.
Step 3: For each window of X΄, check whether the central pixel Mij is corrupted or not.
Step 4: If Mij is uncorrupted, goto Step 3.
Step 5: If Mij is corrupted, then read the pixels (CDP) of Diagonal 5-point formula and
check if those pixels are corrupted or not.
a. If corrupted
i. Label those pixels as CP
ii. Compute: UMED ← Median ( UP), where UP represents the
uncorrupted pixels in W
iii. CP ← UMED
iv. DMED ← Median [ Mi-1,j-1, Mi+1,j-1, Mi-1,j+1, Mi+1,j+1]
v. Mij ← DMED
vi. Goto Step 3.
b. If uncorrupted,
i. DMED ← Median [ Mi-1,j-1, Mi+1,j-1, Mi-1,j+1, Mi+1,j+1]
ii. Mij ← DMED
iii. Goto Step 3.
Step 6: Stop.
380 P. Shanmugavadivu and P.S. Eliahim Jeevaraj

5.2 Flowchart

The Pictorial representation of MPATS is depicted in Fig. 3.

Fig. 3. Flowchart of MPATS

6 Results and Discussion

The proposed noise filter is developed in Matlab 6.5. The proposed filter was tested
on many standard images like lena, cameraman, rice, saturn and mandrill. The
 Modified Partial Differential Equations Based Adaptive Two-Stage Median Filter 381

window size was chosen as 3 × 3 and the noise probability was varied between 10%
and 90%. The performance of the proposed filter was recorded in terms of PSNR and
human visual perception.

Table 1. Comparison of PSNR values for Lena image

Method Noise Density (%)

10 20 30 40 50 60 70 80 90
Corrupted 15.5 12.4 10.7 9.4 8.5 7.7 7.0 6.4 5.9
SM 28.7 26.4 22.6 18.3 15.0 12.2 9.8 8.1 6.5
CWM 29.7 24.1 19.5 15.7 13.0 10.8 8.9 7.6 6.3
PSM 30.7 28.7 26.9 23.7 20.0 15.2 11.1 8.3 6.4
IMF 27.2 26.7 26.1 25.1 23.9 21.2 16.6 12.1 8.0
SDROM 30.3 26.7 22.0 17.6 14.4 11.7 9.4 7.8 6.4
SDROMR 30.5 28.5 25.7 23.5 20.7 17.8 14.2 10.6 6.9
ACWM 30.9 27.2 22.4 18.1 14.8 12.1 9.7 8.1 6.5
ACWMR 31.4 28.8 25.8 23.3 20.8 18.1 15.0 11.8 8.1
RUSSO 31.0 27.6 24.9 22.7 20.3 17.6 14.7 11.7 8.6
ZHANG 32.8 28.2 23.3 18.6 15.3 12.5 10.0 8.3 6.7
SUN 31.0 27.5 23.0 18.4 15.0 12.2 9.8 8.1 6.5
IRF 30.2 27.0 22.5 18.2 14.9 12.2 9.7 8.1 6.5
TSM 30.3 24.4 19.6 15.5 12.7 10.4 8.4 7.1 6.0
MPATS
(Proposed Filter) 32.3 28.9 26.9 24.8 23.5 21.6 19.6 17.9 15.0

Table 2. Comparison of PSNR values for Mandrill image

Noise Density
Method
10 20 30 40 50 60 70 80 90
Corrupted 15.7 12.6 10.9 9.6 8.6 7.9 7.2 6.6 6.1
SM 23.8 23.0 20.7 17.7 14.7 12.3 10.9 8.3 6.8
CWM 25.5 22.7 18.9 15.6 12.9 10.9 9.1 7.7 6.6
PSM 27.8 26.4 25.0 22.9 19.5 15.3 11.0 8.9 6.5
IMF 23.0 22.8 22.6 22.3 21.5 20.0 16.0 12.0 8.3
SDROM 26.0 24.0 20.8 17.4 14.3 11.8 9.6 7.9 6.6
SDROMR 25.8 24.6 23.2 21.8 19.7 17.6 14.2 10.8 7.1
ACWM 27.1 24.8 21.4 17.9 14.8 12.3 9.9 8.2 6.8
ACWMR 27.5 25.7 23.9 22.3 20.2 18.2 15.4 12.4 8.8
RUSSO 29.5 26.7 24.3 22.4 20.2 17.7 15.0 11.9 8.9
ZHANG 28.8 26.1 22.1 18.5 15.2 12.7 10.2 8.4 6.9
SUN 26.1 24.4 21.3 18.0 14.8 12.3 9.9 8.2 6.7
IRF 25.8 24.1 21.2 17.8 14.7 12.3 9.9 8.2 6.8
TSM 25.4 22.9 18.9 15.4 12.6 10.5 8.6 7.2 6.2
MPATS 29.6 27.0 24.8 23.8 22.4 21.1 19.9 18.2 17.1
(Proposed Filter)
382 P. Shanmugavadivu and P.S. Eliahim Jeevaraj

To ascertain the performance of MPATS, the PSNR values of lena and mandrill are
combined with few high performing filters reported in the literature [2]. The PSNR
values of lena and mandrill images are given in Table I and II respectively. It is
apparent from these tables the PSNR values of MPATS for the lena image is higher
than all the filters considered for comparison, except ZHANG for 10% of noise and
IMF for 40% and 50% of noise, whereas the PSNR values of MPATS for mandrill are
higher for all the noise probabilities. Moreover, the merits of the filter as reflected by
the PSNR values are duly confirmed by the visual perceptions of the filtered images
shown in Fig. 4.

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

(k) (l) (m) (n)

Fig. 4. (a) Original image of Lena (b) Lena image with 50% noise; (c) Filtered Image of (b);
(d) Lena image with 60% noise; (e) Filtered Image of (d); (f) Lena Image with 70% noise; (g)
Filtered Image of (f); (h) Original Image of Mandrill (i) Mandrill Image with 50% noise; (j)
Filtered Image of (i); (k) Mandrill Image with 60 % noise; (l) Filtered Image of (k); (m)
Mandrill Image with 70% noise; (n) Filtered Image of (m);

7 Conclusions
The proposed filter MPATS eliminates the fixed–value impulse noise remarkably
from the corrupted images and ensures higher degree of restoration which is recorded
 Modified Partial Differential Equations Based Adaptive Two-Stage Median Filter 383

in terms of PSNR and visual perceptions. Hence, this filter can be widely used in
document imaging, medical imaging, bulk digitization of black and white films and
image scanning. This filter also finds application in eliminating noise from Scanning
Electron Microscope (SEM) images, used for the study of surface morphology,
because SEM images are invariably degraded by noise.

References
1. Gonzalez, R.C., Woods, R.E.: Digital Image Processing, 3rd edn. Pearson Prentice Hall,
India (2009)
2. Besdok, E., Emin Yuskel, M.: Impulsive Noise Suppression for Images with Jarque – Bera
Teat based Median Filter. International Journal of Electronics and Communications 59,
105–110 (2005)
3. Somasundaram, K., Shanmugavadivu, P.: Adaptive Iterative Order Statistics Filters.
Journal of ICGST – GVIP 09, 23–32 (2009)
4. Ghandeharian, B., Yazdi, H.S., Homayouni, F.: Modified Adaptive Center Weighted
Median Filter for Suppressing Impulsive Noise in Images. International Journal of
Research and Reviews in Applied Sciences 01(03) (2009)
5. Satpathy, S.K., Panda, S., Nagwanshi, K.K., Ardill, C.: Image Restoration in Non – Linear
Domain using MDB approach. IJICE 06(01) (2010)
6. Mansoor, M., Roomi, S., Lakshmi, I.M., Abhai Kumar, V.: A Recursive Gaussian
Weighted Filter for Impulse Noise Removal. Journal of ICGST – GVIP 06(03), 33–37
(2006)
7. Somasundaram, K., Shanmugavadivu, P.: Impulsive Noise Detection by Second Order
Differential Image and Noise Removal using Nearest Nieghbourhood Filter. International
Journal of Electronics and Communications, 472–477 (2007)
8. Windya, P.S.: Fast Impulse Noise Removal. IEEE Trans. on Image Processing 10(01),
173–179 (2001)
9. You, Y.L., Kaveh, M.: Fourth order PDE for noise removal. IEEE Trans. on Image
Processing 09(10), 1723–1730 (2000)
10. Shanmugavadivu, P., Eliahim Jeevaraj, P.S.: Fixed – Value Impulse Noise Suppression for
Images using PDE based Adaptive Two -Stage Median Filter. In: ICCCET 2011 (IEEE
Explore), pp. 290–295 (2011)
11. Shanmugavadivu, P., Eliahim Jeevaraj, P.S.: Selective 8-Neighbourhood based Medan
Filter for Images Corrupted with High Density Fixed – Value Impulse Noise. In: 3rd IIST
(2011)
12. Grewal, B.S.: Higher Engineering Mathematics, 2nd edn., pp. 1129–1130. Khanna
Publishers, New Delhi (2007)
13. Bertozzi, A.L., Greer, J.B.: Travellling Wave Solutions of Fourth Order PDEs for Image
Processing. SIAM, Journal of Mathematical Analysis 36(01), 36–68 (2004)
 Enhanced Classification Performance Using
Computational Intelligence

Indrajit Mandal and N. Sairam

School of Computing
Sastra University, Tamil Nadu, India
{Indrajit,sairam}@cse.sastra.edu

Abstract. This paper presents a computational intelligence technique for

enhancing the performance of classifier using a proposed algorithm called
Modified Genetic Search Algorithms (MGSA) that avoids local bad search
space with merit and scaled fitness variables, detecting and deleting
bad candidate chromosomes, thereby reducing the number of individual
chromosomes from search space and subsequent iterations in next generations.
It addresses the strength of Modified Genetic Search algorithm combined with
the Artificial Neural Network (ANN). In this work dynamic Backpropagation
Neural Network is used. For training purpose, dynamic learning rate is used that
causes the learning rate to decrease in subsequent epoch.
The combined MGSA-ANN is used for the classification of diabetes patients
to identify positive and negative cases. It also discusses the main findings and
concludes with promising result of the proposed model. The experimental
results obtained by synergistic combination of Modified Genetic Search
Algorithm with ANN surpass the performance of ANN by 1.4322%.

Keywords: Computational intelligence, Modified Genetic Search Algorithm,

Backpropagation Neural Network, dynamic learning rate.

1 Introduction
The improvement in the performance of Artificial Neural Network has been discussed
by Jiansheng and Mingzhe, 2005 [8] in different perspective. Here Back-propagation
neural networks use multiple layer perceptron as system framework with MGSA is
used and Back-propagation Algorithm as training rule. It helps to avoid network from
diverging away from the targeted output and improves general performance. The
experimental setup exploits the use of MSGA-ANN to classify patients as having
diabetes using java programming. The real world data Pima Indian Diabetes database
available and is adapted from UCI Machine Learning Repository that contains 768
instances. Besides, it contain eight numeric inputs and a target attribute class which
takes on values 0 (tested-negetive) and 1 (tested-positive). The list of attribute (all
numeric-valued) are listed below
1. 'preg'- Number of times pregnant
2. 'plas' -Plasma glucose concentration a 2 hours in an oral glucose tolerance test

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 384–391, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Enhanced Classification Performance Using Computational Intelligence 385

3. 'pres'-Diastolic blood pressure (mm Hg)

4. 'skin'-Triceps skin fold thickness (mm)
5. 'insu'-2-Hour serum insulin (mu U/ml)
6. 'mass'-Body mass index (weight in kg/(height in m)^2)
7. 'pedi'-Diabetes pedigree function
8. 'age'-Age of patients (years)
9. Class variable- tested_negative, tested_positive.

2 Artificial Neural Network

In this work, dynamic Backpropagation Neural Network with single hidden layer is
used trained via sigmoid function in which the network parameters is possible to be
monitored and modified while training time. For training purpose, dynamic learning
rate is used that will cause the learning rate to decrease in subsequent epoch. The
starting learning rate is divided by the epoch number, to determine the current
learning rate. It helps avoiding the network from diverging away from the target
output, as well as improves general performance. All the training data is normalized
to enhance the network performance between 1 and -1. Brief summary of
backpropagation algorithm as discussed by I. Mandal, 2010 [1].
The activation function used in the neural network training is sigmoid function
mentioned in (1) where k is integer constant and x is the input value.

f(x) = 1 / √(1 + e(-kx)) (1)

Other network parameters

Number of neurons in hidden layer= (attribute+ classes)/2= (9+2)/2=6
Learning rate = 0.3
Momentum = 0.2
Number of epochs = 500.

3 Modified Genetic Simple Algorithm

As GA finds solution through evolution process, it is not inclined to good solution but
move away from bad solutions. There is a chance that it may lead solution to a dead
end. Also the number of iterations or generations required is very high. In the
proposed MGSA, the idea is to identify and eliminate the bad chromosomes using
merit and scaled variables so that the search space is reduced. Once the search space
is minimized, containing prospective chromosomes then it leads to better optimization
in search process. Here the merit and scaled variables refer to the classification error
rates of chromosomes. Bad chromosomes indicate those individuals which lead to a
dead end. Here is a brief summary of MSGA:
386 I. Mandal and N. Sairam

1) Initial population generation. The parameters of the model to be optimized are

considered as chromosomes. A randomly chosen set of population of individuals is a
binary string with a fixed length.
2) Fitness function: In each generation for which GA is run, the fitness of each
individual is determined which is close to optimality. Here in the attribute selection
problem, we have defined linear function f `= af + b, where f `, f are scaled and raw
fitness values of chromosomes and a,b are constants.
3) Detection and Elimination: Those chromosomes having greater classification
error rate determined by merit and scaled variables are determined and eliminated so
that it is not allowed to be a candidate chromosome in the next generation and thereby
filtering out the bad chromosomes and hence the number of iterations in the
subsequent generations gets reduced.
4) Selection: A pair of best fit chromosomes or individuals which has least
classification error rate is selected from the residual population.
5) Crossover: It is the reproductive stage where two new individuals are crossed with
probability Pc to generate new pair of offspring.
6) Mutation: A single point alteration of bit string from zero to one and vice-versa
with probability Pm is applied to selected individuals to generate a new pair of
offspring and avoids premature convergence.

4 Experimental Design and Results

4.1 Case 1

The BPNN as said in Haykins, 1999 [9] is applied to the dataset to classify with 10-
fold cross validation as mentioned by Picard and Cook,1984[7] and shows that it
achieves a very impressive 75.7813 % classification accuracy. Here the data is
divided into equally ten parts and nine are used for classification and one is used for
testing. This is repeated for ten times and average accuracy is considered. Details are
shown in table1.
A confusion matrix as discussed by Marom, Rokach, Shmilovici, 2010[5] is shown
in table2 that contains the analytical details of classifications where eight attributes
are input to it. Its performance is evaluated based on the data in the matrix for two
class classifier. Accuracy is measured by Received Operator Characteristics (ROC)
area as discussed by Kohavi and Provost, 1998[4], area under graph with TP as Y-
axis and FP as X-axis and ranges from zero to one. For class tested-negative ROC
plot =0.793 is shown in fig1 and Cost/benefit analysis is shown in fig 2 where gain is
141.21. The following tables 1, 2 show the confusion matrix and accuracy parameters
respectively.

4.2 Case 2

In case 1 all the eight attributes are considered as input but in real world data
irrelevant, redundant or noisy attributes are common phenomena, which impairs the
 Enhanced Classification Performance Using Computational Intelligence 387

classifier’s result. The learning scheme Wrapper subset evaluation as discussed by

Mallik, 2007[7] with ANN is now integrated with Modified Genetic Search
algorithm. After filtration process only five attributes are selected as relevant.
Attributes preg, skin, insu are eliminated.
The attributes of MGSA that includes
Population size: 30
Number of generations: 30
Probability of crossover: 0.6
Probability of mutation: 0.033
Now reclassifying the records ANN with 10-fold cross validation; however, this time
only five attributes are input to the classifier. Selected attributes are plas, pres, mass,
pedi, age.
In table3, every subset is a chromosome and merit is the fitness score reported by
MGSA, which is equal to the corresponding classification error rate. Also, each
chromosome’s scaled fitness is shown in the scaled column where we use linear
scaling technique to scale the values. By definition, the raw fitness and scaled fitness
values have the linear relationship

f`= a.f + b (2)

where f ` and f are the scaled and raw fitness values respectively. The constants a and
b are chosen where

f`avg = favg and f`max= K f`avg (3)

The constant K represents the expected number of copies of the fittest individual in
the population. Thus, by computing the average fitness values from table5, we obtain
f`avg =0.055755 and favg =0.055753.
Solving equations (2) we get a= 5.012392 and b= -0.883779. We use equation (3)
to determine K i.e. K= f`max / f`avg =0.27273/0.238443= 1.143795
Observe in fifth row in table3 f`=0. The raw fitness value of 0.28516 corresponds
to the largest classification error in the population produced by chromosome {3 6 7
8}, and as a result, f` is mapped to zero to avoid the possibility of producing
negatively scaled fitness.
The improved classifier has accuracy level of 77.2135 %, which indicates that the
second model outperforms the first model by 1.4322 % where the input to the later
model is only five attributes indicated in ROC area. For class tested-negative ROC
plot =0.809 is shown in fig3 and Cost/benefit analysis is shown in fig 4 where gain is
159.76 and observe the smoothness achieved in the graph. Moreover reducing the
number of attributes in second model has increased the gain in cost benefit analysis.
That is, the classification accuracy has improved positively where only five of eight
attributes are specified in the input. Though the increased accuracy is not a dramatic
improvement, it shows the strength of Modified Genetic Search algorithm is an
appropriate algorithm for attribute selection process.
388 I. Mandal and N. Sairam

Fig. 1. ROC area=0.793 for case1 Fig. 2. Cost/benefit analysis for gain=141.21
in case 1 with 9 attributes

Table 1. Classification Accuracy by neural Network

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

Accuracy
0.832 0.392 0.798 0.832 0.815 0.793 tested_negative
by Class
0.608 0.168 0.66 0.608 0.633 0.793 tested_positive
Weighted
0.754 0.314 0.75 0.754 0.751 0.793
Average

Fig. 3. ROC area=0.809 for case2 Fig. 4. Cost/benefit analysis for gain=159.76
in case2 with 5 attributes
 Enhanced Classification Performance Using Computational Intelligence 389

Table 2. Confusion Matrix for neural network

Predicted
tested_negative tested_positive
416 84 tested_negative
105 163 tested_positive Actual

Table 3. Initial population characteristics for the 30 chromosomes

merit scaled subset

0.23073 0.27273 23678
0.23073 0.27273 23678
0.23073 0.27273 23678
0.2349 0.25185 2368
0.23073 0.27273 23678
0.23984 0.22706 2367
0.2388 0.23228 235678
0.2349 0.25185 2368
0.2474 0.18922 12578
0.23073 0.27273 23678
0.2388 0.23228 235678
0.23073 0.27273 23678
0.23073 0.27273 23678
0.23073 0.27273 23678
0.2388 0.23228 235678
0.24193 0.21662 2378
0.23073 0.27273 23678
0.23073 0.27273 23678
0.23073 0.27273 23678
0.24193 0.21662 2378
0.25755 0.13832 234
0.2388 0.23228 235678
0.2388 0.23228 235678
0.23073 0.27273 23678
0.24323 0.2101 2678
0.23073 0.27273 23678
0.2349 0.25185 2368
0.28516 0 3678
0.23438 0.25446 25678
390 I. Mandal and N. Sairam

Table 4. Classification Accuracy for neural network and genetic algorithm

TP FP F- ROC
Precision Recall Class
Rate Rate Measure Area
Accuracy
0.846 0.377 0.807 0.846 0.826 0.809 tested_negative
by Class
0.623 0.154 0.684 0.623 0.652 0.809 tested_positive
Weighted
0.768 0.299 0.764 0.768 0.766 0.809
Average

Table 5. Confusion Matrix for neural network and genetic algorithm

Predicted
tested_negati tested_positive
ve
423 77 tested_negative
Actual
101 167 tested_positive

5 Conclusion
MGSA can perform global search rather than local search efficiently. It addresses the
strength of proposed Modified Genetic Search algorithm combined with the Artificial
Neural Network (ANN). For training purpose, dynamic BPNN trained with sigmoid
function with learning rate to decrease in subsequent epoch and improve general
performance. The proposed model MGSA addresses the local optima problem using
merit and scaled variables in which the bad individual chromosomes are detected and
deleted, reducing the search space and further iterations thereby improving efficiency.
We have shown that MSGA is used for reducing the number of input parameters
without sacrificing the performance of the classifier.
It shows synergistic performance of MGSA-ANN in the classification problem
with 1.4322 % demonstrated with Pima Indian Diabetes dataset. The computational
intelligence obtained from the MGSA helps to eliminate the irrelevant variables from
the real world data set so it acts as subset selection evaluator.
Although in this paper there are only eight attributes, there are still 255 possible
attribute subsets that can be given as input. If the numbers of attributes increases say
hundred then there are 1.27x 1030 possible attribute subsets to choose. In such a hard
situation like this MGSA may prove helpful in determining the optimal subset. From
the experimental results, it is concluded that GA is an effective algorithm that can be
exploited in complex situations. It can be used in various applications such as
optimization of weights of links in neural networks. Further the performance of GA
can be improved by parallel implementation using MPI and other techniques and can
be applied to vast field of emerging applications.
 Enhanced Classification Performance Using Computational Intelligence 391

References
1. Mandal, I.: Software reliability assessment using artificial neural network. In: International
Conference and Workshop on Emerging Trends in Technology, pp. 698–699. ACM, New
York (2010)
2. Pima Indians Diabetes dataset, National Institute of Diabetes and Digestive and Kidney
Diseases, UCI Machine Learning Repository,
http://archive.ics.uci.edu/ml/datasets.html
3. Goldberg, D.E.: Genetic algorithms in search, optimization and machine learning.
Addison-Wesley, Reading (1989)
4. Kohavi, Provost (1998),
http://www2.cs.uregina.ca/~hamilton/courses/831/notes/confus
ion_matrix/confusion_matrix.html
5. Marom, N.D., Rokach, L., Shmilovici, A.: Using the Confusion Matrix for Improving
Ensemble Classifiers. In: 2010 IEEE 26th Convention of Electrical and Electronics
Engineers in Israel (2010)
6. Mallik, R.K.: The Uniform Correlation Matrix and its Application to Diversity. IEEE
Transactions on Wireless Communications 6(5), 1619–1625 (2007)
7. Picard, R.R., Cook, R.D.: Cross-Validation of Regression Models. Journal of the American
Statistical Association 79(387), 575–583 (1984),
http://www.jstor.org/stable/2288403
8. Wu, J., Liu, M.: Improving Generalization Performance of Artificial Neural Networks with
Genetic Algorithm. In: 2005 IEEE International Conference on Granular Computing,
vol. 1, pp. 288–291 (2005)
9. Haykin, S.: Neural Networks-A Comprehensive Foundation, 2nd edn. Prentice Hall, Inc.,
New Jersey (1999)
10. Shi, H., Zhang, S.: Improving Artificial Neural Networks Based on Hybrid Genetic
Algorithms. In: 5th International Conference on Wireless Communications, Networking
and Mobile Computing, WiCom 2009, pp. 1–4 (2009)
 Feature Modeling of the Evolving Access Control
Requirements

K. Shantha Kumari1 and T. Chithralekha2

1
Research Scholar,
Department of Banking Technology,
Pondicherry University
[email protected]
2
Reader,
Department of Computer Science,
Pondicherry University

Abstract. Access control mechanisms are an integral part of most modern

software systems, but it is still not considered as an explicit part in the
development process. Access control mechanisms and policies are generally
added to existing systems as an afterthought, with all the problems of
unsatisfied security requirements, integration difficulties, and mismatches
between design models. In order to increase the overall system security, access
control requirements should be taken into account early in the software
development process. Due to the integration of Access control requirements
early in the development process, potential security breaches can be detected
and removed earlier. But, on the other hand, this integration results in spreading
of access control functionalities across the software systems and making them
tightly cohesive. When there is a need arises for updating the existing access
control requirement, the changes would impact on several functional
requirements. Moreover the access control requirements tend to evolve
continually. Accordingly, the design phase should support both the static and
dynamic changes of the access control requirements without making much
impact on the functional requirements. To address this issue at the design level,
we need an approach that could support to model the access control
requirements along with the constraints and also capture the changes, without
affecting the functional design. In this paper, we propose a Feature Conclave
Model based on feature modeling, for analyzing and designing the evolving
access control requirements along with the system design. The Feature
Conclave Model views the access control requirements as “Features” and also
provides step-by-step integration with the functional domain. Moreover, the
new updates of the access control requirements can be represented as the
variants and can easily be adapted within the system design.
Keywords: Access control requirements, Role based access control, Feature
Modeling, Assembling, Verification.

1 Introduction
The need for Access control mechanisms in large scale software systems has become
a major priority for all organizations. The increasing number of threats and

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 392–403, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Feature Modeling of the Evolving Access Control Requirements 393

vulnerabilities has repeatedly shown that the design and engineering of these systems
from a security stand point is often lacking. As a general practice, Access control
policies are often added in an ad-hoc basis during the later part of the product
development resulting in errors and vulnerabilities that provide a potential for
exploitation. However, given the complexity and pervasiveness of today's software
systems, this kind of approach might not be completely satisfactory. As such,
engineering software security is essential. In order to increase the overall system
security, access control requirements should be taken into account early in the
software development process. An efficient Access control system design must
support system evolution, such as changes in security policies, user population and
their roles, and also the changes in the applications.
From the design perspective, access control policies give rise to security concerns
that must be addressed in a design [1]. Prior to this view, many researchers have
recommended for specifying and maintaining Access control requirements separately
from the application design. This allows the Access control requirements to be
analyzed clearly and the policies can be changed independently of the application. But
this kind of separate policy specification has a problem associated with it-The
scalability and late integration problems.
To properly address security risks and vulnerabilities without jeopardizing speed or
cost, companies must bring security into the development process. This results in a
tightly cohesive system with the scattering of the access control functionality across
the various functional modules design [2]. This intervention of access control features
with the behavior of the functional specifications makes it very difficult to understand
and analyze the working of the access control policies. Moreover the access control
requirements may evolve continuously. For example, Current and former employees
and contractors who have or had authorized access to their organization's system and
networks are familiar with internal policies, procedures, and technology and can
exploit that knowledge to facilitate attacks and even collude with external attackers .
Addressing this kind of pervasive access control requirements for complex
information systems, using an improvised approach, would lead to a design with
inconsistencies in satisfying the security constraints. This will result in security
failures and violations [2].
In this paper, the access control requirements are modeled using Feature conclave
Model on the basis of Feature modeling to support a systematic and planned way of
integrating with the functional requirements in the design. This Model would reduce
the developmental complexities and increase the effectiveness in the integration.
Using the abstraction of ‘Features’ results in an ease of adaptation of the changes in
the access control requirements and further integration with the application
requirements. UML design principles are used for modeling the access control model
as feature conclave model that would enhance the factor of reusability.
The paper is organized as follows. Section 2 gives an overview of current modeling
techniques and its limitations. In Section 3 we present Feature conclave model and
also explain the associated feature modeling method. Section 4 briefs about the formal
basis of design principles for RBAC features. Section 5 demonstrates how RBAC
features can be configured for an example. Section 6 concludes the paper and
discusses future work.
394 K. Shantha Kumari and T. Chithralekha

2 Modeling of Access Control Requirements

The main objective of access control is to provide confidentiality of sensitive
information, one of the main characteristics of security. Access control policies
determine how sensitive information and computing resources are to be protected.
Olden year software systems, by their very nature, have simple and generic access
operations and are not concerned with the meaning of the resources they handle. On
the other hand, modern Information sensitive systems and applications handle quite
complex access operations due to their specific user purposes and high functionality.
This complexity needs to be translated by generic models and specific access control
requirements into design phase understandable processes.
Approaches for specifying and analyzing the access control requirements are based
on sophisticated mathematical concepts, that formally stated, allows one to check the
developed access control requirements enforce required level of protection. In
practice, however, applying mathematically-based formal specification techniques can
be difficult because of the high degree of mathematical skill needed. Therefore, a
representation that can be analyzed without sacrificing understandability and usability
is desirable. Generally the access control requirements are formally modeled to an
Access control model like Mandatory Access control, Discretionary Access control or
Role based Access control. Role-based access control (RBAC) models are becoming a
de facto standard, greatly simplifying management and administration tasks. Role-
based access control (RBAC) received considerable attention as an alternative to
traditional mandatory and discretionary access control policies.
In this paper Role-Based Access control is used for modeling the access control
requirements. Once requirements are presented as Role-based access control model,
further modeling of RBAC is required to facilitate the smooth integration with the
functional requirements. The Unified Modeling Language (UML) is considered to be
the industry de-facto standard for modeling software-based systems. Use of the UML
to specify RBAC policies ease the task of incorporating the policies into UML
application models. The following table shows the works done in modeling the RBAC
using UML and UML Profiles- the Extension mechanism of UML:

Using UML UML Profiles

UML class diagrams are used to A UML profile called UMLsec is
illustrate the structure of RBAC in [3] proposed for modeling [8] and
and the Object Constraint Language evaluating security aspects for
(OCL) [4] to define RBAC constraints distributed systems based on the multi-
level security model described in [9].
Object diagrams are used to visualize SecureUML [10] is the UML profile
the RBAC constraints in [5] proposed for defining security concepts
based on RBAC.
In the work of [6] an access control Security aspects of RBAC and MAC
model is viewed as a design pattern are incorporated into UML model
and uses the Gang-of-Four (GoF) elements in the work proposed in [11]
pattern template [7] to describe RBAC.
 Feature Modeling of the Evolving Access Control Requirements 395

The effectiveness of the UML as a standard is predicated. Among other things, on

there being a clear, precise, and pragmatic semantics for its notations. But the semi-
formal semantics of UML has ambiguity and inconsistency issues. Moreover UML
captures the static requirements and represents their behavior using anyone of the
diagram. But the pervasive nature of the access control requirements has to be
represented, so that evolving new access control requirements can be addressed.
Hence there is a need for a new level of abstraction to represent the evolving access
control requirements [14]. Features represent the variants- Changes.
Moreover the Access control requirements may scatter over the functional design.
This will lead to many problems listed as follows: (1) Altering the access control
requirement needs making changes in more number of places in the design, (2)
Assessing different ways for enforcing the access control requirements as model is
difficult when they crosscut a system design. These limitations can result in software
that cannot be expected to protect sensitive information.Hence the impact resulting in
updating an access control requirement on various functional requirements should be
minimized by reducing the tightly cohesive nature of the access control requirements
with the domain. This is not presented in UML kind of representations.
The above -mentioned observations have given rise to the need for developing a
new modeling approach that could support the evolving access control requirements
and also minimizing the impact by recognizing the integration in a step-by-step
manner [13]. The new approach proposed is based on feature modeling that facilitates
the reusability and comfortable composition of access control policies with the
functional requirements. Reusability is advocated by creating feature models for new
and evolving access control requirements by lining up the features from an available
and continuously growing warehouse of features. The next section will explain the
framework of Feature Conclave and feature warehouse.

3 Feature Conclave
Software Product Lines has emerged as a new technology to develop software product
families related to a particular domain. Software product line engineering (SPL)
involves developing the requirements, architecture, and component implementations
for a family of systems, from which products (family members) are derived and
configured. One of the multiple views of the product line is the feature modeling
view. The feature model is essential for both variability management and product
derivation, because it describes the product line requirements in terms of commonality
and variability, as well as defining the product line dependencies. A feature is a
logically cohesive piece of functionality and is present in all phases of software
development. Features may occur at any level, for example high-level system
requirements, architectural level, subsystem and component level. Features are basic
building blocks that satisfy intuitive user formulated requirements on the software
system. Features are main abstractions in design and implementation. They reflect
user requirements and incrementally refine one another.
The Proposed Feature Conclave consists of Feature Association rules and also a
Feature warehouse that houses all the features related with the domain. The Feature
Association is based on Feature Assembly modeling [14]. This aims to bring new
396 K. Shantha Kumari and T. Chithralekha

features as well as the previously defined features. The Feature conclave framework
consists of a warehouse that houses all the features related with the domain. In the
previous works of feature modeling, a Feature is viewed with its intention and also its
application in developing a product. But the contribution of the feature for creating
different products depends upon the requirements. Hence, it can be separated from the
definition. This enables the feature definition to be independent of its applicability in
various products. Therefore Features can be classified as “Concrete feature” that
provides the ample definition of its capabilities and as “Optional feature”, that
represents the Specialization of specific features. Feature association supports
reusability of features, by the usage of Feature warehouse. The Feature Association
rule – indicates the level of contribution by a particular feature on a product that gives
the variation. And the Feature Warehouse- 1.stores the features with annotations,2.
Association relations of features and 3. Variants of features and feature dependencies.
It doesn’t stores the Context of Use of the particular feature. The following figure
shows the Feature Conclave.

Feature Feature
Association Warehouse
Rules

Fig. 1. Feature Conclave framework

3.1 Need for Meta-Modeling

Generally the requirements for large scale soft wares provide different scope. The
requirements may scale with one another or may also produce conflicts. Hence the
modeling the requirements of a particular domain, requires a multiple stand for
analyzing and designing. The features of the domain may contribute differently in
multiple standpoints. Hence there is a need for a Meta Model of feature that could
facilitate the multiple standpoints. The Meta Model of feature will be realized
according to the perspective taken by the product line at the instance. The feature
warehouse is realization of this meta-modeling approach.

3.2 Feature Warehouse

Feature Warehouse is repository to store the feature definitions, relations between

features, dependencies among features and conditions for using the features. Features
in Warehouse actually represent the conceptual representations. They are stored
together along with the metadata information like owners, date of creation. The
relations, dependencies and conditions of the features are stored independent of the
context of application. Hence the reusability of the features is possible with or without
 Feature Modeling of the Evolving Access Control Requirements 397

its full definition. The Concrete feature can be used with or without its descendents
and options, where as the Optional feature can be used only when the options are
defined. By separating the definition of the relations and its implications, higher
degree of reusability is achieved. Thus the Meta-Model for the complete view of a
feature association is given as follows:

Feature

Capability Relations

Inheritance
Mandatory Elective
Concrete Optional
Association

Conditions
Dependencies

Pre Post
Excludes Extends

Consists Requires

Fig. 2. Meta-Model of Feature Definition

3.2.1 Relations and Dependencies between Features

The Relations and Dependencies between features are stored as a separate property
irrespective of any application in a Feature warehouse. Whenever a new product line
has to be built, it’s a very simple search operation through the warehouse for features
that matches with the requirements. Once the selection of features are done, they need
to composed together to define a product. This feature composition is supported by
“Relations”, either as Mandatory or Optional. A Mandatory relation depicts an
“ALL” Composition that indicates the compulsory and conclusive nature. The
Elective relation gives the “Choice based” composition depending upon the
requirements. This Elective relation is associated with the Optional features to
produce the variants. This is depicted in the following figure
398 K. Shantha Kumari and T. Chithralekha

Concrete
Mandatory Relation

Optional
feature 1
Elective Relation

Optional
feature 2

Fig. 3. Meta-Model of Feature Relation

In addition to the relations, the features may be defined with dependencies also.
They specify how the selection of one feature has impact with other features. They
also visualize the interactions between features. The different kinds of dependencies
are as follows:
1. Feature A extends Feature B
2. Feature A consists Feature B
3. Feature A requires Feature B
4. Feature A excludes Feature B

3.2.2 Conditions
Along with the above explained the constraints for associating a feature with another
feature is also mentioned in feature definition. They are as follows:
1. Pre-Conditions for a Feature Association
2. Post-Conditions for a Feature Association.

Pre-conditions give the prerequisites for selecting the feature for a product design.
Unless the conditions are satisfied, the corresponding feature selection will not
provide its functionality. Whereas, the Post-conditions express the conditions needed
after the feature composition. These post scenario work is very important to maintain
the stability of the developing product line.

3.3 Feature Association Rules

The Feature Association rule indicates the level of contribution by a particular feature
on a product that gives the variation. This information is a specific part of this rule, by
which the specific software product line can be generated. The feature association
 Feature Modeling of the Evolving Access Control Requirements 399

process is referred as the feature composition in Feature modeling techniques. It is as

follows:

Step -1: Analyze the requirement in the Domain  Capture the features [both
mandatory and Optional]  Populate Warehouse  Create the Product Line through
feature Relations and dependencies  Develop the feature Model
Step -2: New Requirement  Capture the features [both mandatory and Optional] 
Check the WarehouseIf existing extract from warehouse  else define new
features  Create feature Relations/dependencies/conditions  add the new defined
features to the warehouse.

Feature Association is the process of composing the features based on feature

association rules. This Association supports modeling of very large scalable software
during analysis and design. The Feature association rule thus defines the process of
feature association via feature definition, feature relations, dependencies and
conditions.

4 Formulating RBAC as Feature Model

Access control requirements which tend to be pervasive can be modeled using feature
modeling as the features reflect user requirements and incrementally refine one
another. In this paper, RBAC is chosen, as it is an efficient and scalable access control
model that governs access based on user roles and permissions. RBAC consists of
group of components (Features), each providing a different access control function.
From modeling perspective, The NIST RBAC standard is explored in this work.
By NIST standard [11], RBAC Core defines the minimum collection of RBAC
elements, element sets, relations in order to completely achieve a role-based access
control system. It includes: 1.user-role assignment 2.permission-role assignment. The
Hierarchal RBAC adds relations for supporting role hierarchies. And the SOD
component is used to enforce conflict of interest policies that organizations may

Fig. 4. RBAC with Core, Hierarchy and SOD Components

400 K. Shantha Kumari and T. Chithralekha

employ to prevent users from exceeding a reasonable level of authority for their
position. The Static SOD places restrictions on sets of roles. If a user is assigned to
one role, the user is prohibited from being a member of a second role. The Dynamic
SOD, places constraints on the roles that can be activated within or across user’s
sessions. It supports each user has different levels of permission at different time. It is
often referred as timely revocation of trust. The following figure shows the Core
RBAC with Hierarchy and Separation of duty components.

4.1 RBAC – A Candidate for Feature Modelling

In this paper, the different components of RBAC are considered as a new feature to be
added with the RBAC Core, which is the essential collection of RBAC elements.
Hence, three features are available and they are: role hierarchy (RH), static constraints
(SSD), dynamic constraints (DSD).These three features can be considered as stepwise
refinement of the RBAC core feature. With this Feature oriented view, the
development of RBAC models will be incremental and compositional. For example,
RBAC that has Role hierarchy and dynamic constraints will be built by integrating the
base model RBAC core with features RH and DSD. Therefore this approach will
greatly enhance the reusability of the base model and features, as well as provide
great flexibility for RBAC evolution to meet new system needs. Thus we have a nice
and elegant solution to issue of addressing the evolving access control needs. RBAC
feature is designed in the way that it encapsulates those and only those properties that
pertain to the functions of the features [19]. . From the design guidelines, the RBAC
feature has the following:

1. Core feature – Concrete feature consists of

a. Users – Abstract view
i. Roles- Concrete
b. Session- Optional
c. Permissions- Optional and Pre-conditions should be checked
i. Operations- Optional
ii. Objects- Concrete and Post conditions for analysis
2. Hierarchy feature – Optional Feature that requires
a. EITHER General -
b. OR Limited
c. Advanced – Optional
3. SOD feature – Optional that requires
a. Static SOD OR
b. Dynamic SOD

The following figure depicts the feature association model of RBAC:

 Feature Modeling of the Evolving Access Control Requirements 401

RBAC Feature

SOD

Core
Hierarchy

Dynamic Static

Users Session Permissions General Limited

Legend:
Excludes

Mandatory
Roles Operations Objects
Optional

Check Check Precondition

Roles consistency
Postcondition

Fig. 5. RBAC with Core, Hierarchy and SOD Components in Feature Association model

This model can be extended with consideration of other features such as temporal
access control and privacy aware policies.

5 Example

An example for integrating the access control requirements with the functional
requirements using feature conclave model is explained in this section. Consider a
Banking product line application that is supposed to have various types of services
and also evolving requirements based upon the services. Let us assume the
Transaction component of a Banking service. The features identified are:
1. Accountant, customer, Manager- Role feature
2. Pass book, Recorder, Transaction register, Cash Box- Object Feature
3. Update entries, Monitor, Cash handling – Permissions feature after checking
the Roles
4. Manager is superior to Accountant – Hierarchy feature that requires General
These features are then defined in the Feature Warehouse along with the relations and
dependencies. The dependencies are:
4. Feature Role consists of Permission features
5. Object feature requires Permission features
6. Feature Role extends Hierarchy feature
402 K. Shantha Kumari and T. Chithralekha

The Pre and Post conditions to be checked are:

7. Permission feature has to be added to the role feature before realizing with
the object feature
8. Hierarchy feature has to be added before permission feature is associated
with role feature.
9. Object feature should be the last feature after a role feature is associated with
the permission feature and hierarchy feature.
The following figure shows the partial view of the Banking domain requirements
integrated with Access control requirements.

Banking Feature

RBACCore
Hierarchy

Users Session Permissions

General

Roles Objects
Operations

Accountant
Update
Cash book Register
Manager

Customer

Fig. 6. Partial view of RBAC applied to Banking services using Feature Association model

5.1 Evolving Requirements

The term “Evolving requirements”, can be considered as either an extension of the
existing requirement or an entirely new requirement. As explained in section 3, both
these situations can easily be handled by the feature conclave framework. In this
example, let a new role called senior manager be introduced in the above explained
banking model. This role can easily be modeled, just by extending the Manager
feature and Hierarchy feature and defining a new Permission feature. Whenever this
feature association has to be applied for developing a new product, for example, to
support the ATM services, new features have to just define and can be combined with
existing features in Warehouse in ease of manner.
 Feature Modeling of the Evolving Access Control Requirements 403

6 Conclusion
Many related works on modeling the access control policies is done previously. This
paper discusses the Access control policies for software systems that need continuous
evolution of access control policies on need-basis. RBAC is chosen as the Access
control model, along with static and dynamic separation of duty constraints. The
Banking system is taken as case study and the composed model is produced, which is
a model with access control policies integrated along with the functional
requirements. This work can be further preceded by including the modeling
requirements for unpredicted changes of access control policies and the verification
mechanism using Propositional Logic for checking whether the composed model
enforces the access control continuously.

References
[1] Ray, I., France, R., Li, N., Georg, G.: An aspect-oriented methodology for designing
secure applications. Information and Software Technology 51, 846–864 (2009),
doi:10.1016/j.infsof.2008.05.004
[2] Georg, G., Ray, I., France, R.: Using Aspects to Design a Secure System. In: Proceedings
of Eighth IEEE International Conference on Engineering of Complex Computer Systems
2002. IEEE Computer Society, Los Alamitos (2002)
[3] Shin, M., Ahn, G.: UML-Based Representation of Role-Based Access Control. In: Proc.
of IEEE Int. Workshop on Enabling Technologies, pp. 195–200
[4] Warmer, J., Kleppe, A.: The Object Constraint Language, Getting Your Models Ready for
MDA, 2nd edn. Addison Wesley, Reading (2003)
[5] Kim, D.-K., Ray, I., France, R.B., Li, N.: Modeling role-based access control
using parameterized UML models. In: Wermelinger, M., Margaria-Steffen, T. (eds.)
FASE 2004. LNCS, vol. 2984, pp. 180–193. Springer, Heidelberg (2004)
[6] Priebe, T., Fernandez, E., Mehlau, J., Pernul, G.: A Pattern System for Access Control. In:
Proc. of Conf. on Data and Application Security, pp. 22–28
[7] Gamma, E., Helm, R., Johnson, R., Vlissides, J.: Design Patterns: Elements of Reusable
Object-Oriented Software. Addison-Wesley, Reading
[8] Jürjens, J.: UMLsec: Extending UML for secure systems development. In: Jézéquel, J.-
M., Hussmann, H., Cook, S. (eds.) UML 2002. LNCS, vol. 2460, pp. 412–425. Springer,
Heidelberg (2002)
[9] Harrison, M., Ruzzo, W., Ullman, J.: Protection in Operating Systems. Communications
of the ACM 19(8), 461–471 (1976)
[10] Lodderstedt, T., Basin, D., Doser, J.: SecureUML: A UML-based modeling language for
model-driven security. In: Jézéquel, J.-M., Hussmann, H., Cook, S. (eds.) UML 2002.
LNCS, vol. 2460, pp. 426–441. Springer, Heidelberg (2002)
[11] Ferraiolo, D.F., Sandhu, R., Gavrila, S., Richard Kuhn, D., Chandramouli, R.: Proposed
NIST Standard for Role-Based Access Control, NIST. ACM Transactions on Information
and System Security 4(3), 224–274 (2001)
[12] Fernandes, P., Werner, C., Teixeira, E.: An Approach for Feature Modeling of Context-
Aware Software Product Line. Journal of Universal Computer Science 17(5), 807–829 (2011)
[13] Kim, D.-K., Lu, L., Kim, S.: A verifiable modeling approach to configurable role-based
access control. In: Rosenblum, D.S., Taentzer, G. (eds.) FASE 2010. LNCS, vol. 6013,
pp. 188–202. Springer, Heidelberg (2010)
[14] Zaid, L.A., Kleinermann, F., De Taylor, O.: Feature Assembly framework: Toward
Scalable and reusable feature Models. In: VaMos 2011. ACM, New York (2011),
doi:978-1-4503-0570-9/01/11
 Fault Diagnosis of Pneumatic Valve Using PCA and
ANN Techniques

P. Subbaraj1 and B. Kannapiran2

1
Sri Nandhanam College of Engineering & Technology,
Tirupattur-635601, Vellore District, TamilNadu, India
2
Department of Instrumentation & Control Engineering,
Arulmigu Kalasalingam college of Engineering,
Anand Nagar, Krishnankoil-626190
Srivilliputhur, Virudunagar District, Tamilnadu, India

Abstract. Detection and Diagnosis of faults in pneumatic valve used in cooler

water spray system in cement industry is of great practical significance and
paramount importance for the safe operation of the plant. In this paper the
dimensionality reduction techniques such as principal component analysis
(PCA) are used to reduce the input features is proposed. PCA is used to extract
the primary features associated with the pneumatic valve used in cooler water
spray system. The training and testing data required for the dimensionality
reduction technique such as PCA model were created at normal and faulty
conditions of pneumatic valve in a real time laboratory experimental setup. The
performance of the developed PCA model is compared with the MLFFNN
(Multilayer Feed Forward Neural Network) trained by the back propagation
algorithm. From the simulation results it is observed that the performance of
PCA had the best classification properties when it is compared with the
performance of ANN.

Keywords: Fault diagnosis, Principal component analysis, Artificial Neural

Network, Back propagation, Pneumatic valve.

1 Introduction
The cooler water spray system plays an important role in cement industry. This is an
important task to reduce the temperature of the clinker at the output of the kiln. The
valve is one of the most important key component used in cooler water spray system.
Defects and malfunctions (simply called faults) of pneumatic valve will result in
significant economic loss. There is a crucial need for checking and monitoring the
condition of the pneumatic valve precisely since they are mostly subject to hazardous
environments, such as severe shocks, vibration, heat, friction, dust, etc When a
possible fault is detected in pneumatic valve, diagnosis is carried out to pinpoint the
fault. That is to say, diagnosis is a process of locating the exact cause(s) of a fault.
Once the fault has been detected, the maintenance engineer is to identify the
symptoms, analyze the symptomatic information, interpret the various error message
and indications and come up with the right diagnosis of the situation in terms of

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 404–413, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Fault Diagnosis of Pneumatic Valve Using PCA and ANN Techniques 405

which components may have caused the fault and the reason for the fault of the
components. Since the valve has many components built inside the chamber is highly
complex in nature. Fault diagnosis usually requires technical skill and experience. It
also requires extensive understanding of the pneumatic valve structure and operation,
and some general concepts of diagnosis. This requires an expert engineer to have a
domain-specific knowledge of maintenance and know the ‘inputs and-outputs’ of the
system [1]. In order to better equip with a non-expert to carry out the diagnosis
operations, it would be wise to present an approach to define the cause–symptom
relationship for quick comprehension and concise representation. Presently, many
diagnosis methods have been proposed to help maintenance engineer to do diagnosis
analysis. For example, expert system [2], neural network [3], soft-computing
technology (fuzzy logic, rough set and genetic algorithm) [4, 5] and their integration
method [6] are the popular approaches. Unfortunately, because of the bottleneck of
knowledge acquisition, the application of expert system is limited. The drawback of
these methods used for fault diagnosis is that the information used in the study is
limited in scope and correlation of factors is neglected.
Many researchers have developed diagnostic methods by using ANNs to detect the
problems of march-motors, electric motors [7], rotating machine parts [8], gears and
gearboxes [9], Fisher Rosemount valves [10, 11], and compressors. But the neural
network approach needs lot of data to develop the network before being put to use for
real time applications. There are usually many irrelevant variables in the sampled data
from real life fault diagnostic measurements. Irrelevant and redundant attributes in the
input not only complicate the network structure, but also degrade the performance of
the networks. By selecting only the relevant variables as input features and excluding
irrelevant ones, higher performance is expected with smaller computational effort.
Therefore, efficient methods for dimensionality reduction must be developed. A
feasible option for solving this problem is to use feature extraction technique such as
principle component analysis (PCA). PCA [12, 13] is a method based on the
projection of the data into a low dimensional space that accurately characterises the
state of the system. It produces a lower dimensional representation in a way that
preserves the correlation structure between the process variables. This paper deals
with the design and development artificial neural network based model for the fault
detection in pneumatic valve used in cooler water spray system in cement industry,
with PCA for reducing the input dimension.
The paper is organized as follows: in the next section, the description of the
experimental system for this study is outlined. Section 3 describes the fault detection
in pneumatic actuator. Sections 4 describe the development of ANN model for fault
diagnosis. Sections 5 demonstrate the Principal component analysis and section 6
describe the result and discussion and finally, in Section 7, conclusions are drawn
from the work.

2 System Description
In the cement industry the output of kiln is the clinker product which is of high
temperature up to 1100 ºC. It is necessary to reduce the temperature of the clinker by
using cooler fans and the water spray system. The functions of the clinker cooler
406 P. Subbaraj and B. Kannapiran

process are to maximize the heat recovery to the kiln process and to minimize the
ultimate clinker temperature.
The measurement of clinker temperature is done by using thermocouple sensor.
The direct measurement of clinker temperature is not possible on a continuous basis.
Hence, the temperature of cooler vents gas temperature is measured. The sensor
output is proportional to the temperature of the clinker. When the temperature of the
clinker product at the outlet of kiln is up to 1100 ºC, then the temperature is reduced to
750 ºC at the grate cooler section by using external blower fans. It is necessary to
reduce the temperature of the hot gas up to 250 ºC before given to the ESP
(Electrostatic Precipitator) section. If the temperature of the hot gas exceeds beyond
250 ºC, then the anode and cathode plates of ESP gets damaged.

)ORZ RI +RW

*DVWR(63
9 

5
9 

)=
9 
 *
5
9 
39 )7
9  )=

9 5
9 
$)5
3 67
)=
&$/

3 

 67 6WRUDJH7DQN
 9 5HWXUQ9DOYH
 999 'HOLYHU\9DOYHIRUZDWHUIORZ
 999 'HOLYHU\9DOYHIRUDLUIORZ
 $)5 $LU)LOWHU5HJXODWRU
 &$/ &RPSUHVVHG$LUOLQHIURP3ODQW
 33  3XPS
 55 )ORZ5HGXFHU
 39 3QHXPDWLF9DOYH
)7 )ORZ7UDQVPLWWHU

 )=)= )ORZ1R]]OH%DQNV
 * 3UHVVXUHJDXJH
 67 6WRUDJH7DQN

Fig. 1. Schematic layout of the cooler water sprays system

In order to prevent this damage to the ESP, cooler water spray system is used in
cement industry. The schematic layout of the cooler water spray system setup is
shown in Figure 1. The hot air recovered from this cooling process is recycled back
 Fault Diagnosis of Pneumatic Valve Using PCA and ANN Techniques 407

into the kiln or preheated system to maintain the thermal energy of the hot gas. The
type of valve used in the water spray system is automatic equipment designed to
regulate the flow rate in a pipe system. Then the water is injected through the
pneumatic valve using pump and the water reaches the flow nozzle for water spray
process inside the hot gas flow duct line. The description about this system is well
explained in [14]. Next section deals with fault detection in pneumatic actuator in
cement industry.

3 Fault Detection in Pneumatic Actuator

A system that includes the capability of detecting and diagnosing faults is called the
‘fault diagnosis system’. Fault detection and diagnosis are important tasks in cement
industry. It deals with the timely detection, diagnosis and correction of abnormal
condition of faults in the plant.

Table 1. List of various types of faults


1DPHRIWKH 6\PERO
7\SHVRIIDXOW 1DPHRIWKHIDXOW 6\PEROV
IDXOW V
YDOYHFORJJLQJ H[WHUQDOOHDNDJH
) EXVKLQJFRYHUV )
WHUPLQDOV
YDOYHRUYDOYHVHDW LQWHUQDOOHDNDJH
) )
&RQWUROYDOYH HURVLRQ YDOYHWLJKWQHVV
IDXOWV YDOYHRUYDOYHVHDW PHGLXP
VHGLPHQWDWLRQ ) HYDSRUDWLRQRU )
FULWLFDOIORZ
LQFUHDVHGRIYDOYH 
) 
RUEXVKLQJIULFWLRQ
WZLVWHGVHUYR VHUYRPRWRU
V
PRWRU
VSLVWRQURG ) GLDSKUDJP )
3QHXPDWLF SHUIRUDWLRQ
VHUYRPRWRUIDXOWV VHUYRPRWRU
V VHUYRPRWRU
V
KRXVLQJRUWHUPLQDOV ) VSULQJIDXOW )
WLJKWQHVV
HOHFWURSQHXPDWLF SUHVVXUHVHQVRU
) )
3RVLWLRQHU WUDQVGXFHUIDXOW(3 IDXOW37
IDXOWV URGGLVSODFHPHQW 
) 
VHQVRUIDXOW'7
3RVLWLRQHUVXSSO\ IXOO\RUSDUWO\
SUHVVXUHGURS ) RSHQHGE\SDVV )
YDOYHV
*HQHUDO LQFUHDVHRI IORZUDWHVHQVRU
IDXOWVH[WHUQDO SUHVVXUHRQYDOYH IDXOW)7
) )
IDXOWV LQOHWRUSUHVVXUHGURS
 RQYDOYHRXWSXW
SUHVVXUHGURSRQ 
YDOYHDWLQOHWRU
) 
LQFUHDVHRISUHVVXUH
RQYDOYHRXWSXW
408 P. Subbaraj and B. Kannapiran

Early detection and diagnosis of plants while the plant is still operating in a
controllable region can help avoid abnormal event progression and reduce
productivity loss. In order to accomplish these tasks, information that reflects a
change from the normal behaviour of the process has to be obtained. This is generally
called symptom generation. Any method of fault diagnosis must characterize how
abnormal symptoms (measurements) are related to faults (malfunctions). The major
categories of faults in the pneumatic valve include Control valve faults, Pneumatic
servo-motor faults, Positioner faults, General faults / External faults. The individual
fault of each category along with their symbol is listed in Table 1. The sources of each
fault are given in [15]. A complete description of the faults and the way they affect
the valve can be found in [16].

4 Development of ANN Models for Fault Diagnosis

The proposed methodology for fault detection in pneumatic actuator is based on using
back propagation for detecting the normal and abnormal conditions of the given
parameters, which leads to various faults. The main purpose of selecting ANN as a
tool is inability to form a mathematical relationship due to the nonlinearity between
the inputs and the outputs, good generalization ability, fast real time operation and to
perform the complicated mapping without functional relationship.
The following issues are to be addressed while developing the neural network
based model for fault detection in pneumatic actuator in cooler water spray system.

4.1 Selection of Input and Output Variables

For the application machine learning approaches, it is important to properly select the
input variables, as ANN’s are supposed to learn the relationships between input and
output variables on the basis of input-output pairs provided during training. In neural
network based fault detection model, the input variables represent the operating state
of the pneumatic actuator, and the output is the type of fault present in the valve.

4.2 Training Data Generation

The generation of training data is an important step in the development of ANN
models. To achieve a good performance of the neural network, the training data
should represent the complete range of operating conditions of the pneumatic actuator
which contains all possible fault occurrences.

4.3 Data Normalization

If the generated data are directly fed to the network as training patterns, higher valued
input variables may tend to suppress the influence of smaller ones. Data normalization
compresses the range of training data between 0 and 1 or -1 to +1 depending on the
type of transfer function. The input and output data are normalized using the
expression,
X n = ( X − X min ) ( X max − X min ) +starting value (1)
Where Xn is the normalized value of the data and Xmin and Xmax are the minimum and
maximum values among all the values of the data.
 Fault Diagnosis of Pneumatic Valve Using PCA and ANN Techniques 409

5 Principal Component Analysis

Dimensionality reduction can be achieved either by eliminating data closely related
with other data in the set, or combining data to make a smaller set of features. The
feature extraction techniques used in this study are principal components analysis.
Using the example of projecting data from two dimensions to one, a linear projection
requires the optimum choice of projection to be a minimization of the sum-of-squares
error. This is obtained first by subtracting the mean x of the data set. The covariance
matrix is calculated and its eigenvectors and eigen-values are found. The eigenvectors
corresponding to the M largest eigen-values are retained, and the input vectors x n are
subsequently projected onto the eigenvectors to give components of the transformed
vectors z n in the M-dimensional space. Retaining a subset M < d of the basis vectors
μi so that only M coefficients z i are used allows for replacement of the remaining
coefficients by constants bi. This allows each x vector to be approximated by an
expression of the form:
−
 
m d
x= z μ + bμ
i −1 i i 1− m +1 i i (2)
Where μ i represents a linear combination of d orthonormal vectors

6 Results and Discussion

This section presents the development and testing of MLFFNN and PCA models for
fault detection on Pneumatic valve used in cooler water spray system. Using real time
simulation on laboratory experimental setup of pneumatic valve in cooler water spray
system, the required data was generated.
The parameters are collected directly from the real time system which is interfaced
with personal computer using Data Acquisition (DAQ) card. The Photograph of the
pneumatic actuator system interfaced with PC using DAQ is shown in Figure 2.The
observations are made for normal operating condition and the procedure for observing
the parameters are repeated for different faults.

Fig. 2. Shows photograph of the pneumatic actuator system interfaced with PC using DAQ
410 P. Subbaraj and B. Kannapiran

6.1 Effect of Artificial Neural Network

The data contains 5 input features, which is given in Table 2 and one output that is
labeled as either normal or as a fault, with exactly one specific fault. The total number
of data generated is 1000, which contain 25% normal patterns and 75% of patterns
with faults belonging to the various types of faults listed in Table 1. Among them, 750
patterns are used for training and 250 patterns are used for testing. The testing data
comprises of both normal and abnormal (faulty) data, which are totally different from
the training data.

Table 2. Name of the input features in the experimental system

Input Feature Name

code
1. Differential Pressure Transmitter output(mA)
2. Temperature of water form RTD output (outlet)
3. Pressure gauge reading(Inlet Pressure)
4. Potentiometer (Stem movement)
5. Input and output power of the pump

Table 3. Network performance for each fault category of control valve faults

Type of fault Name of the Fault Symbol Training Testing Efficiency

pattern pattern
Valve Clogging F1 75 25 100
Valve seat erosion F2 75 25 97.6
Valve seat sedimentation F3 75 25 100
Control valve
Bushing friction F4 75 25 99.4
faults
External leakage F5 75 25 99.1
Internal leakage F6 75 25 99.3
Critical flow F7 75 25 99
Considering all fault class in control
Valve faults(total classification accuracy) 75 25 98.6

Table 4. Various parameters of the ANN model

Description of parameters values

Number of training patterns 75%
Number of testing patterns 25%
Transfer function used Logarithmic sigmoidal, Linear
Training Time 12.7344 seconds
Mean Square Error in Training 0.0018
Mean Square Error in Testing 0.0017
Percentage of fault classification 98.6%
 Fault Diagnosis of Pneumatic Valve Using PCA and ANN Techniques 411

Fig. 3. Training Performance of the Network

After training, the generalization performance of the network is evaluated with the
250 test data that contain the combination of both normal as well as all types of fault
categories. The trained neural network classified 246 data correctly, which shows an
overall detection rate 98.6%. Table 3 shows the network performance for each fault
category of control valve faults. The network is trained with least mean square
algorithm until it reaches the mean square error of 0.001. The mean square error
achieved during training is 0.0018. During testing, the mean square error achieved by
the network is 0.0017. With 9×10 hidden nodes, the network took 12.7344 s to reach
the error goal. This is shown in Table 4 it is found that the network has correctly
classified all the data during the testing stage. This shows that the trained ANN is able
to produce the correct output even for the new input.
The neural network model was trained using the back propagation algorithm with
the help of MATLAB neural network toolbox. At the end of training process, the
model obtained consists of the optimal weight and the bias vector. The performance
of the ANN during training is shown in Figure 3.

6.2 Effect of Feature Extraction

The performance comparison between the various parameters of ANN and PCA
model is shown in Table 5. On comparing the performance of without dimensionality
reduction, it is found that the mean square error in testing and training of using PCA
is incredibly low when compared with ANN and the accuracy of fault classification is
high in using PCA when compared with ANN. The classification rate using the
original parameters is 98.6%. The classification rate with feature extraction through
PCA is 100%.
Also, the obtained results proved that the network with PCA technique performed
better than ANN without PCA. The reduced performance of the classifier without
feature extraction is due to the existence of irrelevant and redundant feature
parameters existing in the data set. The irrelevant features burden the classifier and
tend to decrease its performance. By using PCA the useful features are extracted from
the original input variables.
412 P. Subbaraj and B. Kannapiran

Table 5. Performance comparison of ANN model with and without PCA

Description of parameters Without dimensionality With dimensionality

reduction reduction
Number of training patterns 75% 75%
Number of testing patterns 25% 25%
Functions logarithmicsigmoidal, logarithmicsigmoidal,
Linear Linear
MSE in Training 0.0018 9.9831 e-004
MSE in Testing 0.0017 8.1919 e-004
Training Time 12.7344 seconds 11.7188
Percentage of fault classified 98.6 % 100

7 Conclusion
This paper has presented a neural network based approach with principal component
analysis based feature extraction for fault diagnosis of pneumatic actuator used in
cooler water spray system in cement industry. The data required for the development
of neural network model have been obtained through the laboratory grade
experimental setup. Totally 19 faults from the pneumatic actuator were considered in
the developed model. A key issue in neural network based approach is identifying a
representative set of features from which to develop the network for a particular task.
This problem has been addressed by feature extraction through principal component
analysis. Simulation results show that the PCA based approach is very much
effective in identifying the important key features of the network. Based on the results
obtained, it is observed that the performance of the neural network model has been
significantly improved by reducing the input dimension. The effectiveness of the
proposed method has been demonstrated through different fault condition in
pneumatic actuator system.

References
1. Patel, S.A., Kamrani, A.K.: Intelligent decision support system for diagnosis and
maintenance of automated systems. Computers Industrial Engineering 30(2), 297–319
(1996)
2. Yang, B.S., Lim, D.S., Tan, C.C.: VIBEX: an expert system for vibration fault diagnosis of
rotating machinery using decision tree and decision table. Expert System with
Application 28, 735–742 (2005)
3. Chen, C.Z., Mo, C.T.: A method for intelligent fault diagnosis of rotating machinery.
Digital Signal Processing 14, 203–217 (2004)
4. Gayme, D., Menon, S., Ball, C., et al.: Fault detection and diagnosis in turbine engines
using fuzzy logic. In: 22nd International Conference of Digital Object Identifier, vol. 24-
26, pp. 341–346. North American Fuzzy Information Processing Society (2003)
5. Wang, Y.G., Liu, B., Guo, Z.B., et al.: Application of rough set neural network in fault
diagnosing of test-launching control system of missiles. In: Proceedings of the Fifth World
Congress on Intelligent Control and Automation, Hangzhou, PR, China, pp. 1790–1792
(2004)
 Fault Diagnosis of Pneumatic Valve Using PCA and ANN Techniques 413

6. Xiang, X.J., Shan, J.F., Fuchs, F.W.: The study of transformer failure diagnose expert
system based on rough set theory. In: The Fourth International IPEMC Power Electronics
and Motion Control Conference, vol. 2, pp. 533–536 (2004)
7. Bayir, R., Bay, O.F.: Serial wound starter motor faults diagnosis using artificial neural
network. In: IEEE ICM 2004, Istanbul, pp. 194–199 (2004)
8. Rajakarunakaran, S., Venkumar, P., Devaraj, D., Rao, K.S.P.: Artificial neural network
approach for fault detection in rotary system. Applied Soft Computing 8, 740–748 (2008)
9. Chen, D., Wang, J.W.: Classification of wavelet map patterns using multilayer neural
networks for gear fault detection. Mechanical Systems and Signal Processing 16, 695–704
(2000)
10. Karpenko, M., Sepehri, N.: Neural network classifiers applied to condition monitoring of a
pneumatic process valve actuator. Engineering Applications of Artificial Intelligence 15(3-
4), 273–283 (2002)
11. Karpenko, M., Sepehri, N., Scuse, D.: Diagnosis of process valve actuator faults using a
multilayer neural network. Control Engineering Practice 11(11), 1289–1299 (2003)
12. Chang, Y.-W., Wang, Y.-C., Liu, T., Wang, Z.-J.: Fault diagnosis of a mine hoist using
PCA and SVM techniques. China Univ. Mining and Technology (18), 327–331 (2008)
13. de Moura, E.P., Souto, C.R., Silva, A.A., Irmao, M.A.S.: Evaluation of principal
component analysis and neural network performance for bearing fault diagnosis from
vibration signal processed by RS and DF analyses. Journal of Mechanical Systems and
Signal Processing (2010)
14. Subbaraj, P., Kannapiran, B.: Artificial Neural Network Approach for Fault Detection in
Pneumatic Valve in Cooler Water Spray System. Int. Journal of Comp. Applications 9(7),
0975–8887 (2010)
15. Koj, J.: The Fault Sources of Pneumatic Servo-Motor-Control Valve Assembly. In:
Proceedings of the III Polish National Conference on Diagnosis of Industrial Processes,
Jurata, Poland, pp. 415–419 (1998)
16. Louro, R.: Fault Diagnosis of an Industrial Actuator Valve. MSc Dissertation, Technical
University of Lisbon, Portugal (2003)
 A Generalized Classification Approach for Indian Sign
Language Datasets Using Instance Based Classifiers

M. Krishnaveni and V. Radha

Department of Computer Science,

Avinashilingam Deemed University for Women,
Tamilnadu, India
[email protected]
[email protected]

Abstract. Sign language is the primary means of communication in the deaf

community. The importance of sign language is emphasized by growing
research activities and various funding agencies which deal with the technical
aspects of deaf people's communication needs. Signing is primarily done with
the hands. In technical terms the hands are used to describe the so-called
manual parameters, which are pose, posture, position and motion. General view
based sign language recognition systems extract these parameters by a single
camera view because it seems to be user friendly and hardware complexity;
however it needs a high accuracy classifier for classification and recognition
purpose. The decision making of the system in this work employs statistical
classifiers namely Navie bayes and K-NN to recognize the sign language
isolated signs. Indian sign language datasets is used for the training and
performance evaluation of the system.It compares the ability to deploy the K-
NN and the Naïve Bayesian classifiers in solving the sign classification
problem. The impact of such study may reflect the exploration for using such
algorithms in other similar applications such as text classification and the
development of automated systems.

Keywords: Sign Language, boundary based, K-NN classifier, Navie Bayes,

classification accuracy.

1 Introduction
Sign language is important in humankind that is showing an increasing research
interest in eradicating barriers faced by differently abled people in communicating
and contributing to the society [1]. Automation over sign language recognition
systems can greatly facilitate the vocal and the non-vocal communities which can be
equivalently best and successive as speech-recognition systems [2]. Despite common
misconceptions, sign languages are complete natural languages, with their own syntax
and grammar. However, sign languages are not universal. As is the case in spoken
language, every country has got its own sign language with high degree of
grammatical variations. The sign language used in India is commonly known as
Indian Sign Language (henceforth called ISL)[6].Linguistic studies on ISL were

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 414–422, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 A Generalized Classification Approach for Indian Sign Language Datasets 415

started around 1978 and it has been found that ISL is a complete natural language,
instigated in India, having its own morphology, phonology, syntax, and grammar. ISL
is not only used by the deaf people but also by the hearing parents of the deaf
children, the hearing children of deaf adults and hearing deaf educators. Therefore the
need to build a automation system that can associate signs to the words of spoken
language, and which can further be used to learn ISL, is significant. This paper
proposes a system for recognizing Indian sign language using the view based
approach and the statistical classifier algorithm as a learning tool. These classifiers
take the advantage of accuracy, ease of implementation and operational robustness.
The paper is arranged as follows. Section 2 gives a brief overview of Indian sign
language. Section 3 describes the feature extraction methodology, by using boundary
detection techniques. Section 4 deals with the short description of classifiers used.
Section 5 gives the performance evaluation of the classifiers using ISL datasets.
Finally, section 6 summarizes the framework and future work that can be adopted.

2 Overview of Indian Sign Language

Indian Sign languages are rich, faceted language, and their full complexity is beyond
current gesture recognition technologies[6]. The interpersonal communication
problem between signer and hearing community could be resolved by building up a
new communication bridge integrating components for sign[5]. Several Significant
problems specific to Automatic Sign Recognition are i) Distinguishing gestures from
signs ii) Context dependency (directional, verbs, inflections, etc) iii) Basic unit
modeling (how do we describe them?) iii) Transitions between signs (Movements)


,QGLDQ6LJQV


 KŶĞ,ĂŶĚĞĚ EŽŶDĂŶƵĂů
dǁŽ,ĂŶĚĞĚ

^ƚĂƚŝĐ DŽǀĞŵĞŶƚ
 ^ƚĂƚŝĐ DŽǀĞŵĞŶƚ
 
DĂŶƵĂů EŽŶͲDĂŶƵĂů



dLJƉĞϬ dLJƉĞϭ



DĂŶƵĂů EŽŶ DĂŶƵĂů

Fig. 1. Framework for ISL Dictionary

416 M. Krishnaveni and V. Radha

iv) Repetition (Cycles of movement). The ISL dictionary is been build by Sri
Ramakrishna mission Vidyalaya College of Education, Coimbatore which as split the
ISL into five parameters namely handshapes, locations, Orientation, Location,
Movements and Facial Expression. Figure 1 depicts the framework for ISL dictionary.

3 Feature Extraction Methodology

The great variability in gestures and signs, both in time, size, and position, as well as
interpersonal differences, makes the recognition task difficult [11]. By extracting
features from image processing sequence classification can be done by a
discriminative classifier. Gestures, particularly in sign language, involve significant
motion of the hands[5]. Thus, in developing a sign language recognition system, it is
important to model both the motion (temporal characteristics) and shape (spatial
characteristics) of the hand. In this research work only the spatial characteristics of
the hand are of concern. Figure 2 shows the schematic view of gesture recognition
process.

Fig. 2. Schematic view of gesture recognition process

Feature extraction is the process of generating a set of descriptors or characteristic

attributes from a binary image. Most of the features used in existing sign language
recognition systems focus only on one aspect of the signing like hand movements or
facial expressions. Figure 3 shows the feature extraction concept.

Fig. 3. Hand feature extraction

The strategy used for segmentation is boundary segmentation method which traces
the exterior boundaries of objects, as well as boundaries of holes inside these objects,
in the binary image [3]. The two important need for this method of sign extraction is
First, it keeps the overall complexity of the segmentation process low. Secondly, it
eliminates candidates that may produce smooth continuations, but otherwise are
inconsistent with the segments in the image. This simplifies the decisions made by the
segmentation process and generally leads to more accurate segmentation[7]. The
 A Generalized Classification Approach for Indian Sign Language Datasets 417

feature extracted from the database used is mean intensity, area, perimeter, diameter,
centroid.
Mean Intensity: The mean intensity μ in the selected region of interest (ROI) is
given in eqn (1):
1
μ =
N  x,y
R O I ( x, y )dxdy (1)

Area: The area of an object is a relatively noise-immune measure of object size,

because every pixel in the object contributes towards the measurement. The area can
be defined using eqn (2)

A(S ) =  I ( x, y )dydx (2)

Where, I(x,y) = 1 if the pixel is within a shape,(x,y) ∈ S,

0 otherwise.
Perimeter: The perimeter measurement gives the smooth relative boundaries and the
perimeter of the region is defined as by the eqn (3)

P (S ) = t
x 2 (t ) + y 2 (t )d t (3)

Diameter: The distance around a selected region is called the circumference. The
distance across a circle through the center is called the diameter. is the radius of the
circumference of a circle.. Thus, for any circle, if you divide the circumference by the
diameter, you get a value close to . This relationship is expressed in the following
eqn (4):

(4)

Where, C is circumference and d is diameter.

Centroid: It specifies the center of mass of the region. Centroid is the horizontal
coordinate (or x-coordinate) of the center of mass, and the second element is the
vertical coordinate (or y-coordinate) and it is written as given in eqn (5)

(5)

4 Short Description on Classifiers

Classification (generalization) using an instance-based classifier can be a simple
matter of locating the nearest neighbour in instance space and labelling the unknown
instance with the same class label as that of the located (known) neighbour [4]. This
418 M. Krishnaveni and V. Radha

approach is often referred to as a neighbour classifier. Classifier always

tries to improve the classification rate by pushing classifiers into an optimised
structure[10].

4.1 K-NN Classifier

The K-Nearest Neighbor classifier is a supervised learning algorithm where the result
of a new instance query is classified based on majority of the K-nearest neighbor
category. The purpose of this algorithm is to classify a new object based on attributes
and training samples[8]. The Nearest neighbor classifier is based on learning by
analogy. The training samples are described by n-dimensional numeric attributes.
Each sample represents a point in an n-dimensional pattern space. More robust
models can be achieved by locating k, where k > 1, neighbours and letting the
majority vote decide the outcome of the class labelling. A higher value of k results in
a smoother, less locally sensitive, function. The Nearest Neighbor classifier can be
regarded as a special case of the more general K-Nearest Neighbors classifier,
hereafter referred to as a K-NN classifier.
The data for KNN algorithm consists of several multivariate attributes names Xi
that will be used to classify the object Y. The data of KNN can have any measurement
scale from ordinal, nominal, to quantitative scale, this study deals only with
quantitative Xi and binary (nominal) Y. All training samples are included as nearest
neighbors if the distance of this training sample to the query is less than or equal to
the Kth smallest distance. In other words, the distances are sorted of all training
samples to the query and determine the Kth minimum distance. The unknown sample
is assigned the most common class among its K-Nearest Neighbors. These K training
samples are the closest k nearest neighbors for the unknown sample. Closeness is
defined in terms of Euclidean distance, where the Euclidean between two points, X =
(x1, x2,….xn) and Y= (y1,y2,…yn) is given in eqn (6)

, 6

The drawback of increasing the value of k is when k approaches n, where n is the size
of the instance base, the performance of the classifier will approach that of the most
straightforward statistical baseline, the assumption that all unknown instances belong
to the class most frequently represented in the training data. The high degree of local
sensitivity makes nearest neighbor classifiers highly susceptible to noise in the
training data.

4.2 Naive Bayesian Classifier

A Naive Bayes classifier assigns a new observation to the most probable class,
assuming the features are conditionally independent given the class value [9]. It can
outperform more sophisticated classification methods by categorizing incoming
objects to their appropriate class. The Naive Bayesian classifiers can handle a random
number of independent variables whether continuous or categorical.
 A Generalized Classification Approach for Indian Sign Language Datasets 419

It classifies data in two steps:

1. Training step: Using the training samples, the method estimates the parameters of a
probability distribution, assuming features are conditionally independent given the
class.
2. Prediction step: For any unseen test sample, the method computes the posterior
probability of that sample belonging to each class. The method then classifies the test
sample according the largest posterior probability.
Bayes theorem used, takes the eqn as given in (7) and (8)

P (H|X) = P(X|H) P(H) / P(X) (7)

It can also be expressed as

|
| (8)

Where P(X) is constant for all classes, only P(X|Ci)P(Ci) need be maximized.
The class-conditional independence assumption greatly simplifies the training step
since estimation can be done using one-dimensional class-conditional density for each
feature individually. This assumption of class independence allows the Naive Bayes
classifier to better estimate the parameters required for accurate classification while
using less training data than many other classifiers. This makes it particularly
effective for datasets containing many predictors or features. Since the build process
for Naive Bayes is parallelized it can be used for both binary and multiclass
classification problems.

5 Performance Evaluation

The proposed classifier Navie bayes is used to justify the objects using new methods
to get a maximum accuracy comparing to K-NN classifier. The features are the
parameters extracted from the sign images which are taken from normal camera. In
each hand posture, a measure of properties like area, mean intensity, centroid,
perimeter and diameter are taken; the classifier then uses these properties to determine
the sign in different angles. They estimate the probability that a sign belongs to each
of the target classes that is fixed. In the training phase, the training set is used to
decide how the parameters must be weighted and combined in order to separate the
various classes of signs. The classifier used for classification is validated using
sensitivity, specificity error rate, predictive value, likelihood value, plotting the
classification and misclassification error rate according to the sample datasets.
To keep track of the performance during the validation of classifiers Table 1
mentioned values are the measures taken into consideration. Conducted experiments
with Indian sign language datasets shows that the Navie bayes classifier has better
performance than K-NN and the current implementation for both classifiers yield
good results when compared to other implementations by other researchers.
420 M. Krishnaveni and V. Radha

Table 1. Performance evaluation of NB and KNN Classifiers

S.No Performance Measures NB Performance KNN Performance

1. Classified rate 1.0000 0.9500

2. Sensitivity 0.7143 0.5714
3. Specificity 0.8571 0.6857
4. Error rate 0.3429 0.6857
5. Inconclusive rate 0 0
6. Positive predictive value 0.5556 0.3077
7. Negative Predictive Value 0.9231 0.8636
8. Negative likelihood 0.3333 0.6316
9. Positive Likelihood 5.000 1.7778
10. Prevalence 0.2000 0.2000

NB Class Performnace
5

target
class
4.5

3.5
Target Classes

2.5

1.5

1
0 5 10 15 20 25 30 35
Number of Samples

Fig. 4. Navie Bayes class Performance

Figure 4 and 5 depicts the classifier performance based on the datasets taken and
the target class is less achieved by K-NN classifier than the proposed Navie bayes
classifier.
 A Generalized Classification Approach for Indian Sign Language Datasets 421

KNN Class performance

5
Target
Class
4.5

3.5
Target Classes

2.5

1.5

1
0 5 10 15 20 25 30 35
Number of Samples

Fig. 5. KNN class Performance

6 Conclusion
The classifier methods described in this paper cover a large variety of practical
applications. This help in understanding the current state of knowledge in Indian sign
language research area. Additionally attempts are made to identify the main directions
leading towards the standardised procedure of classifier system design. However, the
results in this work are biased by the size of the database: on the one hand, the lack of
training data and the large amount of singletons leads to a very difficult task. The
methods used in this work are focused towards good recognition accuracy and not
toward real-time performance and thus it meets the sample experimental
requirements. This paper is also a contribution to the research in Indian Sign
Language, which had very limited research in its automated recognition.

Acknowledgments. The authors thank RamaKrishna Mission Vidyalaya,

International Human Resource Development Centre (IHRDC for the disabled) for
providing Indian sign language dictionary.

References
1. Phu, J.J., Tay, Y.H.: Computer Vision Based Hand Gesture Recognition Using Artificial
Neural Network. Faculty of Information and Communication Technology. Universitiy Tunku
Abdul Rahman (UTAR), MALAYSIA. [email protected], [email protected]
2. Salleh, N.S.M., Jais, J., Mazalan, L., Ismail, R., Yussof, S., Ahmad, A., Anuar, A.,
Mohamad, D.: Sign Language to Voice Recognition: Hand Detection Techniques for
Vision-Based Approach. Current Developments in Technology-Assisted Education (2006)
422 M. Krishnaveni and V. Radha

3. Kovář, J.P., Vlček, M.: Still Image Objective Segmentation Evaluation using Ground
Truth. In: 5th COST 276 Workshop, pp. 9–14 (2003)
4. Elmezain, M., Al-Hamadi, A., Appenrodt, J., Michaelis, B.: A Hidden Markov Model-
Based Isolated and Meaningful Hand Gesture Recognition. International Journal of
Computer Systems Science and Engineering 5(2) (2009)
5. Hassanpour, R., Wong, S., Shahbahrami, A.: VisionBased Hand Gesture Recognition for
Human Computer Interaction: A Review. In: IADIS International Conference Interfaces
and Human Computer Interaction (2008)
6. Dasgupta, T., Shukla, S., Kumar, S.: A Multilingual Multimedia Indian Sign language
Dictionary Tool. In: The 6th Workshop on Asian Language Resources (2008)
7. Haralick, R.M., Shapiro, L.G.: Survey: Image Segmentation Techniques. Computer
Vision, Graphics, and Image Processing 29, 100–132 (1985)
8. Laaksonen, J., Oja, E.: Classification with learning k-nearest neighbors. In: Proceedings of
ICNN 1996, USA, pp. 1480–1483 (June 1996)
9. Ruta, D., Gabrys, B.: An Overview of Classifier Fusion Methods. Computing and
Information Systems, 1–10 (2000)
10. Danesh, A., Moshiri, B., Fatemi, O.: Improve text classification accuracy based on
classifier fusion methods. In: Proceeding of the 10th International Conference on
Information Fusion, July 9-12, pp. 1–6. IEEE Computer Society, USA (2007),
doi:10.1109/ICIF.2007.4408196.
11. Xu, L., Krzyzak, A., Suen, C.Y.: Methods of Combining Multiple Classifiers and Their
Applications to Handwriting Recognition. IEEE Transactions on Systems, Man, and
Cybernetics 22, pt. 3, 418–435 (1992)
 Agent Based Adaptive Learning System

R. Gowri1, S. Kanmani2, and T.T. Sampath Kumar3

1
Sri Manakula Vinayagar Engineering College,
Pondicherry India
[email protected]
2
Pondicherry Engineering College Pondicherry
[email protected]
3
Sri Manakula Vinayagar Engineering College
[email protected]

Abstract. This paper proposes the agent based intelligent, adaptive learning or
tutoring system (IATS). The proposed architecture of the agent based system
provides the framework for the intelligent tutoring system. We define many
agents like tutoring agent, student agent, student clustering agent, resource
classification agent. The system model is developed using the O-MaSE
development model. The clustering is done by the combination of Evolutionary
clustering methods and fuzzy c-means clustering. The classification of the
resources is representation of the resources in the tree based hierarchical
structure from which the adaptive resources are generated. The system provides
the adaptive course materials, adaptive user interface, and adaptive
examinations to the learner.

Keywords: Agent, Intelligent tutoring system, adaptive course content, Jade,

fuzzy c-means clustering.

1 Introduction
Due to the increasing population of the Internet users, distance education has become
a popular choice for various areas. Learning is no longer restricted to specific time
slices or locations. In the e-learning environment, the design of learning materials
requires more considerations for individual learners. The traditional arrangement of
fixed content and learning sequence for all learners may not be suitable for all
learners in the distributed environment. Intelligent tutoring systems mostly emphasize
the importance of adaptive learning process. Intelligent tutoring systems are the
outgrowth of the earlier computer-aided instruction models. Most of Intelligent
tutoring systems are expected to be domain independent and task independent as well.
Among various categories of intelligent tutoring systems, the mostly known function
is to monitor students’ learning process and provide suitable learning path.
Because of the adaptive characteristics of intelligent agents, researchers have tried
to apply the capability of agents in the world of intelligent tutoring systems. Multi-
agent systems (MAS) [1] [3] [10] bring several advantages to the implementation of

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 423–431, 2011.
© Springer-Verlag Berlin Heidelberg 2011
424 R. Gowri, S. Kanmani, and T.T. Sampath Kumar

intelligent tutoring systems since they deal well with distribution, cooperation, and
integration of computer software components. The interaction protocol of the agent
community would be another important issue to be considered.
Clustering is the process of grouping a data set in a way that the similarity between
data within a cluster is maximized while the similarity between data of different
clusters is minimized. Clustering is used for pattern recognition in image processing,
and usually requires a high volume of computation. This high volume computation
requires considerable amount of memory which may lead to frequent disk access,
making the process inefficient. With the development of affordable high performance
parallel systems, parallel algorithms may be utilized to improve performance and
efficiency of such tasks. Several Clustering Algorithms are proposed for the
Clustering of students [12-14] with several learner’s behaviors and parameters.

2 Intelligent Tutoring Systems

The traditional intelligent tutoring systems model contains four components,
including the domain model, the student model, the teaching model, and a learning
environment or user interface. The domain model is also called expert model, which
is used to determine the learning status of the student. The teaching (instructional)
model is the central part of the architecture and interacts with all other models. The
student model contains all learning related information used by the expert model.
Learning path planning is one of the most important tasks in the intelligent tutoring
systems. Y. Caseau [2] defines the routing strategies for message handling that
maximize the ability of the EAI system to meet these requirements, in a self-adaptive
and self-healing manner,

2.1 Agent-Based Solutions for ITS

In this research we propose a generic framework for the Tutoring System. We make
this tutoring system as Intelligent Tutoring system by giving the adaptive property.
This framework deals with the level of adaptation of the study materials to the
different categories of the agents. The level of adaptation depends on the learner
behavior, likeliness, the degree of learning and different learning parameters. The
adaptation extends to the adaptation of the user interface for the specific learners. We
define different types of agents, roles and their goals. Each agent performs a specific
action based on the external parameters.

3 Proposed Work
In our approach the intelligent Tutoring system is modeled using the O-MaSE [11]
model which specifies the agent oriented software engineering methodology. We also
define agents, their roles and behaviors. These agents are independent agents; they
can act independently in accordance with the environment and the system
 Agent Based Adaptive Learning System 425

characteristics. The Similarity between the several learners are identified by the
clustering of the students and analyzing their behaviors. The classification of the
learning resources is done using the tree based hierarchical classification and present
the resources accordingly to the student’s learning capability and student’s behavior.
The implementation of the agent based ITS is done using the JADE [6] platform.

4 Agent Oriented Software Engineering

One of the major problems faced by agent-oriented software engineering is the failure
to achieve a strong industry acceptance. One of the reasons hindering this acceptance
is a lack of an accepted process-oriented methodology for developing agent-
based systems. An interesting solution to this problem is the use of approaches that
allow us to customize processes based on different types of applications and
development environments. One technique that provides such a rational approach for
the construction of tailored methods is Method Engineering [16].
Method Engineering is an approach by which process engineers construct
processes (i.e., methodologies) from a set of method fragments instead of
trying to modify a single monolithic, “one-size-fits-all” process. These fragments are
generally identified by analyzing these “one-size-fits-all” processes and extracting
useful tasks and techniques. The fragments are then redefined in terms of a
common meta-model and are stored in a repository for later use. To create a
new process, a process engineer selects appropriate method fragments from the
repository and assembles them into a complete process based on project
requirements[16].

4.1 O-MaSE Methodology

The goal of the O-MaSE [15] Process Framework is to allow process engineers
to construct custom agent-oriented processes using a set of method fragments,
all of which are based on a common meta-model. To achieve this, we define O-
MaSE in terms of a meta-model, a set of method fragments, and a set of guidelines.
The O-MaSE meta-model defines a set of analysis, design, and implementation
concepts and a set of constraints between them. The method fragments define how a
set of analysis and design products may be created and used within O-MaSE. Finally,
guidelines define how the method fragment may be combined to create valid O-
MaSE processes, which we refer to as O-MaSE compliant processes. We apply
the O-MaSE methodology for developing the Intelligent Tutoring system. This has
paved a way for us to create a stable agent oriented system using the O-MaSE model.
The first process we derive is appropriate for a small agent-oriented project in
which reactive agents achieve goals that have been assigned at design time.
Essentially, the only products required for this type of system are the system
goals, agent classes, agent plans, and inter-agent protocols. This type of process
leads to a rigid MAS but is very easy and fast to develop. This process may also be
suitable for prototyping, where a simple and rapid process is needed.
426 R. Gowri, S. Kanmani, and T.T. Sampath Kumar

5 Agents, Behaviors, Communication, Service Discovery

5.1 Types of Tutoring Agents

In our research, we propose several types of Agents, capabilities, roles and

responsibilities. The different categories of agents are profile analysis agent, tutor
agent, learner agent, classification agent, clustering agent, communication agent,
service register/discover agent. The profile analysis agent and the learner agents are
used to analysis the several characteristics of the learner, their likeliness, technical
capabilities etc. The tutor agent performs the role of creating the course,
announcement to students, analysis of student’s path of learning and evaluating the
learning process etc. The classification agent performs resource classification based
on the hierarchical tree structure where the resources are represented in the tree like
structure based on different courses. The student clustering agent helps to achieve the
collaborative learning where similar learners are identified and they are allowed to
communicate with each other. The communication agent performs the role of setting
up the communication path between the learners and the tutors and between different
types of agents.
The learner agents have three sub-types of agents. The evaluation agents perform
all tests, including pre-learning tests and post-learning tests. Based on the evaluation
results, the path planning agents analyze the learning profile and propose the course
learning path. The problem mitigation agents have the responsibilities of learning
problem analysis and assisting the learners to resolve the problems. The instructor
agents can be course management agents or portfolio analysis agents. The course
management agents assist the teachers to construct, arranges, and adjusts the
course content. The portfolio analysis agents collect student learning portfolio and
provide statistical analysis to the teacher for adjustment of course content or unit
sequencing.

5.2 Agent Communication

Agent communication is defined by the FIPA [7] specification where the agent
communicates using the Agent Communication Language (ACL). The
communication architecture offers flexible and efficient messaging, where
JADE creates and manages a queue of incoming ACL messages, private to each
agent; agents can access their queue via a combination of several modes: blocking,
polling, timeout and pattern matching based. The full FIPA communication
model has been implemented and its components have been clearly distinct and fully
integrated: interaction protocols, envelope, ACL, content languages, encoding
schemes, ontologies and, finally, transport protocols. All agent communication is
performed through message passing, where FIPA ACL is the language to represent
messages.
 Agent Based Adaptive Learning System 427

5.3 Service Discovery

In our work, all the learning process is mapped into different types of services. If an
agent wants to look for the available courses which is registered in the directory by
the tutoring agent. All the available functionalities are mapped into services. We
adopt the publisher/subscriber [4] methodology for the registering and retrieving the
service details of the services which is implemented via JADE platform.

6 Evolving Clustering Method

Evolving clustering method [5] is a clustering algorithm which increases the number
of clusters or adjusts the centers and the radius real-time dynamically accordingly to
the data points. It is primarily used to solve the initial cluster center problem which is
faced by FCM algorithm. In any cluster, the maximum distance between the cluster
centre and data points is always less than the threshold value‘d’. The threshold
selection has a direct impact on the clustering numbers. Evolving clustering method is
a clustering algorithm of kind of evolution, on-line and bounding by a maximum
distance. It increases the number of cluster or adjusts the centers and the radius real-
time dynamically as they entered sample data increasing. In any one cluster, the
maximum distance between the example of cluster points and the corresponding
maximum distance are less than the threshold Dthr; Dthr selection will have a direct
impact on the clustering numbers. Examples of clustering process from a data stream,
the whole process start clustering from an empty set. Some of the created clusters to
be updated through depend on the location of the current example in the input space
as well as changing the location of the cluster centers and increasing the radius of the
cluster with the new examples appearance, it will no longer to be updated when its
radius meet the threshold.

7 Fuzzy C-Means Algorithm

Fuzzy C- means [8] [9] is a method of clustering which allows one piece of data to
belong to two or more clusters. It is based on the optimization of objective function.
By iteratively updating the cluster centers and membership grades for each data point,
Fuzzy C means moves the cluster centers to the right location within the dataset.
Fuzzy C-means is a clustering algorithm that used membership degree to determine
each data point belongs to a certain cluster. FCM divided the n vectors xi (i = 1, 2,…,
n) into c fuzzy group, and computing the cluster center of each group, making value
function of non-similarity index to achieve the minimum. FCM algorithm making
each of the given data points with values between 0, 1 membership to determine its
degree of belonging to various groups through fuzzy partition. And to suit the
introduction of fuzzy partition, the membership matrix ‘U’ allowed the element value
between 0, 1. The above algorithm can also initialize cluster center firstly, and run the
iterative process after. The algorithm does not guarantee an optimal solution to the
428 R. Gowri, S. Kanmani, and T.T. Sampath Kumar

convergence, so its performance depends on the initial cluster centers. Therefore,

Another fast algorithm can be used to determine the initial cluster centers, or start the
algorithm with different initial clusters center every time, running FCM repeatedly.

8 Agent Based Intelligent Tutoring System Model

The system is modeled through the O-MaSE which is done using eclipse. The system
model consists of the agent model diagram, domain diagram, goal diagram, and plan
diagram. The different agent’s services, components are explained in fig. 1.

Fig. 1. Agent Diagram for ITS

The plan diagram for the intelligent adaptive tutoring system is explained in the
fig 2. All the functions, parameters are also explained.

Fig. 2. Plan Diagram for ITS

 Agent Based Adaptive Learning System 429

The goal diagram describes about the various goals in the hierarchical tree
representation which is explained in the figure 3. In fig. 3 the different forms of
adaptation is explained. Adaptive user interface, adaptive content delivery during the
course, adaptive examinations questions is divided into different sub-goals and when
each sub-goal is achieved the mail goal is achieved.

Fig. 3. Goal Diagram For ITS

In the fig. 4 the domain Diagram is explained in which each agent’s behaviors,
relationships, different objects are explained.

Fig. 4. Domain Diagram for ITS

430 R. Gowri, S. Kanmani, and T.T. Sampath Kumar

9 Implementation

The system is developed with the jsp, servlets, ajax, agents developed in JADE,
tomcat as web-server, mysql as database. Using the fuzzy clustering method the
different types of students are identified and the different types of course contents are
presented to the user according to their position in the cluster. We construct a tree like
structure for the representation of the course contents. Whenever the student is in
particular level of the course only the contents corresponding to the tree are
only presented to the user. The resources can be classified into html, video, audio,
flash or combination of those types. These contents have to be delivered adaptively
to the user accordingly to the user’s likeliness. The user’s activity has to be monitored
and in accordance to the user the adaptive user-interface is delivered by the
system.

Fig. 5. Architecture Diagram

10 Conclusions and Future Work

The Intelligent Tutoring system is systematically developed applying the Agent

Oriented Software Engineering approach. The modeling of the Intelligent Tutoring
System is performed using the O-MaSE approach and system is analyzed using
several diagrams .The adaptive behavior of the Intelligent Tutoring system is
achieved by clustering of the similar students based on the likeliness, skill levels and
various parameters which is performed by the ECM and FCM. The course materials
are adaptively presented to the user by representing the resources in a hierarchical tree
structure.
In our future work the system will be implemented in the distributed environment
analyzing various parameters such as replication, fragmentation of resources
consolidation of resources due to increase in the size of the course materials such as
video etc and this large amount of data has to be stored and retrieved for the students
in an optimum strategy.
 Agent Based Adaptive Learning System 431

References
[1] Badjonski, M., Ivanovic, M., Budimac, Z.: Intelligent Tutoring System as Multiagent
System. In: IEEE International Conference on Intelligent Proccssing Systems, pp. 871–
875 (1997)
[2] Caseau, Y.: Self-adaptive and self-healing message passing strategies for process-oriented
integration infrastructures. In: Proceedings of 11th IEEE International Conference and
Workshop on the Engineering of Computer-Based Systems, pp. 506–512 (May 2004)
[3] Chang, Y.H., Lu, T.Y., Fang, R.J.: An Adaptive E-learning System Based on Intelligent
agents. In: 6th Conference on WSEAS International Conference on Applied Computer
Science, pp. 200–205 (2007)
[4] Eugster, P.T., Felber, P.A., Guerraoui, R., Kermarrec, A.M.: The Many Faces of
Publish/Subscribe. ACM Computing Surveys (CSUR) 35(2), 114–131 (2003)
[5] Du, G.-Y., Miao, F., Tian, S.-L., Liu, Y.: A Modified Fuzzy C-means Algorithm in
Remote Sensing Image Segmentation. In: International Conference on Environmental
Science and Information Application Technology, ESIAT 2009, pp. 447–450 (2009)
[6] Jade-Java Agent Development Framework, http://jade.tilab.com/
[7] Foundation for Intelligent Physical Agents, http://www.fipa.org/
[8] Turksen, L.B.: Fuzzy data mining and expert system development. In: IEEE International
Conference on Systems, Man, and Cybernetics (1998)
[9] Zhang, H.-Z., Chen, H., Bao, L.-X.: An improved Fuzzy C Means clustering algorithm
and its application in traffic condition recognition. In: Seventh International Conference
on Fuzzy Systems and Knowledge Discovery, FSKD (2010)
[10] Nargesian, F., Nazemi, E.: An Agent-Based Architecture with Centralized Management
for a Distance Learning System. In: International Conference on Computational
Intelligence and Multimedia Applications (2007)
[11] Jia, Y., Huang, C., Cai, H.: A Comparison Of Three Agent-Oriented Software
Development Methodologies: Mase, Gaia, And Tropos. In: IEEE Youth Conference on
Information, Computing and Telecommunication, YC-ICT 2009 (2009)
[12] Zhang, K., Cui, L., Wang, H., Sui, Q.: An Improvement of Matrix-based Clustering
Method for Grouping Learners in E-Learning. In: 11th International Conference on
Computer Supported Cooperative Work in Design (2007)
[13] Arenaz, P., Fisher, W., Kubo, C., Della-Piana, C.K., Gomez, R., Dahl, E.: CircLES: a
retention program for entering students in engineering, mathematics and science. In: 29th
Annual Frontiers in Education Conference (1999)
[14] Wu, F.: Apply Data Mining to Students’ Choosing Teachers Under Complete Credit
Hour. In: Second International Workshop on Education Technology and Computer
Science, ETCS (2010)
[15] Garcia-ojeda, J.C., Deloach, S.A., Oyenan, W.H.: O-MaSE: A Customizable Approach to
Developing Multiagent Development Processes. In: The 8th International Workshop on
Agent Oriented Software Engineering (2007)
[16] Brinkkemper, S.: Method Engineering: Engineering of Information Systems Development
Methods and Tools. Jnl. of Information and Software Technology 38(4), 275–280 (1996)
 GPGPU Implementation of Parallel Memetic Algorithm
for VLSI Floorplanning Problem

Subbaraj Potti1 and Sivakumar Pothiraj2

1
Theni Kammavar Sangam College of Technology,
Theni-625 534,
Tamil Nadu, India
[email protected]
2
Arulmigu Kalasalingam College of Engineering,
Krishnankoil,
Tamil Nadu-626190,
India
[email protected]

Abstract. The VLSI Physical design floorplanning is the process where circuit
description is converted into geometric description. This NP-Optimization
problem with non-slicing blocks is very hard to solve. Memetic algorithms are
used to solve NP-Optimization problems. Memetic algorithms are a family of
meta-heuristic search algorithms in which a rule-based representation of Local
Search (LS) is co-adapted alongside the candidate solutions within a hybrid
evolutionary system. However, they may execute for a long time for difficult
problems, because of several fitness evaluations. A promising approach to
overcome this limitation is to parallelize these algorithms. In this paper, we
propose to implement a parallel Memetic Algorithm (MA) on Graphical
Processing Units. The General Purpose Graphical Processing Unit (GPGPU) is
the complete parallel hardware which is best used for the Parallel Computing.
The Parallel Memetic Algorithm we followed for an application perspective is a
modified or hybrid genetic algorithm with extensive attributes to local search.
The Parallel Memetic Algorithm operator gives a perfect exploration in the
available search area and it is typical for addressing a modular assignment
problem. It reduces the time efficiently for execution which is a boon for VLSI
industry where we don’t have to spend more weeks for Physical design to be
completed. We perform experiments to compare our parallel MA with an
ordinary MA and demonstrate that the former is much more effective than the
latter. Both fitness evaluation and genetic operations are spread over the GPU
cores, allowing for the efficient processing of very large datasets. The parallel
nature of Memetic algorithm can be best harnessed with help of GPGPU and
significant changes in execution time and optimization are shown for VLSI
Floorplanning problem.

Keywords: Floorplanning, Graphical Processing Unit, Memetic algorithm, very

large scale integrated circuit (VLSI).

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 432–441, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 GPGPU Implementation of Parallel Memetic Algorithm 433

1 Introduction
The Parallel Memetic Algorithm is developed as a mathematical tool for getting
solution to various problems where we need to take into account the transferring of
cultural information from one generation to another. The path for finding this is
realized using a sequence of steps using genetic algorithm and local search. The
inspiring earlier work carried out by Maolin [1] has enabled to look for further
investigation into the optimization and parallelization. The parallel Memetic
Algorithm frame work will surely give its potential to a parallel computing sequence
of execution and is further termed as the application boosting.
Graphics Processing Units (GPUs) are fast, highly parallel processor units. In
addition to processing 3D graphics, modern GPUs can be programmed for more
general purpose computation [10]. This facilitates the chance of integrating Parallel
Memetic Algorithm and GPGPU to give better Platform of execution in terms of time
and optimization [3]. The potential of GPGPU is carried out with help of Compute
Unified Device Architecture (CUDA). The CUDA model is a collection of threads
running in parallel. CUDA consists of a test of C language extensions and a runtime
library that provides APIs to control the GPU[4].
Because capable hardware and software are new, there is relatively little previous
work on using GPUs for evolutionary computation. For example [6] implements a
evolutionary programming algorithm on a GPU, and finds that there is a 5-fold speed
increase. Work by [16] expands on this, and evaluates expressions on the GPU. There,
all the operations are treated as graphics operations, which make implementation
difficult and limit the flexibility of the evaluations. Yu et al [17], on the other hand,
implement a Genetic Algorithm on GPUs. Depending on population size, they find a
speed up factor of up to 20. Here both the genetic operators and fitness evaluation are
performed on the GPU [5, 7].
Memetic Algorithm covers one area of Memetic Computing, in particular dealing
with areas of evolutionary algorithms that use other deterministic refinement
techniques for solving optimization problems. We demonstrate a method for using the
GPU as an evaluator for Memetic programming expressions and show that there is a
considerable increase in speed of execution.

2 VLSI Floorplanning Problem

The VLSI physical design contains large number of steps which are highly complex
and tentative. Floorplanning is one of them. The modular assignment problem is
considered as the basic problem in Floorplanning. Consider a rectangular space where
we need to arrange n modules. The search space for this is given by n!Cn [1]. The
rectangular space is limited with width ‘W’ and height ‘H’.
We can anticipate that the blocks must be arranged such that they should not
overlap each other. Moreover even if we arrange the blocks it should give minimum
wasted area. The ultimate goal is to reduce area and wire length. As a part of modular
assignment with non-slicing, the efficient representation is considered as O-Tree
434 S. Potti and S. Pothiraj

representation. The time-complexity for the transform between an O-tree

representation and its corresponding floorplan is only O(n)[14]. The representation
not only covers all sub-optimal floorplan but also has a smaller search space. For a
VLSI floorplanning problem of n modules, the search space of the representation is
(n!)*cn, where,

(1)

The representation gives the geometrical relations among modules which are valuable
for identifying meaningful building blocks when designing genetic operators.
A widely used global search method for VLSI floorplanning problems is genetic
algorithm (GA)[2][15]. GA has been successfully applied to solve non-slicing VLSI
floorplanning problems. Since the encoding scheme does not capture any topological
information of the floorplans, the performance of the GA is not satisfactory. Although
GA is effective in solving NP hard problems, they may execute for a long time to find
better solutions, because several fitness evaluations must be performed. In this paper,
we present a parallel Memetic Algorithm for a non-slicing and hard-module VLSI
floorplanning problem.

2.1 The Problem Statement

Generally we consider a VLSI floorplanning problem as NP-hard problem. Even

though floorplanning is highly tentative the assignment is carried out with some
successful approaches. A module mi is a rectangular block with fixed height hi and
width wi, M = {m1, m2, . . mn} is a set of modules, and N is a net list specifying
interconnections between the modules in M. A floorplan F is an assignment of M onto
a plane such that no module overlaps with another. Here the cost of the floorplan is
defined with an equation.

Cost (F) = w1 × ∗ + w 2× ∗ (2)

In the above equation, Area∗ and Wirelength∗ represent the minimal area and the
interconnection costs, respectively. Since we do not know their values in practice,
estimated values are used[1]. The constants w1 and w2 are weights assigned to the
area minimization objective and the interconnection minimization objective,
respectively, where 0 ≤w 1, w 2 ≤ 1, and w1 + w2 = 1. The interconnection cost is the
total wire length of all the nets, and the wire length of a net is calculated by the half
perimeter of the minimal rectangle enclosing the centre of the modules that have a
terminal of the net on it[13]. The minimization of the cost function is achieved by
keeping the floorplan as an admissible one. A floorplan is admissible if no modules
can be shifted left or bottom without moving other modules in the floorplan. In order
to minimize the area, a module may be rotated 90 degrees. The area of a floorplan F,
Area (F), is measured by the area of the smallest rectangle enclosing all the modules.
Given M and N, the objective of the floorplanning problem is to find a floorplan F
such that the cost function minimized.
 GPGPU Implementation of Parallel Memetic Algorithm 435

3 Parallel Memetic Algorithm Framework

3.1 Representation of Genetic Population

In this section we describe the internal representation of population. If the GPU is to

perform GA operations, the first thing we need to do is to represent population data in
a format that is accessible by the GPU. The general idea is to store population in a set
of texture maps and to exploit fragment programs to implement genetic operators. In
our representation, the chromosome of each individual is sequentially divided into
several segments which are distributed in a number of textures with the same position
(see Fig. 1). Every segment contains four genes packed into a single pixel’s RGBA
channels. We call those textures population textures. Another 2D texture map, called
fitness texture, is used to hold the fitness scores of each individual in the population.
The mapping is done between the fitness of a particular individual and the position of
the individual in the population [17]. The proposed representation enables the
efficient computation of genetic operations and operators. It has several advantages:
First, it naturally keeps 2D grid topology of the population. Second, for each
individual, fragment programs of genetic operators only need lookup considered pixel
or near neighbourhood pixels in each texture map. Thus it keeps the memory footprint
of each fragment as low as possible to efficiently utilize texture cache. Third, packing
of four concessive genes in one Texel makes use of the wide SIMD computations
carried out by the GPU [17].

Fig. 1 . The representation of chromosome in a population as a set of 2D textures as shown

3.2 Genetic Operators

The genetic operations are defined in such a way that subtrees are transmitted and
evolved through crossover operator and mutation operator. For Crossover, given two
parents, both of which are admissible floorplan represented by an O-tree, the
crossover operator transmits the significant structural information from two parents to
child [14]. By recombining some significant structural information from two parents;
it is hoped that better structural information can be created in the child.
436 S. Potti and S. Pothiraj

The mutation operations randomly select a single existing individual and generate
a new one by changing one or more of its genes. They are crucial to push the
algorithm towards unexplored regions of the search space. The differences between
the new and the existing individual change from one operation to the other. Three
mutation operations have been defined: the first is the simplest one: a gene is
randomly selected, and its value is changed; the second affects in a deeper way the
solution: an integer k is generated, and the genes k, 2k, 3k, 4k are randomly changed;
the third one radically changes the whole individual: For each gene the next
implementation in the implementation list is chosen; as the list is ordered with respect
to the aspect ratio of the implementations, the operator executes a rotation of all the
modules.

3.3 Fitteness Value Evaluation

Fitness value evaluation determines the “goodness” of individuals. It is one of the

core parts of GA. After each evolution, the fitness value of each individual in the
current population is calculated. The result is then passed to the later stage of GA
process. Each individual returns a fitness value by feeding the objective function with
the genomes of the individual. This evaluation process usually consumes most of the
computational time. Since no interaction between individuals is required during
evaluation, the evaluation is fully parallelizable. Recall that the individuals are broken
down into quadruples and stored in the tiles within the textures. The evaluation shader
hence looks up the corresponding quadruple in each tile during the evaluation. The
fitness values are output to an output texture of size w × h, instead of 4w ×2h,
because each individual only returns a single value[16].

3.4 Competition and Selection

Replacing the old generation is the last stage of each evolution. There are two major
processes involved: competition and selection. GA employs a stochastic selection
(soft selection) through the tournament schema. Each individual in the union set of
parent and offspring population takes part in a q-round tournament. In each round, an
opponent is randomly drawn from the union set of parents and offspring. The number
of opponents defeated is recorded by the variable win. After the tournament, a
selection process takes place and chooses the best µ individuals having the highest
win values as parents for next generation.
Exploitation is the process of using information gathered to determine which
searching direction is profitable. In GA, exploitation is realized by competition.
Individuals in the population compete with q randomly drawn opponents, where q is
the predefined tournament parameter. The considered individual wins if its fitness
value is better than (in our case smaller than or equal to) that of the opponent. The
times of winning is recorded in win. It tells us how good this individual is.
Competition can be done either on GPU or CPU. For GPU implementation, q textures
of random values have to be generated by CPU and loaded to GPU memory for
picking q opponents within each evolution. In the shader, q fitness comparisons are
performed by looking up the opponent’s fitness in the fitness textures obtained
 GPGPU Implementation of Parallel Memetic Algorithm 437

previously. Such look-up is achieved by regarding the random values as indices to

fitness textures. The times of winning, win, is recorded and output to a texture. On the
other hand, for CPU implementation, only the fitness textures of both parent and
offspring population have to be transferred from GPU memory to main memory. Then
the tournament is CPU performed by referring to these fitness textures. It seems that
GPU implementation should be faster than the CPU one, as it parallelizes the
competition. However, our experiments show that using GPU to implement the
competition is slower than that of using CPU. The major bottleneck of GPU
implementation is the transfer of q textures towards GPU memory. It evidences the
limitation of slow data transfer of current GPU. As the competition does not involve
any time-consuming fitness evaluation (it only involves fitness comparison), the gain
of parallelization does not compensate the loss due to data transfer [8]. Hence, we
suggest to perform the tournament in CPU unless the data transfer rate is significantly
improved in the future GPU.

3.5 Local Search Method

The local search method is based on a deterministic algorithm proposed by Guo et al.
[1].Given an initial floorplan encoded in an O-tree (T, π), the local search method
finds a local optimal solution through systematically examining those O-trees that can
be obtained by removing a module from, and then putting it back to, the O-tree. When
a module is added, its orientation may be changed if it leads to a smaller cost. The
algorithm is shown below:
1) For each node m i in (T, π):
a) Delete mi from (T, π);
b) Insert mi in the position where we can get the smallest cost value among all
possible insertion positions in (T, π) as an external node of the tree;
c) Perform (a) and (b) on its orthogonal O-tree.
2) Output (T, π).

4 Experimental Results
GPUs are considered as SIMD processors, when evaluating a population of programs
in parallel, the common approach so far is to implement some form of interpreter on
the GPU. The interpreter is able to execute the evolved programs (typically machine
code like programs represented as textures) in a pseudo-parallel manner.
The NVIDIA® Tesla™ C1060 computing processor enables the transition to
energy efficient parallel computing power by bringing the performance of a small
cluster to a workstation. With 240 processor cores and a standard C compiler that
simplifies application development, Tesla scales to solve the world’s most important
computing challenges—more quickly and accurately.
The population and genetic algorithm run on the CPU, with evaluations run on the
GPU. The CPU converts arrays of test cases to textures on the GPU and loads a
438 S. Potti and S. Pothiraj

shader program into the shader processors. The Accelerator tool kit compiles each
individuals GP expression into a shader program. The program is then executed, and
the resulting texture is converted back in to an array. The fitness is determined from
this output array.
In this section, we experimentally evaluated the quality of the floorplan produced
by our GPU based parallel Memetic algorithm on a large number of non-slicing and
hard blocks that are part of the widely used MCNC circuit floorplanning benchmarks
suite. All experiments were carried out on Pentium Quad core 2 processor 2.6 GHz
with NVIDIA Tesla C1060 computing processor, with 8GB main memory and
512MB GPU memory.
The following parameters were used in the experiments:
• population size: µ =40, 80, 100
• tournament size: q= 10
• maximum number of generation: G= 100
The two objectives are mutually constrained with each other [1]. The weights for the
area minimization and the interconnection minimization objectives were set to 0.8 and
0.2, respectively (w1 = 0.8 and w2 = 0.2), which means the MA focused more on
minimizing the area of the floorplan. GPU based implementation was compared with
software implementation running on single CPU with different benchmark circuits.
Table 1 shows the best value out of 100 runs of Parallel Memetic Algorithm on CPU.
Table 2 shows the best value out of 100 runs of PMA on GPGPU. Table 3 shows the
statistics for the test results on the performance of the Parallel MA.

Table 1. Best value of 100 generation on without GPU

MCNC Module Net Width Height Floorplan Total Dead Wire Time Cost
Circuit Count (mm) (mm) Area Modules Area(%) (mm) (s) (F)
(mm) Areas(mm2)
ami49 49 408 10.6370 3.7150 39.5165 35.4454 10.3021 780.4 251 1097597
ami33 33 123 1.8160 0.7320 1.3293 1.1564 13.0039 72.9 245 1192814
Hp 11 83 10.4830 1.0080 10.5669 8.8306 16.4314 175.3 61 1167594
xerox 10 203 3.2070 7.6440 24.5143 19.3503 21.0653 490.6 62 1248478
apte 9 97 5.3780 10.4530 56.2162 46.5616 17.1741 430.4 60 1170876

Table 2. Best value of 100 generation on without GPU

MCNC Module Net Width Height Floorplan Total Dead Wire Time Cost
Circuit Count (mm) (mm) Area Modules Area (mm) (s) (F)
(mm) Areas (%)
(mm2)
ami49 49 408 9.9360 3.9850 39.5950 35.4454 10.4800 720.2 40 1040067
mi33 33 123 1.8590 0.7300 1.3571 1.1564 14.7834 64.6 15 1179438
Hp 11 83 3.4400 2.6520 9.1229 8.8306 3.2040 163.4 30 1052091
xerox 10 203 8.8570 2.4380 21.5934 19.3503 10.3878 470.2 10 1137913
apte 9 97 12.2180 3.9850 48.6887 46.5616 4.3688 410.4 12 1051531
 GPGPU Implementation of Parallel Memetic Algorithm 439

Table 3. Statistics for the test result on the performance of the Parallel MA

MCNC With GPU Without GPU

Circuit Best Mean Std.dev Best Mean Std.dev
ami49 1040067 1192655 79400.68 1097597 1519900 232887.3
ami33 1179438 1193769 22010.5 1192814 2758732 22649.53
Hp 1052091 1617788.5 2238.42 1167594 1873492 4729.34
xerox 1137913 2083469.4 10622 1248478 2513694 10987
apte 1051531 1492840 49203 1170876 1683491 87349

Table 4. Speedup for different population size for ami49

POPULATION PMA WITH PMA SPEEDUP

SIZE GPU(S) WITHOUT
GPU(S)
40 20 50 2.5X
80 35 140 4X
100 40 250 6.25X

To study the dynamic performance of Parallel Memetic Algorithm (PMA) on GPU

and PMA on CPU, we captured the dynamically changing minimal cost obtained in
each run for each of the benchmarks. Fig. 3 displays the dynamic performance of the
PMA on GPU and PMA on CPU for benchmark ami49. It can be seen that the
dynamic performances of PMA on GPU for ami49 problem are consistent.
The execution time of PMA without GPU and PMA with GPU with population
sizes of 40, 80, and 100 are shown in Fig. 2. It is interesting to notice that, the GPU
based implementation is much faster than former approach. The speedup obtained is
on an average from 2.5 times to 6.25 times. We see that the speedup increases as we
increase the population size. When the population size is small, the speedup of GPU
based implementation is marginal compared to implementation without GPU. This
happens because, first time GPUs see a new piece of code, and it spends some time
analyzing it and compiling various instruction sequences for faster lookup on
subsequent runs. That first run will be a little slower, but for long-running
computations there won’t be any noticeable lag.

Time Vs Generation for ami49

400
Time (Sec.)

200 With GPU

0
40 80 100 Without
GPU
Population size

Fig. 2. Execution time of the PMA with GPU and without GPU
440 S. Potti and S. Pothiraj

3.00E+06

Cost 2.50E+06
ami49
2.00E+06 without gpu

1.50E+06
ami 49
1.00E+06 1
with gpu
9
17
25
33
41
49
57
65
73
81
89
97
Fig. 3. Dynamic performance comparison of Parallel MA running on GPU and running CPU on
benchmark ami49

5 Conclusion
In this research, an implementation of Parallel MA to solve VLSI floorplanning
problem based on GPU was presented. The parallel MA on GPU is a hybrid of
master-slave and fine-grained models. Competition and selection are performed by
CPU while fitness evaluation, mutation, and reproduction are performed by GPU
which is essentially a massively parallel machine with shared memory. GPU allows
processors to communicate with any other processors directly; thus more flexible
fine-grained MAs can be implemented on GPU. We have done experiments to
compare our parallel MA on GPU and a sequential MA on CPU. The experimental
results show that GPGPU has good accelerated performance on larger population
sizes. A couple of other important factors can help to get the best performance out of
our transformation to GPUs such as avoiding of excessive memory transfer, inherent
parallelism and computation dependency between CPU and GPU. Note that the
commercial tool can be made to run on several CPUs using a distributed option. If
each of these CPUs had a GPU on board, then the GPU approach could also exploit a
distributed option, and the above speedup numbers would be increased. More
sophisticated cluster approaches to parallelisation will have to be examined in the
future.

References
1. Tang, M., Yao, X.: A Memetic Algorithm for VLSI Floorplanning. IEEE Trans. on
Systems, Man, and Cybernetics 37, 62–68 (2007)
2. Rebaudengo, M., Reorda, M.: GALLO: A genetic algorithm for floorplan area
optimization. IEEE Trans. Computer-Aided Design Integrated Circuits System 15(8),
943–951 (1996)
3. Cohoon, J., Hegde, S., Martin, W., Richards, D.: Distributed genetic algorithms for the
floorplan design problem. IEEE Trans. Computer-Aided Design Integrated Circuits
System 10(4), 483–492 (1991)
 GPGPU Implementation of Parallel Memetic Algorithm 441

4. Liang, S., et al.: Design and evaluation of a parallel k-nearest neighbor algorithm on
CUDA-enabled GPU. In: Proc. IEEE 2nd Symposium on Web Society (SWS), pp. 53–60.
IEEE Press, Los Alamitos (2010)
5. Harding, S., Banzhaf, W.: Fast genetic programming on GPUs. In: Proceedings of the 10th
European Conference on Genetic Programming, pp. 90–101 (2007)
6. Wong, M.-L., Wong, T.-T., Fok, K.-L.: Parallel evolutionary algorithms on graphics
processing unit. In: IEEE Congress on Evolutionary Computation, vol. 3, pp. 2286–2293
(2005)
7. Harding, S., Banzhaf, W.: Fast Genetic Programming and Artificial Developmental
Systems on GPUs,”. In: Proc. 21st International Symposium High Performance Computing
System and Applications (HPCS 2007), p. 2. IEEE Computer Society Press, Los Alamitos
(2007)
8. Robilliard, D., Marion-Poty, V., Fonlupt, C.: Population Parallel GP on the G80 GPU. In:
O’Neill, M., Vanneschi, L., Gustafson, S., Esparcia Alcázar, A.I., De Falco, I., Della
Cioppa, A., Tarantino, E. (eds.) EuroGP 2008. LNCS, vol. 4971, pp. 98–109. Springer,
Heidelberg (2008)
9. Jiening, W., Chufeng, Z., Jiankang, D.: Implementation of Ant Colony Algorithm based on
GPU. In: Proc. of the IEEE Sixth Int. Conference on Computer Graphics, Imaging And
Visualization (CGIV 2009), pp. 50–53 (2009)
10. Wu, E., Liu, Y.: Emerging technology about GPGPU. In: IEEE Asia Pacific Conference
Circuits and Systems (APCCAS 2008), pp. 618–622 (2008)
11. Sutanthavibul, S., Shragowitz, E., Rosen, J.B.: An Analytical Approach to Floorplan
Design and Optimization. IEEE Transactions on Computer Aided Design 10(6), 761–769
(1991)
12. Deng, Y., Mu, S.: The potential of GPUs for VLSI physical design automation. In: 9th
International Conference on Solid-State and Integrated-Circuit Technology (ICSICT
2008), pp. 2272–2275 (2008)
13. Tang, X., Tian, R., Wong, M.D.F.: Minimizing wire length in floorplanning. IEEE Trans.
on Computer-Aided Design of Integrated Circuits and System 25(9), 1744–1753 (2006)
14. Takahashi, Guo, T., Cheng, P.N., Yoshimura, C.K.: Floorplanning using a tree
representation: a summary. IEEE Circuits and Systems Magazine 3(2), 26–29 (2003)
15. Rebaudengo, M., Reorda, M.: GALLO: A genetic algorithm for floorplan area
optimization. IEEE Trans. Computer-Aided Design Integrated Circuits System 15(8), 943–
951 (1996)
16. Fok, K.-L., Wong, T.-T., Wong, M.-L.: Evolutionary Computing on Consumer Graphics
Hardware. IEEE Intelligent Systems 22(2), 69–78 (2007)
17. Yu, Q., Chen, C., Pan, Z.: Parallel Genetic Algorithms on Programmable Graphics
Hardware. In: Wang, L., Chen, K., S. Ong, Y. (eds.) ICNC 2005. LNCS, vol. 3612, pp.
1051–1059. Springer, Heidelberg (2005)
18. Chitty, D.M.: A data parallel approach to genetic programming using programmable
graphics hardware. In: Proceedings of the 9th Annual Conference on Genetic and
Evolutionary Computation (GECCO 2007), vol. 2, pp. 1566–1573. ACM Press, New York
(2007)
19. Tang, M., Lau, R.Y.K.: A Parallel Genetic Algorithm for Floorplan Area Optimization. In:
IEEE Seventh International Conference on Intelligent Systems Design and Applications,
pp. 801–806 (2007)
 Improving Cloud Security through Virtualization

B. Loganayagi1 and S. Sujatha2

1
Research Scholar (Full Time) in the Faculty of ICE
2
Head of the Department, Department of MCA
Anna University of Technology Tiruchirappalli (Main Campus), Tiruchirappalli, India
[email protected], [email protected]

Abstract. Cloud Computing is accessing Services through Internet based on

pay per usage model. Software as a Service (SaaS), Infrastructure as a Service
(IaaS) and Platform as a Service (PaaS) are available in Cloud. Cloud based
products will eliminate the need to install and manage client rich applications.
Cloud Service providers are helping companies to reduce the high cost
infrastructure installation and maintenance cost. The customer is charged only
for the resources consumed like utility based computing. Virtualization is a
technique to implement cloud computing resources such as platform,
application, storage, and network. The beauty of virtualization solutions is that
you can run multiple virtual machines (VM) simultaneously on a single
machine. This paper provides details about cloud computing issues, how
virtualization techniques can be used to overcome those issues.

Keywords: Cloud Computing, SaaS, IaaS, PaaS, Virtualization.

1 Introduction
Cloud Computing is defined by a large-scale distributed computing paradigm that is
driven by economies of scale, in which a pool of abstracted, virtualized, dynamically-
scalable, managed computing power, storage, platforms, and services are delivered on
demand to external customers over the Internet.[1]
Cloud computing has the features of grid computing, cluster computing, service
computing, and utility computing [8][9]. End user has to run one time application,
like tax calculation, can use cloud application as the best choice. End users can do
their work or access services with thin client, no need to invest on resources. Instead
they can rent/lease the resources. More over they can choose the best available
computing platform, since many companies are going to offer cloud based
applications in near future, like mobile services.
Small companies, whose main business is not software related, can outsource their
work to cloud providers, so that they can concentrate on such as application
development only, instead of concentrating their focus on hosting environment. It is
easy to scale if the application is deployed in cloud. It takes away all the risks of
managing resources.
Cloud is providing preconfigured infrastructure at lower cost, which generally
follows the Information Technology Infrastructure Library, can manage increased
peak load capacity and moreover uses the latest technology, provide consistent

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 442–452, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Improving Cloud Security through Virtualization 443

performance that is monitored by the service provider. Dynamic allocation of the

resources as and when is needed.
Furthermore cloud provides secured computing platform, and scalability won’t be
a problem with the use of cloud. Overall giving good performance at lower cost
instead of making capital investment. The rest of the paper is organized as follows:
Section 2 discusses about the need for Cloud Computing. Section 3 discusses about
the Cloud providers. Section 4 discusses about the Cloud Computing and
Virtualization. Section 5 discusses about the Cloud Issues and Virtualization
Solutions and Section 6 discusses about the Conclusion and Future direction.

2 Need for Cloud Computing

Software industry is everyday changing. What you get/invest for your company/need
today is going to become obsolete within a decade itself. As an end user if you need
some services, instead of getting all the software, hardware straightway user can run
their services in cloud, which provides necessary hardware, software support.
Cloud computing reduces capital expenditure (CAPEX) and it offers high
computing at lower cost [4]. Upgrading your hardware/ software requirement also
easy with the cloud, without disturbing the current work. Scalability and maintenance
is easy in the case of cloud. Easily user can rent/lease the services offered by cloud
computing vendors. User will be charged as pay per usage like utility based services.

3 Cloud Providers
Cloud computing [7] [10] [14][6]systems generally falls into three course grain
categories. Infrastructure as a service (IaaS), Platform as a Service (PaaS), Software
as a Service (SaaS). Many companies are offering services.

A. Infrastructure as a Service(IaaS)
Infrastructure as a Service (IaaS) provisions hardware, software, and equipments to
deliver software application environments with a resource usage-based pricing model.
Infrastructure can scale up and down dynamically based on application resource needs.
Typical examples are Amazon EC2 (Elastic Cloud Computing) Service and S3
(Simple Storage Service) where compute and storage infrastructures are open to
public access with a utility pricing model. This basically delivers virtual machine
images to the IaaS provider, instead of programs, and the Machine can contain
whatever the developer want.

Example:
Full virtualization (GoGrid, Skytap), Grid computing (Sun Grid), Management
(RightScale), Paravirtualization (Amazon Elastic Compute Cloud).

B. Platform as a Service(PaaS)
Platform as a Service(PaaS) offers a high-level integrated environment to build, test,
and deploy custom applications. Generally, developers will need to accept some
444 B. Loganayagi and S. Sujatha

restrictions on the type of software they can write in exchange for built-in application
scalability.
An example is Google’s App Engine, which enables users to build Web
applications on the same scalable systems that power Google applications, Web
application frameworks, Python Django (Google App Engine), Ruby on Rails
(Heroku), Web hosting (Mosso), Proprietary (Azure, Force.com).

C. Software as a Service (SaaS)

User buys a Subscription to some software product, but some or all of the data and
codes resides remotely. Delivers special-purpose software that is remotely accessible
by consumers through the Internet with a usage-based pricing model. In this model,
applications could run entirely on the network, with the user interface living on a thin
client.
Salesforce is an industry leader in providing online CRM (Customer Relationship
Management) Services. Live Mesh from Microsoft allows files and folders to be
shared and synchronized across multiple devices.
Identity (OAuth, OpenID), Integration (Amazon Simple Queue Service), Mapping
(Google Maps, Yahoo! Maps), Payments (Amazon Flexible Payments Service,
Google Checkout, PayPal), Search (Alexa, Google Custom Search, Yahoo! BOSS),
Others (Amazon Mechanical Turk).

Amazon EC2, Developer can install anything they want in their virtual machines.
EC2 gives developer lots of control, but developer still needs to install OS, Web
server, Web or Application container, Database and the whole nine yards that you do
today for application deployment. It provides a virtual machine, and developer is
responsible for managing everything inside the VM. It offers very flexible system
along with the risk of managing everything.

Google App Engine, here developer doesn’t have any virtual machine and are not
responsible for installing/managing OS, software, database. Right now developers are
restricted to use Python to code the app and they need to use DataStore (persistent
service) to get/put the data. Developer is responsible for developing and deploying the
application. User also gets the access to Google accounts, Gmail etc. No OS, server,
database headaches with the disadvantage of being restricted to the Python and
DataStore. It's not for Java apps right now, but they may support more languages in
the future. Google seems to be more interested in Python than Java.

Microsoft Azure is offering PaaS, mainly for Microsoft related technologies. Azure
helps the user to develop, host and manage environment with tools for Microsoft
Visual Studio to enable the creation, building, debugging, running, and packaging of
scalable services.

Force.com is offered by Salesforce.com CRM applications and restricted to

proprietary Apex language and Visual force presentation layer.Other than the listed
above companies, many companies started offering cloud computing services.
 Improving Cloud Security through Virtualization 445

4 Cloud Computing and Virtualization

Virtualization is a technology which exists for more than three decades. This can be
used as a key technology to create cloud computing platform. Unused resources can
be effectively utilized, based on the requirement easily resources can be scaled using
this technology compared to physical resource scaling. Entire existing OS along with
all its application also can be moved to a virtual machine or vice versa also possible.

End User

End User

Cloud Providers
End User End User

End User

Application Developer

Fig. 1. Cloud Application deployment and Access through Cloud Providers

Cloud computing can use virtualization as a key technology, so everything can be

accessed through internet. Whether resources might be platform, network, server,
application or anything. Depending upon the computational need, easily it can be
scaled also.
In IaaS, customers can select their infrastructure such as server, Operating System to
run their application, network, and storage capacity likewise the entire necessary
infrastructure to provide their service [4] by using these infrastructures as a service
(Fig.2).For their application/service they are the owner at the same time maintenance of
infrastructure to a third party, a trusted service provider or in other words IaaS Provider.

Fig. 2. Running Services on Cloud

446 B. Loganayagi and S. Sujatha

Since infrastructure is accessed as a service, managing peak load also easy by

scaling the resources. Customers will be charged only for their usage of resources at
the same reducing capital expenditure to provide their services.
The Proposed Cloud Architecture is depicted as in fig.3. First layer consists of
resources, second layer is the virtual management layer, third layer is cloud
management layer, fourth layer consists of different type of web services and last one
is where the end user connects with the cloud.

Fig. 3. Cloud Architecture

Cloud Resources includes hardware, software, legacy application and network

resources. Virtual management layer consists of virtual machines, VM management
and deployment, Xen supervisor. Cloud Management layer consists of Management
Console, QoS Negotiation, Admission Control, Pricing, SLA Management,
Monitoring, Execution Management, Metering, Accounting, and Billing.Fourth layer
consists of Cloud Programming tools, Compute, Database, Queuing, Storage, Identity
Services, Map Reduce Service, Web 2.0, Mashups, Concurrent and Distributed
Programming. Topmost layer is the one, where user applications can be run. Security
and Compliances applicable to all the layers.
Virtualization technologies, Middleware frameworks, Grid architecture, Load
balancing, Optimal deploy configuration, Consistency models, Software as a Service
(SaaS), Hardware as a Service (HaaS), Data grid & Semantic web, Web services,
Security and Risk, Fault tolerance and reliability, Auditing, monitoring and
scheduling, Utility computing, High-performance computing, Peer to peer computing
are used as components in Cloud Computing.
To Deploy Web services, JDKTM 6 can be used. JDKTM 6 _Java EE 5 Development
using GlassFish Application Server. GlassFish, allow application developers to
develop and deploy Java EE-compliant applications such as Deploying Enterprise
Applications, Web Applications, EJB module, Connector module, Life cycle Module,
Application client module, Custom MBeans.

5 Cloud Issues and Virtualization Solution

Cloud Computing has some issues besides its nice features. The fig.4 shows the cloud
challenges/issues, a survey taken by IDC in Aug 2008.
 Improving Cloud Security through Virtualization 447

Fig. 4. Cloud Challenges/Issues by IDC Survey

Virtualization is not a new technique. It has emerged since from 1960s. IBM
introduced this concept to increase the resource utilization. Before that Mainframe
was used for one process at a time basis. After the virtualization concept, many virtual
machines were introduced and each can be used to run for a separate
process/application , so that effective utilization of resources were achieved.
Simply we can say virtualization is simulating hardware and/or software to run on
top of another hardware and/or software. Full Virtualization deals with running one or
more virtual machines directly on underlying hardware where as in Para
Virtualization, hypervisor act as a intermediate layer and it will co ordinate the virtual
machines and underlying host infrastructure.
Virtualization not only offers effective utilization of resources. The same feature
can be used to reduce the resource cost, since in one physical machine, many virtual
machines [3],[12],[13] can be run (Fig.5). Managing and monitoring these virtual
machines also easier work for network administrators.

Fig. 5. Many VM’s Running on a Single Physical Machine

448 B. Loganayagi and S. Sujatha

Running different OS in a Single Physical Machine solves the compatibility

problem (fig.5 & 6) . Entire Physical machine can be moved to VM and Vice Versa
also Possible. Cloning of VM’s Easy, which in turn offers the feature of reduced
downtime. Fault tolerance handling of Virtual Machines are easy. For example, with
the help of images taken, snapshots taken, system can be brought to its original state
in case of problems [5].

Virtual Virtual Virtual Virtual

OS OS OS OS
E.g. E.g. E.g. E.g.
Fedora Fedora FreeDOS DOS
Client Client

Virtual Machine Monitor

(Hypervisor)

Host Machine Resources

Memory, CPU, Disk, Network

Fig. 6. Layered Representation of VMs running on Host machine

If suppose any VM is intruded, or any kind of error, it can be isolated separately,

without affecting other VMs, since isolation is possible in virtual machines. By
carefully monitoring (fig.7,8 & 9) the resource utilization like CPU, memory,
network, port, Isolating the intruded Virtual Machine is possible. By taking
appropriate steps to recover or it can be destroyed for security purposes.
Easily system can be rebuilt with the images taken already. Since Many VMs
running on a single machine, effective utilization of resources are possible. Moving
the VM from one Physical Machine to Another Physical Machine is Possible. Green
Computing is achieved with the help of Virtualization technology. Some services and
application may conflict with each other, that can be run on different VMs. With the
help of Virtualization we can achieve Security, Reliability, Performance, Availability,
Portability, Less Cost, Easy Manageability, Less downtime, etc.
Besides having nice features, single physical machine break down will bring down
all the VMs running in that. By properly updating the VMs Operating system,
installing and updating the firewalls in VMs like the physical machine. A single VM
can be specifically used for all the incoming request whether it is genuine request or
not then it can be forward.
 Improving Cloud Security through Virtualization 449

Fig. 7. Monitoring of Network Usage

Fig. 8. Monitoring of Services and Ports

450 B. Loganayagi and S. Sujatha

Fig. 9. Topology showing system running with unknown services

Fig. 10. Report showing system running with unknown service

If there is any hidden/unknown request (Fig.9,10), along with proper resource

monitoring and port scanning, intruded VM can be easily identified. Upon error,
necessary steps can be taken down to overcome by isolating the intruded VM. Proper
isolation is important. Since resources allocated for each machine is different, that
cannot be accessed through other virtual machine. By giving restricted access/ by
devising proper privilege level all the VMs to access the physical resource, security of
Virtualization can be fully achieved. With Secured Virtualization technologies, cloud
security can be improved.
 Improving Cloud Security through Virtualization 451

6 Conclusion
Cloud computing is going to be one of the venture technology in future, even though
it has some serious issues like security, privacy, social and political issues[11]. Cloud
user should understand their own network, system, applications and data are moving
to an unknown network which poses serious threat to security and privacy [2]. So they
have to vigilantly select Cloud Service Providers by checking their security service
standards and compliances.
Nowadays, Most of the vendors provide free usage to deploy their services in their
cloud with certain limits like hours of usage, or hard disk space, storage or data
transfer or number of end users. As a cloud user or developer, they have to choose the
vendor based on their Service Level Agreements.
In future, more emphasis is going to be in the area of cloud security by defining
new cryptographic methods and key management techniques, since service can be
deployed anywhere in the Cloud , sometimes against the privacy policy depends upon
the regional location where service is deployed that has to be taken in to
consideration.
Using virtualization server, resources, network, desktop, application, operating
system and storage can be virtualized. One of the major concerns in future is
computing with less power. With virtualization, apart from flexibility, scalability,
security, utilizing underutilized resources/idle resources, manageability, cost effective
cloud computing with virtualization technology takes less power since more than one
virtual machines can be run on a single physical machine. In future like IaaS, PaaS
and SaaS, Anything as a Service (XaaS) is going to be possible and can be achieved
through Virtualization.

References
1. Foster, I., Zhao, Y., Raicu, I., Lu, S.: Cloud Computing and Grid Computing 360-Degree
Compared. In: Grid Computing Environments Workshop, GCE 2008, November 12-16
(2008)
2. Harauz, J., Kaufman, L.M., Potter, B.: Data Security in the World of Cloud Computing.
IEEE Security and Privacy 7(4), 61–64 (2009)
3. Loganayagi, B., Sujatha, S.: Creating virtual platform for cloud computing. In: 2010 IEEE
International Conference on Computational Intelligence and Computing Research
(ICCIC), December 28-29, pp. 1–4 (2010), doi:10.1109/ICCIC.2010.5705744
4. Loganayagi, B., Sujatha, S.: Cloud Computing in Stax Platform. In: IEEE International
Conference on Computer Communication and Electrical Technology (IEEE- ICCCET
2011), March 18-19, pp. 1–5 (2011)
5. Zhao, S., Chen, K., Zheng, W.: The Application of Virtual Machines on System Security.
In: 2009 Fourth ChinaGrid Annual Conference, Chinagrid, pp. 222–229 (2009)
6. Zhang, L.-J., Zhou, Q.: CCOA: Cloud Computing Open Architecture. In: IEEE
International Conference on Web Services, ICWS 2009, July 6-10, pp. 607–616 (2009)
7. Viega, J.: Cloud Computing and the Common Man. Computer 42(8), 106–108 (2009)
452 B. Loganayagi and S. Sujatha

8. Cunsolo, V.D., Distefano, S., Puliafito, A., Scarpa, M.: Volunteer Computing and Desktop
Cloud: The Cloud@Home Paradigm. In: Eighth IEEE International Symposium on
Network Computing and Applications, pp. 134–139 (2009)
9. Mei, L., Chan, W.K., Tse, T.H.: A Tale of Clouds: Paradigm Comparisons and Some
Thoughts on Research Issues. In: IEEE Asia-Pacific Services Computing Conference, pp.
464–469 (2008)
10. Reese, G.: Cloud Application Architectures. O’Reilly, Sebastopol (2009)
11. Mather, T., Kumaraswamy, S., Latif, S.: Cloud Security and Privacy. O’Reilly, Sebastopol
(2009)
12. Ritinghouse, J.W., Ransome, J.F.: Cloud Computing: Implementation Management and
Security. CRC Press, Boca Raton (2010)
13. Virtualbox Information,
http://www.virtualbox.org/wiki/Documentation
14. Cloud General Information,
http://en.wikipedia.org/wiki/Cloud_computing
 Selection of Software Testing Technique:
A Multi Criteria Decision Making Approach

Monisha Victor and Nitin Upadhyay

BITS Pilani, K.K. Birla Goa Campus,

Near NH17B Bypass road, Zuari Nagar,
Goa-403726, India
{h2009038,nitinu}@bits-goa.ac.in
http://www.bits-goa.ac.in

Abstract. The appropriate usage of efficient testing technique at any stage of

the Software Development Life Cycle (SDLC) is still in infant stage. There are
a number of testing techniques at various phases of testing. Selection of the
right testing technique at any stage is one of the critical problems. The selection
method should not only take the subjective knowledge but also incorporate
objective knowledge for the efficient selection of testing technique as per the
requirements. Selection of testing technique at every stage of SDLC depends on
many factors such as resources, schedule, cost of the project, etc. Thus we
formulate the selecting of testing technique as a multi criteria decision making
problem and propose an efficient solution.

Keywords: Software Testing, SDLC, Testing Criteria, Selection of testing

techniques, TOPSIS, AHP, MCDM.

1 Introduction
Software Testing is one of the most crucial and important phases of the Software
Development Life Cycle (SDLC). The quality of the delivered product can be assured
by testing the software as early in the SDLC as possible. Testing is done at every
stage of the SDLC with different objective. For example, when a unit or module is
implemented, the module is tested in isolation to find errors in the module [1]. Unit
tested modules are integrated and the objective of integration testing is to test the
integration among the unit tested modules and ensure that the newly added module
does not introduce any new errors [2]. Software Testing evolved with time from
testing being a part of debugging to testing requirements, design and implementation.
For high confidence before deploying a system, the software has to be tested
exhaustively. Since it is not feasible, the system is tested for a particular set of inputs
designed based on testing techniques and checked against the expected outputs. There
are numerous testing techniques of different types for testing requirements,
architecture, design, code, integration of components, system [3][4]. The testing
technique to be chosen at any stage depends on various factors like the software
development method, test objective, cost of the project, schedule of the project,

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 453–462, 2011.
© Springer-Verlag Berlin Heidelberg 2011
454 M. Victor and N. Upadhyay

human resources, computational resources etc [5]. Research has been done to improve
the effectiveness of testing based on reducing the number of test cases [6],
experimental evaluation of testing techniques [7][8], based on data flow relationship
[9], fault coverage [10], prioritizing the test cases based on certain objective function
[11][12], rate at which faults are detected [13], etc. There is a need for a selection
method that considers not only the subjective knowledge but also the objective
knowledge and multiple criteria in selecting the best possible alternative during
various stages of SDLC like unit testing, integration testing or system testing.
The approach in this paper considers the selection of the best testing technique at
any phase of SDLC as a Multi Criteria Decision Making (MCDM) problem and uses
Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) along with
Analytic Hierarchy Process (AHP) to solve it. AHP is used to determine the weights
of the criteria. The weights calculated using AHP are used in TOPSIS, a technique for
solving decision problems and is used to calculate the relative advantages of the
alternatives. The method is based on the concept of TOPSIS that the best alternative
chosen is at a shorter distance from the positive ideal solution and is the farthest from
the negative ideal solution.
The methodology used to solve the problem is discussed in Section 2 and an
illustrative example is given in the Section 3 of the paper.

2 Methodology
The selection of a testing technique can be considered as a multi criteria decision
making problem. The method proposed in this paper uses TOPSIS along with AHP
[14]. TOPSIS is based on the idea that the chosen alternative has the shortest distance
from the positive ideal solution and is the farthest from the negative ideal solution
[15]. AHP is used to calculate the weights of the criteria used for evaluation. AHP
uses consistency ratio as a measure to verify the consistency of the weights obtained
for the criteria. This ensures that the weights of the criteria are consistent.
Figure 1 depicts the flow chart of the methodology proposed.
The following steps need to be performed to choose the best alternative when n
criteria are used to compare m alternatives.

Step 1 - Decision Matrix [D] m × n

In this step, a decision matrix of criteria and alternatives based on the available
information is developed. The value dij of the matrix is the actual value of the ith
alternative in terms of jth decision criteria i.e., the rating of the ith alternative with
respect to the jth criteria. The order of the decision matrix is m × n.

d11 d12 … d1n

Dm×n = d21 d22 … d2n (1)

… … … …
dm1 dm2 … dmn
 Selection of Software Testing Technique 455

Fig. 1. Flow chart of the proposed methodology

Step 2 - Normalized decision Matrix [R] m×n

The decision matrix [D] is normalized so that ratings given in different scales and
units to various criteria are transformed into common measurable units to make
comparisons easy. The order of the normalized decision matrix is m × n. The value rij
of the matrix [R] is calculated using equation (2).
r =d /[ )2]0.5 (2)
ij ij ij
Step 3 – Attribute Weights [W] 1×n
AHP [16] is used in this step to calculate the weights of the criteria used to evaluate
the alternatives. A pair wise comparison of the criteria is done and the absolute
weights of the criteria are calculated using AHP. The pair wise comparison is done on
a scale of 1-9 and the values along with its explanation are shown in detail in Table 1.
456 M. Victor and N. Upadhyay

Table 1. AHP rating scale

Value Explanation
1 Equal importance
3 Moderate importance
5 Strong importance
7 Very Strong importance
9 Extreme importance
2,4,6,8 Intermediate values

If A is the pairwise comparison matrix, the relative weights of the criteria [W] is
calculated using the equation (3).

( A - λmaxI ) × W = 0 (3)
where λmax is the largest eigen value of the matrix A and I is the n × n unit matrix.
The value of consistency ratio while calculating the weights is considered as a
factor which judges the consistency of the absolute weights given to the criteria. Any
value less than 0.1 implies that the judgment is consistent. The consistency ratio is
calculated using the equation (4).
CR = CI / RI (4)
where CI is consistency index and RI is the random consistency index. The value of
the consistency index is calculated using the equation (5).

CI = (λmax – N) / (N – 1) (5)
where N is the number of criteria.
The values of the random consistency index for different values of N are given in
Table 2.

Table 2. Random consistency index

N 1 2 3 4 5 6 7 8 9
RI 0 0 0.58 0.9 1.12 1.24 1.3 1.41 1.45

Step 4 – Weighted Normalized Matrix [V] m×n

The weighted normalized matrix [V] is calculated by multiplying each column of the
normalized decision matrix [R] with the respective column of the [W] matrix. The
value vij of the matrix [V] is calculated using equation (6).
vij = Wj × rij (6)
Step 5 – Positive ideal solution [Is+] n×1 and Negative ideal solution [Is-] n×1
The positive ideal solution is calculated as the best value each criterion can attain. The
negative ideal solution is calculated as the least/worst value each criterion can attain.
They are produced as follows:
 Selection of Software Testing Technique 457

Is+ = { (max vij / j Є τ), (min vij/ j Є τ') for i=1,2,3,…,m}

= {V1+, V2+, … , Vn+} (7)
Is- = { (min vij / j Є τ), (max vij/ j Є τ') for i=1,2,3,…,m}
= {V1-, V2- , … , Vn-} (8)
Where τ is associated with the benefit or positive criteria and τ' is associated with the
cost or negative criteria.
Step 6 – N dimensional Euclidean distance
The N dimensional Euclidean distance is calculated from each alternative to the
positive and negative ideal solution as follows:

Si+ = {∑ j=1 N (Vij – Vj+) } 0.5, i = 1,2,…,m (9)

and
Si- = {∑ j=1 N (Vij – Vj-) } 0.5, i = 1,2,…,m (10)

Step 7 – Relative Closeness [Ci*] m×1

Relative closeness of each alternative with respect to the positive and negative ideal
solution is calculated as follows:
Ci* = Si- / (Si+ + Si-) (11)
Step 8 - Ranking
The alternatives are arranged in the descending order of their relative closeness (Ci*)
from the ideal solutions. The alternative with the maximum value of Ci* is chosen as
the best alternative.

3 Illustrative Example
A software to maintain the Research Information at a University was developed. The
integration among the unit tested modules had to be tested. A list of alternatives
(testing techniques) and criteria were identified and the proposed methodology was
applied to find the best integration testing technique among the alternatives.
The integration testing techniques identified for testing the software are:
A = {Incremental Integration, Top-Down integration testing, Bottom-up
integration testing, Sandwich, Big bang, End-to-End testing, High Frequency testing,
Smoke testing}.
The set of criteria considered for evaluating and selecting the best alternative is
C = {Time when the basic functionality is exposed, Reusability of the test cases,
End user view, Time of fault detection, Effort required in terms of the additional work
to be done, Ease of fixing errors, Frequency of running the tests, Ease of writing the
test cases, and Possibility of automating the testing}.
The following are the steps followed in finding the best alternative among the above
mentioned ones. The decision matrix [D] which contains the rating of the ith
alternative with respect to the jth criteria is shown in Table 3.
458 M. Victor and N. Upadhyay

Table 3. Decision Matrix D

C1 C2 C3 C4 C5 C6 C7 C8 C9
A1 3 4 1 3 4 5 3 4 3
A2 5 4 4 5 3 4 3 4 1
A3 1 1 1 2 3 4 3 5 1
A4 5 3 2 4 3 4 2 3 1
A5 5 2 4 4 3 1 1 3 1
A6 5 5 5 5 4 2 3 3 4
A7 3 3 1 5 4 5 5 3 5
A8 5 3 5 5 3 4 4 4 3

Table 4 gives the decision matrix after normalization [R].

Table 4. Normalized Decision Matrix R

C1 C2 C3
A1 0.250000 0.423999 0.106000
A2 0.416667 0.423999 0.423999
A3 0.083333 0.106000 0.106000
A4 0.416667 0.317999 0.212000
A5 0.416667 0.212000 0.423999
A6 0.416667 0.529999 0.529999
A7 0.250000 0.317999 0.106000
A8 0.416667 0.317999 0.529999
C4 C5 C6
A1 0.249136 0.414781 0.458349
A2 0.415227 0.311085 0.366679
A3 0.166091 0.311085 0.366679
A4 0.332182 0.311085 0.366679
A5 0.332182 0.311085 0.091670
A6 0.415227 0.414781 0.183340
A7 0.415227 0.414781 0.458349
A8 0.415227 0.311085 0.366679
C7 C8 C9
A1 0.331295 0.383131 0.377964
A2 0.331295 0.383131 0.125988
A3 0.331295 0.478913 0.125988
A4 0.220863 0.287348 0.125988
A5 0.110432 0.287348 0.125988
A6 0.331295 0.287348 0.503953
A7 0.552158 0.287348 0.629941
A8 0.441726 0.383131 0.377964
 Selection of Software Testing Technique 459

The weight matrix [W] is calculated using AHP [17] and is shown in Table 5. The
consistency index calculated when finding weights of the criteria in the step 3 is 0.06
which is less than 0.1. Hence the weights calculated are consistent and can be used
further.
Table 5. Weight Matrix W

C1 C2 C3
0.183400 0.112800 0.085500
C4 C5 C6
0.056700 0.115000 0.030200
C7 C8 C9
0.161600 0.200400 0.054300

Weighted Normalized Decision matrix [V] is calculated using the equation (6) and
is shown in Table 6.

Table 6. Weighted Normalized Decision Matrix V

C1 C2 C3
A1 0.045850 0.047827 0.009063
A2 0.076417 0.047827 0.036252
A3 0.015283 0.011957 0.009063
A4 0.076417 0.035870 0.018126
A5 0.076417 0.023914 0.036252
A6 0.076417 0.059784 0.045315
A7 0.045850 0.035870 0.009063
A8 0.076417 0.035870 0.045315
C4 C5 C6
A1 0.014126 0.047700 0.013842
A2 0.023543 0.035775 0.011074
A3 0.009417 0.035775 0.011074
A4 0.018835 0.035775 0.011074
A5 0.018835 0.035775 0.002768
A6 0.023543 0.047700 0.005537
A7 0.023543 0.047700 0.013842
A8 0.023543 0.035775 0.011074
C7 C8 C9
A1 0.053537 0.076779 0.020523
A2 0.053537 0.076779 0.006841
A3 0.053537 0.095974 0.006841
A4 0.035691 0.057585 0.006841
A5 0.017846 0.057585 0.006841
A6 0.053537 0.057585 0.027365
A7 0.089229 0.057585 0.034206
A8 0.071383 0.076779 0.020523
460 M. Victor and N. Upadhyay

Positive and Negative ideal solutions [Is+] and [Is-] are calculated from the
weighted normalized decision matrix using the equation (7) and (8) and is shown in
Table 7.
Table 7. Positive and Negative Ideal solution Is+ and Is-

C1 C2 C3
+
Is 0.076417 0.059784 0.045315
-
Is 0.015283 0.011957 0.009063
C4 C5 C6
+
Is 0.023543 0.47700 0.013842
-
Is 0.009417 0.035775 0.002768
C7 C8 C9
+
Is 0.089229 0.095974 0.034206
-
Is 0.017846 0.057585 0.006841

The N-dimensional Euclidean distance from each of the alternative to the Positive
and Negative ideal solution is calculated using equation (9) and (10) and is shown in
Table 8.

Table 8. Euclidean distance from Positive and Negative Ideal solution (IS+, IS-), relative
closeness and ranking of the testing techniques

A1 A2 A3 A4
S+ 0.395931 0.343459 0.486917 0.435656
S- 0.403377 0.448899 0.287030 0.360109
C* 0.504658 0.566536 0.370866 0.452532
Rank 5 3 8 6
A5 A6 A7 A8
S+ 0.458015 0.298710 0.359335 0.298881
S- 0.331204 0.479840 0.436294 0.479734
C* 0.419661 0.616325 0.548363 0.616138
Rank 7 1 4 2

The option which has the highest value of relative closeness is end-to-end testing
technique. The characteristics of end-to-end testing like exclusive focus on end user
view, reusability of test cases [18] which are some of the criteria used to evaluate the
testing techniques makes it the best choice among the available alternatives. The
option with the least value of relative closeness is Bottom-up testing technique. Some
of the characteristics of bottom-up integration testing which makes it not the best
choice among the alternatives are non-reusability of the test cases, no end-user view
from the beginning of the integration, effort in terms of coding drivers [2]. This shows
that the methodology proposed considers all the criteria to evaluate the testing
techniques and gives the appropriate result.
This method can be used in any software organization for the right selection of
testing technique at any stage of the SDLC. Criteria like resources required (human or
 Selection of Software Testing Technique 461

computational), previous use of a testing technique in the organization, ease of fixing

the defects by the developers, training required before the use of the technique by a
tester can also be used by the testing team according to the schedule, cost or resources
requirement of the project in the process of decision making. Thus the method
considers not only the subjective knowledge but also practical aspects of the testing
techniques in choosing the best testing technique to be used.

4 Conclusion
Selecting the right testing technique is very critical for the success of the project. The
best suitable testing technique at various phases of the software development life
cycle can be chosen based on various factors. The problem of choosing the best
suitable testing technique is formulated as a Multi Criteria Decision Making problem
and a method using TOPSIS and AHP is proposed for the same. This method can
select a testing technique from among M alternatives available based on N selection
or decision criteria. An illustrative example is demonstrated to show the applicability
of the proposed approach.

References
1. Ntafos, S.C.: A comparision of some structural testing strategies. IEEE Transactions on
Software Engineering 14(6), 868–874 (1988)
2. Naik, K., Tripathy, P.: Software Testing and Quality Assurance. Wiley, Chichester (2008)
3. Amar, M., Shabbir, K.: Systematic Review on Testing Aspect oriented Programs -
Challenges, Techniques and Their Effectiveness. Blekinge Institute of Technology. In:
MSE-2008, 17 (2008)
4. Juristo, N., Moreno, A.M., Vegas, S.: State of the Empirical Knowledge on Testing
Techniques. In: Proceedings of the International Conference on Software Maintenance
(1999)
5. Vegas, S.: Characterisation schema for selecting software testing techniques (2002)
6. Mahapatra, R.P., Singh, J.: Improving the Effectiveness of Software Testing through Test
Case Reduction. World Academy of Science, Engineering and Technology 37, 345–350
(2008)
7. Roper, M., Miller, J., Brooks, A., Wood, M.: Towards the Experimental Evaluation of
Software Testing Techniques. In: Proc. EuroSTAR 1994 (1994)
8. Frankl, P.G., Weiss, S.N.: An Experimental Comparison of the Effectiveness of Branch
Testing and Data Flow Testing. IEEE Transactions on Software Engineering 19(8), 774–
787 (1993)
9. Clarke, L.A., Podgurski, A., Richardson, D.J., Zeil, S.J.: A comparison of data flow path
selection criteria. In: Proceedings of the Eighth International Conference on Software
Engineering, pp. 244–251 (1985)
10. Basili, V.R., Selby, R.W.: Comparing the effectiveness of software testing. IEEE
Transactions on Software Engineering SE-13(12), 1278–1296 (1987)
11. Rothermel, G., Untch, R.H., Chu, C., Harrold, M.J.: Test Case Prioritization: An Empirical
Study. In: Proceedings of the International Conference on Software Maintenance, pp. 179–
188 (1999)
462 M. Victor and N. Upadhyay

12. Williams, L., Snipes, W., Meneely, A.: On Increasing System Test Effectiveness through a
Test Case Prioritization Model Using Static Metrics and System Failure Data. In:
Reliability Analysis of System Failure Data Workshop (2007)
13. Ma, Z., Zhao, J.: Test case prioritization based on analysis of program structure. In: Asia-
Pacific Software Engineering Conference, pp. 471–478 (2008)
14. Upadhyay, N., Deshpande, B.M., Agrawal, V.P.: Integrated decision approach for COTS
selection. Int. J. Decision Sciences, Risk and Management 2(3/4), 165–177 (2010)
15. Jadidi, O., Hong, T.S., Firouzi, F., Yusuff, R.M., Zulkifli, N.: TOPSIS and fuzzy multi-
objective model integration for supplier selection problem. Journal of Achievements in
Materials and Manufacturing Engineering 31(2), 762–769 (2008)
16. Saaty, T.L.: The Analytic Network Process. RWS Publications, Pittsburgh (1996)
17. James, T.: Whalen & Associates, LLC, http://www.whalen.ws/
18. Tsai, W.T., Bai, X., Paul, R., Shao, W., Agarwal, V.: End-to-end Integration Testing
Design. In: Computer Software and Applications Conference, pp. 166–171 (2001)
 A Hybrid Evolutionary Algorithm for the
Page Number Minimization Problem

Dharna Satsangi, Kamal Srivastava*, and Gursaran

Department of Mathematics,
Dayalbagh Educational Institute,
Agra, India
{dharna.satsangi,kamalsrivast,gursaran.db}@gmail.com

Abstract. In this paper we present a hybrid evolutionary algorithm (HEA) for

solving the pagenumber problem. In HEA, random depth first search of the
graph is used for placing the vertices on the spine and an edge embedding
heuristic is designed to distribute the edges on a minimal number of pages. The
results show that the algorithm achieves the optimal pagenumber for most of
the standard graphs tested by us. HEA performance is also compared with the
GA described by Kapoor et al. [1] on select instances of standard and random
graphs. It is observed that HEA gives a better solution for most of the instances.

Keywords: Pagenumber problem, Hybrid Evolutionary Algorithm, Depth First

Search.

1 Introduction

The book embedding problem consists of embedding a graph in a book with its
vertices in a line along the spine of the book and its edges on the book in such a way
that edges residing on the same page do not cross. The pagenumber of a graph G is
the minimum number of pages of the book into which G can be embedded. This
problem abstracts layout problems arising in direct interconnection networks, fault-
tolerant processor arrays, fault tolerant VLSI design, sorting with parallel stacks and
single row routing [1]. The problem of obtaining pagenumber is NP-complete for
general graphs [1]. However, this problem has been studied for some standard graphs
and their page numbers have been found. Pagenumbers of FFT (Fast Fourier
Transforms), Benes, Barrel Shifter networks and t-partite graphs [2], complete graphs,
grids, hypercubes, trees, X-trees, pinwheel graphs [3], planar graphs, deBruijn and for
shuffle-exchange graphs ([4], [5]) are known in literature. Shahrokhi et al. [6] have
given polynomial time algorithms to generate near optimal drawing of graphs on
books. Berhart et al. [7], showed that the pagenumber is less than or equal to one if
and only if the graph is outerplanar. Swaminathan et al. [8] have shown that
bandwidth-k graphs can be embedded in a book of k-1 pages, though it is not a tight

*
Corresponding author.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 463–475, 2011.
© Springer-Verlag Berlin Heidelberg 2011
464 D. Satsangi, K. Srivastava, and Gursaran

bound when considered on standard graphs with known bandwidth. Upper and lower
bounds on the pagenumber of k-ary hypercubes are given by Bettayeb et al. [9]. In
order to deal with the problem of obtaining the pagenumber for general graphs, a
genetic algorithm (GA) for finding the pagenumber has been proposed by Kapoor et
al. [1]. However, they have reported the results for only a few standard graphs of
small sizes. In this paper a Hybrid Evolutionary Algorithm (HEA) has been designed
for the pagenumber problem (PNP). In devising a solution to PNP, one has to focus
on two main factors: 1) placing the vertices on the spine and 2) distributing the edges
over the pages of the book. In HEA, vertex labels generated using the random depth
first search (RDFS) are used for placing the vertices on the spine and a distribution
heuristic motivated by the method for obtaining the pagenumber of complete graphs
[3] is used for placing the edges on the pages of the book. Experimental results show
that HEA outperforms genetic algorithm proposed by Kapoor et al. [1] for PNP on the
graphs on which the two were compared.
In this paper, the proposed hybrid evolutionary algorithm is described in Section 2
and the implementation details are described in Section 3. In Section 4 the
experiments and results obtained are described. Section 5 presents the conclusions of
the work described in the paper.

2 Hybrid Evolutionary Algorithm for PNP

2.1 Notations

Let , be an undirected graph with vertex set and edge set . A labeling of
the vertices of an undirected graph with | | vertices is a bijective
function : 1,2, … . A pair of edges and cross in a drawing iff
1 and both lie on the same page.

Fig. 1. Edge crossing condition

2.2 Motivation

A depth first search (DFS) based technique for placing the vertices on the spine
(a particular placement of vertices on the spine is also referred to as a layout of
vertices in this paper) seems to be helpful in minimizing the number of crossings
when edges are placed in a single page as it helps in generating layouts in which there
are edges between adjacent nodes. In order to confirm this, we carried out an
experiment on random connected graphs to explore the number of crossings on a
 A Hybrid Evolutionary Algorithm for the Page Number Minimization Problem 465

single page obtained by randomly generated layouts and by DFS based layouts. For
the experiments, 50 graphs of 50 vertices having an edge density 30%, where edge
density of a graph with of vertices and edges is taken as 2 1 , were
generated. Mean crossings and standard deviation for the DFS based layouts were
found to be 17147.32 and 2050.042 as compared to 20943.42 and 2470.513 for the
random layouts. Keeping in view the results of the experiment, we have generated an
initial population of layouts for HEA using random depth first search (RDFS) where
the root vertex and the neighbors of a visited vertex are chosen randomly.
PNP remains a difficult problem even when the layout is fixed [10]. Given a layout
there is no guarantee that a particular embedding of edges onto the pages is minimal.
Chung et al. [3] have given an edge embedding strategy that provides the optimal
pagenumber for complete graphs. The edge density of complete graphs is 100% and
the layout does not play a role in determining the pagenumber. Since any simple
graph that is not a complete graph will have an edge density that is less than 100%,
the approach proposed by Chung et al. [3] can be used to develop a strategy for
distributing edges on pages. The edge embedding heuristic (EEH) has been developed
by us based on the strategy proposed by Chung et al. [3] to determine a distribution of
edges using minimal pages. It is described in detail in Section 3.1.

2.3 Solution Representation

In HEA, for a graph , , , ,…, , each solution, , is

represented in two components:
1) An array of length that represents a labeling . Vertices are placed on
the spine in the label sequence 1, 2, … , .
2) An | | 3 array , where each row contains the end vertices of an edge in
the first two columns and the third column stores the page index in which the
edge is placed.
The pagenumber of a solution is max 3 and is
denoted _ , . We will use the term ‘optimal pagenumber’ for the
optimal value of page number.

2.4 The Algorithm

Evolutionary algorithms (EAs) are the approximate optimization algorithms that are
based on evolutionary principles. They are inspired by nature's capability to evolve
living beings well adapted to their environment. At the core of EA is a population of
individuals. In each iteration in the algorithm, reproduction operators are applied to the
individuals of the current population to generate the individuals of the next generation.
Recombination operators may also be used to recombine two or more individuals to
produce new individuals. The algorithm may also use mutation operators that cause
self-adaptation of individuals. The driving force in the EAs is the selection of
individuals based on their fitness (which might be based on the objective function or
some kind of simulation experiment). Individuals with higher fitness have a higher
probability to be chosen as members of the next generation (parents) which corresponds
466 D. Satsangi, K. Srivastava, and Gursaran

to the principle of survival of fittest in natural selection. Simulated annealing (SA)

proposed by Kirkpatrik et al. [11] is another powerful technique used to solve
optimization problems. The algorithm is of sequential nature and converges slowly. The
robustness of EAs is usually enhanced when they are hybridized with other meta
heuristics such as tabu search [12], SA etc. Yip and Pao [13] proposed to incorporate the
process of SA into the selection process of EA and to utilize its power to guide local
search. We first give a detailed outline of the algorithm involving the hybridization of
EA with SA in the context of PNP, which is followed by its pseudo code.
HEA starts with building an initial population of parents consisting of pop_size
number of solutions produced by RDFS and then applying EEH to distribute the
edges on the pages. The SA parameters namely, initial temperature , final
temperature , and cooling ratio are initialized. Also a non-zero positive integer
is assigned to _ ,1 _ , where _ indicates the
number of children to be produced by parent in the generation. It means
initially each parent is given an equal chance to produce children. Clearly, the total
number of children to be produced by the parents is _ . After
generating the initial population of parents the iterative process (steps 7 to 23)
starts in which each parent generates _ number of children using IDFS
(see section 3.3) and EEH. We refer to the population of children as , where
denotes the child of parent . Now for each , 1 _ ,
child with lowest pagenumber ( _ ) is determined for finding the parents for
the next generation. This child becomes the parent for the next generation if its
pagenumber is less than that of parent; otherwise there is a probability that the child
will be selected to become a parent (Steps 12 to 20). Next important step is to decide
_ for the next generation of parents . This decision is based on the
competition between parents wherein the performance of a parent is measured by
comparing the pagenumber of each of its children with the _ ,
where _ is the lowest pagenumber obtained so far in this process.
Pseudo code for computing _ is given in the procedure _ _ at the
end of this section. The iterative process is repeated until no improvement in the
_ is observed for a specified count ( ) of generations or when
final temperature is reached.
The pseudo code for HEA
1. Initialize , number of children to be produced by each parent of the
initial generation , cooling ratio , initial temperature , final
temperature .
2. Set 0; T;
3. for 1 _
4. _ ;
5. endfor
6. Generate initial population of parents using RDFS and then
apply EEH to distribute the edges on the pages
7. Repeat
8. for 1 _
 A Hybrid Evolutionary Algorithm for the Page Number Minimization Problem 467

9. Apply IDFS on , _ times to create children

,1 _ and then apply EEH to distribute the
edges on the pages
10. endfor
11. Apply mutation on × _ × number of children from .
12. for 1 _
13. Find child with lowest pagenumber i.e. _
14. _ , _ _ );
15. if 0 or exp / // is a random number lying
between 0 and 1
16. _ ;
17. else
18. ;
19. endif
20. endfor
21. Update
22.
23. Until (stopping criteria is satisfied)
Procedure
1.
2. for
3.
4. for
5. ;
6. if > or β/T >
7.
8. endif
9. endfor
10.
11. endfor
12. for
13. )
14. endfor

3 Implementation of Hybrid Evolutionary Algorithm

3.1 The Edge Embedding Heuristic (EEH)
Steps:
• Consider a labeling of vertices in graph .
• Now for a complete graph with vertex labels , place the edges
( denotes an edge with endpoints and ) of in a sequence
according to the procedure below:
468 D. Satsangi, K. Srivastava, and Gursaran

Initialize to the empty sequence;

for

endfor

where += means concatenate with the sequence obtained from the previous
step and addition and subtraction are modulo operators assuming labels are
placed in a circle. Note that S includes all the edges in a complete graph of n
vertices with some edges possibly included more than once.
• Now . Insert edges in in the same sequence as they
appear in in edge matrix , without duplicating the edges.
• Place the edges in one by one, in sequence, starting on page one using the
following rule:, Assign page one to the first edge. For each in ,
, assign the lowest numbered page to if adding to the page
does not result in a crossing. Otherwise assign a new page to .
• Pagenumber=total number of pages assigned.

3.2 Initial Population

An initial population consists of solutions. For

each is generated using RDFS and EEH is used to generate .
In this section we take an example, shown in Fig. 2a, to illustrate the components of a
solution.

(a) (b)

Fig. 2. (a) A graph and (b) its DFS tree

Starting from vertex 1, RDFS of the graph (DFS tree is shown in Fig 2(b)) provides
the following labeling:

= 1 2 11 5 4 3 6 8 9 10 7
 A Hybrid Evolutionary Algorithm for the Page Number Minimization Problem 469

In Fig.2b, for each vertex, the number in the parenthesis corresponds to its depth
first index, i.e., the label of the vertex.
EEH gives the distribution of edges in the array as

= 1 2 5 7 7 2 1 7 5 2 10 8 8 4 4 6 1 9 5 5 6
Edges
2 3 9 8 11 6 6 9 6 4 11 10 9 5 6 7 8 10 7 11 11
Page # 1 1 1 1 1 1 2 1 1 3 2 2 1 1 2 2 3 1 1 4 2

The pagenumber for this solution is 4.

3.3 Generating Children

We use a unary reproduction operator, intermediate depth first search (IDFS) to

produce children from the parents, which is explained as follows:
Let the parent and the child are denoted by arrays
and respectively. A vertex is selected randomly.
The label of this vertex is . Labels of vertices with labels less than or equal
to are copied from to at the respective positions.
These vertices are referred to as and remaining ones are referred to
as . This process is shown in pseudo code below:

1. for
2. if
3.
4. endif
5. endfor

Now starting from the vertex , DFS of the vertices of the graph is
carried out. The labeling of the vertices starts from and
the labels are stored at the respective positions of array . This ensures that a
new obtained child solution also has a DFS based labeling. Thus IDFS helps to
explore various possible DFS trees of the graph starting with the same root vertex.
Fig 3 shows a DFS tree obtained from that shown in Fig 2(b) when the RDFS is
carried out after the randomly chosen vertex 6.
470 D. Satsangi, K. Srivastava, and Gursaran

Fig. 3. A DFS tree obtained by applying IDFS on the tree of Fig. 2.

Application of EEH on this layout provides the following edge distribution over
the pages. The pagenumber for this solution is 3.

= 1 2 11 10 9 3 4 7 8 6 5

EEH gives the distribution of edges in the array Q as

= 1 2 4 5 5 9 8 2 1 8 6 5 7 6 1 4 10 7 7 5 2
Edges
2 3 6 7 11 10 10 6 6 9 7 9 11 11 8 5 11 8 9 6 4
Page # 1 1 1 1 1 1 1 1 2 1 1 1 1 2 2 1 1 3 3 1 1

3.4 Mutation

In HEA, insert mutation is used to perturb the population of children . In this, a

vertex is selected randomly and is placed at another randomly chosen position on the
spine. This facilitates the entry of non-DFS solutions in the population. Mutation is
applied on number of randomly selected children, where is
mutation rate.

4 Experiments and Results

In this section, we first describe the classes of graphs that we use in the experiments.
We then present the results of our experiments for HEA and the genetic algorithm
(GA) given in Kapoor et al. [1]. The code is written in MATLAB 7.0 and the machine
specifications are: Intel Core 2 Duo E8400 @ 3.00 GHz, 2.87 GB of RAM.
 A Hybrid Evolutionary Algorithm for the Page Number Minimization Problem 471

4.1 Test Sets

We use two types of graphs to test HEA: standard graphs and random connected
graphs (RCGs).

Standard graphs: Following classes of standard graphs are used:

• Complete graphs ( )
• Hypercubes : The hypercube, of dimension , is a -regular graph with
vertices and edges. Each vertex is labeled by a distinct d-bit binary
string, and two vertices are adjacent if they differ in exactly one bit.
• Cube-connected cycles : The Cube-connected cycle of dimension is
formed from by replacing each vertex with a -cycle of vertices in
and then joining each cycle vertex of the corresponding neighbor of in . It
has vertices, edges.
• Complete binary tree : The binary tree is a tree structure in which each
node has at most two child nodes. The binary tree is complete if the nodes are
well balanced.
• Pinwheel : The depth-n pinwheel graph, has vertices
and and edges connecting each pair of vertices of the form

where
• Star : The star graph is a complete bipartite graph , a tree with one
internal node and leaves. It has vertices and edges.
• Triangulated triangle : The Triangulated triangle graph, is a graph whose
vertices are the triples of non-negative integers summing to l, with an edge
connecting two triples if they agree in one coordinate and differ by one in the
other two coordinates.
• Complete bipartite graphs : A complete bipartite graph, with
is a simple bipartite graph such that two vertices are adjacent if
and only if they are in different partite sets.
• Shuffle Exchange : Its vertices are 0-1 vectors of length . There is an
edge between any vertex and vertex .
Also, there is an edge between any vertex and the vertex
, .
472 D. Satsangi, K. Srivastava, and Gursaran

Random Connected Graphs (RCGs): We generate RCGs of sizes 40 and 50, and
edge densities varying from 10% to 50%. For each combination of size and edge
density, 20 different graphs are randomly generated for the experiments. We use
RCGs (1) to study the effect of edge density on pagenumber and (2) to compare HEA
with GA.

4.2 GA Settings and Parameter Tuning

To compare HEA with the GA of Kapoor et al [1], we coded the GA with cyclic
crossover and insert mutation operator as no specific operator is mentioned in their
paper. Furthermore, as other GA parameters such as population size, mutation rate
etc. are also not specified, the various parameters involved in HEA and GA are set
after performing preliminary experiments on the randomly generated graphs. For GA,
the size of initial population ( ) is set as whereas for HEA,
=10, = 10, which is the number of children generated initially, are fixed.
The SA parameters, namely, initial temperature ( ), final temperature ( ), and
cooling ratio ( ) are chosen as 1, .01 and 0.99 respectively. Mutation rate ( ) of 0.2
is used in both the algorithms. In both the algorithms, the iterative process is repeated
until no improvement is observed in minimum pagenumber for 50 consecutive
iterations.

4.3 Results

The results of HEA and GA on various graphs are shown in Table 1 through Table 4.
The results of standard graphs with known optimal pagenumbers are given in Table 1.
HEA performs better than GA in most of the cases. The results in bold indicates the
cases for which an algorithm has better performance than the other. HEA attains
optimal pagenumber for most of the graphs. The pagenumber for complete graphs are
all optimal as expected since EEH is based on edge distribution strategy for the
optimal pagenumber of these graphs. GA does not perform well, for the hupercubes as
the pagenumbers obtained by it are substantially higher than the optimal pagenumber
which is . HEA attains these values easily when tested for . Both the
algorithms attain the optimal pagenumber 1 for the complete binary trees and star
graphs which are obvious as these graphs are outerplanar. HEA is able to obtain
optimal pagenumbers of pinwheel graphs ( ) tested for = 8, 16 and 32.
Table 2 contains the results of those classes of standard graphs for which optimal
pagenumbers are not known and for which some upper bounds are available for them.
The results for shuffle exchange and triangulated triangles are much less than their
known upper bounds. HEA and GA both give large pagenumbers of complete
bipartite graphs. The reason for this may be attributed to the fact that the optimal
pagenumber is obtained with a non-DFS layout.
Experimental results for random graphs of sizes 40 and 50 are shown in Table 3
and Table 4 respectively. Results (Fig. 4) clearly demonstrate that HEA outperforms
GA in terms of solution quality though the elapsed time of HEA is more than that of
GA. The difference in the page numbers obtained by HEA and GA is more significant
in the case of higher density graphs.
 A Hybrid Evolutionary Algorithm for the Page Number Minimization Problem 473

Table 1. Comparison of GA and HEA for the standard graphs with known optimal
pagenumbers

Graphs
GA 7 7 8 9 86 135 2 3 6 2 3 7 13
HEA 6 7 7 8 50 75 2 3 3 2 3 4 5
Optimal 6 7 7 8 50 75 2 3 3 2 3 4 5

Table 2. Comparison of GA and HEA for the standard graphs with the upper bounds on the
pagenumbers

Graphs
GA 4 7 1 2 3 3 3 3 3 4 4 3 4 5 6 7 8 9
HEA 3 5 1 2 3 3 3 3 2 3 4 4 5 6 7 8 9 10
Upper 5 5 3 3 3 3 4 5 6 7 8 3 4 5 5 6 7 7
Bound

Table 3. Comparison of GA and HEA for the random graphs ( )

Edge density Pagenumber Elapsed Time (in sec.)

(%) GA HEA GA HEA
10 6.6 6.6 147.278 202.65
20 11.2 10.8 244.8094 347.581
30 14 13.2 391.4998 492.0376
50 18.6 16.4 736.2366 795.784

Table 4. Comparison of GA and HEA for the random graphs ( )

Edge density Pagenumber Elapsed Time (in sec.)

(%) GA HEA GA HEA
10 9.4 9.2 220.4186 490.2936
20 14.8 14 455.9404 845.9158
30 19 16.8 1.0004e+003 1.20776e+003
50 24.2 20.8 1.86306e+003 1.9547e+003
474 D. Satsangi, K. Srivastava, and Gursaran

Fig. 4. Edge density vs Pagenumber graph for HEA and GA of sizes (a) 40 and (b) 50

5 Conclusion
We have presented a hybrid evolutionary algorithm for the pagenumber problem in
which a random depth first search of the graph is employed for placing the vertices on
the line. An edge embedding heuristic is designed to distribute the edges on a minimal
number of pages. The results show that the algorithm achieves the optimal
pagenumber for most of the standard graphs tested by us. HEA outperforms the GA
described by Kapoor et al. [1] in terms of solution quality for most of the instances
tested.
 A Hybrid Evolutionary Algorithm for the Page Number Minimization Problem 475

Acknowledgement. This work is supported by University Grants Commission

(UGC), New Delhi vide letter no 36-66/2008 (SR).

References
1. Kapoor, N., Russell, M., Stojmenovic, I.: A Genetic Algorithm for finding the Pagenumber
of Interconnection Networks. J. Parallel and Distributed Computing 62, 267–283 (2002)
2. Games, R.A.: Optimal Book Embeddings of the FFT, Benes, and Barrel Shifter Networks.
Algorithmica 1, 233–250 (1986)
3. Chung, F.R.K., Leighton, F.T., Rosenberg, A.L.: Embedding Graphs in Books: a Layout
Problem with Applications to VLSI design. Siam J. Alg. Disc. Math. 8(1), 33–58 (1987)
4. Hasunuma, T., Shibata, Y.: Embedding deBrujin, Kautz and Shuffle-exchange Networks
in Books. Discrete Applied Mathematics 78, 103–116 (1997)
5. Obrenic, B.: Embedding deBrujin and Shuffle-exchange graphs in 5 pages. In: Proc. ACM
SPAA, pp. 137–146 (1991)
6. Shahrokhi, F., Shi, W.: On crossing sets, disjoint sets and Pagenumber. J.
Algorithms 34(1), 40–53 (2000)
7. Bernhart, F., Kainen, P.C.: The Book Thickness of a Graph. J. Combin. Theory Ser. B 27,
320–331 (1979)
8. Swaminathan, R.P., Giriraj, D., Bhatia, D.K.: The Pagenumber of the class of Bandwidth-k
Graphs is k-1. Information Processing Letters 55, 71–74 (1995)
9. Bettayeb, S., Hoelzeman, D.A.: Upper and Lower Bounds on the Pagenumber of the Book
Embedding of the k-ary Hypercube. J. Digital Information Management 7(1), 31–35
(2009)
10. Garay, M.R., Johnson, D.S., Miller, G.L., Papadimitriou, C.H.: The Complexity of
Coloring Circular Arcs and Chords. SIAM J. Algebraic Discrete Methods 1, 216–227
(1980)
11. Kirkpatrick, S., Gelatt, C.D., Vechhi, M.P.: Optimization by Simulated Annealing.
Science 220(4598), 671–680 (1983)
12. Glover, F., Laguna, M.: Tabu Search, 2nd edn. Kluwer, Boston (1997)
13. Percy, P.C., Yip, Pao, Y.-H.: Combinatorial Optimization with use of Guided Evolutionary
Simulated Annealing. IEEE Transactions on Neural Networks 6, 2 (1995)
 Trust-Enhanced Recommendation of Friends in Web
Based Social Networks Using Genetic Algorithms to
Learn User Preferences

Vinti Agarwal and Kamal K. Bharadwaj

School of Computer and Systems Sciences,

Jawaharlal Nehru University, New Delhi, India
{vinti.agarwal3,kbharadwaj}@gmail.co.in

Abstract. Web-based social networks (WBSNs) are a promising new paradigm

for large scale distributed data management and collective intelligences. But the
exponential growth of social networks poses a new challenge and presents
opportunities for recommender systems, such as complicated nature of human
to human interaction which comes into play while recommending people. Web
based recommender systems (RSs) are the most notable application of the web
personalization to deal with problem of information overload. In this paper, we
present a Friend RS for WBSNs. Our contribution is three fold. First, we have
identified appropriate attributes in a user profile and suggest suitable similarity
computation formulae. Second, a real-valued Genetic algorithm is used to learn
user preferences based on comparison of individual features to increase
recommendation effectiveness. Finally, inorder to alleviate the sparsity problem
of collaborative filtering, we have employed trust propagation techniques.
Experimental results clearly demonstrate the effectiveness of our proposed
schemes.

Keywords: Web-Based Social Networks, Friend Recommender System,

Collaborative Filtering, Profile Similarity, Real-Valued Genetic Algorithm,
Trust Propagation.

1 Introduction
Web-based Social Networks (WBSNs) are growing dramatically in size and number.
The huge popularity of sites like Facebook, LinkedIn, Delicious, YouTube and others
has drawn in hundreds of millions of users, and the attention of scientists and the
media[1]. The public accessibility of WBSNs offers great promise for researchers
interested in studying the behavior of users and how to integrate social information
into applications. The Social Web therefore provides huge opportunities for
recommender technology.
With the explosive growth of resources available through the Internet,
information overload has become a serious concern. Web based recommender
systems(RSs) are the most notable application of the web personalization to deal
with problems of information and product overload[3][4]. Our Friend RS based on
collaborative filtering (CF), the most widely used and commercially applied

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 476–485, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Trust-Enhanced Recommendation of Friends in Web Based Social Networks 477

recommendation technique. Recommendation of friends in WBSNs poses a different

set of challenges w.r.t. traditional recommending techniques such as those for movies,
books and restaurants etc.
Here, we proposed a friend recommender system that suggests new friends to users
in WBSNs. The main contributions of the work presented in this paper are as follows:
• Identification of appropriate attributes in a user profile to find similarity with other
users in WBSNs and suggestion of suitable similarity computation schemes.
• Learning user preferences, through real-valued GA algorithm, based on individual
features in an intuitive and efficient manner to increase recommendation
effectiveness.
• Use of trust propagation techniques towards handling sparsity problem of
collaborative filtering.
The rest of the paper is organized as follows: Section 2 summarizes related studies
about social recommender systems and collaborative filtering. Section 3 introduces
Friend Recommender System and set out the procedure used for profile building,
similarity computation, weight learning through real-valued GA and at the end
methodology used to handle sparsity problem based on trust propagation. The
experimental evaluation of various scheme is presented in section 4. Finally, last
section discusses the findings and presents our conclusions.

2 Social Recommender Systems(SRS) and Collaborative Filtering

Social Recommender Systems (SRSs) aim is to alleviate information overload over
social media users by presenting the most attractive and relevant content, often using
personalization techniques[4],adapted for the specific user. CF is a classical method,
in we rely on recommendation about a user based on making prediction about a user
preferences of a group of a user that are considered similar to this user. The most
popular technique used to measure the similarity between users based on Pearson’s
correlations identified the list of nearest neighbors of user.
The key challenging factors that extensively affect the quality of recommendations
are privacy and reciprocity [5] and attract major attention of several researchers to
find better way to connect people. Other studies have incorporated social network
elements into traditional CF [6],[7] but in our study we incorporate CF in WBSNs.

3 Proposed Friend Recommender System Framework

Recommendation systems for social networks discussed in [8] are different from other
kinds of systems, since the item recommended here are rational human beings rather
than goods. Thus, the ‘Social’ factor has to be accounted for when making a
recommendation. In the following section, details of proposed friend recommender
system are discussed.

3.1 User Profiling and Friendship Trends

In FRS a user profile is important because it records relevant information such as
preferences and behavior pattern of individual users [9], [10]. Some factors are most
478 V. Agarwal and K.K. Bharadwaj

important and strongly affect the relationship in social networks e.g. Geographic
location is strongest factor affecting how relationship shapes up in social network.
Presence of attributes such as hometown, city and state higher up in the testimony to
that fact; relationships are more likely to develop between those belonging to same
age group etc. Based on the popular WBSNs, we have identified list of 13 attributes to
maintain user profile and divided into two groups i.e. Basic attributes and Behavioral
attributes associated with preferences (Table 1 and Table 2).This grouping of
attributes allows us to better understand the trend immerging in social networks.

Table 1. Basic attributes Table 2. Behavioural attributes

Similarity Measures
One of the important steps in CF is to identify the users with similar inclinations,
which leads to the formation of neighborhood set. In order to identify them several
similarity measures have been proposed in the literature [11],[12],[13]. Here, we used
a quantity called as weight associated with individual attributes similarity to reflect
their importance while making recommendations. These weights are generated offline
to every attribute for each user and determined in such a way that sum of all weights
is equal to 1, i.e. ∑ 1. Initially we considered all weights are equally distributed,
indicating that user u is giving equal importance to all features.
The proposed method combines weights with other similarity of each feature to
emphasize the total similarity of target user with other who shares appreciation on
several features with active users. Overall similarity between target user and other
user computed as in [11], aggregating all the similarities of individual attributes,
which is given by
, , , …. , (1)
Where
. overall similarity of user u with other users
w1,w2……..w13 weights for each attributes
. similarity function for each attribute.

Determination of sim() Function for.

• Attribute A1: Gender Biasedness
One plausible way of doing so is to use rating scheme assigned by user u to its friend.
We keep list of persons how many user u has male and female friends and then
 Trust-Enhanced Recommendation of Friends in Web Based Social Networks 479

compute male(or female) ratio by dividing total number of friends of the given user u
as shown in Table 3.
The formula to compute the gender similarity for user i has in j is defined as:
M MJ F FJ
, 1 and can also be written as

, 1 M MJ or 1 |F FJ | (2)
• Attributes A2, A3 ,A4 and A5: Geographical separation, Language , Relationship
status and religion
, 1, 0 (3)

Table 3. Computation of male and female ratio

Male Female Total Male Female

friends friends friends ratio ratio(F)
(M)
User i Nmi Nfi Ni Nmi/ Nfi/ Ni= Fi
Ni=Mi
User j Nmj Nfj Nj Nmj/ Nj= Nfi/ Nj= Fj
Mj

• Attribute A6: Socio-Economic Status(SES)

Attribute A6 could be a better dimension towards the efficient friend recommendation
because for all above attributes users are provided space to identify them. Rather than
omit this variable we have included it also in overall similarity computation. Here we
divide the monthly income into three categories: Low, Medium and High. Then after
categorization, we aggregate the difference between their class with the father’s
occupation to compute final similarity of SES.
• Attributes A7, A8, A9, A10, A11 , A12 and A13: Here for, Career, Common music,
Movies ,Activities, Books and Nature respectively.
While computing the similarity for these attributes we are taken into consideration the
preferences order. The preferences information corresponding to each feature for a
particular user is represented by a vector called user profile preference
vector(UPPV).In this way UPPV of a given feature A for a user u captures the
preferences given as : UPPVU( A)=(P1, P2, P3 ).
In our experiment, we assume the maximum length(L) of UPPV vector as 3.It may
vary depending on the number of preferences on common items. It could be zero if
there is no common choice between two users. Here Pi assumes numeric value.
Consequently, we formalize the similarity of these attributes for active user a with
others based on the difference of preferences. As the difference between preferences
decreases similarity value gets increased.

, ∑ 1 (4)
480 V. Agarwal and K.K. Bharadwaj

The various cases generated during similarity computation based on preferences are
tabulated as follows:
Case 1: if all items are common, the number of possible cases for the order of
preferences between users would be C(3,3) Χ C(3,1) Χ C(3,3) Χ C(2,1) Χ C(1,1)=6.
Case 2 : Similarly ,for two common items : C(3,2) Χ C(3,1) Χ C(2,1) Χ C(1,1)=18.
Case 3 : And for one : C(3,1) Χ C(3,1)=9.

3.2 Learning Weight by Using Real Valued Genetic Algorithm

The above defined similarity measures treats all attributes equally, taking all wi’s
equal, while trying to evaluate recommendation to a user. This may not truly reflect
the contribution of each attribute towards similarity. The fact that “how much” a user
liked an attribute is important than he has liked it. To make it possible we proposed
real valued GA method to learn these weights in an offline learning process and then
used to compute the similarity value during prediction.
Traditional GA where chromosome representation technique is bit strings suffers
from slow convergence due to the large chromosome structure, especially in
optimization problems involving several parameters. We have used real valued GA, to
learn the weights for each of the similarity measures as real numbers and therefore a
floating point representation of chromosomes is most suited for our purpose. The
chromosome is a set of 13 weights where the value of each weight lies between two
real values which indicate maximum and minimum value as discussed in [14]. The
chromosome search space for real valued GA is defined as
Ω = { w Є Rm | ≤ ≤ ,… ……. ≤ ≤ } (5)
Genetic operators being used here are different from traditional operators described
in[15]. GA terminates either a maximum number of generations elapsed or a desired
level of fitness is reached.

Fitness Function
Fitness function validates the diversion of process towards its optimization goal by
allowing best individuals to breed that lead to good recommendation [3]. Recursively
the individuals that achieve a higher fitness are more likely to be selected as parents
for next generation. For this, every set of weights employed in RS rerun on entire set
including weights generated after crossover and mutation operators in order to
compute the predicted value of rating for each user. The average difference between
actual and predicted rating for all users in training set is used as fitness score for that
set of weights.
∑ | | (6)

3.3 Handling Sparsity Problem Using Trust

One of the biggest challenges of designing a collaborative filtering system is how to

provide accurate recommendations with the sparse user profile. Trust has been
 Trust-Enhanced Recommendation of Friends in Web Based Social Networks 481

successfully used to enhance neighborhood set in order to provide accurate

recommendations [16] ,[17],[18] with sparse data. The following subsections discuss
how to utilize trust value to obtain more accurate recommendation when data is
highly sparse.

Trust Computation
Trust metric compute an estimate of what degree a user should trust on other, if are
unknown to each other. It can be computed in several ways such as using similarity or
rating etc. In our study we are computing trust between user a and b based on
difference of rating assigned by a and b to their common friends [19]. The formula to
compute trust between them is given as :
∑
, ,
, 1 (7)

Here, is the rating assigned by a and assigned by b and n is the total number
of common friends and rmax is the max value of rating scale i.e. 5.

Trust Propagation

By formalizing trust relationship, trust between strangers can be established by

analyzing path linking parties together. In this paper we use the max-min based
approach described in [20] to express transitive trust relationship. Let us illustrate this
with Fig.1.user a trusts b and b trusts c then it is reasonable to say that a trusts c to a
certain degree but in real life trust is not always transitive [16]. However trust could
not completely transfer without decay through distance. Its values reduce as path
length increases. So propagation distance is a constraint in transitive relationship. As
shown in Fig.1 [17] there are two paths between a and c .Hence their associated
inferred trust is given by formula (14).

Fig. 1. Trust Propagation

,
max min , , , , min , , , , , , ,
(8)
482 V. Agarwal and K.K. Bharadwaj

3.4 Main Steps of the Proposed Friends Recommender Systems

Step 1 : Similarity computation at the attribute level.
i.e. , , , ,………… , between user u and ui .
Step 2 : Weight learning through real-valued GA.i.e. w1, w2…………, w13.
Step 3: Generating enhanced neighborhood sets based on trust propagation.
Step 4 : Making recommendation using CF.

4 Experimental Results and Discussion

4.1 The Dataset

To demonstrate the effectiveness of our proposed techniques experiments on

simulated dataset of 20 users is conducted, which forms a small social network with
each node representing people or users and ties between them denoting relationship
along with a rating score from 1-5. In the first step of our experiment, we used hybrid
of content and traditional CF for recommendation on both equal and learned weights.
The actual ratings of an active user are randomly divided into two sets: Testing set
and Training set. Training dataset used to construct neighborhood set for rating
prediction. In the next experiment we demonstrate method for alleviating the sparsity
problem using trust inferences and result shows that trust has potential to improve the
way recommendations are made.

4.2 Results and Discussions

Fig. 2 illustrates the deviation of CFEW (collaborative filtering with equal weights)
predicted rating score from actual rating score, compared to the deviation of CFLW
(collaborative filtering with learned weights) for the same neighborhood size N=8
over 30 runs of the experiment. It is clear from the Fig. 2 that the lower value of MAE
corresponds to the better performance. During the experimentation of CFLW we
consider initial population size 20 where weights are real numbers and range between
0 and 30. Then 20 new individuals being produced in each generation after applying
crossover and mutation operators.

Fig. 2. Comparison of MAE of two Fig. 3. Comparing MAE based on

neighborhood based methods: CFLW and different neighborhood size
CFEW plotted over 30 runs
 Trust-Enhanced Recommendation of Friends in Web Based Social Networks 483

In our experiment GA followed by an elitist approach such that the process ends
when the fitness function value for the best chromosome remains the same in 10
consecutive iterations. Fig. 3 represents the variation in MAE as neighborhood size N
varies. The evaluation results show that, instead of improving the prediction results,
increasing the neighborhood size might add some bad neighbors, which lead to bigger
prediction error.
Table 4 presents the MAE comparison of various schemes. The results show that
introduction of trust significantly improve the accuracy of recommendations. Hence,
trust based rating prediction obviously reduced the deviation from actual rating score
as compared to the deviation of CFEW and CFLW. MAE of trust based rating
prediction is 0.0652 while that of CFEW and CFLW are 0.9417 and 0.4226
respectively.

Table 4. MAE for different prediction algorithms

Prediction Algorithms MAE(N=8)

CFEW 0.9417
CFLW 0.4226
Trust based CFEW 0.0778
Trust based CFLW 0.0652

5 Conclusion and Future Work

We have presented a Friend Recommender System, that exploits web information to

recommend relevant friends on social-networking websites. Recommendation are
based on social network information using traditional CF. The method of evolving
individuals weights by real-valued GA according to their preferences has significantly
enhanced the performance. To deal with sparsity problem we employed a method
based on trust inferences which provide additional information to CF algorithm in
terms of enhanced neighborhood sets. The experimental results clearly show that the
trust based CFLW outperforms rest of the schemes i.e. CFEW, CFLW and trust based
CFEW on sparse datasets. The experiments in this work are performed on a simulated
dataset and we plan to conduct experiments on larger dataset obtained through
survey/WBSNs.
Our future work will focus on applying different techniques to compute and
propagate trust [16]. It is to be seen how two level filtering based on both trust and
reputation [21] and utilization of various sparsity measures [13] can be incorporated
in the proposed system to further enhance its performance.

References
1. Guy, I., Ronen, I., Wilcox, E.: Do You Know? Recommending People to Invite into Your
Social Network. In: Proceedings of the 14th International Conference on Intelligent User
Interfaces, IUI 2009. ACM, New York (2009)
484 V. Agarwal and K.K. Bharadwaj

2. Chen, J., Dugan, C., Muller, M., Guy, I.: Make New Friends, but Keep Old-
Recommending People on Social Networking Sites. In: Proceedings of the
27th International Conference on Human Factors in Computing Systems, CHI 2009,
pp. 201–210 (2009)
3. Al-Shamri, M.Y.H., Bharadwaj, K.K.: Fuzzy-Genetic Approach to Recommender Systems
Based on a Novel Hybrid User Model. Expert Systems with Applications 35, 1386–1399
(2007)
4. Adomavicius, G., Tuzhilin, A.: Personalization Toward the next generation of
recommender systems: A survey of the state-of-the-art and possible extensions. IEEE
Transaction on Knowledge and Data Engineering 17(6), 734–749 (2005)
5. Garcia, R., Amatriain, X.: Weighted Content Based Methods for Recommending
Connections in Online Social Networks. In: Recsys 2010, ACM, Barcelona (2010), 0-
12345-67-8
6. Liu, F., Lee, H.J.: Use of Social Network Information to Enhance Collaborative Filtering
Performance. Expert Systems with Applications 37(7), 4772–4778 (2009)
7. Massa, P., Avesani, P.: Trust Aware Recommender Systems. In: RecSys 2007:
Proceedings of 2007 ACM Conference on Recommender Systems (2007)
8. Karkada, U.H.: Friend Recommender System for Social Networks. SI583 Term Paper,
School of Information, University of Michigan (2009)
9. Hsu, W.H., King, A.L., Paradesi, M.S.R., Pydimarri, T., Weninger, T.: Collaborative and
Structural Recommendation of Friends using Weblog-Based Social Network Analysis. In:
AAAI Spring Symposia 2006 on Computational Approaches to Analysing Weblogs (2006)
10. Banks, L., Wu, S.F.: All Friends are NOT Created Equal: An Interaction Intensity based
Approach to Privacy in Online Social Networks. In: International Conference on
Computational Science and Engineering, IEEE.CSE 2009, pp. 970–974 (2009)
11. Anand, D., Bharadwaj, K.K.: Adapting User Similarity Measures for Recommender
Systems:A Genetic Programming Approach. In: 2010 3rd IEEE International Conference
IEEE Computer Science and Information Technology, ICCSIT (2010)
12. Anand, D., Bharadwaj, K.K.: Enhancing Acuuracy of Recommender System through
Adaptive Similarity Measures Based on Hybrid Features. In: Nguyen, N.T., Le, M.T.,
Świątek, J. (eds.) Intelligent Information and Database Systems. LNCS, vol. 5991,
pp. 1–10. Springer, Heidelberg (2010)
13. Anand, D., Bharadwaj, K.K.: Utilizing Various Sparsity Measures for Enhancing
Accuracy of Collaborative Recommender Systems based on Local and Global Similarities.
Expert Systems with Applications 38(5), 5101–5109 (2011)
14. Corcoran, A.L., Sen, S.: Using Real Valued Genetic Algorithm to Evolve Rule Sets for
Classification. In: IEEE-CEC (1994)
15. Michalewicz, Z.: Genetic algorithms + data structures = evolution programs. AI. Springer,
New York (1992)
16. Josang, A., Hayward, R., Pope, S.: Trust Network Analysis with Subjective Logic. In:
Proceedings of the 29th Australasian Computer Science Conference (ACSC 2006), vol. 48
(2006)
17. Victor, P., Cock, M.D., Cornelis, C.: Recommender Systems Handbook. In: Kantor, P.B.,
Ricci, F., Rokach, L., Shapira, B. (eds.) Chapter in Recommender Systems Handbook, pp.
645–676. Springer, Heidelberg (2010)
18. Papagelis, M., Plexousakis, D., Kutsuras, T.: Alleviating the Sparsity Problem of
Collaborative Filtering Using Trust Inferences. In: Herrmann, P., Issarny, V., Shiu, S.C.K.
(eds.) iTrust 2005. LNCS, vol. 3477, pp. 224–239. Springer, Heidelberg (2005)
 Trust-Enhanced Recommendation of Friends in Web Based Social Networks 485

19. Fazeli, S., Zarghami, A., Dokoohaki, N., Matskin, M.: Mechanizing social trust-aware
recommenders with T-index augmented trustworthiness. In: Katsikas, S., Lopez, J.,
Soriano, M. (eds.) TrustBus 2010. LNCS, vol. 6264, pp. 202–213. Springer, Heidelberg
(2010)
20. Luo, H., Niu, C., Shen, R., Ullrich, C.: A Collaborative Filtering Framework Based on
Both Local User Similarity and Global User Similarity. Machine Learning 72(3), 231–245
(2008)
21. Bharadwaj, K.K., Al-Shamri, M.Y.H.: Fuzzy Computational Models for Trust and
Reputation Systems. Electronic Commerce Research and Applications 8(1), 37–47 (2009)
 Modeling and Verification of Server Aided
Verification Protocol Using NuSMV

Vikram Saralaya, Kishore J.K., Sateesh Reddy, and Sanjay Singh

Department of Information and Communication Technology,

Manipal Institute of Technology,
Manipal University, Manipal-576104, India
[email protected]

Abstract. Cryptographic algorithms are useful for signing and verifying

the authenticity of sender of the message. The verifier may not have the
required computational ability and relies on a powerful server to aid
the verification process. The server and the illegitimate prover/signer
may collaborate and try to cheat the verifier. A legitimate prover can
also repudiate the message sent by himself. In this paper we model the
scenario where the legitimate, cheating or repudiating prover sign’s the
message. The verifier then authenticates the message via an untrusted
server. Specifications are written using CTL(Computational-Tree Logic).
NuSMV(extension of Symbolic Model Verifier) is the tool used to verify
the specifications.

1 Introduction
Cryptographic protocols have rarely been bug-free in its initial design. These
programs are critical, for they work on the security aspect of a communication
system. An organization would expect a certain degree of assurance for the
designed protocol, before implementing it for their internal security. Extensive
testing may expose some design flaws, but its only a best effort service. We need
to formally model and verify these protocols.
Formal methods can help us detect bugs as well as improve the quality of our
design. Initially a model of the protocol is developed which has all the features of
its real time counterpart. The properties expected from the protocol is specified
and the conformance of the model to the specification is verified. The protocol
which goes through these three stages is expected to be more reliable and robust
than that merely based on testing.
There has been a lot of research focused on developing better models for
Server-aided verification. Girault et. al. [1] gave a clear picture about server-
aided verification in theory and practice. Chow et. al. in [2] has discussed a
generic pairing based server-aided verification protocol and claims it to provide
better security compared with Girault et. al. in [1].
In server-aided verification there are three major entities - prover, verifier
& server. The prover initially sign’s a message to prove its authenticity to the
verifier. Various computational intensive algorithms exist to sign and verify the

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 486–495, 2011.

c Springer-Verlag Berlin Heidelberg 2011
 Modeling and Verification of Server Aided Verification Protocol 487

messages. These schemes may have to be incorporated in smart cards and other
low cost chips with low computational capability. Such verifiers which cannot
carry out computations on its own requests some powerful but probably un-
trusted server to carry out major part of the computation. Wu et. al. in [3] has
further worked on [1] and introduce the existential unforgeability concept. They
claim their work to be more secure but computationally more intense. More work
on similar lines were carried out by Wu et. al. in [4], Pfitzmann et. al. in [5] and
Begiun et. al. in [6]. To the best of our knowledge, no work has been reported in
literature to formally specify and verify the Server-Aided Verification Protocol.
In this paper we model and verify the properties specified in [1].
The server being untrusted gives rise to two major problems, first the illegit-
imate prover and server can collaborate and make the verifier accept an invalid
signature. The second problem is when a legitimate prover tries to repudiate
the message sent by himself with the help of the server. In this paper we ad-
dress these problems by formalizing the server-aided verification protocol. The
properties we expect from the model, such as non-repudiation etc. are given as
specifications.
NuSMV model verifier [7][8] is used to verify whether the model indeed meets
its specification. The prover module is designed such that it can assume the
role of a legitimate prover, cheating illegitimate prover or the one which tries to
repudiate. The verifier module is modeled using two approaches slightly different
from one another. The NuSMV outputs whether the prover succeeds or f ails in
authenticating itself to the verifier.
This paper is organized as follows. Section 2 gives an introduction to Server
Aided Verification process followed in the paper. Section 3 discusses the detailed
formalization and modeling of the protocol. Section 4 is about the properties any
authentication protocol should satisfy and how these can be formally specified
using CTL. Section 5 briefly explains about the verification results obtained.
Finally section 6 concludes the paper.

2 Server Aided Verification Protocol Using RSA Scheme

The message signed by the prover is authenticated by the verifier with the help
of the server as shown in Fig. 1. The signing can be done using any of the existing
popular schemes. Let us consider the RSA(Rivest, Shamir and Adleman) scheme
[9]. The functionalities of the three entities are as described below:

2.1 Prover/Signer
The prover can use its own private key (n,d) and sign the message. The encrypted
message f(m) is used to produce the signature. A legitimate prover will produce
the signature as shown below:

σ = f (m)d mod n (1)

488 V. Saralaya et al.

Fig. 1. Basic structure of Server Aided Verification protocol

If the prover tries to cheat then it will choose a value Ȳ and generate a suitable
signature using this value as shown below:

σ = f (m)d Ȳ −2 mod n (2)

If instead the prover tries to repudiate the message sent by itself, then it will
generate the signature normally, but send σ̄ which is co-related to σ.

2.2 Verifier
Verifier uses the public key (n,e) of the prover to authenticate the signature. Its
job is to check whether the following equality holds:

σ e mod n = f (m) (3)

If this equality holds then the verifier can assume that the prover is indeed the
one he claims to be. Due to computational constraint the verifier delegates a
part of the computation to the server. The delegation can be carried out using
two approaches. In the first approach the verifier asks the server to compute:

Y = σ (e−1)/2 mod n (4)

The value of Y, so obtained from the server can be used as shown below:

Y 2 σ mod n = f (m) (5)

Evaluating the above equality requires less computational resources and hence
can be performed by the verifier easily. In the second approach verifier chooses
two values e1 and e2 such that,

e = e1 ∗ e2 , e1 << e2 (6)
 Modeling and Verification of Server Aided Verification Protocol 489

Now, the server has to compute:

Y = σ e2 mod n (7)

The value of Y, so obtained from the server can be used as shown below:

Y e1 mod n = f (m) (8)

In section 5 we will further see the pros and cons of the two approaches with
respect to the given specifications.

2.3 Server
A sincere server which has not collaborated with any misbehaving prover will
return the value of Y as explained in the previous section. On the other hand if
the server has collaborated with a cheating illegitimate prover then the precom-
puted value Ȳ will be returned by the server. Finally, if the legitimate prover
has collaboration with the server, then the server returns the suitable value of
Y computed from the values provided by the repudiating prover : σ, σ̄.

3 Modeling the Server Aided Verification Protocol

The model of the protocol involves three modules : prover module, verifier mod-
ule and the server module. Each module has its own set of state variables to
hold the values computed by them.

3.1 Prover Module

This module maintains its current state represented by the variable prover.state,
its type of signing represented by the variable prover.sign and the signature of
the encrypted message represented by the variable prover.signature. The state
space of these state variables are given in Table 1.

Table 1. State Variables in the Prover Module

Variable State Space

prover.state idle, signed, success, failure
prover.sign none, legitimate, correlated, cheat
prover.signature positive integer

Initially before computing the signature it will be in the idle state. Once the
message is signed, a transition to signed state takes place. The prover remains in
this state until the verifier completes the authentication process. If the authen-
tication is successful then the prover moves to success state else to the f ailure
state.
490 V. Saralaya et al.

3.2 Verifier Module

The current state information is held in the variable verif ier.state. In this paper
we model both the approaches discussed in section 2.2 for the verification process.
The state space of the state variables are depicted in Table 2.

Table 2. State Variables in the Verifier Module

Variable State Space

verifier.state idle, get y1 from server, get y2 from server
verifier.result none, success, failure

If the verifier decides to use the first approach then it transits to the
get y1 f rom server state. Instead to opt for the second approach it will have
to move to the get y2 f rom server state. Once the server returns the Y value,
the state of the verifier will be verif y.
In the verif y state the verifier has to use the value returned by the server
and verify whether the prover is authentic. The verification result is stored in a
separate state variable verif ier.result.

3.3 Server Module

The server will remain in the idle state as long as the verifier does not issue a
request to help it in the verification process. This is represented using the state
variable server.state. If the verifier had chosen the first approach then the server
computes the value for the variable server.y1 else the value of server.y2 will be
set. These state variables and their state space are given in Table 3.

Table 3. State Variables in the Server Module

Variable State Space

server.state idle, compute y1, compute y2, computed y1, computed y2
server.y1 -1 or positive integer
server.y2 -1 or positive integer

If the first approach is chosen then the value of server.y2 will remain -1
because it will not be used by the verifier for its computation. On the other
hand, if the second approach is chosen then the computed value is assigned
to server.y2 and hence server.y1 will remain -1. In order to track the cur-
rent computation taking place in server module we maintain the compute y1
and compute y2 to represent the ongoing computation. Once the computa-
tion gets over the state changes to either computed y1 or computed y2. This
change acts as an indicator for the verifier to carry out its final verification
process.
 Modeling and Verification of Server Aided Verification Protocol 491

3.4 State Machine Representation of the Protocol

The Finite State Machine(FSM) representation of protocol P can be written as

P = {V, Q, π, δ, q0 }

where,

V = { prover.state, prover.sign, prover.signature,

verifier.state, verifier.result,
server.state, server.y1, server.y2
}

Q = set of states in the model, with values assigned to each v ∈ V

π: { Q → D(V) | D(V) contains range of each v ∈ V }

δ : Transition Function

q0 : Initial state ∈ Q

Initially before the prover signs the message and the verifier verifies the same,
each of these has to be initialized to a stable state. From this stable state we
can systematically model the signing and verification process. The initial values
assigned to each of the state variables are as shown in Table 4.

Table 4. Initial State of the Protocol

Variable Value Variable Value Variable Value

prover.state idle verifier.state idle server.state idle
prover.sign none verifier.result none server.y1 -1
prover.signature -1 server.y2 -1

From the initial state the prover assumes the role of either a legitimate, cheat-
ing or repudiating(by generating a correlated signature) prover. In Fig. 2 a pos-
sible transition sequence is shown. It can be seen that once the prover signs the
message and enters the signed state in q2 , the verifier has an option to choose
one of the two approaches, i.e. either q3 or q4 .
Based on the chosen approach the further computations take place and finally
the verifier completes the verification and the result will be either success or
f ailure. In Fig. 2, since the prover is legitimate the final state q9 will have
verif ier.result = success.
492 V. Saralaya et al.

Fig. 2. A possible sequence of transitions from the idle state to final verif y state of
verifier

3.5 Example

In the previous sections we have defined the different state variables which are
required to model the protocol. To generate and verify the signature we will
have to further define some variables whose values remain constant during a
complete verification cycle. Here cycle refers to the transition from idle state of
each module to final state and back to the idle state again. The values we need
to define are given in Table 5.
Now, let us consider p = 2, q = 5, then n = 10. Let public key be e = 5 then
as per the RSA algorithm the private key is d = 1. Let f (m) = 9. The legitimate
prover will generate the signature(σ) as σ = 91 mod 10, i.e. σ = 9. In the rest

Table 5. Defined Variables in the Protocol

Variable Value Variable Value

p prime number f(m) encrypted value for message
q prime number sig f (m)d mod n
n p∗q sig correlated correlated value to sig
d private key of prover sig cheat generated using random Y value
e public key of prover
 Modeling and Verification of Server Aided Verification Protocol 493

of the paper we represent σ as sig. If the illegitimate prover wants to cheat

instead, then it will randomly choose Ȳ , say Ȳ = 3 and generate the signature
as sig cheat = 9 ∗ 3−2 mod 10, i.e. sig cheat = 1. When the server gets request
from verifier using the first approach it just sends the known value Ȳ = 3. The
verifier finally computes whether 32 ∗ 1 mod 10 = f (m), i.e. 9 = 9 and hence the
signature is accepted. Thus this approach fails when a illegitimate prover tries
to cheat the verifier.

4 Specification and Model Checking of Server Aided

Verification Protocol
A model is never right or wrong in itself. It either satisfies the property desired
out of it or it does not. To verify this, we need to specify the properties in a
way that is understandable by the system. Consider our model P with the states
qi . Now we have to check whether the model satisfies a given property φ in a
particular state as shown below.

C, si |= φ

The property φ can be specified in a suitable logic. For our model we need to
quantify over different paths rather than different states in a chosen path that
change with time. To accomplish this temporal aspect we prefer the CTL. The
specifications in this paper are coded using the syntax of CTL [10].
The desired properties of a good authentication protocol are :
– φ1 : Illegitimate prover should not cheat the verifier
– φ2 : Non-repudiation property has to be ensured
– φ3 : Legitimate provers signature should always succeed to prove his identity
The first property φ1 can be specified using CTL as shown below:

!EF(prover.sign = cheat & verifier.result = success)

It means if the provers intention is to cheat then the verifier must not produce
the result as success. In order to ensure non-repudiation we need to make sure
that the model will not accept a correlated signature generated by a legitimate
prover. This property φ2 can be specified using CTL as shown below:

!EF(prover.sign = correlated & verifier.result = success)

The third property ensures that the model correctly authenticates a valid prover.
This property φ3 can be specified using CTL as shown below:
prover.state = legitimate →

A[prover.state = legitimate U verifier.result = success]

It states that if the prover is legitimate, then in every path(A) prover continues to
remain in legitimate state until(U) verifier succeeds in its authenticating process.
494 V. Saralaya et al.

5 Verification Results and Discussion

The properties specified in the section 4 has to be evaluated against the model de-
veloped in the section 3. To verify it we use the symbolic model verifier NuSMV.
The model developed in section 3 is given as one of the inputs to the system
and the properties are stated using the keyword ’SPEC’. All the three properties
discussed in the section 4 has been checked in the model.
The first two properties φ1 and φ2 fails in the model and hence NuSMV
produces a counterexample for the same. The property φ1 (marked as 1 inside the
circle in Fig. 3) fails because in the first approach deployed by the verifier, if the
server returns the precomputed value by collaborating with the prover(marked
as 2 inside the circle in Fig. 3), then the verifier wrongly accepts (marked as 3
inside the circle in Fig. 3) the invalid signature. The complete state transition is
shown by the screen shot in Fig. 3.
The second property φ2 fails for a similar reason. In this case though the value
of Y is not precomputed, the server knows σ and the correlated value of σ. Thus
when the verifier sends the request, it can easily help the prover in its attempt
to repudiate the message.
The third property φ3 holds in true in the model. This is because when the
prover is legitimate every path will lead to a state wherein verifier successfully
authenticates the prover. Thus both the approaches chosen by the verifier satisfy
φ3 .

Fig. 3. Screenshot : Verification of Property φ1 developed in 4 which fails to hold in

the model
 Modeling and Verification of Server Aided Verification Protocol 495

6 Conclusion

Designing cryptographic algorithms and communication protocols is a complex

process. Formal methods are increasingly being used in the design of these sys-
tems. It provides the stake holders a certain degree of assurance that the product
is bug free. In this paper we have tried to model and verify the basic Server Aided
Verification Protocol. Some of the possible loopholes in the approach were de-
tected using NuSMV as the model verifier tool. This is just a starting work in
the direction and a lot of scope for improvement exists. We considered only the
prover being legitimate, cheating and repudiating. There are other attacks that
the prover or the server itself can come up with. From the verifier point of view
we have considered two approaches to deal with the server. Further work can be
carried out to develop better approaches to delegate work to the server.

References
1. Girault, M., Lefranc, D.: Server-aided verification: Theory and practice. In: Roy,
B. (ed.) ASIACRYPT 2005. LNCS, vol. 3788, pp. 605–623. Springer, Heidelberg
(2005)
2. Chow, S.S.M., Au, M.H., Susilo, W.: Server-aided signatures verification secure
against collusion attack. In: Proceedings of the 6th ACM Symposium on Infor-
mation, Computer and Communications Security, ASIACCS 2011, pp. 401–405.
ACM, New York (2011)
3. Wu, W., Mu, Y., Susilo, W., Huang, X.: Server-aided verification signatures: Def-
initions and new constructions. In: Baek, J., Bao, F., Chen, K., Lai, X. (eds.)
ProvSec 2008. LNCS, vol. 5324, pp. 141–155. Springer, Heidelberg (2008)
4. Wu, W., Mu, Y., Susilo, W., Huang, X.: Provably secure server-aided verification
signatures. Comput. Math. Appl. 61, 1705–1723 (2011)
5. Pfitzmann, B., Waidner, M.: Attacks on protocols for server-aided rsa computation.
In: Rueppel, R.A. (ed.) EUROCRYPT 1992. LNCS, vol. 658, pp. 153–162. Springer,
Heidelberg (1993)
6. Begiun, P., Quisquater, J.J.: Fast server-aided rsa signatures secure against active
attacks. In: Coppersmith, D. (ed.) CRYPTO 1995. LNCS, vol. 963, pp. 57–69.
Springer, Heidelberg (1995)
7. Cavada, R., Cimatti, A., Jochim, C.A., Keighren, G., Olivetti, E., Pistore, M.,
Roveri, M., Tchaltsev, A.: Nusmv 2.4 user manual (2005)
8. Cavada, R.: Nusmv: a new symbolic model checker (2010), http://nusmv.fbk.eu/
9. Stallings, W.: Cryptography and Network Security: Principles and Practice, 5th
edn. Prentice Hall, Englewood Cliffs (2010)
10. Huth, M., Ryan, M.: Logic in Computer Science, Modelling and Reasoning about
Systems, 2nd edn. Cambridge University Press, Edinburgh (2005)
Modeling and Verification of Inter Realm Authentication
in Kerberos Using Symbolic Model Verifier

Punit Mundra, Madhavi Sharma, Shobhit Shukla, and Sanjay Singh

Department of Information and Communication Technology

Manipal Institute of Technology, Manipal University, Manipal-576104, India
[email protected]

Abstract. In open distributed environment several users accesses the network

resources on server, server will allow only authenticated users to access these
resources. So it has become of prime importance for nodes communicating over
a non-secure network to prove their identity to one another in a secure manner.
Various authentication protocols is used for the this purpose, Kerberos protocol
is one of them. Several versions of the protocol exist. The aim of this paper is
modeling and verification of the Inter Realm Authentication and User to User
Authentication in Kerberos protocol through NuSMV model checker. It also
demonstrate that when presence of intruder in the system, and make use of
service, it will generate the counter example.

1 Introduction

Authentication protocol are security mechanism whereby each party is assured its
identity to one another. One of the authentication protocol which is commonly used is
Kerberos. Kerberos was developed in the Athena Project at the Massachusetts
Institute of Technology (MIT) [1]. It is a computer network authentication protocol,
which allow nodes communicating over a non-secure net- work to prove their identity
to one another in a secure manner. It is designed as a client/server model and it
provides mutual authentication i.e. both the client and server verify each other's
identity. Several versions of the protocol exist, versions 1 to 3 occurred only
internally at MIT. Steve Miller and Clifford Neuman, the primary designers of
Kerberos version 4, published that version in the late 1980s, although they had
targeted it primarily for Project Athena.Version 5, designed by John Kohl and
Clifford Neuman then appeared with the intention of overcoming the limitations and
security problems of version 4.
Modeling the protocol through model checker is the best way to find the behavior
of the protocol. Modeling checking is an automatic, model-based verification
approach. It evaluates the all possible state and ease to find the concurrent behavior of
the system. Model checking is based on temporal logic. Temporal logic expresses the
dynamic aspect of the system.
For modeling the concurrent aspect Computation Tree logic (CLT) [2] is used.
Computation Tree logic is basically used for Verification of hardware and
communication protocol. For Modeling Kerberos, we use NuSMV model checker.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 496–506, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Modeling and Verification of Inter Realm Authentication in Kerberos 497

This work has been mainly motivated by the previous work on Model Checking of
Kerberos Protocol using New Symbolic Model Verifier (NuSMV) [3][4]. M. Panti
et.al [3] has tried to express the basic version of Kerberos. Shreya et. al [4] also has
used NuSMV for the basic version of Kerberos for authentication. They have
presented the concept of freshness and which helps to find the reply attack. They have
demonstrated a possible means by which the weakness of the Kerberos protocol
causing the replay attack can be overcome hence making the protocol more stronger.
In this paper we have considered the kerberos protocol with two realm and try to
find the reply attack. We have also demonstrated the authentication forwarding in
kerberos 5. In presence of intruder NuSMV will generate the counter example. The
aim of paper is to model and verify the kerberos protocol by means of NuSMV
[5][6][7].
Model Checking, one of many formal verification methods [8], is an attractive and
increasingly appealing alternative to simulation and testing to validate and verify
systems. Given a system model and desired system properties, the Model Checker
explores the full state space of the system model to check whether the given system
properties are satisfied by the model. The Model Checker either verifies the given
properties or generates a counter examples.
When the model checker verifies a given system property, it implies that all
behaviors have been explored, and the question of adequate coverage or a missed
behavior become irrelevant. There are two main advantages of using model checking
compared to other formal verification methods : -
1. It is fully automatic, and
2. It provides a counter example whenever the system fails to satisfy a given
property.
The remainder of the paper is organized as follows. Section 2 discusses about the
theoretical background of kerberos protocol and limitation of kerberos 4. Section 3
describes how to model the Kerberos protocol in NuSMV. Section 4 shows how the
authentication takes place and the results obtained in this work.
Finally conclusion has been drawn in section 5.

2 Theoretical Background
Kerberos provides a centralized authentication server whose function is to
authenticate users to servers and servers to users. Two versions of Kerberos are in
common use version 4 and version 5 [9].

2.1 Kerberos Version 4

This version contains mainly four entities:-

1. Client (C): - An entity which wants to make use of any service hosted on a
server.
2. Server (V): - An entity which hosts different services which clients request
for.
498 P. Mundra et al.

3. Authentication Server (AS): - A trusted third party which knows the

passwords of all users and stores these in a centralized database. Also, the
AS shares a unique secret key with each server.
4. Ticket Granting Server (TGS): - TGS, issues tickets to clients who have been
authenticated to AS. Thus, the client first requests a ticket-granting ticket
from the AS. The client module saves this ticket. Each time the user requires
access to a new service, the client applies to the TGS, using the ticket to
authenticate itself. The TGS then grants a ticket for the particular service.
The client saves each service-granting ticket and uses it to authenticate its
user to a server each time a particular service is requested.
The protocol also makes use of timestamps to verify that sender's clock is
synchronized with that of receiver and to indicate the time at which ticket was
generated, and the lifetime which informs about the lifetime of the ticket, It prevents
replay after ticket has expired. The basic messages involved are shown in Fig.1.

Fig. 1. Authentication Scenarios in Kerberos

2.2 Kerberos Realms

A Kerberos realm is a set of managed nodes that share the same Kerberos database.
Networks of clients and servers under different administrative organizations typically
constitute different realms. However, users in one realm may need access to servers in
other realms, and some servers may be willing to provide service to users from other
realms, provided that those users are authenticated. Kerberos provides a mechanism
for supporting such inter realm authentication [9].
The Kerberos server in each inter operating realm shares a secret key with the
server in the other realm. The two Kerberos servers are registered with each other.
The scheme requires that the Kerberos server in one realm trust the Kerberos server in
the other realm to authenticate its users. Furthermore, the participating servers in the
second realm must also be willing to trust the Kerberos server in the first realm. A
user wishing service on a server in another realm needs a ticket for that server.
Following are the messages exchanged in inter realm authentication:
Here AS is authentication server, CA is client in one realm A who wants to access
service from server VB in second realm B TGSA is ticket granting server in realm A
and TGSB is ticket granting server in realm B.
 Modeling and Verification of Inter Realm Authentication in Kerberos 499

In this message, the client requests a ticket-granting ticket by sending its

identity and password to the AS, together with the identity of TGS,
indicating a request to use the TGS service and a time stamp, so that the AS
knows that the message is timely.

The AS responds with a message, encrypted with a key derived from the
user's password that contains the ticket requested by client. This message
also contains session key Kc,tgs to be shared between TGS and Client. The
client retrieves this key.
The ticket is encrypted with TGS's key Ktgs which is already shared with
AS. The ticket contains session key Kc,tgs which is retrieved by TGS. Like
this, the session key has been securely delivered to both C and the TGS. The
ticket also contains ADc which prevents the use of ticket from workstation
other than one that initially requested the ticket.

The client sends message to TGS stating the identification of the TGS in
another realm, tickettgs, an authenticator which includes the identity of client
C, address of Client ADc and a time stamp. The authenticator is intended for
use only once and has a very short lifetime. The TGS can decrypt the ticket
with the key that it shares with the AS. The TGS uses the session key Kc,tgs to
decrypt the authenticator. The TGS can then check the identity and address
from the authenticator with that of the ticket and with the network address of
the incoming message. If all match, then the TGS is assured that the sender
of the ticket is indeed the ticket's real owner.

The TGS then sends to C as a response to request. This message is encrypted

by session key shared between TGS and client and includes a session key to
be shared between Client and TGS_rem, the identification of TGS in realm B
and the time stamp of the ticket.

The client sends message to TGS_rem stating the identification of the

requested service VB, an authenticator which includes the identity of client
C, address of client ADc and a time stamp. The TGS uses the session key
Kc,tgsrem to decrypt the authenticator. The TGS can then check the identity and
address from the authenticator with that of the ticket and with the network
address of the incoming message. If all match, then the TGS is assured that
the sender of the ticket is indeed the ticket's real owner.
500 P. Mundra et al.

The TGS_rem then sends Ticketvrem as response to C. This message is

encrypted by session key shared between TGS rem and client and includes a
session key to be shared between client and server V, the identification of V ,
and the time stamp of the ticket.

In this message the client sends the ticket received by TGS_rem to server V.
Client also sends an authenticator. The server can decrypt the ticket, recover
the session key, and decrypt the authenticator.

2.3 Limitations of Kerberos Version 4

Kerberos 4 has limitations in two areas environmental shortcomings and technical

deficiencies. The focus of this paper is on environmental shortcomings following are
the environmental shortcomings of kerberos 4 [9]:-
-Encryption system dependence: Version 4 requires the use of DES. Ex-port
restriction on DES as well as doubts about the strength of DES were thus of concern.
-Authentication forwarding: Version 4 does not allow credentials issued to one
client to be forwarded to some other host and used by some other client. This
capability would enable a client to access a server and have that server access another
server on behalf of the client. For example, a client issues a request to a print server
that then accesses the client's file from a file server, using the client's credentials for
access. Version 5 provides this capability.
-Inter realm authentication: In version 4, interoperability among N realms
requires on the order of N2 Kerberos-to-Kerberos relationships, as described earlier.
Version 5 supports a method that requires fewer relationship.

2.4 Kerberos Version 5

Kerberos 5 has many improvement over kerberos version 4. Some new element are
added in kerberos version 5[9] are as follows:-
-Realm: Indicates realm of user.
-Options: Used to request that certain flags be set in the returned ticket
-Times: Used by the client to request the following time settings in the ticket:
-Nonce: A random value to be repeated in message (2) to assure that the response
is fresh and has not been replayed by an opponent.
The flags field included in tickets in version 5 supports expanded functionality
compared to that available in version 4. There are many flags in Kerberos version 5.
But for user to user authentication FORWARDABLE and FORWARDED are used.
 Modeling and Verification of Inter Realm Authentication in Kerberos 501

Fig. 2. User-to-User Authentication

-FORWARDABLE : Tells TGS that a new ticket-granting ticket with a different

network address may be issued based on this ticket-granting ticket.
-FORWARDED : Indicates that this ticket has either been forwarded or was
issued based on authentication involving a forwarded ticket-granting ticket.
In this paper we mainly emphasis on User to User and Inter Realm Authentication.
Authentication in Kerberos happens between a client and server. The client gets a
ticket for a service, and the server decrypts this ticket using its secret key. This works
fine for a physically- secure server, which keeps its secret key on its local disk. But,
storing the server's key on disk does not work for services that run on users' desktop
machines, since no-one should keep a long-lived secret key on an insecure disk drive.
The solution to this problem is called user-to-user authentication, and it is
implemented in Kerberos 5. In the user-to-user protocol, one user acts as a server, and
the other user acts as a client. At the client-user's request, the server- user sends his
TGT (but not his session key) to the client-user, who then gets credentials from the
key distribution center(KDC), encrypted with the session keys of both TGTs. Both
users can decrypt the new session-key and use it to verify each other's identity. The
advantage of the User to User scheme is that the server-user exposes only his short-
lived TGS session-key to theft; he keeps his long-lived secret, his password, in his
biological memory. An attacker is less likely to bother to steal a short-lived server-
key. However, User to User's downside is that the desktop server cannot operate
autonomously; the service-operator has to refresh his TGT in order for the server to
keep accepting clients' requests. Kerberos 5 has many improvement over kerberos 4,
which we discuss in section II. But we focus on one of the short comings of in
kerberos 4, environmental shortcomings (Inter realm Authentication). Here we also
model the scenario in which one of the client can forward his authentication ticket to
other client.

3 Modeling the Protocol

Formalizing and Modeling the inter-realm authentication, we consider the two realm
and we try to _nd out the attack between the two realms. Consider the Two realm A
and B.In each Realm we have considered four entity:
502 P. Mundra et al.

Relam A Relam B
c1 : client a1 : AS c2 : client a2 : AS
t1 : TGS v1 : server t2 : TGS v2 : Server
These entities will communicate with each other and send message to each other as
described in Section II. Each entity exchanges a set of messages with other entities in
the model. Apart from entities there is one more entity Intruder (I) for the purpose of
checking the protocol against any attack Each entity goes through some state during
communication. Table.1 shows the following state's of entity.

Table 1. States Representation

Across each entity a set of counter is associated. These counter count the number
of message exchanged during communication. When ever the entity send or receive
the message, the corresponding counter increment. All the counter are initialize to 0,
the counter will increment to its maximum number of session. counter associated with
each entity are as follows.
c_ as_ send : counts the number of messages sent from C to AS
c_ as_ receive : counts the number of messages C received from AS
c_ tgs_ send : counts the number of messages sent from C to TGS
c_ tgs_ receive : counts the number of messages C received from TGS
c_ v_ send : counts the number of messages sent from C to V
c_ v_ receive : counts the number of messages C received from V
as_send : counts the number of messages sent from AS to C
as_receive : counts the number of messages received by AS from C
tgs_send : counts the number of messages sent from TGS to C
tgs_receive : counts the number of messages received by TGS from C
v_send : counts the number of messages sent from V to C
v_receive : counts the number of messages received by V from C
 Modeling and Verification of Inter Realm Authentication in Kerberos 503

In the Table.2 various state of each entity is shown. Fig.2 shows how the message
can transfered between the state. Number in the given figure will shows the state's of
the entity which is show in Table.1.

Table 2. Counter

4 Authentication

4.1 User to User Authentication

Kerberos 5 introduced the concept of forwardable tickets. Client himself forward his
ticket to other client. In kerberos 5 FORWARDABLE flag are used for user to user
authentication. During the initial TGT acquisition, a client can request that the ticket
be marked forwardable. Once the FORWARDABLE flag is set on a ticket, the user
can use this ticket to request a new ticket. For modeling User to User authentication in
NuSMV we also use the one flag with each client. If the value of flag is TRUE then
user will forwarded his ticket to other when request come, other wise not. As mention
in this protocol that the life time of FORWARDABLE ticket is very less. So if
intruder will capture it and its life time is stale soon. For modeling this we used one
counter with this flag and the value of counter is 1. For FORWARDABLE ticket
server will also have one counter whose value is 1. when client1 forwarded his
authenticated ticket to client2 then the value of counter is 1. When client2 forwarded
ticket to server, server will check its counter value. When counter value 1 then only
server will respond to that service. After granting the service server set its counter
value to 0.

4.2 Inter Realm Authentication

For proving the authentication, corresponding property is used [3]. The corresponding
property state's that :-
If the event E occurs n times, then the event F must have occurred at least n times.
we generalized this property by CTL specification. The correspondence property can
be formalized by the means of a general schema:
504 P. Mundra et al.

where X.counterF and Y.counterE are the counters of processes corresponding to

entities X and Y. AG is the abbreviation for Always Globally CTL operator, and AGx
means that the propositional formula x must be true in every possible state of the
system. The authentication SPEC are as follows :-

-The (i) and (ii) pair of specifications describe that if the client sends n number
of protocol executions with the AS as responder, then in past the AS must have
started at least n protocol executions with the client as initiator.
-The (iii) and (iv) pair of specifications describe that if the client sends n
protocol executions with the TGS as responder, then in past the TGS must have
started at least n protocol executions with the client as initiator.
-The (v) and (vi) pair of specifications describe that if the client sends n
protocol executions with the V as responder, then in past the V must have started
at least n protocol executions with the client as initiator.
Reply Attack. In order to detect the reply attack we make use of counter. For proving
the authentication property each entity maintain the counter as in shown in Table.2.
when client sends request to Authentication Server(AS),Ticket Granting Server(TGS)
or Server(V) it increment its counter:
c˙as˙count˙send, c˙tgs˙count˙send, c˙v˙count˙send respectively as shown in
Table.2, and when either of these request is accepted by Authentication Server(AS),

Fig. 3. Finding the Attack

Ticket Granting Server(TGS) or Server(V) it increment its counter:

as_receive, tgs_ receive, v_receive
 Modeling and Verification of Inter Realm Authentication in Kerberos 505

Fig. 4. Screenshot of NuSMV Result

When an Intruder (I) is present in the system, it will try to capture the ticket. When
Intruder get the ticket and try to access the service on a particular server (either from
Realm A or Realm B), Intruder will send the request to the server (V). In that case
server will increment its receive counter. Consider the Fig.4, in which intruder will
capture the packet and send it to the server, in that case counter of respective server
will increment and following SPEC evaluates to FALSE.
We can also add other CTL specifications, so as to check whether Intruder (I)
receives the fourth message, as given below: Consider the particular specification

This also returns a similar counterexample tracing the path up to the state where it
becomes false and Intruder I receives the seventh message. But this specification is
unrealistic since in real life scenario it is not possible to check the state of an intruder
and whether it has received any messages or not.

Hence, the specification must hold true in all states of the system. When the
present model of the protocol is run with this specification it returns false and
generates a counterexample.

5 Conclusion
In this paper, we have tried to model and verify the Inter Realm Authentication and in
Kerberos Protocol using the model checker tool NuSMV. We have used the
correspondence property to obtain CTL specifications of all the security requirements
of the given protocol with the help of various counters. We also demonstrate the
authentication forwarding between the client. NuSMV will generate the counter
example in presence of intruder. In future this method of modeling and verifying of
the protocol can be extended to Pre-Authentication in Kerberos Version 5.
506 P. Mundra et al.

References
1. Miller, S.P., Neumann, B.C., Schiller, J.I., Saltzer, J.H.: Kerberos authentication and
authorization system (October 1988),
http://web.mit.edu/Saltzer/www/publications/athenaplan/e.2.1.pdf
2. Huth, M., Ryan, M.: Logic In computer Science. Cambridge University Press, Cambridge
(2004)
3. Panti, M., Spalazzi, L., Tacconi, S.: Using the nusmv model checker to verify the kerberos
protocol (2007), http://www.inform.unian.it/personale/spalazzi/repo
4. Adyanthaya, S., Rukmangada, S., Tiwari, A., Singh, S.: Modeling freshness concept to
overcome replay attack in kerberos protocol using nusmv. In: 2010 International
Conference on Computer and Communication Technology (ICCCT), Allahabad, India, pp.
125–129 (September 2010)
5. Cavada, R., Cimatti, A., Jochim, C.A., Keighren, G., Olivetti, E., Pistore, M., Roveri, M.,
Tchaltsev, A.: Nusmv 2.4 user manual (2005),
http://nusmv.fbk.eu/NuSMV/userman/v24/nusmv.pdf
6. Nusmv 2.4 (2011), http://nusmv.fbk.eu/
7. Cimatti, A., Clarke, E.M., Giunchiglia, E., Giunchiglia, F., Pistore, M., Roveri, M.,
Sebastiani, R., Tacchella, A.: Nusmv 2: An opensource tool for symbolic model checking.
In: Brinksma, E., Larsen, K.G. (eds.) CAV 2002. LNCS, vol. 2404, pp. 268–359. Springer,
Heidelberg (2002)
8. Edmund, O.G., Clarke, M., Peled, D.A.: Model Checking. MIT-PRESS, Cambridge (1999)
9. Stallings, W.: Cryptography and Network Security Principles and Practices, 5th edn.
Prentice Hall, New Delhi (2010)
 A New Bluetooth Security Architecture

Mintu Philip and Asha Das

Toc H Institute of Science and Technology, Kerala, India

[email protected], [email protected]

Abstract. The paper proposes a new Bluetooth security architecture based on

user authentication and chaotic image encryption. The algorithm is able to
overcome man in middle attack as well as privacy for image is ensured using a
fast encryption algorithm based on chaotic encryption. Chaotic encryption
produces the key using a pseudorandom number generator. It is a stream cipher
method.

Keywords: cryptography, security, chaotic, logistic map.

1 Introduction
The Bluetooth wireless technology provides short range wireless connectivity
between common devices. Its popularity and widespread usage triggers us into giving
attention to its security issues. The man-in-the-middle attack and non User
Authentication are significant weaknesses in Bluetooth system specification defined
at the link-level.
Chaos theory has been established since 1970s. The distinct properties of chaos,
such as ergodicity, quasi-randomness, sensitivity dependence on initial conditions and
system parameters, have granted chaotic dynamics as a promising alternative for the
conventional cryptographic algorithms.
A new way of image encryption scheme has been proposed which utilizes a chaos-
based cryptographic scheme using the logistic map. A combined image compression
and encryption scheme is proposed. The model is achieved by using discrete wavelet
transform and RLE for compression. Chaos based system is used for encryption. This
algorithm encrypts image pixel by pixel taking consideration the values of previously
encrypted pixels. This system is robust against cryptanalytic attacks. Also a simple
implementation of image encryption achieves high encryption rate on general purpose
computer. In mobile bandwidth constraints and power saving are needed for which
the proposed algorithm is suitable.

2 Proposed Bluetooth Security Architecture

The proposed system consists of four modules: Bluetooth device and service
discovery, Bluetooth User authentication, Image compression and chaotic image
encryption. The first module discovers the Bluetooth devices within the range of the
transmitting device. This is done by device discovery protocol. All the devices are

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 507–515, 2011.
© Springer-Verlag Berlin Heidelberg 2011
508 M. Philip and A. Das

listed and we can select the device with which communication is to be established.
After selecting the device, we select the service provided by the device, which is done
by the service discovery module. After communication is established, the user
authentication process is started. The authentication is done using RSA and public key
certificates provided by the certificate authority. If the user is authentic, the image
transmission is started. The user selects the image to be transmitted and the third
module will compress the image. The compression algorithm is designed based on
discrete wavelet transform and run length encoding. Then the image can be encrypted
using chaotic encryption technique. The cipher can then be safely sent to the channel.
At the receiver side, after authentication, the cipher is decrypted using reverse chaos
encryption. Then inverse DWT is applied to the bits of data to obtain the original
image.

2.1 Bluetooth Device and Service Discovery

2.1.1 Device Discovery

Bluetooth splits the 2.4-GHz band into 79 channels, with the intent that all devices on
a single piconet use exactly one of these channels at a time. Of these 79 channels, 32
are also used for detecting nearby devices and establishing connections. An inquiring
device transmits inquiry messages on these channels, and a discoverable Bluetooth
device periodically listens for and replies to the inquiry messages.

2.1.2 Device Discovery

The server dynamically allocates a port number. A socket is first bound and set to
listening mode, before the server advertises an SDP service. The client searches for a
service record and uses that information to establish a connection to the appropriate
port. The concept is simple: the server advertises a service, and the client searches for
and finds that service. Actual port numbers are allocated automatically.

2.2 User Authentication

User authentication ensures that the communicating parties are trustworthy. User is
authenticated using RSA and public key certificates. The main advantage of this
algorithm is that it overcomes man-in-middle attack. The steps in user authentication
process are as follows:
1. Sender (A) encrypts its private key and own Bluetooth address using the
certificate authority’s public key.
2. A transmits the encrypted information to receiver (B).
3. B decrypts data using certificate authority’s private key and receives the
private key of sender.
4. B encrypts its private key and own Bluetooth address using the certificate
authority’s public key.
5. B transmits the encrypted information to A.
6. A decrypts data using certificate authority’s private key and receives the
private key of B.
 A New Bluetooth Security Architecture 509

7. A calculate C using equation:

C = KEY PKA MOD N (1)
Where N=pq, where p and q are large prime numbers and PKA is the public
key of A.
8. Send C to B.
9. B calculates the KEY from C using equation:
KEY = C KRA MOD N (2)
Now, A and B communicates further using KEY.

2.3 Image Compression

This phase has two sub-phases: (a) discrete wavelet transforms and (b) run length
encoding.

2.3.1 Discrete Wavelet Transform

The image is compressed using discrete wavelet transform. The advantage of DWT is
that most of the coefficients become zero. DWT converts the image pixels into
approximate and detail coefficients. Some of the coefficients are approximately zero
which can be neglected.

2.3.2 Run Length Encoding

The DWT coefficients are then encoded using Run Length Encoding. RLE works by
reducing the physical size of a repeating string of characters. This repeating string,
called a run, is typically encoded into two bytes. The first byte represents the number
of characters in the run and is called the run count. In practice, an encoded run may
contain 1 to 128 or 256 characters; the run count usually contains as the number of
characters minus one (a value in the range of 0 to 127 or 255). The second byte is the
value of the character in the run, which is in the range of 0 to 255, and is called the
run value. Thus the compressed data stream then undergoes encryption.

2.4 Image Encryption

The image is encrypted pixel by pixel using chaotic image encryption. The algorithm
uses a coupled logistic map. It has a pseudorandom key stream generator that
generates a binary stream which used in chaotic encryption. The first logistic map
whose initial parameters are taken from the KEY generated during user authentication
process provides the initial parameters for second logistic map.

Following are the logistic maps:

xn+1 = μ xn (1-xn) (3)
yn+1 = μ yn (1-yn) (4)
510 M. Philip and A. Das

The steps involved in pseudorandom key stream generator are:

1.
xi+1 = μ1 xi (1-xi) (5)
yi+1 = μ1 yi (1-yi) (6)

2. Convert real number to binary equivalent .

3. Divide into three parts and XOR the three parts to obtain .
4. Perform above steps for value of ‘i’ starting from 1 to n.
5. Convert to real value .
6. The value is given as initial value µ2 to second logistic map.
7. Above steps are repeated for second logistic map.
8. The final value is multiplied by 10 and is converted to binary, stored
in s1.
9. Multiply the value of µ2 by 10 and convert to binary and store in s2.
10. Take first 56 bits of s1 and 5th to 15th bits of s2 and combine it to form the
key to encrypt.
11. Perform XOR operation of pixels with the key to obtain the cipher.
12. At receiver side perform XOR of cipher with the key to decrypt data.
13. Decompress the data using inverse DWT to obtain the pixels of the image.
14. Write the pixels to a new image file.
The binary sequence generated by pseudorandom key stream generator is XORed
with the pixel values of the image to obtain the cipher image.

3 Statistical Analysis

The encrypted images should possess certain random properties in order to resist the
statistical attack. Statistical analysis is done by calculating the histograms, the
correlations of two adjacent pixels in the encrypted images and the correlation
coefficient for several images and its corresponding encrypted images of an image
database. A detail study has been undergone and the results are summarized as
followings. Different images have been tested, and similar results are obtained.
However, due to the page limit, only the results for the Lena.

3.1 Difference between the Original and the Permuted Images

NPCR (Number of Pixels Change Rate) is used to test the difference between the
original image P1 and the permuted one C1. NPCR stands for the number of pixel
change rate. Then, if D is a matrix with the same size as images P1 and C1, D (i,j)is
determined as follows:

D(i,j) = 0 , if P1 ≠ c1
= 1, else (7)
 A New Bluetooth Security Architecture 511

NPCR is defined by the following formula:

NPCR = ∑i=0:M-1∑j=0:N-1D(I,j)x100/(MxN) (8)

The NPCR of the new system is found to be 99.42.

3.2 Correlation Coefficients of Intra and Inter - Color -Components

To quantify the dependence between two images, Pearson’s correlation coefficient is

commonly used. Given by equation, this coefficient is obtained by dividing the
covariance between the two images (eq. 13) by the product of their standard
deviations (eq. 12 and eq. 11). E in eq.11 is the expected value operator. P1 (i, j) and
C1 (i, j) are respectively the pixels gray values of the first and the second images.

E(x) = 1/(M x N) ∑i=0:M∑j=0:N P1(i,j) (9)

2
D(P1) = 1/(M x N) ∑i=1:M∑j=1:N [ P1(I,j) – E(P1(I,j))] (10)
Cov(P1,Ci) = 1/M x N ∑i=1:M∑j=1:N [P1(i,j) – E(P1(i,j))] x [C1(i,j) – E(C1(i,j))]
Rp1c1 = Cov(P1,C1)/(√D((P1)√D(C1) (11)

Following results for found for various standard images:

Table 1. correlation coefficient of encrypted image

Image Correlation coefficient

Lena 0.00099
Pepper 0.002
Cameraman 0.0017

3.3 Distribution of Two Adjacent Pixels

Statistical analysis on large amounts of images shows that on average, 8 to 16

adjacent pixels are correlated.
In this section, some simulations are carried out to test the correlation distribution
between two horizontally, vertically and diagonally adjacent pixels, in the original
and permuted images. Fig. 5 shows the correlation distribution of two horizontally,
vertically and diagonally adjacent pixels in the first component of the original image
and the encrypted images. Then, we plot the pixel value on location (x,y+1) over the
pixel value on location (x, y), location (x+1,y) over the pixel value on location (x, y)
and location (x+1,y+1) over the pixel value on location (x, y).
512 M. Philip and A. Das

(a) (b)

(c)
Fig. 1. (a) horizontal, (b) vertical and (c) diagonal correlation matrix of original image

(a) (b)

(c)

Fig. 2. (a) horizontal, (b) vertical and (c) diagonal correlation matrix of encrypted image
 A New Bluetooth Security Architecture 513

It is clear from Fig 2 that there is negligible correlation between the two adjacent
pixels in the encrypted image.

3.4 Histogram Analysis

An image-histogram illustrates how pixels in an image are distributed by graphing the

number of pixels at each color intensity level. We have calculated and analyzed the
histograms of the encrypted image as well as the original colored image.
As we can see, the histogram of the encrypted image is significantly different from
that of the original image.
Moreover, one can observe that the coupled map improves the uniformity of the
histogram for encryption method.

(a) (b)

Fig. 3. Histogram of (a) original image and (b) encrypted image

3.4 Information Entropy Analysis

Entropy is a statistical measure of randomness that can be used to characterize the
texture of an image. It is well known that the entropy H(m) of a message source m can
be calculated as :
H(m) = ∑ i=0:2N-1 p(mi) log (1/P(mi) (12)
Where p(mi) represents the probability of message mi.
When an image is encrypted, its entropy should ideally be 8. If it is less than this
value, there exists a certain degree of predictability which threatens its security.
Table 2. Information Entropy

Image Entropy(original) Entropy(Encrypted)

Lena 6.95 7.7719

Pepper 6.775 7.7

Cameraman 6.775 7.762
514 M. Philip and A. Das

The obtained results are very close to the theoretical value. This means that
information leakage in the encryption process is negligible.

References
1. A Security Architecture Based on User Authentication of Bluetooth. International Forum
on Information Technology and Applications (2009)
2. Mao, Y.B., Chen, G., Lian, S.G.: A novel fast image Encryption scheme based on the 3D
chaotic baker map. Int. J. Bifurcate Chaos 14, 3613–3624 (2004)
3. Gao, H., Zhang, Y., Liang, S., Li, D.: A new chaotic algorithm for image encryption.
Chaos, Solutions & Fractals 29(2), 393–399 (2006)
4. Maung, S.S., Sein, M.M.: A Fast Encryption Scheme Based on Chaotic Maps. In:
GMSARN International Conference on Sustainable Development: Issues and Prospects for
the GMS (2008)
5. Ahmad, M., Alam, M.S.: A New Algorithm of Encryption and Decryption of Images
Using Chaotic Mapping. International Journal on Computer Science and Engineering 2(1),
46–50 (2009)
6. Wang, F., Zhang, Y., Cao, T.: Research of chaotic block cipher algorithm based on
Logistic map. In: 2009 Second International Conference on Intelligent Computation
Technology and Automation, pp. 678–681 (2009)
7. Yen, J.-C., Guo, J.-I.: A New Chaotic Key-Based Design for Image Encryption and
Decryption. In: IEEE International Symposium on ISCAS 2000, Geneva, pp. IV-49–IV-52
(May 2000)
8. Lee, P.-H., Pei, S.-C., Chen, Y.-Y.: Generating Chaotic Stream Ciphers Using Chaotic
Systems. Chinese Journal of Physics 41(6) (2003)
9. Socek, D., Shujun, L., Magliveras, S.S., Furht, B.: Short Paper: Enhanced 1-D Chaotic
Key-Based Algorithm for Image Encryption. In: First International Conference on Security
and Privacy for Emerging Areas in Communications Networks, pp. 406–407 (2005)
10. Rao, D., Gangadhar, K.: Modified Chaotic Key-Based Algorithm for image Encryption
and Its VLSI Realization. In: International Conference on Digital Signal Processing (2007)
11. Ahmed, H.E.H., Kalash, H.M., Allah, O.S.F.: An Efficient Chaos-Based Feedback Stream
Cipher (ECBFSC) for Image Encryption and Decryption. Presented at Informatical
(Slovenia), pp. 121–129 (2007)
12. Liu, S., Sun, J., Xu, Z.: An Improved Image Encryption Algorithm based on Chaotic
System. Journal of Computers 4(11), 1091–1100 (2009)
13. Awad, A., Saadane, A.: Efficient Chaotic Permutations for Image Encryption Algorithms.
In: Proceedings of the World Congress on Engineering, vol. I (2010)
14. Zhu, A.-h., Li, L.: Improving for Chaotic Image Encryption Algorithm Based on Logistic
Map. In: 2nd Conference on Environmental Science and Information Application
Technology (2010)
15. Chen, G., Mao, Y., Chui, C.K.: A symmetric image encryption based on 3D chaotic maps.
Chaos Solutions Fractals 21, 749–761 (2004)
16. Borujeni, S.E., Eshghi, M.: Chaotic Image Encryption Design Using Tompkins-Paige
Algorithm. Hindawi Publishing Corporation, Mathematical Problems in Engineering,
Article ID 762652, 22 pages (2009)
17. Salleh, M., Ibrahim, S., Isnin, I.F.: Image encryption algorithm based on chaotic mapping.
Journal Teknologi, 1–12 (2003)
 A New Bluetooth Security Architecture 515

18. Liu, S., Sun, J., Xu, Z.: An Improved Image Encryption Algorithm based on chaotic
system. Journal of Computers 4(11), 1091–1100 (2009)
19. Noura, H., El Assad, S., Vladeanu, C.: Design of a fast and robust chaos-based
cryptosystem for image encryption. In: 8th International Conference on Communications
(COMM), pp. 423–426 (2010)
20. Improving Pairing Mechanism in Bluetooth Security. International Journal of Recent
Trends in Engineering 2(2) (November 2009); Security Overview of Bluetooth, Dave
Singelée, Bart Preneel, COSIC Internal Report (June 2004)
21. Two Practical Man-In-The-Middle Attacks on Bluetooth Secure Simple Pairing and
Countermeasures. IEEE Transactions on Wireless Communications 9(1) (January 2010)
 A New Trace Backing Algorithm for
Maximizing Streaming Data Join

G. Hemalatha1 and K. Thanushkodi2

1
CSE Dept, Karunya University,
Coimbatore, India
Tel.: 0422-2614020
[email protected]
2
Akshaya College of Engineering,
Coimbatore, India
[email protected]

Abstract. An increasing number of database queries are executed by interactive

users and applications. Since the user is waiting for the database to respond
with an answer, the initial response time of producing the first results is very
important. The user can process the first results while the database system
efficiently completes the entire query. The state-of-art join algorithms are not
ideal for this setting. Adaptive join algorithms have recently attracted a lot of
attention in emerging applications where data is provided by autonomous data
sources through heterogeneous network environments. The main advantage of
adaptive join techniques is that they can start producing join results as soon as
the first input tuples are available, thus improving pipelining by smoothing join
result production and by masking source or network delays. Since the response
time of the queries places a vital role in adaptive join, the join techniques like
Hash Join, Sort Merge Join cannot be used because they require some prework
before producing the join result. The only possible join technique that can be
used in adaptive join is Nested Loop Join. In Nested Loop Join each row of the
outer relation is compared with each row of the inner relation. The no. of
comparisons done by the nested loop join can be reduced by using a technique
called trace backing. In trace backing technique whenever a miss match occurs,
the next tuple of the outer relation is compared with the mismatched inner
relation tuple, instead of looping all the tuples of the inner relation.

Keywords: Adaptive Join, Trace Back technique, Nested Loop Join.

1 Introduction
Modern information processing is moving into a realm where we often need to
process data that is pushed or pulled from autonomous data sources through
heterogeneous networks. Adaptive query processing has emerged as an answer to the
problems that arise because of the fluidity and unpredictability of data arrivals in such
environments [1]. An important line of research in adaptive query processing has been
towards developing join algorithms that can produce tuples “online”, from streaming,

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 516–522, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 A New Trace Backing Algorithm for Maximizing Streaming Data Join 517

partially available input relations, or while waiting for one or more inputs [4], [7], [9],
[12], [14], [15]. Such non-blocking join behavior can improve pipelining by
smoothing or “masking” varying data arrival rates and can generate join results with
high rates, thus improving performance in a variety of query processing scenarios in
data integration, on-line aggregation and approximate query answering systems.
Compared to traditional join algorithms ( sort-, hash or nested loops-based [13]),
adaptive joins are designed to deal with some additional challenges: The input
relations they use are provided by external network sources. The implication is that
one has little or no control over the order or rate of arrival of tuples. Since the data
source reply speed, streaming rate and streaming order, as well as network traffic and
congestion levels, are unpredictable, traditional join algorithms are often unsuitable or
inefficient. For example, most traditional join algorithms cannot produce results until
at least one of the relations is completely available.
Waiting for one relation to arrive completely to produce results is often
unacceptable. Moreover, and more importantly, in emerging data integration or online
aggregation environments, a key performance metric is rapid availability of first
results and a continuous rate of tuple production. Adaptive join algorithms were
created in order to lift the limitations of traditional join algorithms in such
environments. By being able to produce results whenever input tuples become
available, adaptive join algorithms overcome situations like initial delay, slow data
delivery or bursty arrival, which can affect the efficiency of the join. All existing
algorithms work in three stages. During the Arriving phase, a newly arrived tuple is
stored in memory and it is matched against memory-resident tuples belonging to the
other relations participating in the join.
Since the allocated memory for the join operation is limited and often much
smaller than the volume of the incoming data, this results in tuple migration to disk.
The decision on what to flush to disk influences significantly the number of results
produced during the Arriving phase. The Arriving phase is suspended when all data
sources are temporarily blocked and a Reactive phase kicks in and starts joining part
of the tuples that have been flushed to disk. An importance of this phase is the prompt
handover to the Arriving phase as soon as any of the data sources restarts sending
tuples. Each algorithm has a handover delay which depends on the minimum unit of
work that needs to be completed before switching phases.
This delay has not received attention in the past, but we show that it can easily lead
to input buffer overflow, lost tuples and hence incorrect results. When all sources
complete the data transmission, a Cleanup phase is activated and the tuples that were
not joined in the previous phases (due to flushing of tuples to disk) are brought from
disk and joined. There are two new adaptive join algorithms for output rate
maximization in data processing over autonomous distributed sources. The first
algorithm, Double Index NEsted-loop Reactive join (DINER) is applicable for two
inputs, while Multiple Index NEsted-loop Reactive join (MINER) can be used for
joining an arbitrary number of input sources.
518 G. Hemalatha and K. Thanushkodi

DINER follows the same overall pattern of execution discussed earlier but
incorporates a series of novel techniques and ideas that make it faster, leaner (in
memory use) and more adaptive than its predecessors. More specifically, the key
differences between DINER and existing algorithms are (1) an intuitive flushing
policy for the Arriving phase that aims at maximizing the amount of overlap of the
join attribute values between memory-resident tuples of the two relations and (2) a
lightweight Reactive phase that allows the algorithm to quickly move into processing
tuples that were flushed to disk when both data sources block as show in Fig 1.

Fig. 1. Arriving Phase

The DINER has a higher rate of join result production and is much more adaptive
to changes in the environment, including changes in the value distributions of the
streamed tuples and in their arrival rates. The efficiency of the DINER algorithm can
still be improved by replacing the Nested loop algorithm with a better join technique.
There are many Join techniques which out performs the Nested Loop Join like Hash
Join, Sort Merge Join, XJoin, Double Hash Join etc., But all these join techniques
require some pre-work to produce the resulting Join tuples. As we know in adaptive
environment time is a vital factor and it cannot be spend to do the pre-work. So the
only possible Join technique that can be used is Nested Loop Join.

2 Our Contribution
In Nested loop Join each row of the outer relation is compared with each row of the
inner relation which in turns results in a Cartesian product as show in Fig 2. The other
joins algorithms out performs the nested loop join but still cannot be used in the
steaming join. Because in all the other join algorithms some amount of preprocessing
has to be done, which not possible in adaptive environment as the arrive of input
tuples cannot be predicted and as soon as it is arrived it has to be join before the other
set of input tuples are ready. So the only possible join technique to be used in adaptive
environment is Nested Loop join.
 A New Trace Backing Algorithm for Maximizing Streaming Data Join 519

Fig. 2. Draw Back of Nested Loop Join

In this paper we proposed a trace backing algorithm to be used instead of nested

loop join to maximize the no. of resultant tuples this in turn increased the resultant
tuples produced during a time period. The proposed Trace backing algorithm [as show
in the Fig 3] compares the outer row with the inner row, whenever there is a mismatch
the outer row is incremented instead of inner row, the inner is incremented when there
is match. This reduced the number of comparisons than it is required for Nested Loop
Join.

Fig. 3. Trace backing Technique

3 Algorithm for Trace Backing Technique

Let RA and RB be the two relations to be joined, the smallest relation is taken as the
outer relation and the other is taken as the inner relation. The first row of the outer
relation is compared with first row of the inner relation, if there is a match increment
the inner pointer and compare the first row of the outer relation and the second row of
the inner relation else increment the outer pointer and compare the second row of the
outer relation with the first row of the inner relation. This process is repeated until all
the inner relation tuples are matched.
520 G. Hemalatha and K. Thanushkodi

Input: Tuples from relations RA and RB

Output: Joined tuples, RA >< RB

Set Ro smallest(RA,RB)
Set Ri largest(RA,RB)
PR=0; // rightpointer
Pl=0; // leftpointer
//n is the number of tuples in the relation
While PR <n(Ri) do
While Pl < n(Ro) do
// compare the tuples from two relations
If (tRo[Pl]== tRi[PR]) then
tRi[PR] >< tRo[Pl]
PR++;
Else
Pl++;
End if
End while
if (Pl == n(Ro)) then //check leftpointer status
leftpointer=0;
if(PR== n(Ri)) then //check rightpointer position
break;
End while

4 Performance Evaluation
The following graph shows the no. of comparisons required to perform the Nested
Loop Join and Trace Back technique. The no. of comparison is greatly reduced in

Nested Loop Vs Trace Back

25000000
No. of Comparisions in terms of

20000000
15000000
10000000 Nested Loop Join
rows

5000000
0 Trace Back
Technique
0 10 20 30 40
Table Size in MB

Fig. 4. Comparison between Nested Loop and Trace Back Join Technique
 A New Trace Backing Algorithm for Maximizing Streaming Data Join 521

Trace Back technique. In the worst case scenario the Trace Back technique behaves
similar to Nested Loop Join. In adaptive Join environment the worst case scenario
occurs very rarely.
The performance of this algorithm is evaluated on the basis of throughput [the
number of joined tuples per unit time] and compared with DINER. By using the trace
backing algorithm the number of joined tuples is increased. For example, in
15millisecond, DINER produced 72 joined tuples, where as trace backing technique
produces 108 joined tuples.

600

500
No. of resulting tuples

400

300 DINER
using trace backing
200

100

0
0 20 40 60 80 100 120
Time in ms

Fig. 5. Performance Evaluations for Double Inputs

5 Conclusion and Future Scope

The main advantage of adaptive join techniques is that they can start producing join
results as soon as the first input tuples are available. The only Join technique that can
produce the output without any preprocessing is Nested Loop Join. In Nested Loop
Join each row of the outer relations are compared with the inner relation, which is
same as Cartesian Join. The new technique trace backing reduced the number of
comparisons when it is compared with nested loop join. By using this technique, the
number of comparison are reduced which in turn increased the number of resultant
tuples produced per unit of time. The trace backing technique produced 102 tuples
and DINER produced 72 tuples within 15 millisecond.
In terms of future work, the flushing policy can be extended by considering more
factors to specify which tuples have to be flushed. The flushing policy has an
important role in improving the performance of the join operation.
522 G. Hemalatha and K. Thanushkodi

References
[1] Condie, T., Conway, N., Alvaro, P., Hellerstein, J.M., Berkeley, U.C.: Online
Aggregation and Continuous Query support in MapReduce. In: Proceedings of VLDB
(2004)
[2] Rahman, P., Saad, N.A., Salahi, T.: Early Hash Join: A Configurable Algorithm for the
Efficient and Early Production of Join Results. In: ITSim 2008 (2008)
[3] Avnur, R., Hellerstein, J.M.: Eddies: Continuously Adaptive Query Processing. In:
Proceedings of the 2000 ACM SIGMOD International Conference on Management of
Data, volume 29, May 16-18, vol. 29, pp. 261–272. ACM, New York (2000)
[4] Chen, S., Gibbons, P.B., Mowry, T.C., Valentin, G.: Fractal prefeching B+trees:
Optimizing both cache and disk performance. In: ACM SIGMOD International
Conference on the Management of Data (June 2002)
[5] Hong, W., Stonebraker, M.: Optimization of Parallel Query Execution Plans in XPRS.
Distributed and Parallel Databases 1(1), 9–32 (1993)
[6] Mihaela, B.A., Vasilis, V., Yannis, K., Antonios, D.: Adaptive Join Operators for Result
Rate Optimization on Streaming Inputs. IEEE Transactions on Knowledge and Data
Engineering 22(8), 1110–1125 (2010)
[7] Urhan, T., Franklin, M.J.: XJoin: A Reactively-Scheduled Pipelined Join Operator. IEEE
Transactions on Knowledge and Data Engineering 23(2) (April 2000)
[8] Dittrich, J., Seeger, B., Taylor, D.: Progressive merge join: A generic and non-blocking
sort-based join algorithm. In: Proceedings of VLDB (2002)
[9] Mokbel, M.F., Lu, M., Aref, W.G.: Hash-Merge Join: A Nonblocking Join Algorithm for
Producing Fast and Early Join Results. In: ICDE (2004)
[10] Tao, Y., Yiu, M.L., Papadias, D., Hadjieleftheriou, M., Mamoulis, N.: RPJ: Producing
Fast Join Results on Streams Through Rate-based Optimization. In: Proceedings of ACM
SIGMOD Conference (2005)
[11] Hellerstein, J.M., Haas, P.J., Wang, H.J.: Online Aggregation. SIGMOD, 171–182
(1997)
[12] Haas, P.J., Hellerstein, J.M.: Symmetric Hash Joins. In: Proceedings of ACM SIGMOD
(1999)
[13] Viglas, S.D., Naughton, J.F., Burger, J.: Maximizing the output rate of multi-way join
queries over streaming information sources. In: VLDB 2003, pp. 285–296 (2003)
[14] Babu, S., Srivastava, U., Widom, J.: Exploiting k-Constraints to Reduce Memory
Overhead in Continuous Queries Over Data Streams. In: Proceedings of the 2002 Annual
ACM Symposium on Theory of Computing, pp. 370–379. ACM Press, New York (2002)
[15] Gehrke, J.: Special issue on data stream processing. IEEE Computer Society Bulletin of
the Technical Communication on Data Engineering 26(1) (March 2003)
[16] Arasu, A., Babu, S., Widom, J.: The cql continuous query language: Semantic
foundations and query execution. Technical report, Stanford University Database Group
(October 2003)
A Comparative Analysis of Demand Assignment Multiple
Access Protocols for Wireless ATM Networks

C. Shanmuganathan and P. Raviraj

1
Department of Computer Science &Engineering,
St.Peter’s College of Engg.& Tech., Chennai, Tamilnadu, India
2
Department of Computer Science & Engineering,
SKP Engineering College, Tiruvannamalai, Tamilnadu, India

Abstract. The concept of wireless ATM is now being actively considered as a

potential framework for next-generation wireless communication networks
capable of supporting integrated multimedia services with different QoS
requirements. Several key subsystem design issues for wired ATM and wireless
networks needs to be readdressed in the scope of the wireless ATM, which has
the capability to extend the statistical multiplexing of wired ATM network into
the wireless medium. One of the key subsystem issues is the development of
appropriate medium access control (MAC) protocol. The extension of the ATM
network into wireless environment faces many interesting problems. The
original ATM network was designed for high speed, noiseless, and reliable
channels. None of these characteristics are applicable to the wireless channel.
One of the critical aspects of a wireless ATM network is the Medium or Multi
Access Control (MAC) Protocol used by the Mobile station (MS) to request
service from the BS, which has to consider the Quality of Service (QoS) of the
specific applications. This paper analyzes recently proposed MAC protocols,
particularly those of Demand Assignment Multiple Access protocols using
TDMA technique with Frequency Division Duplex (FDD). It also gives
performance measures of two best suited protocols for wireless network
environment Distributed Queuing Request Update Multiple Access
(DQRUMA) protocol and Adaptive Request Channel Multiple Access
(ARCMA) protocol.

Keywords: MAC, DQRUMA and ARCMA.

1 Introduction

The concept of ATM is end-to-end communication i.e. in a Wide Area Network

environment, the communication protocol will be the same i.e. ATM, and companies
will no longer have to buy extra equipment (like routers or gateways) to interconnect
their LANs (local area networks). Also ATM is considered to reduce the complexity
of the network and improve the flexibility while providing end-to-end consideration
of traffic performance [1][3].

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 523–533, 2011.
© Springer-Verlag Berlin Heidelberg 2011
524 C. Shanmuganathan and P. Raviraj

1.1 ATM Cell

ATM is a fixed size packet-based switching and multiplexing technology designed to

be a connection-oriented transfer mode that supports a wide range of services. Its
transmission scheme can operate at either constant or variable bit rates. The primary
unit (packet) in ATM is called a cell. The ATM standard defines a fixed-size cell with
a length of 53 bytes, which is composed of 5-byte header and a 48-byte payload.
These fixed-sized cells reduce the complexity of ATM switches and multiplexers
allowing cells to be relayed at very high speeds. The typical bit rates at the terminals
are 25 Mbps, 155 Mbps, and 622 Mbps. The conversion of user data into ATM cells,
and vice versa, is performed by the ATM Adaptation Layer(AAL), which resides
directly above the ATM layer of the network[4][5].

2 Multiple Access Protocols

A Multiple access protocol is a scheme to control the access to a shared
communication medium among various users. Access protocols can be grouped
according to the bandwidth allocation, which can be static or dynamic, and according
to the type of control mechanism implemented. Multiple access protocols can be
categorized into fixed assignment, random assignment and demand assignment [2].

2.1 Fixed Assignment

Time-division multiple access (TDMA) and frequency-division multiple access

(FDMA) are fixed assignment techniques that incorporate permanent sub channel
assignment to each user.

2.2 Random Assignment

These techniques allocate the full channel capacity to a user for short periods, on a
random basis. These packet-oriented techniques dynamically allocate the channel to a
user on per-packet basis. Although there are few versions of the ALOHA protocol, in
its simplest form it allows users to transmit at will [6].

3 Demand Assignment Multiple Access Protocols

Most DAMA protocols use time-slotted
channels that are divided into frames.
Depending on the transmission rate
and the type of services, the channel
bandwidth can be represented by a
single or multiple frame(s). Each frame
is divided into uplink and downlink
period (channel). These periods are
further divided into two sub periods
or slots. They can be partitioned on a Fig. 1. Radio Channel Classification: (Slot-By-Slot)
slot-by slot or period basis.
 A Comparative Analysis of Demand Assignment Multiple Access Protocols 525

The uplink channel (mobile-to-

base) is divided into the request
access (RA) and data transmission
access (TA) sub periods. On the
other hand, the downlink is
divided into the acknowledgement
(ACK) and the data downstream
(DD) sub periods.

Fig. 2. Radio Channel Classification by Period

4 Distributed Queuing Request Update Multiple Access Protocol

DQRUMA introduces the concept of a dynamic uplink slot where the uplink slot may
be converted into a whole RA channel filled only with RA minislots. This conversion
occurs when the BS senses that there is significant amount of collisions in the RA
channel. It allows as many as 25 RA minislots in a single time slot where mobile can
send their requests. This method drastically reduces request contention in the RA
channel. Request in the RA channel
use a random access protocol like
slotted ALOHA. The downlink slot
contains the typical DD time slot,
and one or more ACK minislots,
and transmission permission, Perm
slot. In situations where the uplink
slot is converted into a series of RA
minislots, the subsequent downlink
slot is converted into a series of
corresponding ACK minislots.
DQRUMA also introduces an
extra bit called piggyback (PGBK) bit Fig. 3. DQRUMA Timing Diagram
in the uplink channel. Each time a
mobile transmits a packet (uplink), it also includes this PGBK bit to indicate whether
it has more packets in the buffer. This bit serves as a contention-free transmission
request for a mobile transmitting a packet. The BS checks this bit and updates the
appropriate entry in the Request Table (in the BS) accordingly. When this bit is
included, a mobile does not need a request for channel access in the following time
slot.

5 Adaptive Request Channel Multiple Access protocol

This new protocol scheme is based on the DQRUMA protocol while incorporating the
periodic traffic handling of PRMA. [1].
526 C. Shanmuganathan and P. Raviraj

ARCMA is essentially a demand assignment multiple access protocol with

dynamic bandwidth allocation. Its basic architecture is modeled after the DQRUMA
protocol. This scheme is designed to function in a cell-based wireless network with
many Mobile Station (MS) communicating with the Base Station (BS) of their
particular cell. Transmissions are done on a slot-by-slot basis without any frames.

5.1 Description of the Protocol

Since this protocol is modeled after DQRUMA, it is composed of

request/acknowledgement and a permission/transmission phase. Data transmission
from the BS to MS is a straight forward operation where the BS merely broadcasts the
information to the entire cell. The destination MS listens to the broadcast channel and
retrieves the appropriate data packets. If the transmission destination does not reside
in the same cell, the BS will forward the packet to an ATM switch to be routed to the
proper destination.

5.2 Request/Acknowledgement Phase

The request is made in the RA channel. The mobile’s b-bit access ID, assigned during
setup. In addition to the Access ID, the request packet also includes the type of
service being requested. After every random transmission the MS needs to know if
the request was successful.
Since it does not detect
collision by sensing the
channel, the BS has to send a
response to the MS indicating
a successful request. When the
BS receives a request from the
RA channel it immediately
broadcasts an
acknowledgement to the MS.
The BS then inserts the new
request in the request table to
indicate that the MS has
Fig. 4. Timing Diagram For the ARCMA Protocol
packet(s) to transmit.

5.3 Permission/Transmission Phase

General Traffic
The BS is responsible for allocating bandwidth (time slots) to the MS by using a
packet transmission policy. This protocol uses First In, First Out (FIFO) policy. The
MS that makes a request first, is given permission to transmit first.

CBR Traffic
This connection-oriented delay-sensitive traffic is given priority in the request phase.
When a CBR request arrives at the BS, it is inserted in special CBR Request Queue.
The CBR Request Queue also uses FIFO policy but requests in the queue have
precedence over those in the request table.
 A Comparative Analysis of Demand Assignment Multiple Access Protocols 527

5.4 Adaptive Request Channel Strategy

ARCMA also implements a dynamic RA channel strategy similar to that of the

DQRUMA. But ARCMA reduces the idle time slots by considering the probability of
retransmission qr. when the TA channel is first converted into multiple RA minislots,
all new and previously unsuccessful requests are transmitted with the probability of 1.
That is, all requests are sent out immediately regardless of their qr. If no requests are
successfully transmitted the uplink channel remains in the multiple RA mode.
However, in this case, the MS’s retransmit their requests according to their old qr.
Conversely, if successful requests were made during the multiple RA slots, the BS
reverts the next time slot back to normal mode with a single RA minislot. Any
remaining MS’s with unsuccessful requests, retransmit their requests using their
corresponding qr. The following table gives an example illustrating the number of
mobiles handled by each RA channel in normal and multiple RA mode.

Table 1. Number of Mobiles

Normal mode Multiple RA mode

Mobile
ID Channel # # of mobiles per Channel# # of mobiles
(single channel) RA channel (R =3) per RA channel
1 1 5 1 2
2 1 5 2 2
3 1 5 3 1
4 1 5 1 2
5 1 5 2 2

Although the all abo-

ve compared protocols have
their advantages and limit-
ations when it comes to
wireless data packet net-
works, DQRUMA and
ARCMA seem to offer the
most efficient scheme for
wireless ATM services. But
DQRUMA has one major
drawback when compared
to its successor, its inability
to offer any distinction
between different types of
ATM traffic during the
reservation phase. Fig. 5. Timeslot conversion in multiple RA mode
528 C. Shanmuganathan and P. Raviraj

6 Performance Analysis
6.1 Simulated Protocols
In this simulation protocols, ARCMA and DQRUMA are compared under identical
network environments and specification. DQRUMA and ARCMA have proven to be
an effective multiple access protocols that can support integrated traffic in a wireless
packet network. Both these protocols are implemented using the FIFO scheduling
policy in the Request Table and slotted ALOHA with Binary Exponential Backoff
algorithm is used as the random access scheme in the RA channel.

6.2 Simulation Environment

The simulation approach is modeled after the environment of a typical cellular mobile
communication network. The primary characteristics of the environment are
• The network model is a typical cell in a wireless network with one BS and N
active mobiles.
• MSs are buffered portable units that produce data packets according to some
bursty random processes. All data packets are queued in the mobiles buffer until
the BS allocates them a time slot for transmission.
• The uplink and downlink channels are simulated as separate channel. Both
uplink and downlink channels are time slotted.
• The network is assumed to have fixed number of active mobiles throughout each
experiment. The network traffic stays same for the duration of the experiment.
• The slotted ALOHA with BEB algorithm is used as the random access scheme
in the RA channel.
• The length of the time slot is assumed the same for two protocols.
• Network delay is due to access delay and service delay. Propagation delays
between mobiles and BS are assumed negligible.
• Simulations are run for 50 time slots.

6.3 Performance Measures

The simulation involves two different multiple access schemes performing various
traffic conditions. The performance measures are:
• Channel Throughput (TPc): TPc is defined as the ratio of the total number of
transmitted packets and the total time number of time slots. That is, TPc = PT/TTL,
where PT is denoted as the total number of transmitted packets, and TTLis the total
number of time slots.
• Average Transmission Delay (DAVG): DAVG is defined as the ratio of the total
packet transmission delay and number of active mobiles. Hence, DAVG = DTL/M,
where DTL is the total packet transmission delay and M is the number of active
mobile.
• Average Queue Length (LAVG): LAVG is defined as the ratio of the total number
of packets in all the mobiles buffer and the number of active mobiles. Thus, LAVG
= LTL/M, where LTL is the total number of packets in all the buffer, and M is the
number of active mobiles. LTL is the total time taken at the end of simulation. LAVG
is measured by number of packets [7].
 A Comparative Analysis of Demand Assignment Multiple Access Protocols 529

7 Simulation Results
We present the simulation results and the performance analysis of the different
multiple access strategies. We simulate protocols in the following traffic environment:
with CBR traffic, non CBR traffic. As the arrival rate increases the channel
throughput also increases (Fig.8) as the arrival rate increases, more packets are
transmitted. Under heavy traffic condition empty request queue are caused by
contention in the RA channel. Nevertheless, Tpc is very high at high arrival rates as
illustrated in Fig.7.1. Figures 7.1and 7.2 show the performance in a network cell with
100 active mobiles with single packet burst and no CBR traffic.
Figure 7.3 represents the combination of the previous two figures. The exact value
of the arrival rate is not a concern in our analysis. Figure 7.4 and 7.5 show the result
of a simulation in an environment with no CBR traffic and single packet burst. Fig.7.4
illustrates results from a low traffic condition. We can see that as the throughput
increases, so does the average delay (Davg). DQRUMA suffers from a very steep
increase in delay as the throughput increases.
ARCMA follows a similar pattern, but at a much lesser rate. Fig.7.5 shows that the
delay increases at a much higher rate, compared to the previous simulation, for all
three scheme. From the Fig.7.6 we can see that there is an improvement in
performance although there is a twice as many data packets arriving in the channel.
The main reason for this improvement is the piggy-backing strategy .The second
packet in every burst request accesses collision free using the PGBK bit. Figure 7.7
shows the results from a cell with mobiles with CBR traffic. The gap between their
performances grows appreciably as the arrival rate increases. The increase of data
arriving due to the CBR traffic causes a steep increase in the average delay for
DQRUMA .The special CBR handling in ARCMA allows it take advantages of CBR
traffic characteristics.
( No. of packet transmission/ time slot)
Throughput

Arrival rate (Probability of packet arriving every time slot per mobile)

Fig. 6. Channel Throughput VS Arrival Rate ARCMA (M=100, CBR=0, Burst=1)

530 C. Shanmuganathan and P. Raviraj

(Number of time slots)

Average Delay

Arrival rate (Probability of packet arriving every time slot per mobile)

Fig. 7. Average delay VS Arrival rate for ARCMA (M=100, CBR=0,Burst=1)

(Number of time slots)
Average Delay

Throughput (Number of packet transmission per time slot)

Fig. 8. Average Delay VS Channel Throughput for ARCMA (M=100, CBR=0, Burst=1)
(Number of time slots)
Average Delay

Throughput (Number of packet transmission per time slot)

Fig. 9. Average Delay VS Channel Throughput for ARCMA (M=25, CBR=0, Burst=1)
 A Comparative Analysis of Demand Assignment Multiple Access Protocols 531

(Number of time slots)

Average Delay

Throughput (Number of packet transmission per time slot)

Fig. 10. Average Delay VS Channel Throughput for ARCMA (M=100, CBR=0, Burst=1)
(Number of time slots)
Average Delay

Throughput (Number of packet transmission per time slot)

Fig. 11. Average Delay VS Channel Throughput for ARCMA (M=100, CBR=0, Burst=2)
532 C. Shanmuganathan and P. Raviraj

(Number of time slots)

Average Delay

Throughput (Number of packet transmission per time slot)

Fig. 12. Average Delay VS Channel Throughput for ARCMA (M=100, CBR=2%, Burst=1)

8 Conclusion

Wireless ATM is considered as a promising technique for future broadband wireless

networks and is under research prototyping at various institutions. This paper focuses
on the role of the medium access control protocol in supporting multimedia traffic and
QoS. Protocols for multimedia wireless ATM are still evolving. Many protocols have
been proposed, and most of them are only different in details of implementation, or
variant design for specific wireless architectures. The common trend observed in
these protocols is requiring the mobiles to request resources through reservations –
with or without contentions. The base station can be regarded as an ATM switch that
statistically multiplexes the wireless links among the mobiles. This approach will
provide a way to integrate similar packet scheduling strategies on wired ATM
network.
This paper presents comparison for protocols, which contain most characteristics
of current available channel access protocols for wireless ATM. Although all the
compared protocols have their advantages and limitations when it comes to wireless
data packet networks, DQRUMA and ARCMA seem to offer the most efficient
scheme for wireless ATM. Of these protocols compared only DQRUMA and
ARCMA have been simulated.
The simulation results show that ARCMA performs better than DQRUMA
regardless of the traffic load. ARCMA produces significantly higher channel
throughput than DQRUMA because of its ability to differentiate between the services
(CBR, VBR) of ATM. ARCMA brings us one step closer to designing a complete
protocol suite that could be used in the wireless ATM networks.
 A Comparative Analysis of Demand Assignment Multiple Access Protocols 533

References
1. Hac, A., Chew, B.L.: ARCMA – adaptive request channel multiple access protocol for
wireless ATM networks. International Journal of Network Management 11, 333–363
(2001), doi:10.1002/nem.411
2. Kubbar, O., Mouftah, H.T.: Multiple Access Control Protocols for Wireless ATM: Problems
Definition and Design Objectives. IEEE Communications Magazine (November 1997)
3. Ayanoglu, E., Eng, K.Y., Karol, M.J.: Wireless ATM: Limits, Challenges, and Proposals.
IEEE Personal Communications (August 1996)
4. Abramson, N.: Multiple Access in Wireless Digital Networks. Proceedings of the
IEEE 82(9) (September 1994)
5. Handel, R., Huber, M.N., Schroder, S.: ATM networks: Concepts, Protocols, Applications
6. Tanenbaum, A.S.: Computer Networks, 3rd edn.
7. Yuan, R., Biswas, S.K., French, L.J., Li, J., Raychaudhuri, D.: Proc. ”A Signaling and
Control Architecture for Mobility Support in Wireless ATM Networks”. In: ICC 1996,
Dallas, TX, pp. 478–488 (June 1996)
 Perceptual Linear Predictive Cepstral Coefficient for
Malayalam Isolated Digit Recognition

Cini Kurian and Kannan Balakrishnan

Department of Computer Applications,

Cochin University of Science & Technology,
Cochin, 682 022
[email protected],
[email protected]

Abstract. This paper presents an optimum speaker independent isolated digit

recognizer for Malayalam language. The system employs Perceptual Linear
Predictive (PLP) cepstral coefficient for speech parameterization. The powerful
and well accepted pattern recognition technique Hidden Markov Model is used
for acoustic modeling. The training data base has the utterance of 21 speakers
from the age group ranging from 20 to 40 years and the sound is recorded in the
normal office environment where each speaker uttered digits zero to nine
independently. The system obtained an accuracy of 99.5% with the unseen data.

Keywords: Speech Recognition, Malayalam, PLP.

1 Introduction

Humans interact with environment in several ways: sight, audio, smell and touch.
Humans send out signals or information visually, auditory or through gestures [17].
Because of the increased volume data, human has to depend on machines to get the
data processed. Human –computer interaction generally use keyboard and pointing
devices. In fact, speech has the potential to be a better interface other than keyboard
and pointing devices [12].
Keyboard a popular medium requires a certain amount of skill for effective usage.
Use of mouse also requires good hand- eye coordination. Physically challenged
people find it difficult to use computer. It is difficult for partially blind people to read
from monitor. Moreover current computer interface assumes a certain level of literacy
from the user. It expects the user to have certain level of proficiency in English apart
from typing skill. Speech interface helps to resolve these issues. Speech synthesis and
speech recognition together form a speech interface. Speech synthesizer converts text
into speech. Speech recognizer accepts spoken words in an audio format and converts
into text format [14].
Speech interface supports many valuable applications - for example, telephone
directory assistance, spoken database querying for novice users, “hands busy“
applications in medical line, office dictation devices, automatic voice translation into

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 534–541, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Perceptual Linear Predictive Cepstral Coefficient 535

foreign languages etc. Speech enabled applications in public areas such as railways;
airport and tourist information centers might serve customers with answers to their
spoken query. Physically handicapped or elderly people might able to access services
easily, since keyboard is not required. In Indian scenario, where there are about 1670
dialects of spoken form, it has greater potential. It could be a vital step in bridging the
digital divide between non English speaking Indian masses and others. Since there is
no standard input in Indian language, it eliminates the key board mapping of different
fonts of Indian languages
ASR is a branch of Artificial Intelligence (AI) and is related with number of fields
of knowledge such as acoustics, linguistics, pattern recognition etc [2]. Speech is the
most complex signal to deal with since several transformations occurring at semantic,
linguistic, acoustic and articulator levels. In addition to the inherent physiological
complexity of the human vocal tract, physical production system also varies from one
person to another [5, 6]. The utterance of a word found to be different, even when it is
produced by the same speaker at different occasions. Apart from the vast inherent
difference across different speakers and different dialects, the speech signal is
influenced by the transducers used to capture the signal, channels used to transmit the
signal and even the environment too can change the signals. The speech also changes
with age, sex, and socio economic conditions, the context and the speaking rate.
Hence the task of speech recognition is not easy due to many of the above constraints
during recognition [13].
In most of the current speech recognition systems, the acoustic component of the
recognizer is exclusively based on HMM [7,9,10]. The temporal evolution of speech
is modeled by the Markov process in which each state is connected by transitions,
arranged into a strict hierarchy of phones, words and sentences.
Artificial neural networks (ANN) [3,15] and support Vector machines (SVM) [2,7]
are other techniques which are being applied to speech recognition problems. In
ANN, temporal variation of speech can not be properly represented. SVM, being a
binary static classifier, adaptation of the variability of the duration of speech utterance
is very complex and confusing. SVM is a binary classifier while ASR, faces
multiclass issues.
For processing speech, the signal has to be represented in some parametric form.
Wide range of methods exists for parametric representation of speech signals, such as
Linear Prediction coding (LPC) [10], and Mel-Frequency Cepstrum Coefficients
(MFCC) [11] and Perceptual Linear Predictive (PLP) coefficient [10]. Since PLP is
more adapted to human hearing, PLP parameterization technique is used in this work.

2 Speech Processing Using PLP

One of first decisions in any pattern recognition system is the choice of what features
to use. The PLP feature converts speech signal in meaningful perceptual way through
some psychoacoustic process. The various stages of this method are based on our
perceptual auditory characteristics. The different stages of PLP extraction is shown in
fig 1.
536 C. Kurian and K. Balakrishnan

Fig. 1. Stages of PLP extraction

The first block in any feature extraction process is frame blocking. Here the audio
signals which are basically non stationary are cut into fragments which are called
frames. Then frames are passed through Hamming Window and DFT is applied over
it. Then power spectrum of the speech signal is computed as follows

P(ω) = Re( S((ω))2 + Im(S((ω))2 (1)

These steps are common in almost all the feature extraction process. The significant
blocks of PLP analysis are as follows

2.1 Critical Band Integration(Bark Frequency Weighing)

The first step is a conversion from frequency to bark frequency, which is a better
representation of human hearing resolution in frequency. Here power spectrum is
wrapped along its frequency axis into bark frequency. In short, the speech signal is
passed through some trapezoidal filters equally spaced in bark scale.

2.2 Equal Loudness Pre-emphasis

Different frequency components of speech spectrum are pre-emphasized by an equal -

loudness curve, which is an approximation to the unequal sensitivity of human
hearing at different frequencies, at about 40dB level. A filter having a specific transfer
function gives the curve.

2.3 Intensity Loudness Conversion(Cube-Root Amplitude Compression)

Cube-root compression of the modified speech spectrum is carried out according to

power law of hearing [16], which simulates the non-linear relation between the
 Perceptual Linear Predictive Cepstral Coefficient 537

intensity of sound and its perceived loudness. Together with the psychophysical
equal-loudness pre-emphasis, this operation reduces the spectral amplitude variation
of the critical-band spectrum.

2.4 All-Pole Model

Autoregressive modeling represents the final stage of PLP analysis, and consists of
approximating the spectrum by an all-pole model, by using the auto correlation
method. An inverse Discrete Fourier Transformation is applied to the spectrum
samples. For the M-th order all-pole model, only the first M+1 autocorrelation values
are needed.

3 Statistical Speech Recognition and HMM

An unknown speech wave form is converted by a front-end signal processor into a

sequence of acoustic vectors, O = o1,o2,o3,…. The utterance consists of sequence of
words W = w1, w2, w3 ----wn. In ASR it is required to determine the most probable
word sequence, W, given the observed acoustic signal O. Applying Bays’ rule to
decompose the required probability, [11]

S = argwmaxP(W /O) = argwmax(P(O/W)P(W) / P(O))

S = argwmaxP(O/W)P(W)
Posterior prior

Hence a speech recognizer should have two components: P (W), the prior probability,
is computed by language model, while P (O/W), the observation likelihood, is
computed by the acoustic model. Here the acoustic modeling of the recognition unit is
done using HMM.
Since HMM is a statistical model in which it is assumed to be in a Markov process
with unknown parameters, the challenge is to find all the appropriate hidden
parameters from the observable states. Hence it can be considered as the simplest
dynamic Bayesian network [7,9]. In a regular Markov model, the state is directly
visible to the observer, and therefore the state transition probabilities are the only
parameters. However, in a hidden Markov model, the state is not directly visible (so-
called hidden), but the variables influenced by the states are visible. Each transition in
the state diagram of a HMM has transition probability associated with it [13, 15].
These transition probabilities are denoted by matrix A. Here A is defined as A = aij
where aij = P (tt+1 = j │ j = i ) , the probability of being in state j at time t +1 , given
that we were in state i at time t. It is assumed that aij‘s are independent of time. Each
state is associated with a set of discrete symbols with an observation probability
538 C. Kurian and K. Balakrishnan

assigned to each symbol, or is associated with the set of continuous observation with a
continuous observation probability density. These observation symbol probabilities
are denoted by the parameter B. Here B is defined as B = bj(k), where bj(k) = P(vk at
t │it = j ) , the probability of observing the symbol vk , given that it is in the state j.
The initial state probability is denoted by the matrix π, where π is , defined as π = πi
where πi = P (it = 1 ) , the probability of being in state t at t = 1. Using the three
parameters A, B , and π a HMM can be compactly denoted as λ = { A, B , π } An
Example of an HMM with five states is shown in Figure 2.

Fig. 2. Topology of a 5 state HMM

There are three fundamental ASR problems that can be addressed with HMM.
Problem (i) is Scoring and evaluation i.e computing the likelihood of an observation
sequence, given a particular HMM. This problem occurs during the recognition phase.
Here for the given parameter vector sequence (observation sequence), derived from
the test speech utterance, the likelihood value of each HMM is computed using
forward algorithm. The symbol associated with the HMM, for which the likelihood is
maximum, is identified as the recognized symbol corresponding to the input speech
utterance. Problem (ii) is associated with training of the HMM for the given speech
unit. Several examples of the same speech segments with different phonetic contexts
are taken, and the parameters of the HMMs, λ, have been interactively refined for
maximum likelihood estimation, using the Baum- Wetch algorithm [13]. Problem (iii)
is associated with decoding or hidden state determination, where the best HMM state
is decided.

4 System Development and Speech Data Base

The different steps [5] involved in the development of the proposed system is shown
in figure 3.
The database contains 210 isolated spoken words recorded from 21 speakers. Each
speaker uttered numbers form zero to nine separately. The recording is done in
normal environment with a high quality microphone at 16kHz sampling frequency
and quantized at 16 bit . The data is quantized (at 16 kHz) and digitalized (with 16
bit) and the speech data is stored in WAV format.
 Perceptual Linear Predictive Cepstral Coefficient 539

Data collection
Text corpus

Speech corpus

Data preparation

Pronunciation dictionary

Transcription dictionary

Feature Extraction

Preparation of acoustic and language

Training and Testing

Performance evaluation

Fig. 3. System Development

5 Testing and Training

For training and testing the system, the data base is divided into three equal parts- 1,
2, 3 and the experiment is conducted in a round robin fashion. For each experiment,
2/3rd of the data is taken for training and 1/3rd of the data is used for testing. In the
experiment I, part 1 and part 2 of data is given for training. Then the same trained
system is taken for testing the system with part 3 of the database. In experiment II,
part 1 and part 3 of the data base is taken for training and part II of the database is
used for testing. In experiment III, part 2 and part 3 of the database is taken for
training and tested with part 1 of the database. The result obtained from each training
and testing experiment in terms of Word Accuracy, Number of words deleted,
inserted, substituted are detailed in table 1.
540 C. Kurian and K. Balakrishnan

Table 1. Performance Evaluation of the System with Training and Testing Data

TRAINING TESTING

Number of Substitutions
Number of Substitutions
Number of Deletions

Number of Insertions

Number of Insertions
Number of Deletions
Word Accuracy %

Word Accuracy %
Experiment

1 99.29 0 1 0 100 0 0 0
2 99.29 0 1 0 98.57 0 1 0
3 100 0 0 0 100 0 0 0

6 Performance Evaluation and Discussion

Word Error Rate (WER) is the standard evaluation metric for speech recognition. It is
computed by SCLITE [18], a scoring and evaluating tool from National Institute of
Standards and Technology (NIST). The inputs to the SCLITE are the reference text
and the recognized text (output of the decoder). After each training and testing
experiment, recognized text and the reference text are converted into the sclite
compatible form and fed as input in sclite. Then detailed results are obtained in terms
of WER, SER, and number of word deletions, insertions and substitutions. If N is the
number of words in the correct transcript; S, the number of substitutions; and D, the
number of Deletions, then,

WER = ((S + D+ I )N) /100

Sentence Error Rate (S.E.R) = (Number of sentences with at least one word error/
total Number of sentences) * 100
The system is tested with unknown speakers. Five speakers, who are not involved
in the training is selected and recognition performance of their utterance is evaluated.
A word recognition accuracy of 99.5% is obtained. These results prove that the
developed system is speaker independent with negligible errors.

7 Conclusion
The digit recognition system for Malayalam is implemented using PLP cepstral
coefficients and HMM. From the results of the experiment it can be concluded that
PLP and HMM are ideal candidates for spoken digit recognition. The system achieves
almost cent percent (99.5) recognition accuracy.
 Perceptual Linear Predictive Cepstral Coefficient 541

References
1. Ganapathiraju, A., Hamaker, J., Picone, J.: Support Vector machines for speech
Recogntion. In: Proceedings of the International Conferences on Spoken Language
Processing, Sdney, Australia, pp. 292–296 (November 1999)
2. Gold, B., Morgan, N.: Speech and audio signal processing. John Wiley and Sons, N.Y.
(2002)
3. Behrman, Nash, L., Steck, J., Chandrashekar, V., Skinner, S.: Simulations of Quantum
Neural Networks. Information Sciences 128(3-4), 257–269 (2000)
4. Gold, B., Morgan, N.: Speech and audio signal processing. John Wiley and Sons, N.Y.
(2002)
5. Kurian, C., Balakrishnan, K.: Speech Recognition of Malayalam Numbers. IEEE
Transaction on Nature and Biologically Inspired Computing, NaBIC 2009, 1475–1479
(2009)
6. Kurian, C., Firoz Shah, A., Balakrishnan, K.: Isolated Malayalam digit recognition using
Support Vector Machines. IEEE Transaction on Communication Control and Computing
Technologies, ICCCCT 2010, 692–695 (2010)
7. Burges, C.J.C.: A tutorial on support vector machines for pattern recognition. Knowledge
Discovery Data Mining 2(2), 121–167 (1998)
8. Davis, S., Mermelstein, P.: Comparison of parametric representations for Monosyllabic
word Recognition in continuously spoken sentences. IEEE Trans. on ASSP 28, 357–366
9. Dimov, D., Azamanov, I.: Experimental specifics of using HMM in isolated word Speech
recognition. In: International Conference on Computer System and Technologies –
CompSysTech 2005 (2005)
10. Felinek, F.: Statistical Methods for Speech recognition. MIT Press, Cambridge (1997)
11. Huang, X., Alex, A., Hon, H.W.: Spoken Language Processing; A Guide to Theory,
Algorithm and System Development. Prentice Hall, Upper Saddle River (2001)
12. Jurasky, D., Martin, J.H.: Speech and Language Processing: An introduction to natural
language Processing. In: Computational Linguistics, and Speech Recognition, 2nd edn.
(2007)
13. Jyoti, Rakesh, S.: Automatic Speaker Recognition: An Approach using DWT Based
Feature Extraction and Vector Quantization. IETE Technical Review 24(5), 395–402
(2007)
14. Rabiner, L., Juang, B.-H.: Fundamentals of Speech Recognition. Pearson Education,
London (2008)
15. Sperduti, O., Starita, A.: Supervised Neural Networks for classification of Structures.
IEEETransactions on Neural Networks 8(3), 714–735 (1997)
16. Stevens, S.S.: On the psychophysical law. Psychological Review 64(3), 153–181 (1957)
17. Grewal, S.S., Kumar, D.: Isolated word Recognition System for English language.
International Journal of Information Technology and Knowledge Management 2(2), 447–
450 (2010)
18. http://www.icsi.berkeley.edu/Speech/docs/sctk-1.2/sclite.html
 Agent Based Modeling of Individual Voting Preferences
with Social Influence

Vivek Kumar Singh, Swati Basak, and Neelam Modanwal

Department of Computer Science, Banaras Hindu University,

Varanasi-221005, India
[email protected], [email protected],
[email protected]

Abstract. Agent Based Modeling (ABM) is usually referred to as the third way
of doing modeling & simulation in contrast to the historically popular equation-
based macrosimulations and microsimulations. In ABM a system is modeled as
a set of autonomous agents who interact with each other and also with the
environment. The agents represent various actors in the system and the
environment represents agent’s surroundings and the overall context. ABM has
been successfully applied to model and analyze a number of complex social
processes and structures, ranging from habitation patterns to spread of culture.
In this paper, we present our experimental work on applying ABM to model
individual voting preferences. The model explores process of formation of
voting preferences, the factors governing it and its effect on the final voting
pattern. We have incorporated social influence theory in our model and
experimented with various settings. The paper concludes with a short discussion
of the results and their relevance.

Keywords: Agent Based Modeling, Multi-agent Systems, Social Influence,

Social Simulation, Voting Model.

1 Introduction
The development of social simulation over the past half century can be grouped into
three broad periods: macrosimulation, microsimulation and agent-based models.
Sociologists, particularly computational sociologists, first tried macrosimulations to
model problems in supply-chain management, inventory control in a warehouse,
urban traffic, spread of epidemics, migration and demographic patterns.
Macrosimulations consist of a set of differential equations that take a holistic view of
the system. However, taking the complete system as the unit of analysis had inherent
limitations. Microsimulations focused on the use of individuals as the unit of analysis
but continued with macro-level forecasting used in macrosimulations. Though
microsimulations modeled changes to each element of the population but they did not
permit individuals to directly interact or to adapt. The main focus remained
forecasting macro effects of public policies that alter individual behaviours. ABM, the
third methodology, takes a pure bottom-up approach and keeps the individual at the
centre. It allows modeling individual actors in the population and their interactions.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 542–552, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Agent Based Modeling of Individual Voting Preferences with Social Influence 543

This makes it a highly suitable technique for analysis of emergent group behaviours
resulting only from local interactions of individuals. Besides, there are various other
advantages of this approach.
In Agent Based Modeling (ABM) a system is modeled as a set of autonomous
agents, who can perceive the environment and act on the basis of some behavioural
rules. The agents represent the actors in the system; environment represents the
surroundings including neighbouring agents; and the behaviour rules model the
interaction of the agent with other agents as well as with the environment. ABM can
be used to model a number of phenomena in varied domains like Markets &
Economy, Organizations, World Wide Web and Social Systems etc. Availability of
fast and cheap computing power, coupled with other advances in computational
sciences has paved the way for use of ABM as a preferred modeling and simulation
technique. Since last few years ABM has become the frontrunner tool of the
Sociologists and Psychologists who try to model social behaviours, particularly the
behaviour of groups and societies. A large number of researches are now being
carried out using this generative approach to model and analyze social phenomenon,
such as spread of rumors, diffusion of ideas, emergence of cooperation, emergence of
conventions & social norms, evolution and dynamics of spread of religions and
cultures etc
This paper presents our experimental work on using ABM for understanding and
analyzing voting preferences of individuals. We have modeled the process and factors
shaping and affecting individual voting preferences for a particular political party. To
keep the model simple and consequently analysis more accurate, we have taken
individual voting preferences as ‘0’ and ‘1’ representing a two party political system.
Every individual’s final voting decision is shaped and affected by a number of factors,
such as his own perception about parties, his family views & beliefs, perception of his
friends, and his neighbourhood preferences. Our simulation models the role and
relative weights of these factors in shaping an individual’s voting preferences. We
have also incorporated social influence theory in our model which states that similar
persons are more likely to interact, and as a result of frequent interactions they are
likely to become more similar. We experimented with multiple settings and
configurations and obtained interesting results. The rest of the paper is organized as
follows. Section 2 describes the ABM approach for modeling and simulation. Section
3 explains the process and factors believed to shape individual voting preferences.
Section 4 presents the results of our experimental model. The paper concludes
(section 5) with a short discussion of the results and their relevance.

2 Agent Based Modeling

Traditional modeling approaches to social systems relied on equation-based models
that operate with a macro perspective. They operate on population attributes & their
distributions and lack the focus on individual’s role. Equation-based models, though
useful for macro-scale behaviour predictions, fail to model social systems (or
processes) that lack central coordination, systems that are very complex in terms of
their interdependencies and systems which produce novel emergent behaviours in
544 V.K. Singh, S. Basak, and N. Modanwal

absence of a clear understanding of the collective phenomenon. Taking this into

consideration, Axtell [1] suggested that there are three distinct uses of Agent-based
computational models of social systems: (a) when numerical realizations can be
proposed and solved: agent models can be used as social simulations; (b) when a
mathematical model can be formulated but can only be partially solved: agent based
models can be a useful tool of analysis; and (c) when mathematical models are either
apparently intractable or provably insoluble: agent based modeling is perhaps the only
technique available for systematic analysis.
Although technically simple, ABM is conceptually deep. Modeling a complex
social process requires high conceptual clarity and analytical ability. A key issue,
therefore, is to decide when to use ABM for modeling social systems. An indicative
list of situations when it is better to think and model in terms of agents is: (a) when
there is a natural representation of actors as agents; (b) when the interactions between
the agents are complex, non-linear, discontinuous, or discrete; (c) when agents exhibit
complex adaptive behaviours; (d) when the population or topology of interactions is
heterogeneous; and (e) when agents have spatial component to their behaviours &
interactions. ABMs in social sciences involve human agents, whose behaviours are
often complex, irrational and subjective. Therefore, one needs to think carefully about
the social phenomenon at hand before going for ABM. Further, the model needs to be
built at the right level of description, with only the right amount of details.
Unnecessarily adding complexity to a model can make the analysis difficult and the
model useless.

2.1 Steps in ABM

Designing an Agent-based model of a social system require first identifying the

purpose of the model, i.e., the potential question to be answered. The typical steps to
be followed afterward can be summarized as following: (a) identifying the agents and
their behaviour rules; (b) identifying agent relationships and their interaction patterns;
(c) selecting an ABM platform; (d) obtaining the required relevant data; (e) validating
the agent behaviour model; (f) running the model and recording the outputs; (g)
analyzing the outputs with a viewpoint of linking micro-scale behaviours of the
agents to macro-scale behaviour of the system; and (h) validating the model outcomes
and hence the model [2].

2.2 ABM Computational Resources

Many rich and easy to program software platforms and toolkits for ABM are now readily
available. Some of the popular modeling tools (particularly for social sciences) are: Net
Logo (http://ccl.northwestern.edu/netlogo/), Repast (http://repast.sourceforge.net/), Swarm
(http://www.swarm.org) and MASON (http://cs.gmu.edu/~eclab/projects/mason/). Net
Logo is a multi-agent programming language and modeling environment designed in
Logo programming language. It is highly suitable for modeling and exploring emergent
phenomena. Repast is basically an agent-based social network modeling toolkit but has
 Agent Based Modeling of Individual Voting Preferences with Social Influence 545

rich libraries to study dynamics of social processes as well. Swarm is a multi-agent

simulation package to simulate the social and biological interactions of agents and their
emergent collective behaviour. Swarm has two versions, namely Objective-C and Java
versions. MASON is one of the latest Java platforms in this group. In addition to these
platforms, there are many other toolkits & APIs available for modeling and visualizing
social systems.

2.3 Benefits of ABM

Understanding social systems not only requires understanding the individuals that
comprise the system but also how individuals interact with each other resulting in
global behavioural patterns, which can be sufficiently novel. ABM helps researchers
in investigating how large-scale macro effects arise from the micro-processes of
interactions among many agents. Axelrod and Tesfatsion [3] call ABM the third way
of doing science besides induction and deduction. ABM applied to social sciences
take a methodical approach that could permit two important developments: (a)
rigorous testing, refinement, and extension of existing theories that are difficult to
evaluate using mathematical & statistical techniques; and (b) a deeper understanding
of fundamental causal mechanisms in multi-agent systems. Bonabeau [4] goes a step
further in proposing that ABM is a mindset and not merely a technology. He
summarizes benefits of ABM over other modeling techniques in three statements: (a)
ABM provides a natural description of a system; (b) ABM captures emergent
phenomena; and (c) ABM is flexible.

2.4 Validating Agent Based Models

ABM, now an established scientific research practice, needs to incorporate a proper

methodology of validation in order to verify the robustness of its findings and to truly
act as a bridge between disciplines. Since simulation results of ABM are very
sensitive to how agents are modeled, validation of computational model & simulation
results is a critical issue in ABM [5]. Several approaches for validation of agent-based
computational models have been proposed. Carley & Gasser [6] categorizes the
validation approaches into three broad categories: (a) theoretical verification, which
determines whether the model is an adequate conceptualization of the real world with
the help of situation experts; (b) external validation, which determine whether the
results from the model matches the results from the real world; and (c) cross-model
validation, which compares the results of the model with other models. All these
approaches are essentially aimed at validating the model at the macro level. Gilbert
[7], on the other hand, emphasized that in order to completely validate a model it
should be validated not only at macro-level but also at micro-level. Before going for a
macro-level validation it is necessary to confirm that micro-level behaviours of agents
are adequate representation of the actors in the system. Econometric validation and
companion modeling techniques have been advocated to be suitable for empirical
verification of social system models [8]. There are however still some problems in
empirical validation of agent-based models [9], and validation continues to be one of
the central problems of computer simulation methods including ABM [10].
546 V.K. Singh, S. Basak, and N. Modanwal

2.5 ABM Applications in Social Science

ABM is currently being applied to model a variety of complex social phenomenon

where simple local interactions generate emergent system-level behaviours. Some
representative & relevant work can be found in [11], [12] & [13]. Macy & Willer [14]
group most of the work on Agent-based modeling of collective behaviours in two
categories: (a) models of emergent structure which includes works on cultural
differentiation, homophilous clustering, idea diffusion, convergent behaviours and
norms; and (b) models of emergent social order which include viability of trust,
cooperation and collective action in absence of global control. Goldstone & Janssen
[15] also identify three similar themes for Agent-based computational models of
collective behaviour namely: (a) patterns and organizations which include settlement
patterns & segregation, human group behaviours and traffic patterns; (b) social
contagion which include spread of ideas, fashions, cultures & religions; and (c)
cooperation which include evolution of cooperation, trust & reputations and social
norms & conventions.

3 Modeling Individual Voting Preferences

Every individual while deciding his voting preference is directly or indirectly affected
by a number of factors. The final voting decision of an individual can be taken as a
function of these factors. An individual makes a preference about a political party
based not only on his perception about the ideology & programs advanced by it, but
by several other important factors. Individual may have an inherent preference formed
over a substantial period of time from his perceptions. However, family, friends,
neighbourhood (representing social and religious group membership) and mass media
play an important role both in shaping an individual’s perception and deciding his
final vote. We have modeled a population of individuals with their initial preference
for a particular party, and then run the model by varying different factors to observe
how (and possibly why) his initial favour changes to a particular final voting
preference. Since the model deals with a two party system (used for simplicity of
analysis), every individual is believed to have an initial favour for one of the two
parties, represented as ‘0’ or ‘1’. Individuals during course of the time interact with
friends, family members and neighbours. We have also associated feature sets with
every individual to represent his overall mental and behavioural makeup. The friends
of an individual are then defined as those having similar feature set values as that of
the individual. Family of an individual represents the agents close to him and with
whom he interacts very frequently. Family members are known to have a greater
influence over an individual. Neighbours are taken as relatively distant agents on the
grid and are viewed to represent individuals belonging to same social and religious
groups of the concerned individual. An individual is thus believed to be affected not
only by his friends and family members but also by the social and religious groups
he may belong. Mass media also plays role in shaping an individual’s voting
preference. We have devised three versions of the simulation: (i) model with social
influence, (ii) a majority-based model, and (iii) model including friends, family and
neighbours.
 Agent Based Modeling of Individual Voting Preferences with Social Influence 547

3.1 The Theory of Social Influence

The Social influence assumes that individuals (or agents) often imitate good
behaviours through their interactions with other individuals. Its dynamics depends on
familiarity of the interacting individuals, density of the neighbourhood, and popularity
and spatial proximity of the other individuals. Interesting models of social influence
have been proposed by Carley [16], Axelrod [17] and Axtell [18], Coleman [19],
Nowak et al [20] etc. Axelrod, in his social influence model of dissemination of
culture [21], focused on factors of local influence (tendency for people who interact to
become more similar) and homophily (tendency to interact more frequently with
similar agents). The more the agents interact, the more similar they become; and the
more similar they become, the more likely they are to interact. Agents who are similar
to each other interact with each other and become more similar. Axelrod’s model
captures culture as a group of features with multiple trait values per feature. However,
the emphasis is not on the content of a specific culture but on the way in which
culture is likely to emerge and spread. The simulations with varying parameters
regarding grid size, number of features and number of traits per feature resulted in
polarization, despite the only mechanism for change being convergence towards a
neighbour. We have used the basic idea of Axelrod’s culture model to represent
friends of an individual. An individual’s friend is believed to have behavioural feature
set similar to the individual. The feature set similarity represents their like mindedness
of the individuals on various issues and hence possibility of having similar voting
preferences or at least being affected by each other’s preferences.

3.2 Role of Family, Neighbourhood and Mass Media

An individual makes up his voting preference through a complex process involving

non-linear feedback loop of interactions with a number of entities. Individuals interact
more frequently with their family members and relatively less frequently with other
persons belonging to their social and religious groups. An individual’s final voting
value is no doubt a function of these factors. However, what is more important to
understand is that what is the extent of this effect and the by what underlying process
an individual finally makes up his voting preference. We have modeled this in our
second and third simulation versions. The second version is a simple model to
understand the dynamics of voting preference changes due to majority effect. In this
model all individuals are assigned a random initial voting preference value and then
his interactions with his 8-neighbour agents is observed and analyzed. The final
voting preference is a stable value that results by a process of majority sampling.
Since the initial voting preference values seem to play an important role on the final
outcome, we experimented with multiple initializations. This also modeled the
implicit favour (or political wave) introduced by the mass media. The third version of
the simulation is relatively more complex and takes into account voting preferences of
individual’s family and socio-cultural and religious neighbourhood. We have
explored with different weights assigned to the influence of family and neighbours
and observed their effect on the resultant complex set of interactions.
548 V.K. Singh, S. Basak, and N. Modanwal

4 Experimental Setup and Results

We have designed our voting preference simulation using NetLogo [22]. The first
model that we designed can be described as follows. A population of agents is created
on a wrap-around agent grid of size 41 X 41. Every patch on the grid represents an
individual. Every individual has certain attributes that characterize it. These attributes
are represented as a feature set, which can be a 3 to 7 bit long string. Every bit in the
string represents a particular trait of that attribute. The feature set values of two agents
determine how similar they are. Two individuals will be similar if their feature set
string contains common traits. For example, individual A (feature set: 00110) and B
(feature set: 10110) are highly similar. In addition to the feature set, every agent also
has a voting preference value, initialized randomly to either ‘0’ or ‘1’. Agents interact
with various agents in their neighbourhood. According to social influence theory, it is
more likely to interact and be affected by the agents similar to it. Interactions between
similar agents make them more similar, represented by a change in their voting
preference so that they get similar voting preference. The model selects an agent and
samples its neighbourhood. Suppose n out of total m agents in the neighbourhood are
similar to the agent. Then the voting preference of the agent changes to the majority
preference of the similar agents (n). We obtained multiple runs of the model by
varying both (i) the feature size, and (ii) initial voting preference value distribution.
The feature size has been varied from 3 to 7 with feature similarity threshold being
varied from 2 to 3 for feature set size of 5 and 3 to 7 for feature set size of 7. The
initial voting preference assignments has also been varied from 10%-90% to 50%-
50% for vote values of ‘0’ and ‘1’. Figures 1 and 2 show a snapshot of two runs of the
model. With higher similarity threshold (in figure 1) there is less local polarization
towards any vote value despite having initial favour for one vote value. Low feature
set similarity however results in contiguous groups settling to same vote values.

Fig. 1. A snapshot of the run of feature set based simulation. The feature set size is 5 and
similarity threshold is 3. In the figure on the left, the initial values of green = 757 and blue =
924 transform to final values of green = 730 and blue = 951. In the figure on the right, the
initial values of green = 505 and blue = 1176 transform to final values of green = 417 and blue
= 1264.
 Agent Based Modeling of Individual Voting Preferences with Social Influence 549

Fig. 2. A snapshot of the run of feature set based simulation. In the figure on the left, the
feature set size is 7 and similarity threshold is 3 and the initial values of green = 505 and blue =
1176 transform to final values of green = 117 and blue = 1564. In the figure on the right, the
feature set size is 7 and similarity threshold is 7 and the initial values of green = 169 and blue =
1512 transform to final values of green = 169 and blue = 1512.

In the second version of the simulation we did not use feature set concept, rather
we chose to have a simple model where voting preference of an individual is simply
taken as the majority vote value of its neighbours. For example, if more than 4 of the
8 neighbours of an agent prefer the vote ‘1’, then the individual is also likely to vote
‘1’ irrespective of his original preference. Hence the model is driven purely by the
majority. At every tick, an agent sees in its neighbourhood and decides its own vote
value based on the vote values of its 8-neighbours. Therefore, the majority vote value
around an agent persuades him to align with it. In case of tie, that is in the 8-
neighbourhood 4 agents have ‘1’ vote value and the remaining 4 have ‘0’ vote value,
the agent’s original vote value prevails. Since the final voting value of an individual
depends only on the majority voting value present in the 8-neighbourhood and that the
initial vote values are set randomly; we thought it would be interesting to see what
happens when the initial vote values of agents have a non-uniform distribution. We
run the model with multiple initial vote value distributions. The initial vote value
assignments have been varied from 10% agents having ‘0’ vote and remaining 90%
having ‘1’ vote, to many other combinations such as 20%-80%, 30%-70%, 40%-60%,
45%-55% and 50%-50% of ‘0’ and ‘1’ votes. A snapshot of the two interesting runs is
shown in figure 3. The results demonstrate that the final voting preference of the
population in the majority model depend heavily on the initial distribution of voting
preferences. Whenever there is a favour (about 60% against 40%) for initial
distribution of one vote value over the other (say ‘0’ over ‘1’), then the final voting
preference of most of the individuals becomes the favoured one (‘0’ vote). In case of
close to equal distribution of both the vote values (‘0’ and ‘1’ being less than 10%
different), the final voting preference of the majority may tilt relatively favourably on
either of the sides.
550 V.K. Singh, S. Basak, and N. Modanwal

Fig. 3. A snapshot of the run of majority based simulation. In the figure on the left, the initial
values of green = 1561 and blue = 1040 transform to final values of green = 2345 and blue =
256. In the figure on the right, the initial values of green = 1171 and blue = 1430 transform to
final values of green = 659 and blue = 1942.

Fig. 4. A snapshot of the run of modified majority simulation. In the figure on the left, the
initial values green = 505, blue = 1176 & weights own-vote = 80%, family-vote = 15% and
neighbour-vote = 5%; results into final values green = 413, blue = 1268. In the figure on the
right, the initial values green = 841, blue = 840 & weights own-vote = 40%, family-vote = 40%
and neighbour-vote = 20%; results into final values green = 843, blue = 838.

The third version of the model that we explored is a modified majority model. In
this version an individual’s voting preference is determined not simply by the
prevalent majority vote, but trough a weighted average of values of his own vote, his
family members’ vote and his neighbours’ vote. We chose 4-neighbourhood to
represent family and 8-neighbourhood to represent neighbours. Here, 8-neighbours
stand to represent members of individual’s socio-cultural and religious groups. Since
 Agent Based Modeling of Individual Voting Preferences with Social Influence 551

the 8-neighbourhood also contains the 4-neighbourhood on the grid, we have chosen
relatively smaller weight for neighbours. We have run the model by varying the
weights in weighted average computation and also the initial vote value distribution.
Figure 4 presents the results of two such runs. The results demonstrate that despite
giving higher weight to the individual voting preference, the initial bias introduced in
the population results to a certain degree of local polarization. Only very high value of
own vote is able to sustain the other voting preference. In case of approximately equal
initial voting preference also, high own-vote weight value tend to preserve most of the
initial voting pattern. As the weight assignments tend to become relatively more in
favour of the family and neighbourhood combine, local polarization is observed. The
results in this case become somewhat similar to the majority-based version of the
simulation.

5 Conclusion
The ABM methodology is a favorable approach to model and analyze complex social
phenomena that may involve non-linear feedback loops. It has been applied
successfully to model a number of social phenomena involving different social
processes and organizational structures. In this paper, we have demonstrated how
ABM can be applied to model and analyze the voting preference formation and
resultant voting decisions of individuals in a population. The model assumes a two
party system. We designed three versions of the simulation and observed the results
for a large number of runs with different parameter variations. The initial voting
preference distribution in the population, role of friends (likeminded individuals),
affect of family members and neighbours; all produce interesting results to analyze.
Social influence theory incorporated into model shows that individuals’ voting
decisions are affected more by those who are similar to them. However, if the
similarity threshold is relatively lower, the model results in local voting preference
polarizations. When the model is based only on majority, the initial distribution of
voting preferences becomes the key deciding factor. These results taken together
resemble to actual voting patterns observed in closely knit societies and segregations
such as far off villages and tribal groups. The role of implicit favorable waves
introduced by mass media can also be understood by the results of the second version
of the simulation. When the weights of factors playing role in shaping an individual’s
voting preference are varied, the results demonstrate that high weight to individual
vote is able to sustain variations in voting preferences. When the own-weight is
reduced to give more favour to the family and neighbours, polarizing tendency is
seen. Voting behaviour and patterns of relatively informed and well-educated persons
conform to this trend. The experimental simulations attempt to model an interesting
social phenomenon and present interesting results that need to be further analyzed to
draw relevant inferences, analogies and explanations.

References
1. Axtell, R.: Why agents? On the Varied Motivations for AgentCcomputing in the Social
Sciences - working paper 17. Centre on Social and Economic Dynamics, Brookings
Institution, Washington D.C (2000)
552 V.K. Singh, S. Basak, and N. Modanwal

2. Macal, C.M., North, M.J.: Tutorial on Agent Based Modeling and Simulation. In: Kuhl,
M.E., Steiger, N.M., Armstrong, F.B., Joines, J.A., Young, M. (eds.) Proceedings of the
2005 Winter Simulation Conference, The Technical Writer’s Handbook, University
Science, Mill Valley (1989)
3. Axelrod, R., Tesfatison, L.: A Guide for Newcomers to Agent Based Modeling in the
Social Sciences. In: The Handbook of Computational Economics. Agent Based
Computational Economics, vol. 2 (2005)
4. Bonabeau, E.: Agent-based Modeling: Methods and Techniques for Simulating Human
Systems. Proceedings of National Academy of Science 99(3), 7280–7287 (2001)
5. Takadama, K., Kawai, T., Koyama, Y.: Micro- and Macro- level Validation in Agent-Based
Simulation: Reproduction of Human-Like Behaviours and Thinking in Sequential
Bargaining Game. Journal of Artificial Societies and Social Simulation 11(2), 9 (2008)
6. Carley, K.M., Gasser, L.: Computational and Organization Theory. In: Weiss, G. (ed.)
Multiagent Systems - Modern Approach to Distributed Artificial Intelligence, pp. 299–330.
The MIT Press, Cambridge (1999)
7. Gilbert, N.: Open Problems in Using Agent-based Models in Industrial and Labor
Dynamics. In: Leombruni, R., Richiardi, M. (eds.) Industry and Labor Dynamics: the Agent
Based Computational Approach, pp. 401–405. World Scientific, Singapore (2004)
8. Moss, S.: Alternative Approcahes to the Emperical Validation of Agent- Based Models.
Journal of Artificial Societies and Social Simulation 11(1), 5 (2008)
9. Windrum, P., Fagiolo, G., Moneta, A.: Emperical Validation of Agent-Based Models:
Alternatives and Prospects. Journal of Artificial Societies and Social Simulation 10(2), 8
(2007)
10. Kuppers, G., Lenhard, J.: Validation of Simulation: Patterns in the Social and Natural
Sciences. Journal of Artificial Societies and Social Simulation 8(4) (2005)
11. Epstein, J.M.: Generative Social Sciences: Studies in Agent-based Computational
Modeling. Princeton University Press, Princeton (2007)
12. Gilbert, N., Troitzsch, K.: Simulation for the Social Scientist. Open Univ. Press,
Buckingham (1999)
13. Epstein, J.M., Axtell, R.: Growing Artificial Societies: Socail Science from bottom-up.
MIT Press, Cambridge (1996)
14. Macy, M.W., Willer, R.: From Factors to Actors: Computational Sociology and Agent
Based Modeling. Annual Review of Sociology 28, 143–166 (2002)
15. Goldstone, R.L., Janssen, M.A.: Computational Models of Collective Behaviour. Trends in
Cognitive Sciences 9(9) (2005)
16. Carley, K.M.: A Theory of Group Stability. American Sociological Review 56, 331–354
(1991)
17. Axelrod, R.: The Complexity of cooperation. Princeton University Press, Princeton (1997)
18. Axtell, R., Axelrod, R., Epstein, J.M., Cohen, M.D.: Aligning Simulation Models: a case study
and results. Computational and Mathematical Organization Theory 1(2), 123–141 (1996)
19. Coleman, J.S.: The Use of Electronic Computers in the Study of Social Organizations.
European Journal of Sociology 6, 89–107 (1965)
20. Nowak, A., Jacek, S., Latane, B.: From Private Attitude to Public Opinion: A dynamic
theory of social impact. Psychological Review 97, 362–376 (1990)
21. Axelrod, R.: The Dissemination of Culture: A model with local convergence and global
polarization. Journal of Conflict Resolution 41, 203–226 (1997)
22. Wilensky, U.: NetLogo. Centre for Connected Learning and Computer-Based Modeling,
Northwestern University, Evanston, IL (1999),
http://ccl.northwestern.edu/netlogo
 Texture Analysis of Brain MRI and Classification with
BPN for the Diagnosis of Dementia

T.R. Sivapriya1, V. Saravanan2, and P. Ranjit Jeba Thangaiah3

1
Dept. of Computer Science, Lady Doak College, Madurai
2
Dept. of Computer Applications, Dr.N.G.P. Institute of Technology, Coimbatore
3
Dept. of Computer Applications, Karunya University, Coimbatore

Abstract. In this paper, we present an evaluation of diagnosis of dementia us-

ing texture analysis of brain MRI with wavelets and further classification by
BackPropagation Network. The tests were conducted on 3D brain MRI data ex-
tracted from OASIS database. The classification is based on the following
steps: First, the region of interest is extracted from the MRI images by wavelets,
Gray level occurance matrix (GLCM) and Haralick features. Gabor features
were characterized by the distribution of histogram of wavelet coefficients.
These features were segregated into three datasets. The first dataset containing
the GLCM features, the second data set has the Haralick features and the third
dataset has Gabor wavelet based Haralick features. Classification was done by
backpropagation network based on 3 feature vectors. From the analysis it has
been found that the average efficiency of Gabor combined with Haralick fea-
tures is around 97% for all types of datasets, and the average efficiency value
for GLCM is 86 % and that of Haralick features was 90%. From the comparison
of the average efficiency of the wavelet families, statistical features extracted
from Gabor wavlets provides better efficiency than the other two methods.

Keywords: Texture classification, GLCM, wavelet, Gabor, Backpropagation

network.

1 Introduction
Texture analysis is an important issue in image processing with many applications
including medical imaging, remote sensing, content-based image retrieval, and docu-
ment segmentation. Over the last three decades, texture analysis has been widely stu-
died and many texture classification techniques have been proposed in the literature.
Ideally texture analysis should be invariant to translation and rotation. However, most
of the proposed techniques assume that texture has the same orientation, which is not
always the case. In the literature, essentially two different kinds of textures like sur-
face and visual are examined. Texture analysis, which is commonly found in the lite-
rature on image processing and computer vision [1], is expected to provide appropri-
ate features for subsequent studies, such as content-based image retrieval, pattern
recognition, and image segmentation.
Texture is an image feature that provides important characteristics for surface and
object identification from image [2-4]. Texture analysis is a major component of
image processing and is fundamental to many applications such as remote sensing,

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 553–563, 2011.
© Springer-Verlag Berlin Heidelberg 2011
554 T.R. Sivapriya, V. Saravanan, and P. Ranjit Jeba Thangaiah

quality inspection, medical imaging, etc. It has been studied widely for over four dec-
ades. Recently, multiscale filtering methods have shown significant potential for tex-
ture description, where advantage is taken of the spatial-frequency concept to maxim-
ize the simultaneous localization of energy in both spatial and frequency domains [5].
The use of wavelet transform as a multiscale analysis for texture description was first
suggested by Mallat [6]. Recent developments in the wavelet transform [7] provide
good multiresolution analytical tool for texture analysis and can achieve a high accu-
racy rate.
Alzhemier’s disease (AD) is a disease often related to ageing. The financial and
social burdens of AD on the society are becoming overwhelming with the continued
increase of the average life span. As such, considerable efforts have been made to
develop biomarkers for early detection of AD. It is well accepted that brain atrophy is
one of the most important markers of AD. Brain structures such as deep white matter,
corpus callosum, hippocampus, amygdala, caudate nucleus, and thalamus are known
to deteriorate in dementia subject to atrophy. Since brain ventricles are in the center of
the brain, and are surrounded by the brain structures mentioned, any change in these
brain structures will affect the shape of cerebral gray and white matter [8-12]. The
brain volume is significantly changed in AD patients compared to healthy subjects of
the same age group. Visual assessment of ventricle volume or shape change has
shown to be quite reliable [13][14] in detecting AD. Zhu [15] applied Fourier analysis
for image features, Ferrarini[16], Chaplot [17] applied wavelets, Selvaraj etal[18]
applied Haralick components to extract features for brain MRI analysis.
The paper is organized as follows. In Section 2, we briefly review the theory of
wavelets and the traditional approach to texture analysis with wavelets. The backpro-
pagation algorithm and its architecture are reviewed in Section 3. In Section 4, the
proposed methodology, feature extraction and validation are described. Experimental
results and conclusion are presented in Sections 6.

2 Texture Analysis

2.1 Wavelet Transform

Given a real, mother wavelet function ψ (t ) ∈ L2 ( R ) , which satisfies the admissi-

bility condition,

∞ | Ψ (ω ) |2
C ψ =  − ∞ | ω |
dω < ∞ (1)

where Ψ (ω ) is the Fourier transform of ψ (t ) , the scaling parameter p and the shift-
ing parameter q are introduced to construct a family of basis functions or wavelets
{ψ p , q (t )}, p , q ∈ R, p ≠ 0 , where

1 t−q
ψ p , q (t ) = ψ( ). (2)
| p| p
 Texture Analysis of Brain MRI and Classification with BPN 555

If Ψ (ω ) has sufficient decay, the admissibility condition reduces to the constraint

that Ψ (0) = 0 or
∞
−∞
ψ (t )dt = Ψ (0) = 0 (3)

Because the Fourier transform is zero at the origin, and the spectrum decays at high
frequencies, the wavelet behaves as a bandpass filter.
The continuous wavelet transform (CWT) of a function or a signal x(t ) ∈ L2 ( R )
is then defined as
∞
X CWT ( p, q) =  x(t )ψ p , q (t )dt , (4)
−∞

and the inverse wavelet transform is given by

1 ∞ ∞ dpdq .
x(t ) =
Cψ  
−∞ −∞
X CWT ( p, q)ψ p , q (t )
p2
(5)

Since {ψ p , q (t )} is a redundant basis set, it implies that, by discretizing the scaling

and shifting parameters as follows:

p = 2 k , q = 2 k l, k , l ∈ Z , (6)

an orthonormal basis {ψ k ,l (t )} is obtained,

ψ k ,l ( x) = 2 − k / 2ψ (2 − k x − l ), k , l ∈ Z (7)

which comply with the condition,

∞
−∞
ψk,l (t)ψk',l' (t)dt = δ (k − k' )δ (l − l' ), ∀k, k', l, l'∈Z (8)

The wavelet transform is an appropriate tool for multiresolution analysis (MRA) [19].
In MRA, a scaling function φ (x ) and the associated mother wavelet functionψ (x ) ,
which are needed in the construction of a complete basis, must satisfy the two-scale
difference equations:

φ ( x) = 2  h(n)φ (2 x − n)
n
(9)
ψ ( x) = 2  g (n)φ (2 x − n)
n

where the coefficients h(n) and g (n ) satisfy the following:

g (n) = ( −1) n h(1 − n) (10)

556 T.R. Sivapriya, V. Saravanan, and P. Ranjit Jeba Thangaiah

2.2 Gabor Texture

Usually the Gabor features of an image M(Y) can be obtained by convolving the im-
age with the family of Gabor filters

→ →
Gu ,v ( x ) = ( M * ϕ u ,v )( x ) (11)

where Gu,v(x) denotes resultant Gabor representation of image M at orientation u and

scale v. The resultant Gabor feature set thus consists of the convolution results of an
input image M(c) with all of the Gabor filters.

Fig. 1. Magnitude of Gabor wavelet

Fig. 2. Real part of Gabor wavelet

Figures 1 and 2 represent the magnitude and real part of Gabor wavelets.

2.3 GLCM Texture

One of the simplest approaches for describing texture is to use statistical moments of
the intensity histogram of an image or region [20]. Using only histograms in calcula-
tion will result in measures of texture that carry only information about distribution
of intensities, but not about the relative position of pixels with respect to each other in
that texture. Using a statistical approach such as co-occurrence matrix will help to
provide valuable information about the relative position of the neighbouring pixels in
an image. Given an image I, of size N×N, the co-occurrence, matrix P can be defined
as follows:
 Texture Analysis of Brain MRI and Classification with BPN 557

N N  1, ifI( x, y) = i and I ( x + Δx , y + Δ y ) = j
P(i, j) =   (12)
x=1 y =1  0, otherwise

Here, the offset (Δx, Δy), is specifying the distance between the pixel-of-interest and
its neighbour. Note that the offset (Δx, Δy) parameterization makes the co-occurrence
matrix sensitive to rotation. Choosing an offset vector, such that the rotation of the
image is not equal to 180 degrees, will result in a different co-occurrence matrix for
the same (rotated) image. The four matrices are used separately for classification, and
then the final decision is formed by fusing the four decisions. As these matrices are
symmetric, it is more convenient to use the upper or lower diagonal matrix coeffi-
cients in forming the vectors. The second-order statistical features are extracted from
GLCM. In brief, a GLCM is constructed by systematically considering the relation-
ship between pixel pairs and tabulating the frequency of various gray-level com-
binations within an image or within a region of interest The maximum number of gray
levels considered for each image or region of interest is typically scaled down to 32 or
64 levels, rather than using the full dynamic range of, for example, 65,536 gray levels
(16 bits/pixel)[21]. This quantization step is required to avoid a large number of zero-
valued entries in the co-occurrence matrix.

2.4 Haralick Features

Haralick features are extracted from the brain MRI with the formulae given in
Table 1.
Table 1. Haralick Features

Property Formula
Angular second moment G −1 G −1

 [ p(i, j )]
i =0 j = 0
2

Correlation G −1 G −1 ijp (i, j ) − μ x μ y


i =0 j =0 σ xσ x
Inertia G −1 G −1

 (i − j )
i =0 j = 0
2
p (i, j )

Entropy G −1 G −1
−  p (i, j ) log 2 [ p (i, j )]
i =0 j = 0
Absolute value G −1 G −1

 i − j p(i, j )
i =0 j =0
Inverse difference G −1 G −1
p (i, j )
 1+(i − j )
i =0 j = 0
2
558 T.R. Sivapriya, V. Saravanan, and P. Ranjit Jeba Thangaiah

3 Backpropagation Algorithm

The backpropagation algorithm for estimating parameters in neural networks has been
the most popular in the medical literature[21-23]. One of the problems encountered by
researchers utilizing the backpropagation algorithm is that low-frequency patterns
(or rare patterns) may take a long training time to be recognized, because frequent
patterns “dominate” the error. This problem is overcome by faster training using
PSO.
A hidden layer is used for ANN’s internal transformation: hidden neurons trans-
form the “signals” from the input layer and generate signals to be sent to the next
layer, so that MLP is often called the “feedforward” neural network. The input layer
consists of input variables plus an input bias, which adjusts the weighted mean of the
input variables. All input variables are regressed on the next layer, in the same fa-
shion as multiple regression estimation.

Fig. 3. Backpropagation Network

3.1 BPN Architecture Design

BPN architecture is selected based on testing and training of 7 different models. The
steps are as follows:
1. Create a customized BPN model n1.
2. Train and test with new data.
3. Check for the efficiency and ability to recognize new patterns.
4. If the model n1 has comparatively higher efficiency and detects new pat-
terns, it is selected as the ideal model for diagnosis.

4 Proposed Methodology
MRI images are preprocessed with Gaussian filters and normalized before feature
extraction. The overall methodology is depicted in Figure 4.
 Texture Analysis of Brain MRI and Classification with BPN 559

MR image

Image Preprocessing

Feature extraction
(GLCM, GLCM+Haralick, Wavelet+GLCM )

Training by BPN

Testing

Classification & Validation

Fig. 4. Methodology

4.1 Feature Extraction Phase and Classification

For feature extraction, the following steps are done:

1. The voxel-wise texture features of image I (x, y, z) are extracted at each slice of
3D ROI by convoluting with 2D Gabor filters [24], [25]and averaging inside
the ROI. The 2D Gabor filters are mathematically described at location (x, y) as
λ = 1/f is the wavelength, θ the orientation, γ the spatial aspect ratio which
determines the eccentricity of the convolution kernel.
2. A 32X32 gray level co-occurrence matrix with 32 gray are generated for the first
level approximation images of each training image considered for feature
extraction.
3. Haralick features listed in Table-1 is calculated for each training image.
4. For each texture feature compute the average of different angles.
5. Compute the overall mean of each feature of the entire set of images considered
for feature extraction.
6. Feed features as input vectors to BPN models.
7. Test the sensitivity, specificity and accuracy of the classification based on the
texture features.
560 T.R. Sivapriya, V. Saravanan, and P. Ranjit Jeba Thangaiah

4.2 Validation

The ability of the neural network model (a 10 X 7 X 1 structure was used) to perform
the classification was examined by setting aside 20% of the patterns (or observations)
as validation (or testing) data. In this cross-validation approach, the training involved
repeatedly exposing the network to the remaining 80% of the patterns (training data)
for several epochs, where an epoch is one complete cycle through the network for all
cases. (Data were normalized before training.) A network trained in this manner is
considered generalizable, in the sense that it can be used to make predictions.

5 Data Set
OASIS provides brain imaging data that are freely available for distribution and data
analysis [26]. This data set consists of a cross-sectional collection of 416 subjects
covering the adult life span aged 18 to 96 including individuals with early-stage Alz-
heimer’s Disease (AD) . For each subject, 3 or 4 individual T1-weighted MRI scans
obtained within a single imaging session are included [27]. Many studies have been
done to compare manual and computerized diagnosis of dementia [28].

Division of Data into Training and Testing

It is easy to design an optimistically biased (low error, over-trained) algorithm. If, for
example, the same data is used for model selection (i.e. optimization of C and γ or n
in our case), model training, and validation, an obvious risk is the creation of a ma-
chine which does not generalize. Such a model may be of limited effectiveness for
classifying novel data. To eliminate such circular logic for this study validation is
performed on independent/unseen data. The records of 50 patients with Alzhemier’s
disease (AD) and 50 Cognitively Normal (CN) datasets were randomly divided into
two sets of 25 AD, 25 CN.

6 Results and Conclusion

The hybrid wavelet based BPN combined is found to effectively improve classifica-
tion speed from brain MRI images. The BPN model effectively predicted the change
in brain MRI in patients with Dementia as given in Figure 7. From the results of the
above experiments, we can conclude that when wavelet based texture features are
combined with Haralick parameters, classification is very effective and can be used
for real time diagnostic decision making according to Figures 5 and 6 .
Backpropagation network classifier yields high accuracy and sensitivity which is
the need of the hour in day-to-day clinical treatment. It can be observed that Gabor
wavelets are an effective feature extraction method which can improve and clearly
quantify the brain volume in Alzhemier’s disease. Haralick features extracted in
wavelet domain yields accurate results when compared with traditional GLCM fea-
ture extraction. The specificity of the BPN classifier is improved when wavelet based
Haralick features are given as input to the network rather than GLCM features alone
or Haralick features combined with GLCM.
 Texture Analysis of Brain MRI and Classification with BPN 561

120 Effi ciency of Texture analysis methods in Cross-sectional

database
100
Accuracy in %

80

60

40 GLCM

GLCM+Ha ra lick
20
W a velet+Ha ra lick
0

r
en

r
s

te
te
pu
m

at
at
am
ta

m
M
Pu

oc

te
ay

hi
pp

Gr

W
Hi

Region of Interest

Fig. 5. Efficiency of Texture Analysis methods in Cross-sectional OASIS database

Efficiency of Texture features in classification of Longitudinal

Database

100
95
Accuracy in %

90
85 GLCM
80 GLCM+Haralick
75 Wavelet+Haralick
en s r r
pu te te
ta
m at at
Pu am M m
oc ay hit
e
pp Gr
Hi W
Region of Interest

Fig. 6. Efficiency of Texture Analysis methods in Longitudinal OASIS database

562 T.R. Sivapriya, V. Saravanan, and P. Ranjit Jeba Thangaiah

Efficiency of BPN Classifier s

300

250

200

150 Accuracy
Efficiency

Figure 1. Results of Texture analysis for Cross-Sectional database

Specificity
100
Sensitivity
50

0
BPN-Model -1 BPN-Model2 BPN-Model3
Classifier M odels in BPN

Fig.7. Efficiency of BPN classifiers

References
1. Haralick, R.M., Shanmugam, K., Dinstein, I.: Textural Features for Image Classification.
IEEE Transactions on Systems, Man, and Cybernetics 3(6), 610–621 (1973)
2. Smith, G., Burns, I.: Measuring texture classification algorithms. Pattern Recognition Let-
ters 18, 1495–1501 (1997),
http://www.cssip.elec.uq.edu.au/guy/meastex/meastex.html
3. Arivazhagan, S., Ganesan, L., Priyal, S.P.: Texture classification using Gabor wavelets
based rotation invariant features. Pattern Recognition Letters 27(16), 1976–1982 (2006)
4. Vassili, Kovalev, A., Kruggel, F., Gertz, H.-J., Yves, D., von Cramon, D.: IEEE Three-
Dimensional Texture Analysis of MRI Brain Datasets. Transactions on Medical Imag-
ing 20(5) (May 2001)
5. Valkealahti, K., Oja, E.: Reduced multidimensional cooccurrence histograms in texture
classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 20, 90–94
(1998)
6. Mallat, S.G.: Multifrequency channel decompositions of images and wavelet models.
IEEE Trans. on Acoustics Speech And Signal Processing 11(7), 674–693 (1989)
7. Gonzalez, R.C., Woods, R.E.: Digital Image Processing, 2nd edn. Prentice Hall, Engle-
wood Cliffs (2002)
8. Kloppel, S., Stonnington, C.M., Chu, C., Draganski, B., Scahill, R.I., Rohrer, J.D., Fox,
N.C., Jack Jr., C.R., Ashburner, J., Frackowiak, R.: Automatic classification of MR scans
in Alzheimer’s disease. Brain 131, 681–689 (2008)
9. Georgiadis, P., Cavouras, D., Kalatzis, I., et al.: Enhancing the discrimination Accuracy
between metastases, gliomas and meningiomas on brain MRI by volumetric textural fea-
tures and ensemble pattern recognition methods. Magn. Reson. Imaging 27, 120–130
(2009)
10. Provenzale, J.M., Mukundan, S., Barboriak, D.: Diffusion-weighted and perfusion MR im-
aging for brain tumor characterization and assessment of treatment response. Radiolo-
gy 239, 632–649 (2006)
 Texture Analysis of Brain MRI and Classification with BPN 563

11. Sivapriya, T.R., Saravanan, V.: Automatic brain MRI mining using Support Vector Ma-
chine and Decision tree. CiiT International Journal of Artificial Intelligent Systems and
Machine Learning 3(2), 109–116 (2011)
12. Li, G., Yang, J., Ye, C., Geng, D.: Degree prediction of malignancy in brain glioma using
support vector machines. Computers in Biology and Medicine 36, 313–325 (2006)
(PubMed: 16446164)
13. Mahmoud-Ghoneim, D., Toussaint, G., Constans, J., et al.: Three dimensional texture
analysis in MRI: a preliminary evaluation in gliomas. Magn. Reson. Imaging 21, 983–987
(2003)
14. Snyder, A., Girton, L., Morris, J., Buckner, R.: Normative estimates of cross-sectional and
longitudinal brain volume decline in aging and AD. Neurology 64, 1032–1039 (2009)
15. Zhu, H., Goodyear, B.G., Lauzon, M.L., Brown, R.A., et al.: A new local multiscale
Fourier analysis for medical imaging. Med. Phys. 30, 1134–1141 (2003)
16. Ferrarini, Palm, W.M., Olofsen, H., Van Buchem, M.A., Reiber, J.H., Admriraal-Behloul,
F.: Shape Differences of the brain ventricles in Alzheimer’s Disease. Neuroimage (2006)
17. Chaplot, S., Patnaik, L.M., Jagannathan, N.R.: Classification of magnetic resonance brain
images using wavelets as input to support vector machine and neural network. Biomedical
Signal Processing and Control (2006)
18. Selvaraj, H., Thamarai Selvi, S., Selvathi, D., Gewali, L.: Brain MRI Slices Classification
Using Least Squares Support Vector Machine. IC-MED 1(1), 21–33 (2007)
19. Lee, T.S.: Image Representation Using 2D Gabor Wavelets. IEEE Transaction on Pattern
Analysis and Machine Intelligence 18(10), 959–971 (1996)
20. Dunn, D., Higgins, W.E.: Optimal Gabor filters for Texture Segmentation. IEEE Transac-
tion on Image Processing 4(7), 947–964 (1995)
21. Di Luca, M., Grossi, E., Borroni, B., et al.: Artificial neural networks allow the use of si-
multaneous measurements of Alzheimer Disease markers for early detection of the disease.
Journal of Translational Medicine 3, 30 (2005), doi:10.1186/1479-5876-3-30
22. Kyrki, V., Kamarainen, J.K., Kalviainen, H.: Simple Gabor Feature Space for Invariant
Object Recognition. Pattern Recognition Letters 25(3), 311–318 (2004)
23. Laon, Z., Shen, D., Xue, Z., Karacali, B., Resnick, S.M., Davatzikos, C.: Morphological
classification of brains via high-dimensional shape transformations and machine learning
methods. Neuroimage 21(1), 46–57 (2004)
24. Mozer, M., Jordan, M., Petsche, T. (eds.): Advances in Neural Information Processing
Systems, pp. 475–481. MIT Press, Cambridge (1997)
25. Randen, T., Husoy, J.H.: Filtering for Texture Classification: A Comparative Study. IEEE
Transaction on Pattern Analysis and Machine Intelligence 21(4), 291–291 (1999)
26. Marcus, D., Wang, T., Parker, J.M., Csernansky, J., Morris, J., Buckner, R.: Open Access
Series of Imaging Studies (OASIS): Cross-Sectional MRI Data in Young, Middle
Aged, Nondemented and Demented Older Adults. Journal of Cognitive Neuroscience 19,
1498–1507 (2007)
27. Bagci, U., Li Bai, A.: Detecting Alzheimer Disease in Magnetic Resonance Brain Images
Using Gabor Wavelets. In: IEEE 15th International Conference on Signal Processing and
Communications Applications. SIU, pp. 1–4 (2007)
28. Kloppel, S., et al.: Accuracy of dementia diagnosis -a direct comparison between radiolo-
gists and a computerized method. Brain 131, 2969–2974 (2008)
 A Host Based Kernel Level Rootkit Detection
Mechanism Using Clustering Technique

Jestin Joy and Anita John

Rajagiri School of Engineering & Technology, Kochi, Kerala

[email protected],
[email protected]

Abstract. Rootkits are a set of software tools used by an attacker to

gain unauthorized access into a system, thereby providing him with priv-
ilege to access sensitive data, conceal its own existence and allowing him
to install other malicious software. They are difficult to detect due to
their elusive nature. Modern rootkit attacks mainly focus on modifying
operating system kernel. Existing techniques for detection rely mainly
on saving the system state before detection and comparing it with the
infected state. Efficient detection is possible by properly differentiating
malicious and non malicious activities taking place in a kernel. In this pa-
per we present a novel anomaly detection method for kernel level rootkits
based on k-means clustering algorithm.

1 Introduction

In recent years attackers employ a variety of sophisticated methods to gain ac-

cess to the system. Main goal of a rootkit is to conceal the evidence of intruder
activities. Most rootkits need administrative level privileges to install it in the
system. Usually attacker make use of some system vulnerabilities to get ad-
ministrative level privileges. An ordinary user will find it difficult to detect the
presence of a rootkit, since he will not find any discrepancy in the behavior of
the system, even if the system is infected by a rootkit. With the increasing use of
operating system in smart phones and other embedded devices the threat posed
by rootkits[4][16] are of great concern.
The first generation rootkits were easier to detect since they mainly aimed
at modifying user level programs(logs, application binaries . . . ). Methods like
checksum verification could easily detect these type of infections [11]. Rootkits
first appeared in the end of 80’s as method to hide log files and now they pose
a serious threat to computer industry [5]. Great deal of research is going on in
the area of rootkit detection. In the past, system administrators relied heavily
on system utilities like ls, ps . . . to find the presence of rootkits [11]. But new
generation rootkits could easily hide their presence from these utilities. Kernel
level rootkits modifies vital areas of an operating system like its kernel and the
device drivers. Difficulty in detecting them arise out of the fact that they operate
at the same security level as the operating system.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 564–570, 2011.

c Springer-Verlag Berlin Heidelberg 2011
 A Host Based Kernel Level Rootkit Detection Mechanism 565

This paper presents a technique for detecting kernel level rootkits in Linux
based systems. The proposed method works based on detecting the anomaly
exhibited by a malicious kernel.

2 Background
Generally rootkits can be classified into four: (1) Hypervisor Level rootkits [7]
loads the original operating system as “guest”, thereby intercepting all its ac-
tivities. (2) Kernel Level rootkits modify kernel level code, so that security of
the system is compromised (3) Library Level Rootkit replaces system library
with malicious ones (4) Application Level rootkit [5] modifies application level
binaries like ls, ps, netstat . . . to give false information to the user.

2.1 Kernel Level Rootkit

Unlike other rootkits, kernel level rootkit modifies the kernel itself. Kernel being
the lowest level of operating system makes it a good choice for the intruder to
attack, since an attack on it is very difficult to detect. Also being at the kernel
level provides the attacker with complete freedom to access all most all areas of
an operating system.
In most cases rootkits finds its way into the kernel through Loadable Kernel
Modules(LKM). LKM allows extending the functionality of the kernel, without
recompiling the kernel. The code inserted using LKM provides same capabil-
ity as a kernel code. Another main advantage of using LKM is that they can
be added/removed on the fly. Though mainly aimed at debugging the kernel,
/dev/mem is also used by the intruders to attack the system [12]. In Linux based
systems, modules can be inserted through the utilities insmod or modprobe.
Kernel level rootkits occur in different forms. They affect system call table [10],
Virtual File System(VFS)(for example, by adore−ng rootkit), Interrupt Descrip-
tor Table(IDT) [15] . . . . VFS can be thought of as a kernel subsystem which pro-
vides a unified API to user-space applications. There are certain other rootkits
that employ a combination of these techniques(example zk rootkit [11]).
When a user level program access a system resource, it is accomplished through
a system call. The user level application performs a system call, which passes
the control to the kernel. The requested work is done by the kernel and re-
sult is passed back to the user level application. So system call is an important
target for attackers. Kernel level rootkits attack System Call table by different
mechanisms [11]. Three of them are listed below
1. System Call Table modification
In this method, attacker replaces original system call with his own custom
version. Mainly this is done by modifying the system call address by inserting
malicious LKM’s address. Knark [6] rootkit uses this approach.
2. System Call Target modification
In this attack, legitimate code in the target address is modified. This type
of attack does not modify system call table. It works by modifying the flow
566 J. Joy and A. John

of control in a system call. Usually a jump instruction is used to pass the

control to the malicious code.
3. System Call Table Redirection
In this method attacker replaces the call to System Call Table with his
own custom version. This is usually done by overwriting the memory where
System Call Table address is stored. By comparing with System.map file
this type of attacks of could easily be detected.
F ig.1 show the result of a System Call redirection where, the system calls
sys write, sys open, sys getdents are replaced with modified ones; sys rkwrite,
sys rkopen, sys rkgetdents.

Fig. 1. System Call Redirection example

3 Detection Mechanism
Current techniques for detection mainly rely on archived copy of system files for
detection [10]. But this method couldn’t detect all type of attacks and also for
the detection to work properly,the archived copy should be clean.
Because of its easiness and functionality attackers choose system calls as a
target for modification. If kernel level rootkits are used, the method of using
more than one application for getting an information doesn’t work. For exam-
ple we get information about the modules from both /proc/modules and lsmod
 A Host Based Kernel Level Rootkit Detection Mechanism 567

command. /proc filesystem is a special filesystem that is used by the kernel to

export information. Contents of this filesystem are created when they are read.
Many utilities get their information form /proc/. lsmod gets its information
from /proc/modules. If rootkit modifies /proc/modules both /proc/modules
and lsmod command gives same information. So gathering information relying
solely on user space applications doesn’t help.
Cryptographic checksum based methods [11] fail when rootkits use dynamic
directories to store rootkit related information. For example many rootkits rely
on the /proc directory(example knark rootkit) to store its contents.
Detection based on third party integrity monitors is catching up. For example
Copilot[14] uses a coprocessor to check the integrity of kernel code. It uses a PCI-
card to monitor the memory of the host system. Similar approach is followed
by Gibraltar[3]. Like Copilot, Gibraltar executes on an observer machine, and
monitors the execution of the target machine. In virtual machine[8] based rootkit
detection, observer modules fetches data from the guest operating system and
feds it to a rootkit detector, which can then analyze them. The need for third
party mechanism makes these approaches less attractive.
Another method of importance is using cryptographically signed kernel mod-
ules. This method prevents loading unauthorized modules into the kernel space.
Greg Kroah-Hartman proposed a method based on this using RSA encryption to
sign the modules [9]. But this method couldn’t find much support from developer
community because of its resistive nature.

3.1 Proposed Detection Mechanism

The proposed technique detects system call redirection attack which is a widely
employed kernel level rootkit attack. System call redirection can be thought of
as a special case of system call target modification attack. Proposed detection
mechanism doesn’t need prior information about the state of the system to
function. The proposed system make use of statistical information from the
infected system only.
An anomaly is any occurrence that is unusual, or unique. System call redirec-
tion attacks cause anomaly to the host system. System calls could be classified
into either higher level system calls ( such as sys write, sys read) or lower level
system calls ( those constituting VFS layer). Attackers modify higher layer calls
since they provide easier mechanism for modifying the integrity of the system.
During boot time higher level system calls are loaded first. So lower level system
calls tend to have higher memory address as compared to higher layer calls.
Anomaly detection mechanism tries to find out patterns in data that do not
conform to expected behavior. Creation of clusters could guide the system ad-
ministrator in detecting these anomalies. Cluster based estimation can be utilized
for effectively detecting the anomalies. Kernel level rootkit detection mechanism
designed should need less a priori information about the host system. Cluster
analysis being a unsupervised learning method acts as an efficient candidate for
568 J. Joy and A. John

this purpose. Unlike other host based detection mechanisms clustering based
method doesn’t need prior knowledge for dividing the input data into different
clusters. The proposed method uses euclidean distance as metric and variance is
used as a measure of cluster scatter.
First step is to collect data, making using of the current state of the system.
From this essential data are extracted. This involves memory address from sys-
tem call table. Next step involves constructing clusters based on this data. Last
step is analyzing the result to determine whether a rootkit attack occurred or
not. The different steps involved are described in detail in the following para-
graphs.
By default, Linux may access up to 4 Gigabytes of virtual memory, of which
upper portion of this memory is allocated for use by the kernel (Address between
0xC0000000 and 0xF F F F F F F F ). When a LKM based attack using new sys-
tem call occurs, it gets a unusually high memory address as compared to other
system calls. For the detection purpose the running kernel should be disassem-
bled. As our aim is to detect system call related attack, we scan the system
call table and the unconditional/conditional jump instructions are analyzed and
their operands, which are memory addresses are collected. Along with this we
collect system call addresses also. Attacker makes use of higher level system calls
which are loaded earlier in memory for manipulation.
For computational purpose we convert these hexadecimal memory addresses
to decimal values. We then do clustering on these decimal values and two clusters
are formed with the sparse cluster representing the malicious system calls. These
system calls are then checked whether they are higher level system calls. An
uninfected system still shows some members in cluster two even though it is
not affected by a rootkit attack. This is due to the presence of natural outliers
present due to lower level system calls. Analysis of the cluster data could help
to detect these false alarms.

4 Results and Discussion

We have evaluated the effectiveness of the proposed technique against Knark
rootkit[13] and some custom implementations. Running kernel is disassembled[1]
and memory addresses of system calls are retrieved for analysis. Based on this,
clusters are formed to differentiate between normal and malicious system calls.
Our method could detect all the system call modifications.
When tested on a clean system, some false positives are encountered. But
they amount for less than 0.005% of the total analyzed instructions. When these
false positives are analyzed, they turned out to be lower level system calls.
F ig.2 represents the distribution of system calls in a rootkit affected system.
The peaks represents the modified system calls and they are detected by the
clustering approach.
 A Host Based Kernel Level Rootkit Detection Mechanism 569

Fig. 2. System Call Distribution

5 Conclusion and Future Work

Residing in kernel makes kernel level rootkits difficult to detect. The proposed
method provides an efficient host based technique for rootkit detection with
reduced system overhead. Analysis could be done on a machine other than the
host system. This makes the detection approach more flexible.
The presence of natural outliers is a bottleneck of this approach. Even when
the system is not affected by rootkit, it will classify a very small amount of
memory addresses as modified ones. This arises because of the presence of some
low level system calls which are natural outliers. Right now attack based on low
level system calls hasn’t been found. So presence of this can easily be rejected
as a false alarm.
Future work include testing the effectiveness of this method against system call
table redirection attack and finding an efficient mechanism to contain the pres-
ence of false alarms in an uninfected system. As more and more news reports[2]
coming out about the rootkit attack on smart phones, extending this method on
smart phones could also be investigated.

References
1. Gdb documentation (2011), http://www.gnu.org/software/gdb/documentation/
2. Arthur, C.: More than 50 android apps found infected with rootkit malware.
Guardian Technology Blog (2011),
http://www.guardian.co.uk/technology/blog/
2011/mar/02/android-market-apps-malware
3. Baliga, A., Ganapathy, V., Iftode, L.: Detecting kernel-level rootkits using data
structure invariants. IEEE Transactions on Dependable and Secure Comput-
ing 99(PrePrints) (2010)
570 J. Joy and A. John

4. Bickford, J., O’Hare, R., Baliga, A., Ganapathy, V., Iftode, L.: Rootkits on smart
phones: attacks, implications and opportunities. In: Proceedings of the Eleventh
Workshop on Mobile Computing Systems & Applications, pp. 49–54. ACM, New
York (2010)
5. Bunten, A.: Unix and linux based rootkits techniques and countermeasures (2004)
6. Clemens, J.: Intrusion Detection FAQ: Knark: Linux Kernel Subversion (2001)
7. Desnos, A.: Detecting (and creating!) a hvm rootkit (aka bluepill-like). Journal in
Computer Virology, 1–27 (2009),
http://dx.doi.org/10.1007/s11416-009-0130-8
8. Garfinkel, T., Rosenblum, M.: A virtual machine introspection based architecture
for intrusion detection. In: Proc. Network and Distributed Systems Security Sym-
posium, vol. 1, pp. 253–285. Citeseer (2003)
9. Kroah-Hartman, G.: Signed kernel modules. Linux Journal (2004)
10. Levine, J., Grizzard, J., Owen, H.: A methodology to detect and characterize kernel
level rootkit exploits involving redirection of the system call table. In: Proceedings
Second IEEE International Information Assurance Workshop 2004, pp. 107–125.
IEEE, Los Alamitos (2005)
11. Levine, J.G., Grizzard, J.B., Owen, H.L.: Detecting and categorizing kernel-
level rootkits to aid future detection. IEEE Security and Privacy 4, 24 (2006),
http://portal.acm.org/citation.cfm?id=1115691.1115761
12. Lineberry, A.: Malicious Code Injection via/dev/mem. Black Hat Europe (2009),
http://www.blackhat.com/presentations/bh-europe-09/Lineberry/
BlackHat-Europe-2009-Lineberry-code-injection-via-dev-mem.pdf
13. Miller, T.: Analysis of the KNARK Rootkit (2004)
14. Petroni Jr., N.L., Fraser, T., Molina, J., Arbaugh, W.A.: Copilot - a coprocessor-
based kernel runtime integrity monitor. In: Proceedings of the 13th Conference on
USENIX Security Symposium, SSYM 2004, vol. 13, p. 13. USENIX Association,
Berkeley (2004), http://portal.acm.org/citation.cfm?id=1251375.1251388
15. Wang, Z., Jiang, X., Cui, W., Ning, P.: Countering kernel rootkits with lightweight
hook protection. In: Proceedings of the 16th ACM Conference on Computer
and Communications Security, CCS 2009, pp. 545–554. ACM, New York (2009),
http://doi.acm.org/10.1145/1653662.1653728
16. Yan, Q., Li, Y., Li, T., Deng, R.: Insights into Malware Detection and Prevention
on Mobile Phones. Security Technology, 242–249 (2009)
 Fuzzy Clustering for Effective Customer Relationship
Management in Telecom Industry

Gayathri Asokan and S. Mohanavalli

Department of Information Technology,

Sri Sivasubramaniya Nadar College of Engineering,
Kalavakkam, Chennai, India 603 110
[email protected], [email protected]

Abstract. Data mining is the process of extracting interesting patterns from

data. Data mining is recently proving very effective in business decision
making and is becoming a widely used strategy to improve CRM (Customer
Relationship Management). CRM is the process of managing a good
relationship with customer and improving the profitability of their interactions
with the customer. Data mining is widely used particularly in handling large
data sets as in telecom sector. Clustering is a popular mining strategy that
separates those data into subsets called clusters. This research work focuses on
comparing the two main approaches of clustering soft clustering and hard
clustering namely the Kmeans and Fuzzy C Means (FCM) clustering algorithms
on large telecom data to determine the churn ratio as a measure to enhance
CRM. It is observed that FCM outperforms Kmeans in estimating churn ratio
accurately and is more effective in supporting CRM.

Keywords: Data mining; Customer Relationship Management, Clustering,

Kmeans, Fuzzy C Means.

1 Introduction

With the globalization of world economy and market internationalization, the market
competition of the domestic Telecommunications industry has become more
reasonable and increasingly fierce. The fast-growing user group, diversified
operations and competition environment have put forward higher requirements for the
service quality of the Telecom industry. The Telecom industry offers new products
and technologies that creates a series of options and offers to customer in order to
satisfy their needs and requirements. Loss of valuable customers in a telecom
company is a more general problem that telecom company often faces. Retaining
existing clients is the best marketing strategy to survive in industry and it is more
profitable to keep and satisfy existing clients than to constantly attract new ones.
Churn management, as the general concept of identifying those clients most prone to
switching to another company, led to development of variety of techniques and
models for churn prediction. So it is must for the telecom company to classify their
customers into churn and non-churn customers [4].

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 571–580, 2011.
© Springer-Verlag Berlin Heidelberg 2011
572 G. Asokan and S. Mohanavalli

To maximize the reduction in customer loss in the face of the fierce market
competition is to use the most effective and scientific method which is data mining
technology, and is currently the most powerful analytical tool. Nowadays data mining
is mostly used by corporate world to extract useful information from large amount of
databases. Data mining technology uses the known data to find implicit business rules
by establishing a mathematical model. Data mining technology based on rigorous
mathematical calculations can completely change the past status that mobile
enterprise can neither monitor the loss of customers nor achieve customer care after
successfully acquiring customers. The application of data mining to customer
relationship management can help telecom companies win new customers in the mass
market, allow existing customers to create more profits, keep valuable customers,
offer their customers more personalized services, etc, and the enterprise in a more
favourable competitive position. Data mining technology helps telecom companies to
select the right prospects of whom to focus and to identify the existing customers and
the good customers who may about to leave the company [6].
CRM is one of the most important of the information technology and is an
integration of business and technology process to gain the customer satisfaction and
used to satisfy the need of a customer. CRM basically begins with the deep analysis
of customer behavior. A telecom company that wants to create personal interaction
with the customer then they must gather all the customer related data and using this
data they will predict the churn customers.CRM helps telecom companies to increase
the profitability of their interactions with the customer while at the same time making
the interactions more friendlier through individualization. Retaining the customers is
important for telecom company to identify the problem of their customers and to
know about the relationship between customer and organization [1, 11].
Everyday people encounter a large amount of information. One of the vital means
in dealing with these data is to classify them into set of categories or clusters. To
classify the customers into churn and non-churn manually by having this large
number of databases is very complex than by doing data mining algorithms
technically. Classification systems are either supervised or unsupervised. We chose
one of the data mining techniques that comes under unsupervised learning which
groups the customer based on their similarity which is called clustering technique.
Clustering is a statistical technique used to identify both minimize within group
variation and maximize between group variation based on distance or dissimilarity
and to find an optimal set of clusters. A large number of clustering algorithms have
been developed for different type of applications. To cluster the massive amount of
data in telecom industries, we proposed Kmeans clustering (hard clustering) and
fuzzy clustering (soft clustering) from various type of clustering techniques. By using
this both clustering techniques, we can group the telecom customers into churn and
non-churn clusters, so that necessary steps will be taken before the churn customer
switch to a competitor[8].

2 Background Work
Telecommunication companies maintain an enormous amount of information about
their customers and, due to an extremely competitive environment, have great
Fuzzy Clustering for Effective Customer Relationship Management in Telecom Industry 573

motivation for exploiting this information. For these reasons the telecommunications
industry has been a leader in the use of data mining to identify customers, retain
customers, and maximize the profit obtained from each customer. A key component
of modeling a telecommunication customer’s value is estimating how long they will
remain with their current carrier. This problem is of interest in its own right since if a
company can predict when a customer is likely to leave, it can take proactive steps to
retain the customer. The process of a customer leaving a company is referred to as
churn, and churn analysis involves building a model of customer attrition. Customer
churn is a huge issue in the telecommunication industry where, until recently,
telecommunication companies routinely offered large cash incentives for customers to
switch carriers [13].
Data mining technique becomes gradually increasing and important tool for
telecom industry. The various techniques involved in data mining is classification,
estimation, clustering, association and prediction. Clustering is generally used for
market segmentation and classification. Clustering separates a heterogeneous
population into homogeneous subgroups or clusters so that data in each cluster share
some common trait. So it is more efficient for using this clustering technique in
telecom company to group their customers according to the services used by them. In
paper [16], the research workers used RFM (Recency, Frequency, Money) rule to
analyze the customer value. By using this RFM rule prediction of the customer
characteristics can be done. Yi-Hui-Lang defined customer value is fundamental to
CRM. It is a starting point of CRM to understand and measure the true value of a
customer. In paper[2], the author finds the value of a customer using Kmeans
clustering algorithm based on RFM analysis and assessing the value of existing
customer. So by using Kmeans algorithm, author can able to classify the customers
according to their similarity level[3].
In crisp clustering, each data point in the sample space assigned to only one cluster.
We can able to assign each and every customer to a particular service based, in which
a particular service is well used by a set of customers rather than other services. The k
in the Kmeans algorithm stands for the number of cluster seeds initially provided to
the algorithm. The goal of this algorithm is the assignment of each customer into one
and only service. Kmeans is based on the principle which minimizes the clustering
property indicator. A commonly used clustering criterion function is the minimized
variance of all the same points (data or objects) to the central point in the cluster.
Kmeans is a prototype-based, simple partition clustering technique which attempt to
find an user specified k number of clusters. These clusters represented by their
centroids, which is typically the mean of the points in the cluster [5, 9].
Hard clustering technique identifies each customer to a certain service category,
with the nature of “either-or”, so this type of clustering is obvious. But in the
estimation of customers in telecom company is unpredictable. The usage level may
vary day by day and their usage pattern is an intermediary, with the nature of “and
also”, so in these sample uncertain situations, soft clustering called fuzzy clustering
can be used efficiently and easily. Fuzzy Clustering provides a robust and resilient
method of classifying collections of large amount of (data points) customers by
allowing the same customer to reside in multiple service groups (clusters) with
different degrees of membership. In fuzzy clustering, a data point can have a
574 G. Asokan and S. Mohanavalli

membership in multiple clusters (each of different degree).Membership is a measure

of how strongly a data point is a member of a particular cluster [14, 10].
Goal of cluster analysis is the recognition and quantification of the groupings. The
quantification involves 2 processes (identifying membership of the data point and
locating the center of cluster). Fuzzy clustering can able to handle extreme outliers,
either very small or a higher than warranted membership degree in surrounding
cluster. It provides a flexible and robust method of assigning data points into clusters.
Each data point will have a degree of membership for each clusters in the range [0, 1]
indicating the strength of its placement in that cluster [3, 12].

3 Proposed System Design

3.1 Kmeans Clustering

The k in the Kmeans algorithm stands for the number of cluster seeds initially
provided to the algorithm. The goal of the hard Kmeans algorithm is the assignment
of each data point into one and only one cluster. The cluster centers are moved
according to a nearness or similarity function defining the distance from a cluster
center to the data point. The crisp (or hard) Kmeans algorithm itself is fairly
straightforward, involving the basic steps that are listed below [5, 7].

1 (Initialization) Select a set of k starting points. The selection may be done in

a random manner or according to some heuristic.
2 (Distance Calculation) For each data point Xi,1 ≤ i ≤ n, compute its
Euclidean distance to each cluster centroid Cj , 1 ≤ j ≤ k, and then find the
closest cluster centroid .

d = ¦ ( xi − c j ) 2
ji
3 (Centroid Recalculation) For each 1 ≤ j ≤ k, recompute cluster centroid Cj as
the average of data points assigned to it.
4 (Convergence Condition) Repeat steps 2 and 3, until convergence.

The Kmeans algorithm belongs to a family of crisp (or hard) clustering techniques
that uniquely partitions the sample data points into disjoint clusters. The Euclidean
distance metric measures how far away a point is from a cluster center, and thus
points that share common attributes will naturally cohere into the same cluster. All is
well and good if the clusters are distinctly arranged in the n-dimensional sample
space. In practice (that is, in reality), however, cluster perimeters are often quite
ambiguous. Data points in the sample space share attributes from one or more clusters
and should thus be placed in more than one cluster (albeit with a degree of
membership in one cluster that is higher or lower than the degree of membership in
another).
Fuzzy Clustering for Effective Customer Relationship Management in Telecom Industry 575

start

Number of clusters

Initialize Centroids no

No yes
objects end
ŽŵƉƵƚĞƚŚĞĚŝƐƚĂŶĐĞ moved
ďĞƚǁĞĞŶŽďũĞĐƚƐĂŶĚ
ĐůƵƐƚĞƌƐ

Grouping based on
minimum distance

Fig. 1. Flow diagram of Kmeans algorithm

3.2 Fuzzy C Means Clustering

Based on the fuzzy set theory, firstly introduced by Zadeh in 1965. and on the concept
of membership functions, the fuzzy clustering methods have been developed. In fuzzy
clustering entities are allowed to belong to many clusters with different degrees of
membership.The FCM Algorithm is a modified version of an original crisp cluster
methodology. The idea of fuzzification parameter (m) in the range [1, n], which
determines the degree of fuzziness on the cluster. When m > 1 the degree of fuzziness
among points in the decision space increases, when m=1, the clouds do not overlap,
but as you increase the value (m=1.25) the clouds begin to overlap and share many of
the same points [5, 7].
The steps involved in FCM algorithm are as follows:
1 (Initialization) Randomly initialize the cluster center.

2 (Distance Calculation) Creating distance matrix from a point xi to each of the

cluster centers to with taking the Euclidean distance between the point and
the cluster center.

d = ¦ ( xi − c j ) 2
ji
576 G. Asokan and S. Mohanavalli

Input data (Collection of

data sets)

Initialize number of
cluster centroids

Initialize fuzzification
parameter ‘m’

Calculate membership
function

Find the new cluster

centroids

Cluster No
center of
old=new

Yes

Document the obtained

Results

Fig. 2. Flow Diagram for FCM Algorithm

3 (Membership calculation) Finding Membership Calculation: takes the

fractional distance from the point to the cluster center and makes this a fuzzy
measurement by raising the fraction to the inverse fuzzification parameter.
This is divided by the sum of all fractional distances, thereby ensuring that
the sum of all memberships is 1.
1
§ 1 · m −1
¨ ¸
¨ d ¸
µ (x ) = © ji ¹
j i 1
p § 1 ·¸ m − 1
¦ ¨¨
k =1 © d ki ¸¹
Fuzzy Clustering for Effective Customer Relationship Management in Telecom Industry 577

4 (Centroid Recalculation) For each 1 ≤ j ≤ k, recompute cluster centroid Cj as

the average of data points assigned to it.
m
ª º
¦ « µ j ( xi )» xi
i ¬ ¼
C =
j m
ª º
¦ « µ j ( xi )»
i ¬ ¼
5 (Convergence Condition) Repeat steps 2, 3 and 4 , until convergence.

3.3 Advantages of Soft Clustering over Hard Clustering

When compared to Kmeans, FCM gives more accurate and efficient clustering results.
Kmeans performs a crisp clustering and allows each data point to belong to only one
cluster based on the Euclidean distance calculation. In FCM the fuzziness is resolved
using the membership value and is more accurate than hard clustering.

4 Implementation and Experimental Results

The original data was retrieved from Duke University’s churn management unit. The
data contains 65000 records (i.e. customers) and 144 attributes. The attributes were of
two types: categorical and continuous. Some of the categorical attributes such as
name, age of the customer, area where the customer resides etc. seemed to be
irrelevant to the problem under consideration. Hence the categorical attributes were
eliminated. Some of the customer records were with lot of missing attributes and such
records were ignored. Some of the continuous valued attributes like call duration,
number of calls in a month, number of messages etc. which were relevant to the
problem under study. Hence the customer records with the relevant attributes were
taken as input data. Finally an input data set with 1171 records and the most relevant
78 attributes was obtained after applying a data filtering technique.
The preprocessed input dataset wis used in both the hard clustering (Kmeans) and
soft clustering (FCM) algorithms in order to find the churn rate of customers in the
telecom company accurately.The implemented results are shown as follows:
Input given
Number of Customers = 1171, Number of services = 78
Number of Clusters = 2,
Algorithm used = Kmeans & FCM

It was observed that Kmeans clustering with k = 2, resulted 936 customers in group A
and 235 customers group B. On the other hand FCM clustering algorithm resulted in
919 customers in group A and 251 customers in group B.
The results of Kmeans algorithm was analysed and it was found that there were
some misclassifications. On analyzing the results of the fuzzy c means algorithm, the
misclassification was considerably reduced. The attributes which were significant in
producing the maximum deviations in the churn cluster were determined.
578 G. Asokan and S. Mohanavalli

Fig. 3. Input dataset

Misplaced Customers
44,89,184,212,273,283,320,327,354,421,429,520,736,977,1012.
Fuzzy Clustering for Effective Customer Relationship Management in Telecom Industry 579

Comparison of both the results:

From the comparison of clustering techniques, customer churn ratio was observed to
be different as per the clustering result. FCM groups the customers based on the
membership value. The membership value helps in determining the accurate churn
ratio of customers moving from one group to other group. The result of soft clustering
based on membership value was more accurate than Euclidean distance measure used
in hard clustering because the membership value in FCM clustering resolves the
fuzziness in the dataset, if existent. From the experimental results it was observed that
the misclassification was reduced in FCM and it resulted in estimating the churn ratio
accurately.

Conclusion and Future Work

It was observed that FCM gives better support to CRM rather than Kmeans in
clustering telecom data. It is further proposed to improve the scalability of fuzzy
algorithm to handle large real time telecom data and also to develop parallel versions
of the FCM to improve the efficiency.

References
1. Al-Mudimigh, A.S., Saleem, F., Ullah, Z., Al- Aboud, F.N.: Implementation of Data
Mining Engine on CRM – Improve Customer Satisfaction. In: International Conference on
Information and Communication Technologies, pp. 193–197. IEEE, Los Alamitos (2009)
2. Thomas, B., Raju, G.: A Novel Fuzzy Clustering Method for outlier Detection in Data
Mining. International Journal of Recent Trends in Engineering 1(2) (May 2009)
3. Dhillon, I., Modha, D.: A Data clustering algorithm on distributed memory
multiprocessors. In: Zaki, M.J., Ho, C.-T. (eds.) KDD 1999. LNCS (LNAI), vol. 1759,
pp. 245–260. Springer, Heidelberg (2000)
4. Popovic, D., Basic, B.D.: Churn Prediction Model in Retail Banking Using Fuzzy
C-Means Algorithm. Informatica 33 Public Sector Solutions, 243–247 (2009)
5. Cox, E.: Fuzzy modeling and genetic algorithms for Data Mining and Exploration, ch. 7,
pp. 207–206
6. Wies, G.M.: Data mining in Telecommunication Industry, pp. 486–491. International
Publisher of Science and Technology, IGI Global (2009)
7. Gayathri, A., Mohanavalli, S.: Enhanced Customer Relationship Management using Fuzzy
Clustering. In: International Conference on Computer Science and Engineering
Technology (IJCSET), vol. (7), pp. 163–167 (May 2011)
8. Gu, G., He, Y.: Research on the application of Data mining to CRM in Mobile
Communication. IEEE, Los Alamitos (2010)
9. Han, J., Kamber, M.: Data Mining: Concepts and Techniques, 2nd edn. Morgan
Kaufmanm, San Francisco (2006)
10. Zou, K., Wang, Z., Hu, M.: An new initialization method for fuzzy c- means algorithm.
Journal of Fuzzy Optimization and Decision Making 7(4) (December 2008)
11. Levine, S.: The rise of CRM. America’s Network 104(6), 34 (2000)
12. Cannon, R.L., Dave, J.V., Bezdek, J.C.: Efficient Implementation of the Fuzzy c-Means
Clustering Algorithms. IEEE Transactions on Pattern Analysis and Machine
Intelligence PAMI-8(2) (March 1986)
580 G. Asokan and S. Mohanavalli

13. Chen, R.-S., Wu, R.-c., Chen, J.Y.: Data Mining Application in Customer Relationship
Management of Credit Card Business. In: Proceedings of the 29th Annual International
Computer Software and Applications Conference. IEEE, Los Alamitos (2005)
14. Selim, S.Z., Ismail, M.A.: Kmean type Algorithms: A Generalized convergence Theorem
and characterization of Local optimality. IEEE Transactions on Pattern Analysis and
Machine Intelligence 6, 81–87 (1984)
15. Rashid, T.: Classification of churn and non-churn customers for Telecom companies.
International Journal of Biometrics and Bioinformatics (IJBB) 3(5)
16. Liang, Y.-H.: Combining the Kmeans and decision tree methods to promote customer
value for the Automotive Maintenance Industry. In: Proceedings of IEEM. IEEE, Los
Alamitos (2009)
 Dematerialized Deposits Using XSI

V. Radhaa1,*, Surabhi Upender1,2,

N.V.K.D. Ramesh Gorrila1,2, and B. Vijay Kumar1,2
1
Institute for Development and Research in Banking Technology,
Road #1, Castle Hills, Masab Tank, Hyderabad – 500 067 (A.P), India
Tel.: 91-040-23534981-85; Fax: 91-40-23535157
2
Department of Computer and Information Sciences,
University of Hyderabad, Hyderabad, 500 046 (A.P), India
[email protected], [email protected],
[email protected], [email protected]

Abstract. Traditional banking was based on paper documents like ‘deposit

form’, ‘withdrawal form’ etc. The comfort level for the customer on paper
based business was high as they can retain the paper and show as proof in case
of disputes. The present Internet and Core banking systems which are widely in
use are transaction based. One drawback of the present situation is that
customer is not able to keep a copy of the transaction due to various reasons
like multiple channels and lack of suitable infrastructure at client end etc. As a
via media, we tried to bridge the gap and see whether we can give the comfort
and trust levels of the paper document coupled with the intelligence of the
technology.
In this model we propose a method to describe financial product like “Fixed
Deposits” using XML technology. At present, banks give the deposits in paper
form with the deposit amount, maturity date and interest rate prescribed in the
document. The banks system calculates the interest and pays the customer. The
paper itself can’t calculate interest. It just helps the customer as a proof alone.
In our model, we are trying to see whether we can embed the interest
calculation algorithm into XML based deposit and give to the customer, so that
customer himself can find how much he gets using tools like XSI. This will act
as a good verification mechanism, instead of solely depending on banks’
systems. To show case this model as a POC, we used XSD, XML and XSI (an
in house tool with IDRBT).

Keywords: XSD, XML, Fixed Deposit, GUI , XSI.

1 Introduction
Fixed Deposit

Fixed deposit is a money deposit at a banking institution that cannot be withdrawn for
a certain "term" or period of time. When the term is over it can be withdrawn or it can
be held for another term.
*
Corresponding author.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 581–588, 2011.
© Springer-Verlag Berlin Heidelberg 2011
582 V. Radha et al.

Interest rate is paid on deposit accounts by commercial banks and other depository
financial institutions. Fixed interest rate is the rate which is fixed before the
investment of money in the bank and cannot be changed until the maturity of the
deposit is over.

Introduction to XSI

The XSI tool is under development in IDRBT. The XSI dynamically generates GUI
for the customer from composite XML schema document. Customer can enter the
data into the GUI and then XSI generates an XML document conforming to the given
XSD.

Taking input. XSI can create new XML documents or edit existing XML documents.
For new documents, XSI takes schema as input. For editing, XSI takes XML instance
documents as input.
GUI generator. This module presents the user with GUI for the user to enter the data.
Validator. In this module user entered data is verified against its corresponding XSD
document. Users would be alerted for appropriate corrections, in case the data is not
in conformance with the XSD.
XML document generator. This module of XSI generates XML document by taking
the XSD document as the input and user entered data.
Save/Send the document. The ultimate goal of XSI is to send the document to
establish good communication between customer and the bank. The generated XML
is either saved locally or sent through the network.
The following diagram gives the complete idea about the work flow of XSI.

Fig. 1. XSI – Conceptual View

 Dematerialized Deposits Using XSI 583

2 Application Overview
We developed the prototype using XML, XSD and XSI. In this section, we explain
the prototype components and work flow of the application.

2.1 Relationship among Schema, Advertisement and Deposit Document

• There can be one fixed deposit XSD structure to define the Deposit format.
• However, the interest rates keep changing and banks advertise different rates at
different points of time, based on different criteria like Principal amount, term
duration etc. So, each advertisement has to be represented as an XML document.
• There can be many customers depositing money against each advertisement and so
there can be many deposit instances which are again in the form of XML.

Fig. 2. M-M Relationship among schema, advertisement and deposit document

2.2 The Application Components

The application crosses three boundaries in the entire work flow. They are 1. Bank’s
Server, 2. Bank’s Teller and, 3.Bank’s Customer.

2.2.1 Bank’s Server

The Bank’s server hosts XML and XSD files which contain all the information about
bank deposit forms as well as their structure. The banks display advertisements which
show FD-Fixed Deposit interests from time to time. These advertisements are coded
by using XML and XSD.

2.2.2 Bank’s Teller

When the customer approaches the bank’s teller machine, it issues the fixed deposit
certificates to the customers. The bank teller contacts with the bank server to get the
details in the form of XSD and XML and then fills the deposit with the customer
details. Final deposit certificate is generated in the form of XML. Teller sends one
copy of this XML to bank server and another to customer.
584 V. Radha et al.

2.2.3 Bank’s Customer

The customer receives the final deposit certificate in XML format. The customer can
verify the deposit certificate at any point of time using XSI at his/her end.

2.3 Work Flow of the Model

Fig. 3. XSI Deposit Certificate model

1. The teller sends request to bank’s server to get appropriate XSD/XML

advertisement for the customer requirement.

Fig. 4. Fixed deposit Advertisement

2. Bank server provides required XSD/XML advertisement to teller. These

advertisements are coded by using XML and XSD.
 Dematerialized Deposits Using XSI 585

3. The XSI generates GUI for the teller to enter customer specific details.
4. The customer approaches the bank and provides details for taking the fixed deposit
at the teller. Deposit advertisement is created in GUI using deposit XSD.

Fig. 5. XML based advertisement conforming to deposit XSD

5. The teller enters the information in the GUI and submits.

• XSI uses the Deposit XSD to validate the information entered by the customer.
• The Deposit XSD is used to validate the information entered by the customer in the
GUI. If there are any violations, customers are alerted to correct the information.
6. If the validation is successful, the teller sends the XML document to the customer
and one copy is sent to the bank’s server for future reference. This XML file
contains all the information about the deposit money, interest rate and valid or
maturity date of the deposit.
7. The customer can use XSI tool to view the XML based deposit document.
586 V. Radha et al.

Fig. 6. GUI format for above XML

Fig. 7. XSD for Deposit form

 Dematerialized Deposits Using XSI 587

Fig. 8. Final Deposit XML with customer deposit details

3 Advantages

3.1 Advantages to Bank

─ Frequent changes of XML document for representing interest rates is allowed

and so changes are reflected from time to time as documents just like in physical
documents.
─ Generates structured information contained in XML documents. The structured
data is easier to handle, store, process, and analyze by application systems.
Thereby, it helps organizations to make better decisions to improve efficiency,
competitiveness and reduce fraud.
─ Excessive processing on the server is reduced as customers need not approach
the bank always. With tools like XSI, their own systems can participate in
application processing.

3.2 Advantages to Customer

─ We foresee the replacement of traditional browsers in all the financial

transactions with browsers like XSI.
─ The customers can calculate present deposits based on interest rates at any time.
─ This tool is providing GUI, so customer can enter data in an easy way in offline
mode as well, without connecting to Bank’s server and he can keep all copies of
his requests to bank(s) and the corresponding responses from bank(s) in a single
view.
─ Since the deposit is in standard XML format, it will help the customer to track
his transactions in the case of banks’ mergers.
588 V. Radha et al.

─ Since the Deposit Certificate is in standard XML form, new products like
trading the deposits also can become a reality in future.

4 Conclusion
The XSI tool takes both XML documents and XSD as input and generates output in
an XML format. It validates the information without human intervention. Through
this tool we can provide an easy way to calculate the interest rates in the form of
XML and these forms are sent to customers. So, the customer has valid proof for
further verification. Bank servers will update the customer information in an easy
way by changing the XML document.

References
[1] Bray, T., Paoli, J., Sperberg-McQueen, C.M.: Extensible Markup Language,
http://www.w3.org/TR/2008/REC-xml-20081126
[2] The World Wide Web Consortium Document Object Model (DOM),
http://www.w3.org/DOM
[3] XSD tutorial from W3C, http://www.w3schools.com/schema/default.asp
[4] Radha, V., Ramakrishna, S.: XSI- A XML schema based universal customer interface. In:
Enabling Technologies for Smart Appliances (ETSA), January 12-14, IEEE, Hyderabad
(2005)
[5] Radha, V., Ramakrishna, S., Pradeep Kumar, N.: Generic XML Schema Definition (XSD)
to GUI Translator. In: Chakraborty, G. (ed.) ICDCIT 2005. LNCS, vol. 3816, pp. 290–296.
Springer, Heidelberg (2005)
[6] Radha, V., Shiva Kumar, M., Pavan Kumar, K.: SIM-Structured Information Messaging.
In: CSI 2007 Annual Convention, November 28-December 1 (2007)
[7] Garvey, P.: Generating User Interfaces from composite schemas,
http://www.idealliance.org/xmlusa/03/call/xmlpapers/
03-03-04.994/.03-03-04.html
[8] Apparao, V., Byrne, S., Champion, M., Isaacs, S., Jacobs, I., Le Hors, A., Nicol, G., Robie,
J., Sutor, R., Wilson, C., Wood, L.: Document Object Model (DOM) Level 1
Specification. W3C Recommendation (October 1998),
http://www.w3.org/TR/1998/REC-DOM-Level-1-19981001
[9] Bergulund, Boag, S., Chamberlin, D., Fernandez, M.F., Kay, M., Robie, J., Simon, J.:
XML Path Language (XPath) 2.0 Technical Report, W3C Working Draft (2001),
http://www.w3.org/TR/XPath20/
 Classifying Gender from Faces Using
Independent Components

Sunita Kumari and Banshidhar Majhi

National Institute of Technology, India

[email protected],
[email protected]

Abstract. Gender classification algorithms uses a number of dimension-

ality reduction techniques. These techniques are used to remove linear
and non–linear dependency among the data. Most popular techniques
are Principal Component Analysis (PCA) and Independent Component
analysis (ICA). PCA makes the data uncorrelated while ICA gives inde-
pendent data. In this approach, infomax ICA has been used for gender
recognition and it is outperforming PCA. To perform classification using
ICA features, Support Vector Machine (SVM) has been used. For larger
training set SVM is performing with an accuracy of 98%. The accuracy
values are obtained for change in size of testing set and the proposed
system performs with an average accuracy of 96%.

Keywords: PCA, ICA, Gender Classification, Classifier, infomax.

1 Introduction
Face biometric has gained popularity because it is one of the most communica-
tive part of social interactions. Human-Computer Interaction as well needs face
recognition, gender recognition, estimation of age for passive surveillance. How-
ever, during identification of any individual the False Acceptance Rate(FAR)
increases as the database size increases [5]. Further the time required increases
exponentially as we need 1:N comparisons for identification. Hence, more thrust
is put in any biometric system to reduce search time without compromising
with accuracy. In a face recognition system, partitioning the database based on
gender is an important research direction. In this paper, more emphasis is laid
on classification of face database using gender. The major challenge for gender
classification is two-fold, (1) to find discriminating features and (2) an optimum
hyperplane for separating male and female classes. Hence most researchers thrust
on these two issues while attempting gender classification using face. There has
been some investigation done on classification of face database using gender.
By seeing a face image, anyone can say whether it is of a male or female
but machine intelligence needs automation. This is all because the data given to
machine are redundant. It needs an appropriate data/features and an optimum
classifier to take such a decision. Researchers are working in both the directions
equally. Geometric and appearance based features are the two ways to extract

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 589–598, 2011.

c Springer-Verlag Berlin Heidelberg 2011
590 S. Kumari and B. Majhi

such keypoints from an image. The former uses an idea to modal a human face
in terms of salient features like eyes, nose, mouth, etc., and the distance between
the salient feature points. The latter reduces a facial image containing thousands
of pixels to a handful number of pixels. The aim is to capture the distinctiveness
between the two genders from a face. The geometric features are termed as local
features and appearance features are termed as global features. The concept of
geometric keypoints extraction has come from the work of Leonardo Da Vinci.
He worked in the field of human anatomy and drew many pictures of human face
containing different face dimensions. Appearance based keypoints are used by
the researchers recently. But the work done in these fields are very contradictory.
There are several methods for dimensionality reduction like Principal Compo-
nent Analysis (PCA), Factor Analysis (FA), Independent Component Analysis
(ICA) and many others [12]. PCA has been first applied by Kirby and Sirovich
and it is shown that it is an optimal compression scheme that minimizes the
mean squared error between the original and reconstructed images. Later Turk
and Pentland have made it the most popular method for face recognition [13].
ICA is a generalization of PCA. This field is contradictory with respect to the
performance of these algorithms. For face recognition, Bartlett et al. [2], Yuen
and Lai [14], and Liu and Wechsler [8] have claim that ICA outperforms PCA
while Baek et al. [1] claims that PCA outperforms ICA. Moghaddam claims that
there is no significance difference between PCA and ICA [9]. It carries same ef-
fect in case of gender classification also. But all these performances varies across
several factors like the problem domain, distance measure techniques and the
number of basis vectors chosen. The different classifiers developed are multi–
layer neural network, back–propagation neural network, decision tree, support
vector machine(SVM) and many others [7] [10]. In this paper, proposition has
been made with ICA as a feature extraction technique and SVM as a classifier.
The rest of the paper is organized as follows: Section 2 provides a description of
the feature extraction techniques used, Section 3 discusses the classifier. Section 4
has been presented with simulation results and in last Section 5 gives conclusion.

2 Feature Extraction Techniques

This work has been motivated by the keen interest of researchers in face recog-
nition using independent component analysis [2]. A common problem of bio-
metric research is to find a suitable representation for multivariate data. Most
famous transformation methods include PCA, FA, projection pursuit, etc. A
new technique for linear transformation, which has been recently developed is
the independent component analysis. There are two types of dependencies exist
in the data (i) linear (ii) non–linear. Linear dependencies among the data can
be removed by PCA and ICA removes the non–linear dependencies. PCA and
ICA are unsupervised neural network based techniques which are also getting
used for the purpose of dimensionality reduction. As the face image contains large
 Classifying Gender from Faces Using Independent Components 591

redundant data inside it so there is a need to go with such a dimensionality reduc-

tion techniques. PCA decorrelates the data and ICA make the data independent.
In the proposed approach, ICA has been used for the gender classification system
and then it is compared with PCA.

2.1 ICA
ICA gives independent data and independence can be defined in terms of prob-
ability density functions (PDFs). Two variables y1 and y2 are said to be inde-
pendent if joint PDF of y1 and y2 will be factorizable in terms of their marginal
PDF and it can be achieved by either minimizing mutual information or by
maximizing non-gaussianity. This paper adopts the first approach.

Definition: For the given set of input sample x, ICA finds a linear transform
in such a way that

s = Wx

so that the components si are as independent as possible. Here, W is an unknown

separating matrix and it needs to be determined. There are many algorithms to
determine W like Jade algorithm, infomax ICA, fastICA and many others [6].
Infomax ICA is used in the current approach which is proposed by Bell and
Sejnosky [4]. The principle behind this is: when inputs are to be passed through
a sigmoid function maximum information transfer is achieved when the sloping
part of the sigmoid is optimally lined up with high density parts of the input and
this can be achieved by performing stochastic gradient ascent on the entropy of
the output with respect to the weight matrix W. It can be defined as follows:

W = (WT )−1 + E(Y XT )
Here X is the input data matrix, Y represents the output and can be expressed
as Y = f (S) = 1/1+e−s , T indicates transpose and E stands for expected value.
The ratio between the second and first order partial derivatives of the activation
  
function is given as: Y i = f i (Si ) / f i (Si ). It is shown in [3] that by maximizing
the joint entropy of the output Y , the statistical independence between the
outputs are increased or the mutual information are getting decreased.

Preprocessing of the data. It needs two preprocessing technique because it

makes the problem of ICA estimation simpler and better conditioned.
1. Centering: It gives the zero–mean data like for an input vector x subtract
the mean m = Ex from it.
2. Sphering/W hitening: It is a linear transformation method which
gives the whiten signal x̃ from an input signal x such that its compo-
nents are uncorrelated and their variances equal unity. Unit variance can
be achieved by making E(x̃x̃T ) = I(identity matrix). Whitening is done by
592 S. Kumari and B. Majhi

E{xxT } = EDE T . Here, E is the orthogonal matrix of eigenvectors of

E{xxT } and D is the diagonal matrix of its eigenvalues. x̃ is determined
by
x̃ = ED−1/2 E T x
Architecture of ICA. The goal is to find a better set of basis images so that
it can easily estimate an unknown dataset. There are two ways to achieve this
goal:
1. Architecture–1/ ICA–1 : It produces a set of statistically independent basis
images. The data matrix X is constructed in such a way that each row
represents an individual image. This architecture deals about independence
of images so that images are considered as random variables and pixels as
trials as shown in Fig. 1. Two images Xi and Xj are independent if it is
not possible to predict the value of an image Xj from Xi by moving across
pixels. A set of coefficients can be determined by the linear combination of
the independent basis images from which each individual image can be easily
constructed.
2. Architecture–2/ ICA–2 : In ICA–1, the basis images were statistical indepen-
dent from each other but the coefficients were not but in Architecture–2, the
coefficients are statistically independent. It finds a factorial face code for the
set of face images. The data matrix X is constructed in such a way that each
column corresponds to an individual image. This architecture deals with in-
dependence of pixels so that pixels will become random variable and images
as trials. Two pixels i and j are said to be independent, if by moving across
set of images it is not possible to predict the value of j from pixel i as shown
in Fig. 2.

X S

W
.
.
.
Learned Weights .
.
. .
. .
. .
. .
. .
.

Input Face Independent

Images Basis Images

Fig. 1. Architecture-1
 Classifying Gender from Faces Using Independent Components 593

X W S

Coefficient1

Pixel1 Pixel1 ... Pixel1

Coefficient2
Pixel2 Pixel2 ... Pixel2 . .
. .
. .
. . .
.
. . .
.
. . . .
. . .
.
. . .
.
.
.

Pixelk Pixelk ... Pixelk Coefficientk

Input Face Images Learned Filters Independent Components

Fig. 2. Architecture-2

3 SVM
SVM is a learning algorithm for pattern recognition and it found as a better
classifier than others in literature [10]. It designs a hyperplane in such a way that
it will leave more space on either side so that data on both sides can move a bit
more freely with less risk of causing an error [12]. A hyperplane is selected with
generalization performance of the classifier. It is the capability of the classifier,
designed using the training data set, to operate satisfactorily with an unknown
dataset.
Given a training set of instance-label pairs (Xi , yi ), i=1...l, where Xi ∈  N
and yn ∈ {-1,1}.
Proposed SVM is
f (X) = signW · X + b
where, W is the direction and b is the position of the hyperplane in space.
The optimal hyperplane L should satisfy following criteria
y i (W · X + b) ≥ 1 (1)
0 ≤ y i (W · X + b) < 1 (2)
y i (W · X + b) < 0 (3)
Above three cases can be treated under a single type of constraint by introducing
a new set of variables named as slack variable
y i (W · X + b ≥ 1 − ξi
The goal becomes
 N
1
minimize J(w, x, ξ) = W 2 +C I(ξi )
2 i=1
594 S. Kumari and B. Majhi

subject to y i (W · X + b) ≥ 1 − ξi
Where, i = 1, 2 · · · N, ξi > 0

I(ξi ) can be defined as 

1 , ξi > 0
I(ξi ) =
0 , ξi = 0
C is the cost parameter and is a positive constant that controls the relative
influence of above two competing terms. On simplifying above using Primal
Lagrangian and dual [12], we can get the optimal hyperplane as
l

f (X) = sign( y i ai K(X · X i ) + b)
i=1

where K(· , ·) is a kernel function defined by

γ||X i − X j ||2
K(X i , X j ) = exp(− ), γ > 0
2σ 2
Here, Radial Basis Functions (RBF ) kernel is used because it has the property
of symmetry.

Dimensionality Reduction
(ICA)

Training Set
Female
SVM

Male

Dimensionality Reduction
(ICA)

Testing Set

Fig. 3. Gender classification system

4 Classifier Performance
Proposed methodology for gender classification is shown in the Fig. 3. The per-
formance of gender classification system has been evaluated for two ICA archi-
tecture using the FERET database [11](sample images from this database are
 Classifying Gender from Faces Using Independent Components 595

shown in Fig. 5) and it is compared with existing PCA. The dataset consists of
images of 500 individuals. 200 images has been randomly selected which forms
the training dataset. From the remaining 300 images, four group of testing data
has been made. First testset consists of 200 individuals, second has 150 individu-
als, third consists of 100 and fourth has 50 individual images. Performance of the
system has shown on all these testset using same training set. The original size
of the image was 384×256. All these images has been aligned first and then con-
ventional triangularization approach is applied to get region of interest as shown
in Fig. 6. Features has been extracted using ICA from the cropped images of
size 293×241. The size of data matrix is 500×70613, where each row corresponds
to one particular image. The number of independent components(ICs) got by
the ICA algorithm corresponds to the dimensionality of the input. As the perfor-
mance merely depends on the number of separated components so there is a need
to control the number of ICs. To achieve this, ICA can be performed on linear
combination of original images and this has done using PCA. The PCA baseline
system uses 200 eigenvectors. Thus ICA has 200 ICs. The accuracy obtained by
various combinations is shown in Fig. 4 and its associated tabular representation
is given in Table: 1 .

100 ICA−1
ICA−2
PCA
80
Classification Rate

60

40

20

0
50 100 150 200
Different Testdata Size

Fig. 4. Classification rate of genders

By comparing the difference among infomax ICA–1, ICA–2 and PCA, it is

concluded that ICA–1 performs better than ICA–2 and even ICA–2 also out-
performs PCA . Images obtained from PCA is shown in Fig. 7 and of ICA–1
is shown in Fig. 8 and Fig. 9 shows the faces obtained from ICA–2. Litera-
ture entails that ICA–1 and ICA–2 gives local and global features respectively.
Thus, the proposed gender classifier is performing well with local features than
compared to global.
596 S. Kumari and B. Majhi

Fig. 5. Images from FERET database

. .
.
Crppped image
Original image

Fig. 6. Conventional Triangularization Approach

Table 1. Performance results(Accuracy)

Traindata Testdata ICA–1 ICA–2 PCA (%)

size size (%) (%)
200 50 98 96 92
200 100 96 94 90
200 150 93.33 90.67 88.66
200 200 92.25 89 86.5.

Fig. 7. Basis images from Principal Component Analysis

 Classifying Gender from Faces Using Independent Components 597

Fig. 8. Basis images from ICA Architecture-1: Statistical Independent Images

Fig. 9. Basis images from ICA Architecture-2: A factorial Face Code

These accuracies has been obtained using 200 basis vectors. As mentioned
earlier performance also depends on basis vectors, so the variation of basis vectors
on 200 training and 150 testing images is shown in Fig. 10.

95
Recognition rate

90

85

80
20 40 60 80 100 120 140 160 180 200
Number of basis vectors

Fig. 10. Classification rate on varying basis vectors

5 Conclusions
This paper has been presented with the comparison between PCA and ICA
with the problem domain, architecture, algorithm and classifier used. The per-
formance varies with all these parameters. Among all these, The most important
598 S. Kumari and B. Majhi

factor is the nature of task and then used algorithm. The dimensionality reduc-
tion techniques work better with face biometric like face identification, expression
recognition, gender recognition and many others. In current application, results
obtained from ICA are superior than that of the PCA. But there also, its vary-
ing with the features used, like in terms of more local and global appearance of
an image. All these are dealing with the support vector machine because it has
found as an optimum classifier from the literature.

References
1. Baek, K., Draper, B.A., Beveridge, J.R., She, K.: Pca vs. ica: A comparison on the
feret data set. In: Joint Conference on Information Sciences, pp. 824–827 (2002)
2. Bartlett, M., Movellan, J., Sejnowski, T.: Face recognition by independent compo-
nent analysis. IEEE Transactions on Neural Networks 13(6), 1450–1464 (2002)
3. Bell, A., Sejnowski, T.J.: The ‘independent components’ of natural scenes are edge
filters. Vision Research 37, 3327–3338 (1997)
4. Bell, A.J., Sejnowski, T.J.: An information-maximization approach to blind sepa-
ration and blind deconvolution. Neural Computation 7, 1129–1159 (1995)
5. Jain, A.K., Maltoni, D., Maio, D., Prabhakar, S.: Guide to Biometrics. Springer,
Heidelberg (2003)
6. Hyvrinen, A., Oja, E.: Independent component analysis: algorithms and applica-
tions. Neural Networks 13, 411–430 (2000)
7. Jain, A., Huang, J.: Integrating independent components and support vector ma-
chines for gender classification. In: Proceedings of the 17th International Confer-
ence on Pattern Recognition, ICPR 2004, vol. 3, pp. 558–561 (August 2004)
8. Liu, C., Wechsler, H.: Comparative assessment of independent component analysis
(ica) for face recognition. In: International Conference on Audio and Video Based
Biometric Person Authentication, pp. 22–24 (1999)
9. Moghaddam, B.: Principal manifolds and Bayesian subspaces for visual recognition.
In: The Proceedings of the Seventh IEEE International Conference on Computer
Vision, vol. 2, pp. 1131–1136 (1999)
10. Moghaddam, B., Yang, M.-H.: Learning gender with support faces. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 24(5), 707–711 (2002)
11. Phillips, P.J., Moon, H., Rizvi, S.A., Rauss, P.J.: The feret evaluation methodology
for face recognition algorithms. In: In (a) Example 1 (b) Example 1 (c) Right ear)
(d) Mirroed Left ear
12. Sergios Theodoridis, K.K.: Pattern Recognition. Springer, Heidelberg (2006)
13. Turk, M., Pentland, A.: Eigenfaces for recognition. J. Cognitive Neuroscience 3(1),
71–86 (1991)
14. Yuen, P.C., Lai, J.H.: Independent component analysis of face images. In: IEEE
Workshop on Biologically Motivated Computer Vision (2000)
 Experimental Evaluation of Fast Solution
Algorithms for the Simple Motif Problem

Tarek El Falah1,2 , Thierry Lecroq2 , and Mourad Elloumi1

1
Research Unit of Technologies of Information and Communication,
Higher School of Sciences and Technologies of Tunis, 1008 Tunis, Tunisia
2
University of Rouen, LITIS EA 4108,
76821 Mont-Saint-Aignan Cedex, France
[email protected],
[email protected],
[email protected]

Abstract. The Simple Motif Problem (SMP) is: given a set of strings
Y = {y0 , y1 , . . . , yn−1 } built from a finite alphabet Σ, p > 0 an integer
and q ≤ n a quorum, find all the simple motifs of length at most p that
occurs in at least q strings of Y .
This paper presents an experimental evaluation of algorithms dealing
with SMP and using a minimal forbidden pattern approach. The exper-
iments are concerned both with running times and space consumption.

Keywords: algorithms, strings, motifs, SMP, complexities.

1 Introduction
The motif finding problem consists in finding substrings that are more or less
conserved in a set of strings. This problem is a fundamental one in both Com-
puter Science and Molecular Biology. Indeed, when the concerned strings code
biological sequences, i.e., DNA, RNA and proteins, extracted motifs offer to Bi-
ologists many tracks to explore and help them to deal with many challenging
problems. In the literature, several versions of the motif finding problem have
been identified [4]:
– Planted (l,d)-Motif Problem (PMP) [2,11];
– Extended (l,d)-Motif Problem (ExMP) [9,14];
– Edited Motif Problem (EdMP) [12,13];
– Simple Motif Problem (SMP) [12,5].
A simple motif is a string built from a finite alphabet Σ ∪ {?} that cannot begin
or end with ?, where Σ is a set of symbols and ? ∈ Σ is a wildcard symbol, it
can be replaced by any symbol from Σ.
Let Y = {y0 , y1 , . . . , yn−1 } be a set of strings built from an alphabet Σ, p > 0
be an integer and q ≤ n be a quorum, the Simple Motif Problem (SMP) is to
find all the simple motifs of length at most p that occurs in at least q strings
of Y .

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 599–608, 2011.

c Springer-Verlag Berlin Heidelberg 2011
600 T. El Falah, T. Lecroq, and M. Elloumi

Actually, concerning SMP the main existing solutions are:

– the algorithm given by Floratos and Rigoutsos [8],
– the one described by Rajasekaran et al. [12].
– the SMS-Forbid and SMS-H-Forbid algorithms by El Falah et al. [5,6].
In this paper, we focus on SMP and we propose an experimental evaluation of
algorithms SMS-Forbid and SMS-H-Forbid. Both algorithms are based on
a minimal forbidden pattern approach. Algorithm SMS-Forbid uses indexing
structures (either suffix trees or suffix arrays) while algorithm SMS-H-Forbid
uses an hash table. The experiments are concerned both with running times
and space consumption. These experiments were conducted on pseudo-randomly
generated data on two alphabet sizes: 4 (to simulate DNA sequences) and 20 (to
simulate protein sequences).
We organize the rest of the paper as follows: In section 2, we present some
useful definitions and notations. In section 3, we review algorithms SMS-Forbid
and SMS-H-Forbid. Section 4 presents experimental results. In section 5, we
give our conclusion and perspectives.

2 Preliminaries
A simple motif is a string built from an alphabet Σ ∪ {?} that cannot begin or
end with ?, where Σ is a finite set of symbols and ? ∈ Σ is a wildcard symbol,
it can be replaced by any symbol from Σ. Symbols of Σ are said to be solid
while the wildcard symbol ? is said to be non-solid. The length of a simple motif
is the number of the symbols that constitute this motif, including the wildcard
symbols.
A string of  symbols from Σ is called a -mer. A string of  symbols from
Σ ∪ {?} is called a -pattern. A -pattern z1 is equivalent to a -pattern z2
(z1 ∼= z2 ), if at every position where z1 and z2 contains both solid symbols these
symbols are equal. Formally, z1 ∼ = z2 if for 1 ≤ i ≤ : z1 [i] = z2 [i] or z1 [i] = ? or
z2 [i] = ?.
A -pattern z1 is more general than a -pattern z2 if a position in z2 contains
the wildcard symbol implies that the same position in z1 contains the wildcard
symbol and if a position in z2 contains a solid symbol then at the same position
in z1 there could be either the same symbol or a wildcard symbol. Formally
z2 [i] = ? ⇒ z1 [i] = ? and z2 [i] = a ⇒ (z1 [i] = a or z1 [i] = ?) for 1 ≤ i ≤  and
a ∈ Σ.
Let Y = {y0 , y1 , . . . , yn−1 } be a set of strings built from an alphabet Σ and

n−1
let N = |yi |. Let m be the maximal length of the strings in Y . Let σ be the
i=0
size of the alphabet Σ.

3 Minimal Forbidden Pattern Approach

In this section, we present two algorithms SMS-Forbid and SMS-H-Forbid.
 Fast Solution Algorithms for the Simple Motif Problem 601

3.1 SMS-Forbid Algorithm

Given a set Y of n strings, a quorum q ≤ n and an integer p. The algorithms
output the set of motifs of length at most p that occurs in at least q strings.
A pattern z of length at most p is said to be a minimal forbidden pattern if
it occurs in less than q strings but all its proper factors beginning and ending
with a solid symbol occur in at least q strings.
For each position on the input strings, we use all the windows of length 
for 3 ≤  ≤ p. Each window defines an -mer. Each -mer x defines a set of
-patterns X. At each position of each -pattern z of X, the symbol of z is either
the symbol at the same position of x or the wildcard symbol except for the first
and the last symbols of z that are necessarily non-wildcard symbols. Formally,
z[i] = x[i] or z[i] = ? for 1 ≤ i ≤  − 2 and z[0] = x[0] and z[ − 1] = x[ − 1].
These -patterns together with the generality relation form a lattice. The
minimal element of the lattice is x itself and the maximal element is x[0]?−2 x[−
1]. (see example in Fig. 1: the minimal element is x = CAGAT and the maximal
element is C?−2 T).

Fig. 1. A 5-mer x = CAGAT lattice

Each node of the lattice represents an -pattern.

The -patterns are scanned by doing a breadth-first search of the lattice be-
ginning from the minimal element.
When a -pattern z is considered, if it has already been output or it contains
minimal forbidden patterns as factors or it is more general than an output pat-
tern, then it is disregarded otherwise it is searched in the strings of Y . Then if
it occurs in at least q strings it is output and all its successors in the lattice are
not considered since they are more general. On the contrary if it does not occur
in at least q strings it is added to the set of minimal forbidden patterns.
The generation of the -patterns is performed using a breadth-first search
of the lattice for the following reason. When a -pattern is discovered all its
successors in the lattice, that are more general, do not have to be considered.
They are thus marked using a depth-first search of the lattice from the discovered
-pattern. During the remaining of the breadth-first search, marked -patterns
are not considered.
602 T. El Falah, T. Lecroq, and M. Elloumi

Contrary to the algorithm presented in [12], the new approach does not search
for all the -patterns generated from the n strings of Y but it begins by searching
the more specific patterns i.e. the less general patterns which avoids the sorting
step. Moreover it maintains a set of minimal forbidden patterns that do not
occur in at least q strings in order to not search for any pattern that contains a
factor that has already been unsuccessfully searched. This two techniques reduce
the number of patterns to be searched for. The search of one -pattern in one
string y of Y is done using an indexing structure of y which can be done in a
time proportional to . Furthermore the new approach only outputs the more
specific motifs that occur in at least q strings of Y . This should help the biologist
to identify important motifs.
An illustrative example: Let Y = {y0 = ACGAAGACT, y1 = CGTAGCTAC, y2 =
CAGTACTAT, y3 = ACGTACAAA, y4 = CCTACTGCT} be a set of strings built from the
alphabet Σ = {A, C, G, T}. The set of most specific simple motifs having length
at most p = 6 and occuring in at least q = 3 strings of Y is

{TAC, ACT, GTA, CTA, C?TA, CG?A}.

We will give a top-down detailed presentation of the algorithm. It builds the set
Res of searched motifs of length at most p contained in at least q strings and
uses a set T of minimal patterns that are not contained in at least q strings.
It scans the n strings of Y from y0 to yn−1 . For each position of each string it
considers all the windows of length  for 3 ≤  ≤ p (at the end of the strings it
may be less than p). For each -mer x defined by each window of length , the
breadth-first search of the lattice is performed level by level.
During the breadth-first search of the lattice, when a non-marked -pattern x
is considered, it is searched in the set Res of resulting -motifs. For that, the set
Res is implemented using a trie: looking if an -pattern x belongs to Res simply
consists in spelling x from the root of the trie. If x belongs to Res then all it
successors are marked using a depth-first search.
If x does not belong to Res then the algorithm checks if it contains minimal
forbidden patterns. This consists in searching for a finite set of patterns T with
wildcard symbols in a text with wildcard symbols x, where a wildcard symbol in
the text only matches a wildcard symbol in the set of patterns while a wildcard
in the set of patterns may match any symbol in the text.
If x does not contain any minimal forbidden pattern then it is searched in
all the strings of Y from yj+1 to yn−1 since it has not been considered before
dealing with yj and it has at least one occurrence in yj . If it occurs in at least
q strings, it is added to Res and all its successors are marked using a depth-first
search. Otherwise it is added to the set T of minimal forbidden patterns.
The lattice is completely traversed in a breadth-first search in every cases.
Each -pattern x in the lattice is associated with an integer from 0 to 2−2
whose binary representation is given by x[1 . .  − 2] where each solid symbol is
replaced by 0 and each wildcard symbol is replaced by 1. For example ab?ba is
 Fast Solution Algorithms for the Simple Motif Problem 603

associated to 2 whose binary representation is 010. This enables to mark easily

the nodes of the lattice.
The candidate patterns are generated by considering the strings from y0 to
yn−1 . When a pattern is generated from yj it occurs at least once in yj then it
is searched in the following strings: yj+1 , yj+2 , . . . , yn−1 .
This procedure considers a factorization of an -pattern x as follows:

x = u0 v0 u1 , v1 · · · um−1 vm−1 um

where ui ∈ Σ ∗ for 0 ≤ i ≤ m and vj ∈ {?}∗ for 0 ≤ j ≤ m − 1 (remember that

an -pattern begins and ends with a solid symbol).
Then the search of x in y is performed by successively searching for the ui
in y and merging sets of positions. Assume that R contains all the positions of
u1 · · · ui in y, then the procedure computes the set T of all the positions of ui+1
in y. The two sets are merge in order to keep only the positions of R that are
compatible with positions of T . A position g of ui is compatible with a position
of T if there exists a position h in T such that g = h + |u1 · · · vi |. The merge can
be easily realized if the two sets are implemented as ordered arrays. The set of
positions of ui in y can be computed efficiently by using any indexing structures
of y (such as suffix trees or suffix arrays).
A problem remains when a more specific patterns is found after a more general
pattern. Then the set of resulting patterns has to be filtered.

3.2 SMS-H-Forbid Algorithm

In order to easily find the candidate patterns we define a table H for every couple
of solid symbols and every integer k from 0 to p − 3 as follows:

H[a, b, k] = {(i, j) | yi [j] = a and yi [j + k + 2] = b}.

When a candidate -pattern is generated from position j in string yi , if

– it has not already been output or

– it does not contain minimal forbidden patterns as factors or
– it is not more general than an output pattern

its potential occurrences are only searched at the positions in H[yi [j], yi [j +  −
1],  − 2]. In practice, the elements of H[a, b, k] are sorted in decreasing order of
the index of the strings.
The algorithm scans the strings of Y in the same order than algorithm SMS-
Forbid. The breadth-first-search is performed in the same manner. The only
changes appear for counting the number of strings containing an -pattern x
generated from yj . The occurrences of x are searched using the list of pairs in
H[x[0], x[ − 1],  − 3]. Furthermore those pairs (ind, pos) are sorted in decreasing
order thus only pairs where ind > j are considered.
604 T. El Falah, T. Lecroq, and M. Elloumi

3.3 Complexities

We will now give the complexities of both algorithms SMS-Forbid and SMS-
H-Forbid.
The algorithm SMS-Forbid scans all the positions of the n sequences of Y .
For each position it considers all the -patterns defined by the corresponding
-mer for 3 ≤  ≤ p. The number of elements of all the corresponding lattices
is bounded by 2p+1 . Processing one -pattern x consists in looking if x is in
Res, checking if x contains minimal forbidden patterns and searching x in the n
sequences of Y . Looking if x is included in Res can be done in O(|x|) time using a
trie for Res. Checking if x contains minimal forbidden patterns consists in using
an algorithm for searching a single pattern with wildcard symbols in a text with
wildcard symbols for every pattern in T . This can be done in O(|T ||x|). The
search of one -pattern x in one string y of Y consists in spelling all the solid
factors of x into the indexing structure of y. When a solid factor u of length |u|
is searched in y, it can be done in O(|u| log σ) with the suffix tree of y and in
O(|u| + log |y|) with the suffix array and the Longest Common Prefix array of y
(see [3]). However the search for each solid factor can return a list of positions
of size O(|y|).
Thus the time complexity of this step is O(|x| log σ + p|y|) with suffix trees
and is O(|x| + log |y| + p|y|) with suffix arrays.
The time complexity for building the indexing structures for all the n se-
quences of Y is O(N ). Altogether the time complexity of the algorithm SMS-
Forbid is O(N 2p σ p ((p + log m) + pm))) where m is the maximal length of the
sequences of Y .
The algorithm requires to build and traverse all the lattices corresponding
to -patterns. An array of size 2−2 is used to mark the nodes of each lattice.
Thus the space complexity for the lattices is O(2p ). The space complexity of the
indexing structures for all the n sequences of Y is O(N ). In the worst case the
size of Res and T is bounded by |Σ|p . Altogether the space complexity of the
algorithm SMS-Forbid is O(N + 2p + |Σ|p ).
A similar analysis can be done for algorithm SMS-H-Forbid. The complex-
ities are summarized in table 1.

Table 1. Time and space complexities of algorithms SMS-Forbid and SMS-H-Forbid

Algorithm Time Space

SMS-Forbid (suffix trees) O(N 2p σ p (p log σ + pm))) O(N + 2p + σ p )
SMS-Forbid (suffix arrays) O(N2p σ p ((p + log m) + pm)) O(N + 2p + σ p )
SMS-H-Forbid O(N 2p σ p (pm)) O(σ 2 p + 2p + σ p )

Next, we will denote by SMS-ST the SMS-Forbid algorithm using suffix

trees, SMS-SA the SMS-Forbid algorithm using suffix arrays and SMS-Hash
the SMS-H-Forbid algorithm.
 Fast Solution Algorithms for the Simple Motif Problem 605

4 Experimental Results

We implemented our algorithms in C. Experimental results were conducted on

pseudo-randomly generated data using the KISS generator [10] on two alphabet
sizes 4 (to simulate DNA sequences) and 20 (to simulate protein sequences). All
the experiments were performed on a P4.2 GHz machine with 3GB memory.
We measured the computing time and the space consumption of these algo-
rithms for different values of m, n, p and q. All these results have been obtained
through computing an average on 15 draws. Fig. 2 and 3 show the variation of
computing time in function of m and n for p = 7 and q = 10 on DNA alpha-
bet. The X-axis represents m and the Y-axis represents the computing time t in
seconds. We note that we have used a nonlinear method to implement the suffix
arrays for the input strings but which is fast in practice.

Fig. 2. Computing time in seconds of SMS–Forbid using Suffix Trees (left) and using
Suffix Arrays (right)

Fig. 3. Computing time in seconds of SMS-H-Forbid

As we can notice, SMS-H-Forbid performs better than SMS-Forbid in

terms of running time since it uses a hash table instead of an indexing structure
for the strings in Y (see Fig. 4). Indeed, the hash table used in SMS-H-Forbid
makes easier the search of candidate patterns.
606 T. El Falah, T. Lecroq, and M. Elloumi

Fig. 4. Computing time for n = 100 (left) and n = 700 (right) with DNA alphabet

We have also computed the memory consumption in MB of the specific data

structures used by each algorithm (see table 2). This table show that the SMS-
Forbid algorithm using suffix arrays consumes less memory than the other
algorithms.
We have employed our algorithms on various pseudo-randomly generated pro-
tein sequences (see Fig. 5). The experimental study show that SMS-H-Forbid
algorithm is a faster solution to SMP when the input strings are built with pro-
tein alphabet, despite the fact that its worst case time complexity is bigger than
the one of SMS-Forbid.

Fig. 5. Computing time in seconds with protein alphabet for n = 100

Fig. 6 shows the results obtained for SMS-H-Forbid in function of q and n for
m = 100. As we can notice, the curves have a bell-like shape. Indeed, for a low value
of q, the possibility of finding quickly the motifs is higher. When q is getting closer
to n, the number of detected minimal forbidden motifs increases. So, the possibility
to reduce the computing time is also higher. It is also remarkable to note that the
peak of the curves obtained with protein alphabet is for a low value of the quorum.
Indeed, the size of the alphabet makes the possibility of finding simple motifs that
have occurrences in at least q strings decreases rapidly as q increases.
 Fast Solution Algorithms for the Simple Motif Problem 607

Fig. 6. Computing time in seconds of SMS-H-Forbid in function of q and n with

DNA alphabet for p = 7 (left) and with protein alphabet for p = 5 (right)

Table 2. space consumption in MB with DNA alphabet for n = 500

PP
PP m
50 150 250 350
AlgorithmPPP
SMS-SA 0.29 0.73 1.17 1.61
SMS-ST 1.6 4.8 8.02 11.23
SMS-Hash 0.47 1.5 2.52 3.55

5 Conclusion
In this paper, we have proposed an experimental evaluation of algorithms dealing
with SMP. It is remarkable to note that our algorithms have the ability to
identify the most specific motifs earlier in the search process, which allows less
specific motifs to be pruned. By using this technique, we avoid the sorting step
of SMS algorithm presented in [12] and we filter motifs respecting the generality
relation between patterns. Experimental results show that our algorithms are
well performing in practice. To have a more efficient solution in terms of space
consumption, we are studying how to determine a new algorithm using the FM-
Index [7] which is a data structure based on the Burrows-Wheeler Transform [1].

References
1. Adjeroh, D., Bell, T., Mukherjee, A.: The Burrows-Wheeler Transform. Springer,
Heidelberg (2008)
2. Chin, F.Y.L., Leung, H.C.M.: Voting algorithm for discovering long motifs. In:
Proceedings of Asia-Pacific Bioinformatics Conference, pp. 261–272 (2005)
3. Crochemore, M., Hancart, C., Lecroq, T.: Algorithms on Strings. Cambridge Uni-
versity Press, Cambridge (2007)
4. El Falah, T., Elloumi, M., Lecroq, T.: Motif finding algorithms in biological se-
quences. In: Algorithms in Computational Molecular Biology: Techniques, Ap-
proaches and Applications, Wiley Book Series on Bioinformatics: Computational
Techniques and Ingeneering, pp. 387–398. Wiley-Blackwell, John Wiley and Sons
Ltd., New Jersey, USA (2011)
608 T. El Falah, T. Lecroq, and M. Elloumi

5. El Falah, T., Lecroq, T., Elloumi, M.: SMS-Forbid: an efficient algorithm for sim-
ple motif problem. In: Proceedings of the ISCA 2nd International Conference on
Bioinformatics and Computational Biology, Honolulu, Hawai, pp. 121–126 (2010)
6. El Falah, T., Lecroq, T., Elloumi, M.: An efficient motif search algorithm based on
a minimal forbidden patterns approach. In: Proceedings of the 5th International
Conference on Practical Applications of Computational Biology and Bioinformat-
ics. AISC, pp. 175–182. Springer, Heidelberg (2011)
7. Ferragina, P., Manzini, G.: Opportunistic data structures with applications. In:
Proceedings of the 41st IEEE Symposium on Foundations of Computer Science
(FOCS), Redondo Beach, CA, USA, pp. 390–398 (2000)
8. Floratos, A., Rigoutsos, I.: On the time complexity of the teiresias algorithm.
Technical report, Research Report RC 21161 (94582), IBM T.J. Watson Research
Center (1998)
9. Leung, H.C.M., Chin, F.Y.L.: An efficient algorithm for the extended (l,d)-motif
problem, with unknown number of binding sites. In: Proceedings of the Fifth IEEE
Symposium on Bioinformatics and Bioengineering (BIBE 2005), pp. 11–18 (2005)
10. Marsaglia, G., Zaman, A.: The kiss genrator. Technical report, Department of
Statistics, Florida State University, Tallahassee, FL, USA (1993)
11. Rajasekaran, S., Balla, S., Huang, C.H.: Exact algorithms for planted motif chal-
lenge problems. Journal of Computational Biology 12(8), 1117–1128 (2005)
12. Rajasekaran, S., Balla, S., Huang, C.-H., Thapar, V., Gryk, M., Maciejewski, M.,
Schiller, M.: High-performance exact algorithms for motif search. Journal of Clin-
ical Monitoring and Computing 19, 319–328 (2005)
13. Sagot, M.-F.: Spelling approximate repeated or common motifs using a suffix tree.
In: Lucchesi, C.L., Moura, A.V. (eds.) LATIN 1998. LNCS, vol. 1380, pp. 111–127.
Springer, Heidelberg (1998)
14. Styczynski, M.P., Jensen, K.L., Rigoutsos, I., Stephanopoulos, G.N.: An extension
and novel solution to the (l,d)-motif challenge problem. Genome Informatics 15(2),
63–71 (2004)
 Secure Web Services Negotiation

Vakul Mohanty and Chittaranjan Hota

Birla Institute of Technology and Sciences Pilani,

Hyderabad, India
[email protected], [email protected]

Abstract. The work proposed here, suggests a protocol that provides anonymity
to negotiating parties and protects negotiation information without the use of
encryption. The protocol is developed to ensure maximum security while
causing minimum delay. Decentralization of routing and service information
protects the information. The protocol makes use of K-anonymity principle to
ensure security.

Keywords: Security, anonymity, cryptography, negotiation.

1 Introduction
Improvement in Internet technology and its wide reach have made Internet a valuable
tool for conducting business on world wide web. Hence, more and more organizations
are seeking to use web services to forward their business interests. Web services
constitute of three phases: discovery, negotiation[2] and contracting. This work
addresses anonymity and data security issues in the negotiation phase. In our work we
device a protocol that maintains anonymity of negotiating parties and ensures that
negotiation related data is kept hidden. The protocol thus not only ensures anonymity
of negotiating parties but also hides the subject of negotiation, its attributes and
negotiation status. A recent work [3] lists secured communication strategies based on
encryption which is impractical in service sector for high latency. In this paper we
take on this problem using K-anonymity [4] [5]; and the protocol is devised as a
group protocol that can be well crafted into the generic architecture of webservices. In
next section we propose re-engineering web service architecture to support the
proposed protocol that assumes anonymity in service negotiation. The third section
presents a protocol and in the next section it's performance is analysed. Then we
conclude the paper with a comment.

2 Restructuring Web Service Architecture

A web service consists of three entities viz. service-provider, register and consumer
Fig. 1 .Currently, WSDL is enabled for secure communication among services mainly
by sharing keys and encrypting data. But, this could be a hindrance for mass usages
for high latency. We propose to ensure security by adopting K-anonymity for
communication. For that, business entities registered at UDDI are grouped into
several clusters or groups. Each cluster is assigned with a group leader

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 609–618, 2011.
© Springer-Verlag Berlin Heidelberg 2011
610 V. Mohanty and C. Hota

(representative) by the UDDI. The repository of conventional UDDI is augmented

with additional information . The UDDI stores group information: 1. group leader's
address 2. size of group 3. index of group leader.

Fig. 1. Traditional web service architecture

2.1 Groups at UDDI

Initially, business entities (both service consumers and providers) register them selves
at UDDI. And then, according to our scheme the entities are grouped to a predefined
size say Kmax. For each group a group leader is designated and the group members are
uniquely indexed. The information on each group i.e its group leader and members
are stored at UDDI for routing communication to and from the members via the group
leader. The group leader only is aware of physical address of its members. More, the
members in a group may belong to several different business domain to add diversity
to group that camouflage group business domain information to external world. Next,
we will discuss dynamic evolution of a group for movement i.e. joining or leaving of
the group.
Management: The scenarios we may face after initialization of UDDI are:
1. Entry of an entity: As and when a new user wants to carry out business
transactions it first makes a request to the UDDI. The UDDI informs one of the
group leaders to include the user as its member, at the same time updates the
group size. The group leader then allots the user its index hence making it a
group member. Accepting an index and joining a cluster binds a user to group
as a cluster member. If the new member is a service provider it updates the
UDDI with all its service details. Here we would like to mention that we
assume that group leaders are honest and don't participate in activities like
adding users to the groups without being directed by the UDDI.
 Secure Web Services Negotiation 611

2. New cluster formation: In case a new group is to be created the procedure is

the same as the initialization stage. The situation arises due to overgrowth of
a cluster i.e when each group attains the size Kmax.
3. Exit of entity: In case a cluster member wants to leave the cluster it informs
the UDDI and its group leader. The UDDI then updates the cluster size and
in case the user is a service provider its service details are also removed. The
group updates itself deleting the user and its index value. By doing this the
group leader removes the user from the cluster and rendering it incapable of
transacting web services in habitat of the current SOA. On exit, when the
population of a group is below a set minimum threshold value, the group
members are directed to join different groups on dissolving the existing one.
4. Exit of a group leader: Suppose a group leader wants to leave a cluster, then
the above procedure is carried out; in addition the UDDI choose a new group
leader from among the group members and updates its data, i.e. includes the
new group leader, provides a new index, updates group size. Similarly, the
new group leader assumes the information, that was available at previous
group leader .
The dynamic nature of the clusters effect the performance of the protocol suggested.
Hence we assume that the groups engaged in negotiation do not change. We discuss
the consequences of the dynamic nature of the groups later and suggest modifications
that might help to overcome the problems posed.

Table 1. Message structures

Message Message structure Path No (Fig. 2,Fig. 3)

ServiceRequest < servicename >< replyaddress > 1,2
< destinationaddress >
ServiceReply < Servicedetails >< serviceprovideraddress > 3,4
< destinationaddress >
Negotiation {< attributevalues >} < negotiationstatus > 5,6,7
message < replyaddress >< destinationaddress > 8,9,10

3 Protocol
This section presents the proposed protocol designed to carry out an anonymous
negotiation. The basic principle of the protocol is to generate look-real dummy
information that camouflage the real so that any malicious monitor cannot
differentiate the information regarding the negotiation. The identities of the parties
participating in negotiation are made hidden by assigning them indexes instead of
addressing their real address. Its to be noted that only UDDI knows group index and
a group leader knows indexes of its members. We assume that UDDI and group
leaders are trusted so identities of group members and their service informations are
612 V. Mohanty and C. Hota

Fig. 2. Service Discovery phase

confidential. A negotiation message is itself buried by creating a traffic of k dummy

messages from which differentiating the actual message becomes a difficult task.
Further, making message structure i.e. service attributes a pre-knowledge to both the
negotiating parties and avoiding usages of these attribute names at message
transaction helps to hide status of negotiation to a large extent. It is to be noted that
message contains only values so that a malicious monitor can not associate meaning
to these values due to absence of message attribute names. This protects the subject of
negotiation. Further, values in different dummy messages cast a different trends thus
able to obfuscate the status of negotiation as well. Thus the proposed protocol intends
to achieve 1) anonymity of negotiating parties 2) anonymity of context and 3)
anonymity of negotiation status.
Assumptions made: 1. All communications with the UDDI are secure and
observable. 2. all communications in the group are secure and group leaders are
honest. 3. The cluster is like a closed system, they communications taking place
within them are observable. The assumptions imply that negotiation between parties
in the same cluster are always safe and are thus trivial. In case the negotiating parties
are in different cluster we follow the proposed protocol.
Let there be a consumer gc and a provider gp belonging to their respective group
leaders Gc and Gp. A path between them is made of subpaths gc Gc , Gc Gp and
Gp gp. Communications can occur on both the ways on these paths. These
communications can be broadly classified as intra and inter group communications.
Further, we can call messages forwarded by a group leader to a group member as a
delivery message and all other communications as forwarding messages (Fig. 2,
Fig. 3), from the security point of view we need to secure all inter group
communications. The protocol consists of a service finding stage and then the
negotiation stage. The first stage that is the service finding stage occurs once, the
negotiation stage occurs several times until the negotiation process terminates. The
negotiation stage [2]can be further divided into the offer making and the offer
 Secure Web Services Negotiation 613

Fig. 3. Negotiation phase

generation stages. Offer making stage at a party involves the process by which an
offer is submitted to the other party. The offer generation stage involves creating a
counter offer based on a previously received offer. Now, we discuss the protocol by
giving a detailed explanation for each stage. Its to be noted that here we are neither
interested in offer making and counter offer generation strategies. Rather, our interest
includes ensuring anonymity in negotiation.
Service Finding Phase: (Fig:2) Assuming a consumer gc looking for a ser-vice, it
sends a ServiceRequest message to UDDI, with the service name as a message
parameter. We assume that each in SOA uses a common ontology; this assumption
doesn't dilute the topic in the discussion. On receiving the message, the UDDI
searches its repository and replies giving the details of the best service provider
through a message called ServiceReply [Table 1]. The message contains information
on service provider's index i, address of the service providers group leader, service
description with set of feature names and associated initial feature values < f, v >. It is
to be noted that the service details are provided by the service provider when it
registers at the UDDI. This process also occurs with the group leader as the
intermediate node. When the service request is forwarded by the group leader it
appends its index value to it, this allows the UDDI to identify that the service request
is from a group member. but as communication is secure the group leader also has no
idea about the contents of the messages exchanged, thus the group leader is oblivious
about the nature of the service requested. Also the UDDI only entertains service
requests routed through a leader. This ensures that only group members participate in
the web service protocols.
Negotiation Phase: once serviceFinding is completed negotiation is carried by the
exchange of offers and counter offers. A negotiation is shown pictorially in Fig:3.
Offer Making: A party that wants to make an offer sends a negotiation message M to
its group leader. Along with M it also forwards a set of k dummy messages that
614 V. Mohanty and C. Hota

constitute a buffer traffic to hide the negotiation message. The attributes and
construction of these dummy messages is discussed in detail in a Table1. The
message holds the values of the service attributes, address of the destination, reply
address and negotiation status. The status is used to convey to the other party the
status of the ongoing negotiation. On receiving the message the group leader accesses
the destination address and forwards it to the group leader of the destination cluster.(
Destination address contains address of the destination clusters group leader and the
destinations index value. A reply address, of the other end of the negotiating party,
also has a similar structure). When the group leader of the destination cluster receives
the message it access the destination address referring to destination index; And then
forwards.
Offer generation phase: On receiving an offer the party first checks the status field
which shows whether the other party has accepted or rejected the offer, or wants to
continue negotiation. If offer is accepted or rejected negotiation is terminated. If the
party at other end, intends to continue negotiation then negotiation utility value is
calculated. The utility value is calculated from the values it receives for various
attributes in the negotiation. message. Base on utility value, it then accepts or rejects
or decides to continue the negotiation [2] Once this decision is made the party makes
a negotiation message by assigning appropriate values to all the fields. The message is
then sent to the address that is provided in the return address field of the offer
previously received. Hence the process once again enters the offer making stage. A
recursion of offer making and offer generation stage compose the negotiation process.
How the above protocol is able to provide data security and anonymity has been
analysed in next section. An algorithm for the protocol is given at Algorithm 1 and
notations used in it are described in Table 2.

Table 2. Notations
 Secure Web Services Negotiation 615

3.1 Generation of Dummy Messages

When offers are being exchanged during the negotiation phase of the protocol the
sender creates traffic of dummy messages to hide the negotiation message. This
section deals with the generation of the messages according to the properties we
define for them. When creating the dummy messages the following attributes of the
dummy messages are considered:
1. The messages are all of the same length as the negotiation message and have
the same number of fields. The type of the fields and their values are
however different.
616 V. Mohanty and C. Hota

2. The reply address and the destination address of all the dummy messages are
the same as that of the negotiation message.
3. The value in status field is immaterial and is filled with a random value.
The dummy messages are created by varying the combination of field types. If a
message has F fields and we have X data types, we can form FX - 1 different dummy
messages as there are X possible types for each field. A random value generator is
used to generate the values required to populate the fields.

4 Analysis
In order to analyze the protocol we define adversary models. Using these models we
predict the extent to which the protocol is able to provide security. The models are:
1. Passive monitoring: Adversary monitors and makes logs of the traffic.
Traffic analysis allows the adversary to find patterns in the data exchanged
and makes a guess on the data exchanged and on the communicating parties.
2. Active attacker: These adversaries are capable of modifying the messages or
disrupting the proper flow of messages by adding or withholding messages.
3. Cooperative attack: Cooperative attack is an attack where two or more
adversaries collect and share data among themselves to draw conclusions
regarding the identity of the negotiating parties, subject of negotiation and
other details.
As we are not using cryptography we hide data by decentralizing it. Service details
are known only to the UDDI and the negotiating parties. The indexing structure of a
group is known only to the group leader. Hence effectively no single entity has all the
information regarding a negotiation process.
Theorem 1: The probability of identifying negotiating parties is inversely
proportional to sizes of their respective groups.
The proof of the theorem follows from this discussion: Each group is indexed in range
1…Kmax. Say a consumer and a provider with indices i and j respectively are to be
traced from the indices of groups of provider and consumer. Assuming Ps and Cs are
the set of members of provider and consumer groups have respectively, the
probability to trace negotiating parties could be .This can only help if an
attacker can succeed compromise of the group leaders as they only have the identities
of all the group members.

Lemma 1: The probability of tracing a negotiation message is

The lemma follows directly as both the parties transact k dummy messages as
discussed in previous section. And the dummy messages can be sampled from FX
types, making distinguishing the real one pretty difficult.

Theorem 2. Proposed protocol provides both anonymity of negotiating par-ties and

negotiation contents.
 Secure Web Services Negotiation 617

Theorem 1 says of possible maintenance of anonymity of negotiating parties. And

Lemma 1 says the difficulty in tracing the message taking part in negotiation. Further,
not providing field names of a negotiation message, for an attacker understanding of
message context becomes difficult. Now we will discuss of the cases for different
types of attackers. Though a passive attacker who logs transacted messages may have
some time to probe but also it needs service details that are available at UDDI only.
Now, such an attacker needs both group leaders as well as UDDI compromised which
is highly probabilistic. Active attackers can add, subtract, substitute or modify a
message in the traffic. If it either add or delete message contents, than the message
format gets disturbed and this alerts the parties on possible active attack. If it changes
some values in message keeping the message format intact, then negotiation process
may get affected i.e either it is abandoned or false negotiation may set in. In case of
the former, there is loss of secrecy except the time both have spent in negotiating. For
the later case, manipulation of exact field values to force negotiation is highly
unlikely as the negotiation message format is not available in the body transacted
negotiation messages. Even if it happens, this can be identified in contracting stage
when both the parties sign service agreement. Cooperative attackers at both group
leaders can only be aware of occurrence of a negotiation but will have difficulty in
identifying negotiating parties and understanding the contents of negotiation messages
as service details are only known to both the negotiating parties and they are
anonymous in their respective groups.

5 Conclusion
Our work has suggested a secure method for negotiation mostly based on anonymity.
In order to gain time, we have not considered encryption. We argue that many day to
day business transaction needs high response time and are also not so secret. For that
kind of business purpose, the proposed protocol is workable though it needs to
transact some unwanted messages. The analysis indicates that the protocol is secure to
all attacks models we have discussed. However the negotiation is corrupted when an
active adversary modifies the contents of the negotiating packet. Though the
probability of success of the attack is low it cannot be completely avoided. Our future
work is aimed at rectifying this problem. We also seek to simulate the protocol and
attain experimental proof for the results we have predicted during the analysis of the
protocol. We also look to make a comparative study between the protocol we have
suggested and the security protocols generally used now in the business domain.

References
1. Shannon, C.E.: Communication Theory for Secrecy Systems confidential report. A
Mathematical Theory of Cryptography, September 1 (1946)
2. Mohanty, H., Kurra, R., Shymasundar, R.K.: A Framework for Web Based Negotiation.
In: Janowski, T., Mohanty, H. (eds.) ICDCIT 2010. LNCS, vol. 5966, pp. 140–151.
Springer, Heidelberg (2010)
3. Ren, J., Wu, J.: Survey on Anonymous Communication in Computer Networks. Computer
Communication 33(4), 420–431 (2010)
618 V. Mohanty and C. Hota

4. Byun, J.-W., Kamra, A., Bertino, E., Li, N.: Efficient K-Anonymity Using Clustering
Techniques. In: Kotagiri, R., Radha Krishna, P., Mohania, M., Nantajeewarawat, E. (eds.)
DASFAA 2007. LNCS, vol. 4443, pp. 188–200. Springer, Heidelberg (2007)
5. Sweeney, L.: K-Anonymity: A Model For Protecting Privacy. International Journal on
Uncertainty, Fuziness and Knowledge-Based Systems 10(5), 557–570 (2002)
 A Recommender System for Rural and Urban Learners

L. Arockiam1, S. Charles2, V. Arul Kumar3, and P. Cijo4

1
Associate Professor
2
Assistant Professor
3
Research Scholar
4
Student
Department of Computer Science,
St Joseph’s College, Tiruchirappalli

Abstract. In Educational data mining, learning process evaluation reveals

the students’ involvement in learning. It leads to improve the quality and
success rate of the students in educational environment. Students’ intellectual
performance varies in learning environment and evaluated by different criteria.
This study helps to find out the intelligent quotient level of the student and
their performance, which are dissimilar in a classroom. Stanford Binet
Intelligence Scale and Criterion Reference Model are used to evaluate the
intelligence and performance of the students. Machine learning techniques are
employed to find the intelligent quotient and performance of rural and urban
students using students’ dataset. Model based clustering technique is used
to determine the similarity in the students’ dataset based on the nature of the
intelligent quotient and performance. Each cluster reveals the identity based on
students intelligence level. The performance also categorized on descriptive
modeling. Multilayer Perceptron technique classifies the intelligent level of the
students and their performance. The study determines the association
between intelligent level of the rural and urban students and performance
through descriptive and predictive modeling. This analysis recommends the
teaching community to provide right training to the rural students for their
improvement of academic competence.

Keywords: Multilayer Perceptron, Expectation Maximization (EM) Clustering,

Criterion Reference Model, Stanford Binet-Intelligence Scale, Intelligent
Quotient level.

1 Introduction

Educational data mining is an emerging area of research in current scenario. It

identifies three objectives such as educational, management and commercial. This
paper deals with educational objectives, which determines the students IQ level in the
college environment and their performance. Stanford-Binet Intelligence scale has
been used for identifying the intelligence of the rural and urban students in post
graduate course. This scale helps to measure the cognitive abilities of the same age
group of rural and urban students. Criterion – Reference Model is used to
categorize the performance of the students in their studies. The theory and

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 619–627, 2011.
© Springer-Verlag Berlin Heidelberg 2011
620 L. Arockiam et al.

programming marks can be viewed as performance. Cluster analysis is an

unsupervised learning method that is used to investigate the inter-relationships among
a collection of student objects, by organizing them into homogeneous clusters. EM
clustering technique is used to group the student objects based on the intelligent
level into classes of similar students ’objects [1]. Similarly the student performances
are classified as good or fair.
The MLP is used for subcategory classification in the students’ dataset. It is one of
the approaches for discerning a pattern of subcategories. It estimates the useful
information through mining and determines the interesting patterns of the specific
category. Experimental unit variables are used for finding patterns in the students’
dataset. An observational unit variable finds the subcategory of the intelligent level.
Similarly it classifies the performance of students. This paper focuses on
deriving association between intelligent Quotient of rural and urban students and
their performance based on supervised and unsupervised learning process.

2 Motivation

2.1 Stanford Binet-Intelligence Scale

The Stanford-Binet intelligence scale is a test, which assesses intelligence and

cognitive abilities of the age group between 2 and 80. It tests the intelligence
Quotient level (IQL) of the student using the following criteria: general intelligence,
knowledge, fluid reasoning, quantitative reasoning, visual-spatial processing, and
working memory. These criteria are divided by ten subtests that include activities
measuring both verbal and non-verbal intelligence. Activities include verbal
absurdities, picture absurdities, verbal analogies, form patterns, procedural
knowledge, sentence and word memory, position and direction, early reasoning, and
quantitative reasoning. The test score is categorized as 130 and Above: Very High
(VH), 120 – 129: High (H), 110 – 119: Above average (AA), 90 – 109: Average(A),
80-89: Below Average(BA), 70 – 79: Low and 60 or below: Deficient(D) [2,9]
The problem of discrimination and intelligent scale classification is mainly
concerned with differentiating between g (g>=2) mutually exclusive students and
with classifying intelligent level on the basis of Multivariate observations]. The
mathematical problem of discrimination is to divide the observation space RP
into g mutually exclusive and exhaustive regions R1…Rg so that if an observation
space x falls in Ri then it is classified as belonging to the population. The sub
category problem is a mutually exclusive population according to their attributes
and classifying an unseen instance into its sub categories using multiple features.
In this study, the sub-categories are experimental unit variables and observed
features are observational unit variables[3, 4].
Intelligence is characterized based on abstractness and correctness of level of
intelligent the students’ posses. The proposed model determines the intelligent level
and is able to characterize, how the mind functions while learning.
 A Recommender System for Rural and Urban Learners 621

2.2 Criterion-Referenced Model

A criterion-referenced model is an assessment model where the purpose is to assess
the extent to which a student has achieved the goals of a course. The summative
assessment is carried out aligned with specified criteria. Results are expressed in
terms of how well a given student’s performance matches set criteria. The standards
are set before teaching takes place. The result is assigned on the basis of the standard
that the student has achieved on the criteria [5].

2.3 Expectation Maximization (EM) Clustering

The EM algorithm is a popular iterative refinement algorithm that can be used for
finding the parameter estimates. It assigns an object to the cluster with which it is most
similar based on the cluster mean. Instead of assigning each object to a dedicated
cluster, EM assigns each object to a cluster according to a weight representing the
probability of membership. There are no strict boundaries between clusters. Therefore
new means are computed based on weighted measures [6].

2.4 Multi-Layer Perceptron

The Multilayer Perceptron is one of the techniques for classification task. It is a
network of simple neurons called perceptrons. The perceptron computes a single
output from multiple real-valued inputs by forming a linear combination according to
its input weights and then possibly putting the output through some nonlinear
activation function [7].

3 DAIP Approach
In Phase I, the research questions are designed by the expert based on Stanford Binet
Intelligence Scale and pretest the questionnaire. After pretesting a questionnaire, the
Intelligence Scale analyze test has been conducted in Post Graduate Course. The
written and practical marks are obtained from the semester exam, which are viewed
as performance (good (Marks>=60) or fair (Marks<60)) [5, 8, 10] based on
Criterion- Referenced Model. The students’ dataset contains the test results of
Stanford Binet
Intelligence scale and performance of the students in the written and practical
test in programming subject.
In Phase II and III, two data mining methods are employed to find the association
between the IQ level and performance. The methods are EM- clustering and Multi
Layer Perceptron respectively. Each cluster reveals the identity of the students’ IQ level
and designates the cluster as 1) High 2) Above Average 3) Average 4) Below Average
5) Low 6) Deficient. The same method is employed for students’ performance based
clustering [11, 12 and 13]. The MLP technique is used to classify the Stanford Binet
Intelligence level of the students. The five sigmoid nodes are used as inputs, the weights
are assigned to each node, and six output layers are classified based on the category of
intelligent level. According to the performance-based classification, the five sigmoid
nodes are used as input and two output layers are classified based on the skills as good
or fair [14]. In Phase IV, the association between intelligent level and performance of
the students are analyzed using interesting measures.
622 L. Arockiam et al.

'HVLJQLQJWKHTXHVWLRQQDLUHIRU 3KDVH,
6WDQIRUG±%LQHW,QWHOOLJHQFH

3UHWHVWWKH ,QWHOOLJHQFH6FDOH 6WXGHQW3HUIRUPDQFH$QDO\VLV

4XHVWLRQQDLUH $QDO\VLV 3UDFWLFDO 7KHRU\0DUNV

5HVXOW'DWDVHW

&ODVVLILFDWLRQ&OXVWHULQJ &ODVVLILFDWLRQ&OXVWHULQJ
,QWHOOLJHQFH6FDOH 3UDFWLFDO 7KHRU\0DUNV

3KDVH,9 3KDVH,,,
3KDVH,,

$VVRFLDWLRQEHWZHHQ,QWHOOLJHQFH
/HYHODQG3HUIRUPDQFH
5HFRPPHQGDWLRQV

Fig. 1. DAIP Approach

4 Results and Discussions

The dataset containing 327 students’ objects are used to find the association between
intelligent level and their performance of the students.

Table 1. Cluster assignments and Intelligent Level

C0 C1 C2 C3 C4 IQL
56 6 36 6 10 A
6 0 24 0 0 L
17 0 6 0 0 AA
12 6 35 6 0 BA
36 24 12 29 0 D

Table 2. Confusion Matrix for Intelligent level

3UHGLFWHG
&URVVYDOLGDWLRQ
D % F G H ,QWHOOLJHQW/HYHO
 $      $YHUDJH
 %      /RZ
$FWXDO &      $ERYH$YHUDJH
'      %HORZ$YHUDJH
(      'HILFLHQW
 A Recommender System for Rural and Urban Learners 623

Table 1 shows that various clusters and corresponding IL are categorized. It

reveals that Average IQ level of student objects are silhouetted in cluster 0, Low IQ
level student objects are silhouetted in cluster 1, Above Average level student objects
are silhouetted in cluster 2, Below IQ level of student objects are silhouetted in
cluster 3 and Deficient IQ level of student objects are silhouetted in cluster 4.
Similarly the MLP classifies the students’ objects with a support of the class variable
intelligent level. Table 2 reveals that the various classes of observed variables as
Intelligent Level and experimental variables as demographic information of objects
and theory and practical examination mark. The confusion matrix reveals that
various classes using six sigmoid nodes to identify the six IQ levels.

Table 3. Intelligent Level Preference of Student

Intelligent Level Response Agreement %

Above Average 23 7
Average 114 35
Below Average 59 18
Low 30 9
Deficient 101 31

The intelligent level preferences and the percentage of respondents who agreed
with each item are shown in Table 3. Table 3 clearly indicates that the intelligent
level, which majority of Post Graduate Students posses. Approximately 7% of
the students have viewed themselves as above average IQ level. 35% of the
students are in average IQ level, 18% of the students are in below average IQ
level, 9% of the students are in Low IQ level, 31% of the students are in Deficient,
who posses with the different IQ level.

Table 4. Confusion Matrix for theory and practical performance

Validation Data
Predicted Predicted
Actual Category Category(Theory) Category(Practical)
G F G F
G 154 0 303 0
F 0 173 4 20
Mean absolute error 0.0054 0.151228
Root mean squared
0.0158 0.128171
error

Table 4 shows that the categorization of theory and practical performance of

students using the experiment variables. This experiment reveals that 154 student
objects are observed as good in theory subject, 173 students’ objects are observed as
fair in theory. Similarly the Practical marks are categorized using clustering. 303
student objects are observed as good in Practical lab, 24 students’ objects are
observed as fair in Practical lab.
624 L. Arockiam et al.

 

Fig. 2. Classification of theory performance of Fig. 3. Classification of Programming

Rural and Urban Students Lab Performance of Rural and Urban
Students

Fig 2 reveals the classification of theory performance of rural and urban

students, 127 (39%) rural students are observed as good in theory subject, 31(9%)
rural students are observed as fair in theory subject, 133(41%) urban students are
observed as good in theory subject, 36(11%) urban students are observed as fair
in theory subject. Fig 3 reveals the classification of practical lab performance of
rural and urban students, 145 (44%) rural students are observed as good in practical
lab, 13 (4%) rural students are observed as fair in practical lab. 157 (48%) urban
students are observed as good in practical lab, 12 (4%) urban students are observed
as fair in practical lab.

Table 5. Intelligent Level Preference of rural and urban students

AA A BA L D
Area/IL (%) (%) (%) (%) (%)
Urban 5 22 6 2 18
Rural 2 13 13 7 12


Fig. 4. Classification of IQ Level of the Rural and Urban Students

The urban and rural student objects are classified based on the intelligent level and
Performance, 5% of urban students are in above average level, 22% of urban students
are in average level, 6% of urban student are in low level, 18% of urban
students are deficient level. 2% of rural student are in above average level, 13% of
 A Recommender System for Rural and Urban Learners 625

rural students are in average level, 13% of rural students in below average level, 7 %
of rural students in Low level and 13% of rural students in Deficient level. It is
evident that the urban students’ intelligent levels are high compared to rural students
Fig 5 shows that the probability of students having good performance in theory
and practical. The probability calibration chart is used for classification analysis to
find the accuracy. The horizontal axis reveals the predicted probability of
observations. The vertical axis reveals the actual probability of occurrence. The
average predicted probability is calculated, which is lies between 0.62 and 0.65 and
the actual probability is 0.6. The red shaded area reveals the difference between the
predicted and actual probabilities. The average predicted probability error is 0.1056.
The weighted probability error value is low. These error values divulge that the model
is perfect.




Fig. 5. Probability Calibration chart for Good Fig. 6. ROC chart for Good Performance in
Performance in Theory and Practical Theory and Practical

Fig. 6. Sensitivity and Specificity curve for Theory and Practical performance

Fig.6 shows Receiver Operating Characteristic chart and probability threshold

values are elevated. It reveals that the urban students’ performance is good in theory
and practical. Sensitivity and Specificity curve values are equal. Fig.7 shows the
sensitivity curve divulges that the urban and rural students having different IQ levels
are good in Theory and Practical lab. The mapping or function can be formulated as
Y = f (X), where X is the input and the output as Y. This function reveals the
relation between X and Y. The lift measure divulges that correlated value of X
and Y are 1.22, which is positively correlated. It means that the X and Y are
626 L. Arockiam et al.

related with each other. The cosine value of X and Y are 0.712. It reveals that,
positive correlation is existed in X and Y. The observation divulges that IQ has a
strong association with performance.

5 Conclusion
In this Paper, Model based clustering technique and Multi-Layer perceptron are used
to find the association between intelligent quotient and performance. The clustering is
an effective means for classifying the student objects such as Very High, High, above
average, Average, Below Average, Low and Deficient and models each cluster. It reveals
the identity of the IQ level of the students. The performance based clusters are
observed as good and fair. The classifier reveals that, each individual student’s
IQ level related with the performance. More number of urban students having an
average IQ level compared to rural students. Few urban students having an Above
Average IQ level compared to rural students and few rural students having a
below average IQ level compared to urban students. Similarly few rural students
having low IQ level compared to urban students. IQ score gives us a measure of a
students’ academic competence such as reading, learning, problem solving, abstract
reasoning and conceptual learning. This experiment reveals the academic
competence of rural students and provides the recommendation to the rural
learners for improving their skills. The rural learners are need to be improved
their academic competence by taking enough training in various skills. The
correlation analysis divulges that the Intelligent Quotient and performance have
strong association.

References
[1] Khalilian, M., Boroujeni, F.Z., Mustapha, N., Sulaiman, M.N.: K- Means Divide and
Conquer Clustering. In: International Conference on Computer and Automation
Engineering, pp. 306–309. IEEE Computer Society, Los Alamitos (2009)
[2] Becker, K.A.: History of the Stanford-Binet Intelligence Scales: Content and
Psychometrics (Stanford-Binet Intelligence Scales, Fifth Edition Assessment Service
Bulletin No. 1). Riverside Publishing, Itasca (2003)
[3] Choi, S.C., Rodin, E.Y.: Statistical Methods of Discrimination and Classification. In:
Advances in Theory and Applications. Pergamon Press, Oxford (1986)
[4] Duda, R.O., Hart, P.E., Stork, D.G.: Pattern Classification, 2nd edn. John. Wiley & sons
Inc., Chichester (2000)
[5] Khamis, N., Idris, S.: Issues and Solutions in Assessing Object-oriented programming
Skills in the Core Education of Computer Science and Information Technology. In: 12th
WSEAS International Conference on Computers, Heraklion, Greece, pp. 23–25 (2008)
[6] Wu, X., Kumar, V., Quinlan, J.R., Ghosh, J., Yang, Q., Motoda, H., McLachlan, G.J., Ng,
A., Liu, B., Yu, P.S., Zhou, Z.-H., Steinbach, M., Hand, D.J., Steinberg, D.: Top 10
algorithms in data mining. Knowl. Inf. Syst. 14, 1–37 (2008)
[7] Ayers, E., Nugent, R., Dean, N.: Skill Set Profile Clustering Based on Student Capability
Vectors Computed from Online Tutoring Data. In: Baker, R.S.J.D., Barnes, T., Beck, J.E.
(eds.) Proceedings of 1st International Conference on Educational Data Mining, Montreal,
Quebec, Canada, June 20-21, pp. 210–217 (2008)
 A Recommender System for Rural and Urban Learners 627

[8] Pavlik Jr., P.I., Cen, H., Wu, L., Koedinger, K.R.: Using Item-type Performance
Covariance to Improve the Skill Model of an Existing Tutor. In: Proceedings 1st
International Conference on Educational Data Mining, Canada, June 20-21, pp. 77–86
(2008)
[9] Green, T.M., Jeong, D.H., Fisher, B.: Using Personality Factors to Predict Interface
Learning Performance. In: 43rd Hawaii International Conference on System Sciences.
IEEE Computer Society, Honolulu, HI, January 5-8, pp. 1–10. IEEE Computer Society,
Los Alamitos (2010)
[10] Marshall, L., Austin, M.: The relationship between software skills and subject specific
knowledge, theory and practice. Learning and Teaching Projects (March 2002)
[11] McCracken, M., Almstrum, V., Diaz, D., Guzdial, M., Hagen, D., Kollokant, Y., Lazer,
C., Thomas, L.A., Utting, I., Wilusz, T.: A Luti national Study of Assessment of
Programming Skills of First Year CS students. SIGC, SE Bulletin 33, 125–140 (2001)
[12] Chiu, C.: Cluster Analysis for Cognitive Diagnosis: Theory and Applications. Ph.D.
Dissertation, Educational Psychology, University of Illinois at Urbana Champaign (2008)
[13] Ayers, E., Nugent, R., Dean, N.: A Comparison of Student Skill Knowledge Estimates
Educational Data Mining. In: 2nd International Conference on Educational Data Mining,
Cordoba, Spain, July 1-3, pp. 1–10 (2009)
[14] Nghe, N.T., Janecek, P., Haddawy, P.: A Comparative Analysis of Techniques for
Predicting Academic Performance. Paper presented at 37th ASEE/IEEE Frontiers in
Education Conference, Milwaukee, WI, October 10-13 (2007)
[15] Han, J., Kamber, M.: Data Mining Concepts and Techniques, 2nd edn. Morgan Kaufmann
Publishers, San Francisco (2006)
 Study of Organizational Factors Affecting Usability of
Online Helps with Programming Languages

Nilesh Dhannaseth1 and Ganesh Bhutkar2

1
Research Scholar,
Department of Computer Engineering,
Vishwakarma Institute of Technoogy, Pune
[email protected]
2
Assistant Professor,
Department of Computer Engineering,
Vishwakarma Institute of Technoogy, Pune.
[email protected]

Abstract. One of the most common resources for assistance to users while
dealing with programming languages is online help, which is topic-oriented,
procedural or reference information delivered through online software. This
research paper reports observations about organizational factors with online
helps for programming languages. These observations are viz. Organized Index
of Content, Hyperlinks, Grouping of Items, Use of Tables / Charts / figures,
Scrolling Effect, Cascading Effect, Multilingual Support and Help in Video
Format. They are compiled based on authors’ experience, discussions with
active users and literature study about online helps. The users with
programming languages can be classified as tool users, programmers,
developers and architects. We find that organized index, hyperlinks, grouping
of items and use of tables / charts / figures are important for all users with
programming languages. These organizational factors directly affect the
usability of online helps with programming languages. They should be
considered by designers / developers of such online helps so that their users are
encouraged in efficient usage and referencing while dealing with products /
systems like programming languages.

Keywords: Usability, Online Helps, Organizational Factors, Programming

Languages, Programmers.

1 Introduction
As our world is becoming more and more technology-oriented with mounting
operational complexity, the volume of information generated and used is increasing
tremendously. Today, it is estimated that there are approximately 150 websites for
every human being on the earth [11]. Many of these websites support complex online
systems, which have various advantages. The online systems are excellent for brief
descriptions, quick references or even extended explanations. The online systems are
quicker and also provide the ecological benefits like lesser usage and/or wastage of

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 628–637, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Study of Organizational Factors Affecting Usability of Online Helps 629

paper. The other advantages of online systems include available all time, easier
transfer, access, maintenance and updation of information. While dealing with online
systems, the users always require some assistance, which can be given by online
consultants, e-learning systems, online forums or helps [16].

1.1 Online Helps of Programming Languages and Its Users

Online help is a popular system offering assistance and support to online systems. It is
topic-oriented, procedural or reference information delivered through online software.
It can be used to present information on a broad range of subjects. It is required with
commercial, office, home, creative and/or cooperative applications, which may
include military, banking, healthcare or reservation systems and products like
antivirus, operating systems, web browsers or programming languages [16].
Today, hundreds of programming languages have been utilized and more
languages will be developed in future for desired application development. Due to
generalization as well as specialization in application development, the complexity of
programming languages has increased drastically. With the continuing increase in the
variety, functionality and complexity of languages with more widespread use, the
usability of programming languages has become an important issue. The online helps
of programming languages assist users by providing needed assistance and in
acquiring necessary expertise. We have selected programming languages for study of
online helps as most users learn programming languages through online helps and
complain about their usability. For our study, we have selected online helps of Visual
C++ 2010, JAVA SE (Standard Edition) version 6 and HTML version 4.01
provided by vendors - Microsoft Corporation, Oracle-Sun and W3schools
respectively. Their logos are shown in figure 1.

Fig. 1. Selected online helps of Visual C++, HTML and JAVA

The users with programming languages differ in terms of needs, levels of expertise,
usage of tools, time spent online, domain knowledge and/or work experience. These
users can be classified as tool users, programmers, architects and developers as
described below [6]:

Tool users: These users are naive or least experienced and spend more time with
development tools. They have enough knowledge to look through tools like
Dreamweaver, NetBeans and Ajax (Oracle.com). They also utilize GUI simply by
dragging and dropping elements onto a form.

Programmers: They have a good understanding of coding and writing software.

They rely less on tools and prefer writing programs according to the specifications
determined/given. They convert software designs into a logical series of instructions.
630 N. Dhannaseth and G. Bhutkar

Architects: Architects are well-experienced software developers. As they have

domain knowledge, they can think well beyond just technical side of software
development. They identify the appropriate software architecture based on the
requirements and design elements contained in System Requirement Specifications
(SRS). They record design in a software architecture document using use cases and
the Unified Modeling Language (UML).
Developers: These users are expert programmers who understand that there is much
more in developing software than just writing programs. Developers possess a much
better perceptive of concepts and principles of software design. They think about
problems in whole. They need assistance from high-end business tools like
LightSwitch and Eclipse.
Online helps may be studied for various usability aspects viz. accessibility,
consistency, textual elements, visual elements, readability, organization and
navigation. The content of any online help is vital to a success of online system. High
quality content will increase the likelihood of system utilization. But beyond
providing high quality content, online helps also need to organize their content in a
way that makes it accessible, readable and valuable to the users. Organizing
information works well for online helps where there is variety of users. Users of
online helps often experience difficulties when trying to retrieve reference topics.
They have to face many problems like availability of information, textual complexity,
poor layout and language support in the organization of online helps [4]. Online helps
may not offer users a support in navigation through the content due to poor
organization. Taking these problems into consideration, we have studied and collected
several observations with organizational factors in selected online helps.

2 Observations about Organizational Factors in Online Helps

The observations about organizational factors in online helps are compiled based on
authors’ experience, discussions with active users and study of literature about online
helps. Color coding is used in most of the figures in this section for differentiation
among usability features. The positive features in online helps are highlighted with
enclosure in rectangles using green color whereas negative features are in rectangles
using red color. Also, the blue color indicates both positive and negative features.

2.1 Organized Index of Content

An organized index is a list of items in specifically planned order in which items have
to be learnt or referred. In the case of online help, the order of topics is important for
understanding of the users. A well-organized index page is sequentially ordered and is
used as a navigation tool to reach the required topic of interest. Due to rapidly
increasing web pages, it is necessary to include a well-organized index in online help.
An organized index is located preferably in the upper-left quadrant of every web page
of the online help [7]. The index usually includes the titles of the first-level header
such as chapter title and often includes second-level or section titles within the
chapter. The depth of details in an index depends on the availability of the structured
 Study of Organizational Factors Affecting Usability of Online Helps 631

or classified content in online help. A hierarchy containing multi-level topics is used

to link entities either vertically or horizontally.
An organized index available with online help of Visual C++ is given in figure 2
and depicts a sequence of topics to be learned. Figure 2(A) shows the first-level
organization of topics that represents the main content of Visual C++. Figure 2(B)
shows the second-level organization of topics, which represents contents in-depth.
Online help of Visual C++ uses color coding in the index structure of topics. The
color used in this help is blue for unvisited links, orange for links which are in use and
pink for those visited by the user. The first level topics as shown in figure 2(A) have
followed 7+/-2 rule, but second level topics do not follow it [10]. The index arranged
in online help of Visual C++ is in hierarchical fashion and located in upper-left
quadrant of each web page.

A B

Fig. 2. Organized index of online help for Visual C++

2.2 Hyperlinks

A hyperlink is an underlined word or phrase and sometimes, a graphical image or icon

which has been tagged with a particular scripting language command containing the
specific address of another location. By using hyperlink, the user jumps quickly from
one location to another location [3]. Hyperlinks are required in online help to present
a chance for user to stop reading at one place and resume somewhere else. Hyperlinks
are also used to fragment or divide large information into small chunks. The concise
chunks of information i.e. hyperlinks provide a limited view of large documents.
Studies report that about 80% of users scan any new page and only 16% users read
each word. Users spend about 12% of their time trying to locate desired information
on a page [15]. So, hyperlinks are vital elements in online helps. There are four types
of hyperlinks viz. text, image, bookmark and e-mail. The color used for hyperlinks are
blue for unvisited links and red for visited links.

2.3 Grouping of Items

A grouping means all related details are placed together in order to minimize time sp-
nt on searching or scanning of whole web page. With the rapidly increasing
632 N. Dhannaseth and G. Bhutkar

information on the web, there is a great demand for developing methods for
effectively organizing the large amount of relevant details [16]. An appropriate
grouping strategy can help in this regard and may lead to providing useful assistance
for user in getting all information related to a topic. There are various types of
grouping of items such as ‘under one heading’ and ‘closeness by distance’ [8]. Text
items that share the ‘same background color’ typically can be seen as being related to
each other. The information retrieval time is more in case of ungrouped items
provided on web pages [15]. So, online helps should be designed using a suitable
grouping strategy of items.

A B

Fig. 3. Sample screen shots of A) Grouping of editions of JAVA help and B) Download option
for various versions of Java

Figure 3 shows the groupings used in online help of JAVA. The first grouping
depicted in figure 3 (A) uses ‘under one heading’ strategy for categorization of
various editions of Java. Another strategy based on ‘same background color’ for
grouping is shown in figure 3 (B) for popular downloads of various versions of Java.

2.4 Use of Tables / Charts / Figures

A table is a way of representing details like numbers, quantities or items in terms of
rows and columns.
Tables are designed to structure and present details in a systematic fashion. A chart
is a visual representation of numeric values, quantities or items with identified
patterns. A table / chart effectively use a minimum space to communicate a large
amount of details in a small space [17]. Data in tabular form takes only one third of
space to represent information which is in text format [5]. According to Horton
(1991), well-designed graphics used in charts help to structure data and enable the
users to grasp and process information more quickly and efficiently which leads to
increased information retention. There are two types of tables - unidirectional and
bidirectional tables. Unidirectional table is a table in which all details displayed row
wise and bidirectional table is the table whose details are depend on both rows and
columns. There are chart types like bar chart, pie and line chart for representing
information about online helps.
 Study of Organizational Factors Affecting Usability of Online Helps 633

A B

Fig. 4. Screen shots for A) Table for functions used in HTML scripting and B) Color chart used
in HTML scripting shown in online help at www.w3schools.com

Figure 4 (A) shows the unidirectional table for functions used in HTML scripting.
This table is unidirectional because it has only row wise description of HTML
functions. Figure 4 (B) shows the color chart that illustrates shade of colors and their
indices which are used in HTML scripting.

2.5 Scrolling Effect

A scrolling is sliding text, image or video across a display. It does not change the
layout of the text or image. But, it incrementally moves the user's view across what is
apparently a larger text or image that is not wholly seen [14]. Since documents are
generally too long to fit on a screen, users must frequently scroll to see other parts of
the documents. There are two types of scrollings - vertical and horizontal. Sometimes,
both scrollings are provided for user interaction to move viewport in two dimensions.
Most of the users avoid scrolling web pages of online help and therefore, do not refer
information which is present below the fold. In a study, it is found that only 10% of
web users would scroll web page to see any links that are not visible in the initial
display [12]. Generally, the online helps should provide very little scrolling in the web
page not to miss out reference topics which are of importance.

2.6 Cascading Effect

A cascading effect is an effect in which users may have to click hyperlinks more
times to get appropriate details. The process of cascading is really discouraging for
most users [2]. In a study, it is found that when users try to find a reference topic, they
are likely to quit after three clicks [13]. So, the designer of online help should design
help in such a way that the users get required topic of their choice in less than three
clicks. This will save valuable time of users and they don’t have to jump in web pages
of online help to get required details.
Figure 5 shows the graph plotted for number of clicks required to get reference
topics in respective online helps for programming languages under consideration. We
have taken number of clicks on Y-axis and reference topics on X-axis. The reference
634 N. Dhannaseth and G. Bhutkar

topics considered in graph are programming function (e.g. function to set font),
bitwise operator (e.g. ‘&’ operator) and class in language (e.g. basic I/O class). Now,
we consider the example of basic input-output (I/O) class. In general, the online help
of HTML requires about three clicks (average) to get basic I/O function for the users.
But, in case of other helps, number of clicks required is always more than three and it
may consume vital time in search/navigation of information by the users.

Cascading Effect
10
Number of Clicks

8
6
4
2
0
Programming Bitwise Class in
Function Operator Language
Online Helps of Languages
Visual C++ HTML JAVA

Fig. 5. Graph for number of clicks required to get reference topic in online helps

2.7 Multilingual Support

A multilingual support is a property used to mean that there exist forms that can
handle several languages. An appropriate language can be selected by the user while
using multilingual online help. Every online system should provide multilingual
support for all widely-spoken languages like Chinese, Spanish, English, Arabic or
Russian [9].

Fig. 6. Multilingual support in online help for HTML

 Study of Organizational Factors Affecting Usability of Online Helps 635

It should also provide language option like Hindi, which is widely spoken in India
and several countries like Mauritius, South Africa, Singapore, Nepal, and Pakistan
etc. It is also desirable to provide support for major Indian languages like Bengali,
Marathi, Tamil, Gujarati and Malayalam. Layout and design is an important
consideration when building multilingual online helps. The problem arises due to
multilingual support in the organization of web pages of online help. This include
‘direction of text’, ‘format of things like date, time and names’, ‘choice of spelling
convention’, ‘phrases and meaning which cannot be translated’, ‘size of text block’
and ‘increase in web pages of help’. The problem of multilingual support to contents
affects not only users in multilingual regions like Europe or Indian subcontinent, but
also all those who search information sources or databases containing material in
more than one language [1]. Therefore, providing multilingual support helps making
software application designs universal.
As shown in figure 6, the online help for HTML supports total 52 languages. This
is an important feature. As observed, it provides multilingual support for all leading
languages like Chinese, English, Arabic and Russian [9]. It also provides language
option like Hindi, which is widely spoken in India. But, it fails to provide support for
other major Indian languages.

2.8 Help in Video Format

Videos can help users to understand complex concepts and procedures that are
difficult to explain with simply text and graphics. By using visual messages, users can
process the information quicker which helps to encourage their learning acquisition of
the reference topics. In addition, a video can help users to visualize a process or see
how something works. Also, users don’t have to wait for complete download of the
video. With these benefits in mind, every online help should provide video support for
users to convey information very easily, if needed. There are various video formats
like WMV, MP4, 3GP and AVI (Fileinfo.com). Due to these formats, users have to
download additional software component(s) from different websites to watch videos
related to online helps. This may increase the complexity of online help for the users.

A B
Fig. 7. A) Video support for online help of Visual C++ and B) Additional software requires for
watching videos

Figure 7 shows that the online help of Visual C++ provides help in video format to
its users. The chapter ‘Better Understand Using Pointers and References’ is difficult
to learn and complex to understand in text format. So, by using the video of same
chapter, the user can grasp the topic easily. Users have to download software
636 N. Dhannaseth and G. Bhutkar

‘Microsoft Silverlight’ to watch the videos present in this online help. But, it is not
convenient thing to download additional software.

3 Discussion
The observations about organizational factors in online helps discussed consist of 8
organizational factors which can be classified as qualitative and quantitative. The
qualitative set of factors is particularly challenging, as these factors are often described
subjectively. e. g. Organized Index of Content, Scrolling Effect, Multilingual Support etc.
The quantitative set of factors is classified as these factors are measurable.
A useful numerical method for calculation of organizational index based on
organizational factors can be provided. It has two major equations and they are very
easy to compute as all parameters required can be expressed in terms of natural
numbers. The proposed equations/formulae are provided in equations - I & II
respectively. An equation I can be used when help in video format is available;
otherwise equation II can be used. The denominator part in equation II is approximate
result of calculation of denominator of equation I, using studied online helps.

Equation I. Formula for index calculation when videos are available

Equation II. Formula for index calculation when videos are not available
The list of parameters used in formulae are {h = Avg. number of hyperlinks, t =
Number of tables / charts / figures, c = Avg. number of clicks, v = Number of videos}
The study of online helps with programming languages is carried out by
considering organizational factors affecting related online helps. The equations can be
used to comprehend about how well-organized the online helps are. This study may
assist in highlighting organizational design deficiencies in online helps to improve
usability even for tool users. Also, the proposed formulae - OI give a fast and useful
numerical method for calculation of organizational index for online helps. These
formulae also provide an opportunity for comparison among several online helps in a
specific domain. These organizational factors directly affect the usability of online
helps with programming languages and help web designers to develop such online
helps which can encourage better interaction with their users. In future, this study can
be extended to evaluate various online helps in other domains like e-commerce, e-
governance or electronic gazettes.

Acknowledgment. We sincerely acknowledge the help and support provided by post-

graduate research scholars - Mr. Tulsidas Patil and Mr. Dipak Pawar at VIT, Pune during
this research work. We thank Prof. S. B. Karthick, Head, Department of Computer
Engineering at VIT, Pune for his continuous encouragement to our research work.
 Study of Organizational Factors Affecting Usability of Online Helps 637

References
1. Biswas, S.: Multilingual Access to Information in A Networked Environment Character
Encoding and Unicode Standard. In: INFLIBNET’s Institutional Repository, 3rd
Convention PLANNER, pp. 176–186. Assam Univ., Silchar (2005)
2. Bhutkar, G., Rajhans, N., Konkani, A., Dhore, M.: Usability Issues of User Manuals
Provided with Medical Devices. The British Journal of Healthcare Computing and
Information Management (2009)
3. Chaparro, B., Bernard, M., Thomasson, R.: Finding Information on the Web: Does the
Amount of White Space Really Matter? In: Proceedings of the 10th Annual Usability
Professionals’ Association Conference, pp. 1–4 (2001)
4. Chaparro, B., Shaikh, D., Baker, J.: Reading Online Text with a Poor Layout: Is
Performance Worse? Usability News, SURL, 1–5 (2005)
5. Durbin, C., Faarc, J.: Effective Use of Tables and Figures in Abstracts, Presentations, and
Papers. Respiratory Care 49(10), 1233–1237 (2004)
6. Hayes, J.: Types of Programmer (2007), retrieved from
http://ezinearticles.com/?5-Types-of-Programmers—Which-Type-
Are-You?&id=444143
7. Horton, W.: Designing and Writing Online Documentation: Hypermedia for Self-
Supporting Products, 2nd edn. John Wiley & Sons, Chichester (1994)
8. Katre, D., Bhutkar, G., Karmarkar, S.: Usability Heuristics and Qualitative Indicators for
the Usability Evaluation of Touch Screen Ventilator Systems. In: Katre, D., Orngreen, R.,
Yammiyavar, P., Clemmensen, T. (eds.) HWID 2009. IFIP AICT, vol. 316, pp. 83–97.
Springer, Heidelberg (2010)
9. Lewis, P.: Ethnologue: Languages of the World, 16th edn. SIL International, Dallas (2009)
10. Miller, G.: The Magic Number Seven, Plus or Minus Two: Some Limits on our Capacity
for Processing Information. Psychological Review, 81–87 (1956)
11. Momento24. Number of Websites for Every Human Being (2009), retrieved from
http://momento24.com/en/2009/07/31/number-of-websites-
surpasses-world-population-
12. Nielsen, J.: Scrolling and Attention. Jakob Nielsen’s Alertbox Column (2010), retrieved
from http://www.useit.com/alertbox/scrolling-attention.html
13. Porter, J.: Testing the Three-Click Rule (2003), retrieved from
http://facstaff.elon.edu/gibson/COM322/.../Testing_the_Three_
Click_Rule.pdf
14. Sanchez, C., Wiley, J.: To Scroll or Not to Scroll: Scrolling, Working Memory Capacity,
and Comprehending Complex Texts. Human Factors: The Journal of the Human Factors
and Ergonomics Society 51(5), 730–738 (2009)
15. Schriver, K.: Dynamics in Document Design: Creating Text for Readers. John Wiley and
Sons, Inc., Chichester (1997) ISBN: 0-471-30636-3
16. Shneiderman, B.: Designing the User Interface: Strategies for Effective Human-Computer
Interaction, pp. 1–448. Addison-Wesley Longman Publishing Co., Inc., Boston (2005)
17. Stabina, R.: Quantitative Data Graphics: Best Practices of Designing Tables and Graphs for
Use in Not-for-Profit Evaluation Reports. University of Oregon, Applied Information
Management Program, 1–60 (2005)
18. http://msdn.microsoft.com/en-us/visualc/default
(retrieved on December 19, 2010)
19. http://www.oracle.com/us/technologies/java/index.html (retrieved
on January 21, 2011)
20. http://www.w3schools.com/html/default.asp (retrieved on January 17,
2011)
 Finding Reliable Path for Peer to Peer Application in
Mobile Environment

D. Jayashree1 and P. Yogesh2

1
Research Scholar
2
Professor
Dept. of Information Science and Technology,
Anna University, Guindy,
Chennai, India
[email protected],
[email protected]

Abstract. Peer-to-Peer (P2P) as well as mobile ad hoc networks (MANETs)

follows the same idea of creating a network without the help of central entities.
However, P2P and MANETs operate on different network layers. A
combination of both, creates new services and possibilities, but also faces
several problems. As smarter mobile devices are increasingly adopted by the
wider public, to access p2p applications, the question of who to trust while on
the move becomes a crucial one. When operating within ad-hoc type p2p
networks, a trust-based framework is the logical solution to provide ways to
distinguish the most trustworthy neighbors. This promotes reliable information
transfer, and encourages content and resource sharing. This paper proposes a
scheme to set up such trusted paths, based upon aggregating individual peers’
trust, given out by their nearest neighbor.

Keywords: Peer-to-peer networks, mobile network, Trust model.

1 Introduction
The peer to peer network differs from other distributed systems in that its member
nodes undertake to provide routing support themselves. There are no dedicated router
or access nodes. The implication is that trust has to be built from a neutral stand-point.
One can argue that sometimes in distributed systems including ad hoc networks,
concepts akin to trust may be used to decided whether to trust or not. For example, a
system may define a quality of service that a peer must provide [1]. Indeed some
researchers define trust as a “metaphor” that is the direct product of how interactions
within a system evolve [4].
Researchers in the domain of peer to peer networking (MANET) have realised that
in order to enable and sustain reliable networking, trust relationships have to be
formed between interacting peers. Such relationships are acquired over time as the
peer’s experience of its surroundings increases.
In its simplistic form and for the purpose of this paper, trust can be seen as “an
inherent property in which one autonomous component cannot completely control

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 638–646, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Finding Reliable Path for Peer to Peer Application in Mobile Environment 639

another autonomous component but which may need to rely on one another or require
some cooperation from it” Recommendations generally leads to what is known as a
reputation system. Such a system, operating through recommending entities, requires
the presence of a centralized and trusted server to compute, store and display valid
reputation information. This centralized authority is vital to such reputation systems
but cannot be readily assumed to be present in the case of mobile devices. The
reputation system needs to be decentralized [2, 3] such that it is accessible to entities
on the move that may not have access to central repositories.
The case for peer to peer networks can be made in instances where conventional
networking is not available. Conventionally, mobile devices operate by locating the
nearest available trusted wireless point in order to access a known wired backbone.
However, despite advances being made in lean computing and power conservation on
mobile devices, these resources remain constrained and often nodes in a network find
it preferable to operate on low energy in order to access local services. Furthermore,
addressing unreliability through repeating retransmissions in a low energy network
wastes precious bandwidth.
Superimposing trust on the p2p network therefore reinforces the transfer of
information and aids in the delivery of dependable local services. One of the ways in
which the success of a trust framework can be measured is via the quality of service
that it can provide for members adopting it. For the purpose of this paper, the trust
model presented utilizes the success rate of networking requests as one of the criteria
set for achieving quality of service. Investigating quality of service itself is outside the
scope of this paper and will be presented in future work. However, one of the
novelties of the model is the notion that low-level network attributes and behaviour
can be projected to represent higher level abstraction such as trustworthiness, given
the right context. In this paper we investigate the behavior of the model with respect,
to entity feedback provided for network requests, and to its ability to reflect
changing circumstances, thus presenting a changing trust landscape.
In this following section 2 we have discussed the existing methods and their
drawbacks. The next session explains the derivation of trust and the trust path
calculation of p2p application in mobile routing. The experimental discussion have
been given in the last section.

2 Related Work

In recent years, pee-to-peer (P2P) computing has achieved its popularity in many
distributed applications, including file-sharing, digital content delivery, and P2P Grid
computing. However, peer anonymity and autonomy make P2P networks quite
vulnerable to attacks by selfish and malicious peers.
Damiani et al [7] proposed XRep: a reputation-based approach for evaluating the
reputation of peers through distributed polling algorithm before downloading any
information. The approach adopts a binary rating system and it is based on GNutella
[8] query broadcasting method using TTL limit. EigenTrust [9] collects the local trust
values of all peers to calculate the global trust value of a given peer. Additionally,
640 D. Jayashree and P. Yogesh

EigenTrust adopts a binary rating function, interpreted as positive one (representing

satisfactory), or zero or negative one (representing unsatisfactory or complaint).Marti
and Garcia-Molina [10] proposed a voting reputation system aiming at e-commerce
applications that collects responses from other peers on a given peer. The final
reputation value is calculated combining the values returned by responding peers and
the requesting peer’s experience with the given peer. This seems more reasonable
than the model in the work by Damiani et al [7]. However, this work and the work by
Kamvar et al. [9] did not explicitly distinguish transaction reputation and
recommendation reputation. This may cause severe bias in reputation evaluation as a
peer with good transaction reputation may have a bad recommendation reputation
especially when recommending competitors.
Wang and Varadharajan [11] proposed several trust metrics for the trust evaluation
in a decentralized environment (e.g., P2P network), where a trust value is a
probabilistic value in the scope of [0, 1]. Prior to the interaction with an unknown
Peer Px, the end-peer collects other peers’ trust evaluations over Px. A method has
been proposed for trust modification after a series of interactions with Px that a good
value results from the accumulation of constant good behavior leading to a series of
constant good trust values. In another work by Wang and Varadharajan [4], the
temporal dimension is taken into account in the trust evaluation wherein fresh
interactions are weighted more than old ones.
Yu et al. [12] proposed the method of exponential averaging taking into account a
series of interactions of the requesting peer itself. It is similar to the work by Wang
and Varadharajan. In addition, Wang and Varadharajan [13] proposed Trust, a model
for trust evaluation considering recommendation trust. Trust also includes a method to
measure the recommendation trust, based on the requesting peer’s interaction
experience with the target peer and the recommendations of responding peers in
multiple rounds. Lin et al. [14] proposed a method of reputation-based trust
evaluation in service-oriented environments based on a novel architecture of
distributed trust management brokers.

3 Trust Based Mobile Peer-to-Peer Protocol

In this paper we proposed a trust mechanism in MPP (Mobile Peer-to-Peer Protocol)

[5] that include both trust calculation and finding shortest trust path. MPP allows a
realistic operation of Peer-to-Peer services over ad hoc networks. It spans from the
network layer and tries to reuse existing protocols as far as possible. Additionally the
network layer needs knowledge about the trustworthiness of these peers. Providing a
trusted path for resource sharing can help the peers to obtain trustworthy acquaintance
of neighbours without the fear of losing the packets to malicious peers.
To meet the requirements of finding trust path within the mobile network, trust is
extended on EDSR. The general structure of EDSR has been referred in EDSR
protocol [5]. EDSR provides the basis for the realization of trusted P2P applications
in MANETs. The following section explain how trust path is formed and the
calculation of trust on each peer.
 Finding Reliable Path for Peer to Peer Application in Mobile Environment 641

Mobile P2P Service

Network layer services

EDSR
Trust Evaluation model

Fig. 1. Trust Architecture of Mobile peer

This figure shows where the trust evaluation model of trust has been placed in MPP
architecture.

3.1 Trust Evaluation

In a Mobile P2P network, some peers join or leave the network frequently, which
leads to the dynamic topology changes. As a result, a trust management system needs
to frequently update trust ratings, which in turn can increase the communication
overhead. Here we have present an approach, which can easily adopted by any routing
algorithm without much overhead to find a trust path.
Trust relationships are asymmetric. This means that if a node A’s trust in a node B
in performing an action is 0.5 (where 0.5 is the trust metric), it does not automatically
imply that node B has the same amount of trust about node A. Each node holds an
individual trust relationship, defined by a trust metric, with its immediate neighbours
(on a one-hop basis only). All nodes are trustors, with their respective neighbours
being their trustees. The trust metric evolves over time due to two factors: firstly it
may rise or fall depending on new data received by the trustor and secondly, it may
undergo a period of natural “decay” over time if data regarding a certain entity is not
received over time. The “decay” promotes cooperation within the network, as nodes
are eager to gain a good reputation. All nodes are observed by their peers in relation
to their success or failure in carrying out this specific action. The trust metric is
therefore a measure of the reputation of a given node.
This section describes the procedure of the proposed trust ratings aggregation
scheme, This scheme takes into account the trust ratings based on direct and indirect
interactions. In a scenario with two possible paths, this proposed algorithm considers
the best path based on confidence, Ψ. The value of Ψ [0,1] is calculated by
individual peers based on the number of positive and negative interactions with a
peer. We use Ψ for trust ratings aggregation from witness interactions only. Since it
is possible that a peer can act maliciously and provide false trust ratings, the value of
Ψ can determine the behavior of peer in recent interactions. A newly joined peer can
642 D. Jayashree and P. Yogesh

send a trust request to the network. Peers that receive the trust request choose to send
back their trust lists. When the new peer receives this reply, its initial local trust list is
established for its own neighbors. It is possible that a trustworthy peer gets
disconnected due to power shortage or physical location change in the MANET
environment, for this reason, each entry in the t_list has a ttl variable indicating time
to live. If ttl exceeds a threshold and no connection can be established for a node, its
trust ratings are removed from the t_list. If z is a direct peer and rating of x in the
local t_list is higher than threshold then z can be trusted.
If “z” a neighbor peer has poor trust ratings, then a second opinion about
recommendations is sought from other direct neighbors. Any direct neighbor that can
give recommendations about y, while having a higher trust rating compared to z is
considered and the local t_list is updated accordingly. The trustworthiness of
recommendations from a distant node (witness) will be considered if and only if the
value of confidence Ψ is greater than a predefined constant. This is because multiple
ratings are received over multiple hops that can be modified by malicious peers in a
given period.
The weighted average of trustable witness recommendations (Ψ > constant)
provides balanced trust ratings. In witness interactions, the value of Ψ is multiplied to
the calculated trust in order to obtain the witness trust interaction value. If the ttl for a
trust rating stored in the local t_list expires, an update request is made to node z to
provide the latest t_list. If no reply is received, the node is assumed to have been
disconnected, any subsequent trust ratings are therefore removed from the local t_list.
The value of confidence Ψ is determined by the behavior of a peer in terms of
interactions. If the number of completed interactions α is larger than incomplete
interactions β, the value of Ψ is increased. Similarly, as a consequence to a large
number of incomplete transactions this value is decreased. The value of Ψ is given by
(α - β) / α. Frequent disconnections due to mobility in mobile networks cause
disruptions in communication and therefore large number of incomplete transactions.
Although the effect of frequent disconnections on the confidence value due to
mobility is essential to the quality of trust ratings, it is beyond the scope of this paper.

3.2 Path Trust Calculation

Path trust is the trust value associated with the path. This value is defined as the
weighted average of the trust values of the nodes in the path. As trust is considered to
be asymmetric mutual trust between the nodes is are used. Hop count also plays a
prominent role in the selection of the path since the larger the number of nodes, more
is the delay in the network and chances of information modification also increases.
To search for trusted data provider EDSR offers a similar message like the
Gnutella query-message. At the end of each message, the current node stores its own
IP address and the trusted value of the next hop peer. Thus, the route of the reply
message is predetermined.
Request packet: Request packet header is modified by adding a F-trust value. Hence,
that it now contains the following fields: source IP address, destination IP address, a
sequence number and F-trust. Each broadcast packet is modified to include the trust
value of the node from which packet is received. So when Source A broadcasts a
request packet and node B receives it, it updates the F-trust (forward trust) value such as
 Finding Reliable Path for Peer to Peer Application in Mobile Environment 643

F-trust = F-trust + ATB ( trust value of A for B) (1)

At the destination, F-trust contains information about the trust of all nodes involved in
the path.
Reply Packet: Reply packet is modified to accommodate the F-trust and B-trust
values. Each node updates it by including the trust value of the node from which
it received the packet. So when node X receives reply from node Y, it updates B-trust
as:
B-trust = B-trust + XTY ( trust value of X forY ) (2)

When the Reply packet reaches the source node, the most secure path is selected by it.
It calculates the path trust based on the trust values F-trust and B-trust received in the
Reply packet and the number of nodes in the path. The path selected is the one which
has the maximum path trust.
Trust value of ith path is:

Trust Path value = [(F-trust + B-trust) / 2 ] * Wti (3)

Where Wt of ith path is Wti =

1/no. of nodes in ith path

(4)
∑

n is the total number of path discovered from s to d. Wti is the weight assigned to the
ith path. : trust path value is the trust value of the ith path. : trust path value is the trust
value of the path selected as the most trust-worthy path.
As in the case of search requests, the structure of a reply is based on the structure
of a ROUTE REPLY (RREP) of EDSR. Additionally it holds the the transmitting
peer id, destination peer id and the trust value for the next hop peer. Otherwise all the
other packet details have been maintained along with their trust value.

A C
D

E
B
G F

Fig. 2. Path Trust Calculation from peer A to E

The above network model explains the Path Trust value calculation from node A to
E through nodes A, B, C, D, and E.
644 D. Jayashree and P. Yogesh

Trust Path Value A-E

= [{ ATB + BTC+ CTD+ DTE+ ETD+ DTC+ CTB+ BTA } / 2 ] * Wti (5)

Trust Path from A to E contains mutual trust information of all the nodes involved in
the path from A to E.

4 Performance Evaluation

In this section, we discussed the performance on simulation using the following

parameters defined here.
To simulate the trust path calculation in a mobile ad-hoc network environment,
NS2 v2.28, a network simulation tool, is used. EDSR protocol is adopted as a routing
protocol, with 1500m1500m in the size of network area. The simulation is carried out
with a change in the number of peers ranging from 20 up to 100. A mobile scenario
for mobile peers where pause time on average is 2 seconds with average velocity and
maximum velocity varying 3 to 20 m/sec and 5 to 30 m/sec, respectively, is designed
for evaluating their performance. Table 1 shows the values of parameters used in a
simulation.
On top of that, when having found out applicable objects, the number of physical
hops for paths to the objects is reduced, which means that, when an object requester
searches applicable objects, it can quickly find them without going through
various peers and hops count of peers are relatively less from source peer to
destination peer.

Table 1. Simulation Parameters

Parameter Value
Number of Peers 20 to 100
Network size 1500m x 1500m
Mean query generate time(secs) 5

Routing protocol DSDV, AODV

Movement average speed (m/s) 3 to 20

Movement maximum speed (m/s) 5 to 30

Channel Channel / Wireless Channel

Propagation Propagation / TwoRayGround
Network interface Phy / WirelessPhy
Mac Mac / 802_11

The simulation was carried out as required with varying loads of traffic through
various cycles. Figure 3 pertains to the traffic through one node \only within a
network. Data was analyzed at the end of each cycle.
 Finding Reliable Path for Peer to Peer Application in Mobile Environment 645

Fig. 3. The overheads generated at single peer

As far as the node is concerned, it can be seen in Figure 3 that the overhead added
by the trust mode by virtue of its reputation mechanism is not appreciably higher than
the outing overhead.

5 Conclusion
This paper presents a new trust path-finding scheme for mobile P2P networks. This
relies on direct trust ratings and witness recommendations from reliable peers to
determine trust ratings for a node using confidence value. Simulation results
demonstrate that the overall performance of this scheme is reliable and have marginal
increment in their overheads for detecting trusted peers in P2P mobile networks.
In the future, we should do more work to research the discrimination of the
malicious actions and deploy the framework in a real system.

References
1. Poslad, S.: Ubiquitous Computing: Smart, Devices, Environments and Interactions, pp.
309–311. Wiley, Chichester (2009) ISBN: 978-0-470-03560-3
2. Chakravorty, R., Agarwal, S., Banerjee, S., Pratt, I.: MoB: A mobile bazaar for wide-area
wireless services. In: Proceedings of MobiCom 2005, pp. 228–242 (2005); Quercia, D.,
Hailes, S., Capra, L.: MobiRate: Making Mobile Raters Stick to their Word. In:
Proceedings of UbiComp 2008, Seoul, Korea (2008)
3. Quercia, D., Hailes, S., Capra, L.: B-trust: Bayesian Trust Framework for Pervasive
Computing. In: Stølen, K., Winsborough, W.H., Martinelli, F., Massacci, F. (eds.) iTrust
2006. LNCS, vol. 3986, pp. 298–312. Springer, Heidelberg (2006)
646 D. Jayashree and P. Yogesh

4. Wen, D., Huaimin, W., Yan, J., et al.: A recommendation-based peer-to-peer trust model.
Journal of Software 15(4), 571–583 (2004)
5. Schollmeier, R., Gruber, I., Niethammer, F.: Protocol for Peer-to-Peer Networking
inMobile Environments. In: Proc. 12th International Conference on Computer
Communications and Networks (ICCCN 2003), October 20-22, pp. 121–127. IEEE, Los
Alamitos (2003)
6. Damiani, E., di Vimercati, S.D.C., Paraboschi, S., Samarati, P., Violante, F.: A reputation
based approach for choosing reliable resources in peer to peer networks. In: Proceedings of
ACM CCS 2002, p. 207 (November 2002)
7. GNutella, http://www.GNutella.com/
8. Kamvar, S.D., Schlosser, M.T., Garcia-Molina, H.: The eigentrust algorithm for reputation
management in p2p networks. In: Proceedings of the 12th International WWW
Conference, Budapest, Hungary (May 2003)
9. Marti, S., Garcia-Molina, H.: Limited reputation sharing in p2p systems. In: Proceedings
of ACM EC 2004, New York, USA, pp. 91–101 (May 2004)
10. Wang, Y., Varadharajan, V.: Interaction trust evaluation in decentralized environments. In:
Bauknecht, K., Bichler, M., Pröll, B. (eds.) EC-Web 2004. LNCS, vol. 3182, pp. 144–153.
Springer, Heidelberg (2004)
11. Yu, B., Singh, M.P., Sycara, K.: Developing trust in largescale peer-to-peer systems. In:
Proceedings of 2004 IEEE First Symposium on Multi-Agent Security and Survivability,
pp. 1–10 (August 2004)
12. Wang, Y., Varadharajan, V.: Trust2: Developing trust in peer-to-peer environments. In:
Proceedings of 2005 IEEE International Conference on Services Computing (SCC 2005),
Orlando, Florida, USA, pp. 24–31 (2005)
13. Lin, K.J., Lu, H., Yu, T., Tai, C.E.: A reputation and trust management broker framework
for web applications. In: Proceedings of The 2005 IEEE International Conference on e-
Technology, e-Commerce and e-Service (EEE 2005), pp. 262–269 (2005)
 Text Mining for Interpreting Gene

K. Prabavathy1 and P. Sumathi2

1
Department of Computer Science,
PSG College of Arts and Science, Coimbatore-14
Tamil Nadu, India
Tel.: 9942262728
[email protected]
2
Department of Computer Science,
Chikkanna Govt. Arts College , Tirupur-02
Tamil Nadu, India
Tel.: 9150588820
[email protected]

Abstract. Text mining provides methods to retrieve and extract information

contained in free-text automatically. The advances in molecular biology
techniques have given rise to high throughput biology that produces tremendous
data and related publications in the past decades. Here the efficient constituents
of Co-reference resolution incorporating Natural Language Processing (NLP) to
interpret Gene expression are focused. It may overcome the challenges and
limitations of text mining in biological data for resolving unsolved problems
and this paper describes a new phase of text mining process to uncover
interesting term correlations, genomic term identification in curation process ,
identification of biological relations and help the biologists in their analysis of
complex problems. The new phase of text mining process is organized in four
tasks of namely use NLP, find correlated terms, Co-reference resolution and
built a structured database.

Keywords: Bio-NLP, Correlated terms, Co-reference resolution, Gene,

Structured database.

1 Introduction

Data mining refers to automated search for meaningful patterns of data stored in large
databases distributed over the web. Text mining is a young interdisciplinary area
that draws on information retrieval, data mining, machine learning, statistics and
computational linguistics. Text mining generally refers to the process of extracting the
interesting and non-trivial information and knowledge from unstructured text.
The Data mining tools require input data to be fairly homogenous like those found
in conventional database (numbers, characters, etc.). Text mining unlike Data mining
works on unformatted data and semi-structured data. A real text mining discovers new
pieces of knowledge, from approaches that find overall trends in textual data.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 647–653, 2011.
© Springer-Verlag Berlin Heidelberg 2011
648 K. Prabavathy and P. Sumathi

MEDLINE searches crossed 82.07 million being a challenge with approximately 10%
growth every year. This in turn strongly depends on efficiency of database
interoperability and data mining integration. The huge amount of data of this post
genomic era includes genomic, proteomic and literature data. Bioinformatics has
emerged as one and only one solution to handling such complex data.

2 Applications of Text Mining

The text mining is considered as an art than to a science, but contributes very much
to the Concept driven science or conceptual biology. Text mining is one of the tools
now used in fields like Astrophysics, Sequence analysis, Drug Discovery, Gene
identification and disease relation, disease identification, protein interaction studies,
to study protein - protein interactions, protein – gene interactions, metabolomics and
even in the nascent field of Systems biology.
Figure 1 shows the Applications of BioNLP and Text mining .The major topics are
represented by the inner circle of seven approaches, and the corresponding
applications are given in the outer layers of boxes.

Fig. 1. Applications of BioNLP and Text mining

Text Mining Approaches and Tasks

The text mining is a process of finding, organizing and analyzing information. It
creates the new information that is not obvious in a collection of documents. New
information is defined as a pattern, trend, or relationship that can’t be easily gleaned
by reading individual documents. The two approaches of text mining are Bag of
Words that looks for features.
 Text Mining for Interpreting Gene 649

Natural Language Process (NLP) looks for Text analysis by using Syntax and
semantics

3 Natural Language Processing(NLP)

A variety of different systems is available to do natural language processing, but no

system can cope with the complexity of natural language in a spoken language. NLP
is based on the use of computers to process language, and it includes techniques
developed to provide the basic methodology required for automatically extracting
relevant functional information from unstructured data, such as scientific publications.
Information retrieval and NLP systems are soon likely to become important not only
for extracting information but also for assisting in various aspects of research such as
the discovery of new facts, the interpretation of findings, and the design of
experiments. NLP based systems provide solutions on chunks of sentences. The main
advantage of NLP based approaches can be found in the fact that sentence analysis
can be shaped following a hand-crafted linguistically model
Figure 2 represents the NLP and its Steps. The three main tasks of Text mining are
Information Retrieval, Semantics and Information Extraction.


 6;:0 ;
:1: <

 10 

 6;:0 ;
1=: ;



Fig. 2. NLP and its Steps

650 K. Prabavathy and P. Sumathi

Information Retrieval - It is an automatic process of extracting files based on Query

based approach or Document based approach. Document based approach retrieves
documents of similar kinds, whereas, query based approach retrieves documents that
contain certain combination of user defined words or key words.
Semantics - The biomedical texts are often ambiguous inclusive of words or terms
that may have number of synonyms, like, NF-kβ or NF- KB, gene AGP1 or YCC5,
etc. Vector models and other methods are blind enough and very much automatic to
sense such similarity and differences. Semantics identifies the definition to such
terms from the text.
Information Extraction - This automated methods look for facts in the text that may
represent combination of terms or patterns. Selection also considers sequence of the
words and takes into account their morphological changes through the text.
Information extraction is the process of using this information to extract meaning
from a text and regular expressions to produce a structured, machine readable file that
contains essential information.

3.1 Related Work

Text mining is a system for phrase extraction and proximity analysis. Text analysis as
a major new frontier in bioinformatics, whose smashing success in the area of gene
sequence analysis is based, after all, on nothing more than algorithms for finding and
comparing patterns in the four-letter language of DNA. Most of the efforts have been
focused on a few specific problems although that there are many other interesting
areas of potential applications.
Bob Futrelle has organized a large "bio-NLP" information network and enunciated
a radical vision by using coherent theory based on the biological nature of language.
Swanson's, Kostoff's, Tanabe's, and Chowdhury's reliance on human expertise and
manual filtering interfaces between biomedical text mining and the cognitive tradition
in IR. K. Bretonnel Cohen , Lawrence Hunter articulated an in-depth treatment of
statistical approaches to biomedical text mining.

Proposed Phase - Gene Interpretation by Text Mining

Text is "unstructured, amorphous, and difficult to deal with" but also "the most
common vehicle for formal exchange of information." The way to represent the huge
amounts of unstructured text in a compact form is difficult and can do by designed
phase of text mining process. Figure 3 describes the tasks of designed phase that
provides the basis for the biologists in their analysis of complex problems. The tasks
that are undertaken in this paper are
o Use NLP
o Find out the correlated terms
o Co-reference Resolution
o A Structured database
 Text Mining for Interpreting Gene 651

3.2 Use NLP

The first step of NLP approach retrieves the Genes from open access database; the
second step apply semantics for identifying the physical gene structure and by third
step finalizing the characterization of Gene are done. By this the text is indexed into
terms or words that looks for particular keyword or pattern as specified. It helps
researchers to identify the natural behavioral of gene and to find Gene- Disease
relationships.

3.3 Finding Out the Correlated Terms

Define the concepts by the sets of correlated terms rather than by raw term which
can quantify the relations between documents or portions of text. Correlated terms
usually having a more specific meaning are taken and predict the subsequences of
diseases. The parameters were set to discover a thousand combinations of correlated
terms. These sets of terms were further added to the basic index and applied to the
information retrieval problem. The extraction of annotations of genes and proteins
automatically and for detecting protein-protein interactions and regulatory pathways
can be done with correlated terms. Also the extraction of associations between
proteins and their functional features that pose an interesting biological problem are
found out from the co-related terms.

Co-reference Resolution

Two noun phrases referring to the same entity are said to co-reference. Co- reference
resolution is important for many text mining tasks:

• Information extraction

• Summarization

• First story detection

Summarization of Functional genomics can be done by Co-reference resolution of

Gene. A number of text-mining methods have been developed for summarizing
groups of genes.

Structured database

A Biological database can be built from the extracted information. So with this, Gene
expression data (co-regulated genes) along with co-occurrence information can be
predicted which a major resource to study various Gene -Diseases Relationship.
Proteins and genes are characterized within biological databases through unique
identifiers; each identifier is associated with its corresponding protein or nucleotide
sequence and functional descriptions.
652 K. Prabavathy and P. Sumathi

Fig. 3. Tasks of Designed phase

4 Conclusion
This paper describes a new text mining process along with NLP to uncover interesting
term correlations and Gene- Disease relationship. This algorithm comprising these
above said steps identifies combinations of terms that optimize an objective function,
which is the cornerstone of the process. A scientist now may use this information from
the database to know why the genes were co-regulated, what disease may be
 Text Mining for Interpreting Gene 653

associated? , what symptom, or the identification of a gene in conjunction with a

mutation. etc., these can also be linked with other studies like NMR data of
metabolites (metabolomic data) to define causes of disease. Finding meaningful
concepts within a huge collection of corpuses in a reasonable timeframe is a difficult
task to accomplish and that can be accomplished in future.

References
1. Cohen, K.B., Hunter, L.: Getting started in text mining. PLoS Comput. Biol. 4, e20 (2008)
2. Rzhetsky, A., Seringhaus, M., Gerstein, M.B.: Getting started in text mining: part two.
PLoS Comput. Biol. 5, e1000411 (2009)
3. Friedman, C., Kra, P., Yu, H., Krauthammer, M., Rzhetsky, A.: GENIES: a natural-
language processing system for the extraction of molecular pathways from journal articles.
Bioinformatics 17(suppl.1), S74–S82 (2001)
4. Perez-Iratxeta, C., Wjst, M., Bork, P., Andrade, M.: G2D: a tool for mining genes
associated with disease. BMC Genet. 6, 45 (2008)
5. Koussounadis, A., Redfern, O.C., Jones, D.: Improving classification in protein structure
databases using text mining. BMC Bioinformatics 10, 129 (2009)
6. Krallinger, M., Valencia, A.: Text-mining and information-retrieval services for molecular
biology. Genome Biol. 6, 224 (2005)
7. Shatkay, H., Höglund, A., Brady, S., Blum, T., Dönnes, P., et al.: SherLoc: high-accuracy
prediction of protein sub cellular localization by integrating text and protein sequence data.
Bioinformatics 23, 1410–1417 (2007)
8. Natarajan, J., Berrar, D., Dubitzky, W., Hack, C., Zhang, Y., et al.: Text mining of full-text
journal articles combined with gene expression analysis reveals a relationship between
sphingosine-1-phosphate and invasiveness of a glioblastoma cell line. BMC
Bioinformatics 7, 373 (2006)
9. Galvez, C.: Knowledge management for biomedical literature: the function of text-mining
technologies in life-science Research
10. Editorial, K.J.: Briefings in Bioinformatics. (3), 220–221 (2005)
 Improved Input-Delayed Dynamic Neural Network for
GPS/INS Integration

M. Malleswaran1, A. Saravanaselvan2, and V. Vaidehi3

1
Dept of Electronics and Communication Engineering
Anna University of Technology Tirunelveli, Tirunelveli, India
2
Dept of Electronics and Communication Engineering
National Engineering College, Kovilpatti
3
Dept of Computer Technology
MIT Campus, Anna University, Chennai, India
[email protected],
[email protected],
[email protected]

Abstract. At present, real time navigation systems rely on kalman filtering for
integrating global positioning system (GPS) and inertial navigation system
(INS) in order to provide feasible solution for reliable navigation. But there
exist some inadequacies related to the stochastic error models of inertial
sensors, requirement of prior knowledge in fusing data and long design time.
Moreover, recently artificial intelligence (AI) techniques based on static neural
networks were suggested in place of kalman filter. But there also exist
drawbacks like inadequate training during GPS absence, less reliability and
more training time. To solve this problem, a new dynamic architecture based on
Input-delayed model has been developed. This paper, therefore proposes a
scheme based on Improved Input-delayed dynamic neural network (IIDDNN)
to model the samples of INS values with respect to GPS values as target vectors
(Normal mode). This model is capable of providing a reliable and accurate
solution especially for GPS absence (prediction mode). The prediction mode
output obtained using IIDDNN is compared with the performance of normal
mode of IIDDNN. The proposed IIDDNN model is evaluated using test data of
fixed trajectory path for an aircraft vehicle.

Keywords: Global Positioning System, Inertial Navigation System, Artificial

Intelligence, Improved Input-Delayed Dynamic Neural Network, Static Neural
Networks.

I. Introduction

A. Motivation
Nowadays land vehicles are equipped with Global Positioning System (GPS) to
provide accurate position and velocity information [14-17]. However, there are
several situations where GPS experience satellite signal blockage or deterioration of
accuracy (due to multipath effects and clock bias error). Therefore, GPS is usually

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 654–672, 2011.
© Springer-Verlag Berlin Heidelberg 2011
 Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 655

combined with Inertial Navigation System (INS)[18,19]. The integration of GPS and
INS, therefore, provides a navigation system that has superior performance in
comparison with either a GPS or an INS stand-alone system.

B. Navigation using GPS

The last two decades have shown an increasing trend in the use of positioning and
navigation technologies in land vehicle applications. Most of the current vehicular
navigation systems rely on the global positioning system (GPS) that is capable of
providing accurate measurements the GPS needs at least four satellites with good
geometry. In addition, there must be a direct line of sight between the GPS antenna
and those satellites. Unfortunately, this is not always possible since a GPS signal may
be lost when driving around obstacles (through urban areas, overpasses or tunnels on
highways, tree lined streets etc) or when operating in poor weather conditions.
Moreover, the GPS signal is purposefully jammed while an aircraft vehicle on fly.
The satellite signal blockage results in deterioration of the overall position accuracy
[1, 4].

C. Navigation using INS

Inertial Navigation System (INS) is a self-contained system. It incorporates three
orthogonal accelerometers and three orthogonal gyroscopes, which measure three
linear accelerations and three angular rates respectively. A set of mathematical
transformations and integrations with respect to time, known as the mechanization
equations are applied to the raw measurements from the INS sensors to determine
position, velocity and altitude information. Unfortunately, the INS cannot replace the
GPS or operate on a stand-alone basis. During the mechanization procedure, the
accuracy of INS position components deteriorates with time due to the integral of the
sensor errors that exhibit considerable long-term growth. These errors include white
noise, correlated random noise, bias instability and angle random walk. The errors are
stochastic and wide sense stationary process in nature and can cause a significant
degradation in the INS performance over a long period of operation [10, 11, and 12]

D. Existing INS/GPS data fusion techniques

In order to overcome the problems associated with the operation of GPS and INS on
their own, the two systems are often paired together so that the drawbacks associated
with each system are eliminated. The GPS/INS data fusion is commonly performed
real time using a kalman filter (KF) [1,4]. This method requires a dynamic model of
both INS and GPS errors, a stochastic model of the inertial sensor errors and a priori
information about the covariance of the data provided by both systems. Data fusion
employing a KF has been widely used and is considered the benchmark for INS/GPS
integration.
There are however, several considerable drawbacks to its use [2, 5, and 11]. These
include the following: (1) the necessity of accurate stochastic modeling, which may
not be possible in the case of low cost and tactical grade sensors; (2) the requirement
for a priori information of the system and measurement covariance matrices for each
new sensor, which could be challenging to accurately determine; (3) relatively poor
accuracy during long GPS absence; (4) the weak observability of some of the error
states that may lead to unstable estimates of other error states; and (5) the necessity to
656 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

tune the parameters of the stochastic model and a priori information for each new
sensor system. The above drawbacks can be overcome by using intelligent
networks or Artificial Neural Networks (ANN) [2, 4]. The IIDDNN is found to have
optimized architecture when compared to static neural networks in the prediction
mode. [3, 11, 12]

2 Literature Survey
More recently, several techniques based on Artificial Intelligence (AI) have been
proposed to replace KF in order to eliminate some of its inadequacies [1, 2, 13]. The
major inadequacy related to the utilization of KF for GPS/INS integration is the
necessity to have a predefined accurate stochastic model for each of the sensor errors
[3, 12]. Furthermore, prior information about the covariance values of both INS and
GPS data as well as the statistical properties (i.e. the variance and the correlation
time) of each sensor system has to be known accurately [3]. Several AI – based
GPS/INS architectures using Multi-Layer Perceptron Neural Networks (MLPNN) [2],
Radial Basis Function Neural Networks (RBFNN) [5] and Adaptive Neuron-Fuzzy
Inference Systems (ANFIS) were reported for GPS/INS integration [6–7]. The main
idea behind all of these methods is to mimic the latest vehicle dynamics by training
the AI module during the availability of the GPS signals. An empirical model that
processes the INS output (position or velocity) and provides the corresponding INS
position or velocity error can be established during the normal (training) procedure. In
addition, some of the recently published results showed that KF might outperform
some of the AI–based techniques during the first few seconds of the GPS absence
[1,7]. This is due to the inaccurate linearized dynamic model of INS errors utilized by
KF. Such linearized model operates adequately only in the short-term of GPS
absence. In case of long GPS absence, the ignored non-linear and nonstationary parts
of the errors grow to large values and deteriorate the overall positioning accuracy. The
negative impact of the inappropriate stochastic models of inertial sensors also appears
in cases of long GPS absence [4–6]. In this research, we aim at developing an AI
based GPS/INS integration based on IIDDNN model and thus providing a better
positioning accuracy during the absence of GPS signal. Such technique combines the
advantages of IIDDNN model in representing the sequential process in the time
varying input pattern (INS position).

3 Improved Input Delayed Dynamic Neural Network

The INS dynamic error model used by KF is a linearized one. In addition the
stochastic error model for the sensor errors is also linearized in the form of 1st order
difference equations. Such state space model is required for the operation of KF. The
non-linear and the non stationary parts of the INS errors can be modeled using
extended kalman filter, particle filter and unscented kalman filter but they all required
complex modelling process. Moreover, INS position error will grow to large value
using kalman filter thus deteriorating the positioning accuracy, especially in the long
term GPS absence. This leads to large position errors in case of relatively long GPS
 Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 657

outages. That is the reason for opting Artificial Neural Network (ANN) to find the
solution for navigation problem in an aircraft vehicle. The non-linear complex AI –
based modeling capabilities is therefore suggested in this study.
Dynamic networks are generally more powerful than static networks (although
they may be somewhat more difficult to train) [7,20]. Because dynamic networks
have memory, they can be trained to learn sequential or time-varying patterns. In fact,
in order to predict temporal patterns, an ANN requires two distinct components: a
memory and an associator. The memory holds the relevant past information, and the
associator uses the memory to predict future events. The associator is simply a static
multi layer perceptron neural network (MLPNN) and the memory is generated by a
time delay unit (or shift register) that constitutes the tapped delay line. Traditional
MLPNN is a static and memory less network that is effective for complex non linear
static mapping. A common method is to consider a sequence of time varying input
patterns. This approach is suitable to find accurate position of an aircraft vehicle at the
time of GPS absence. Here, the network is provided with a static memory that is
specifically expert knowledge-dependent. A major limitation of the IIDDNN model is
the increase of computational complexity since the input layer has to have a number
of neurons equal to the number of past samples. As a result, both the complexity and
training time will dramatically increase.
A. Proposed Architecture description
In this case, the static neural network is transformed into a dynamic neural network by
incorporating a memory and associator units at the input layer of input delay neural
network. This input delayed model is trained to learn sequential or time varying
patterns of the INS samples. The memory holds past samples of INS values and the
associator uses the memory to predict future events. The model configuration of
Improved Input-delayed neural network is shown in figure1. The case shown in
figure1 considers a tapped delay line that involves the ‘d’ most recent inputs. For a
case of ‘d’ delay elements, the network processes time varying pattern of the INS
samples applied as inputs. This dynamic model neural network considers a cluster
inputs for the training process whereas the static neural networks produces the future
outputs based on current samples position alone. In other words, we can say that the
static neural networks output does not considers the past events. Because static neural
networks are memory less topology that is effective for complex non linear static
mapping. In fact, INS position error prediction is a procedure where previous errors
have to be seriously considered. A fixed number of past events are selected and
introduced to the input layer of the network. Therefore, the network is involved with a
static memory that is specifically expert knowledge-dependent. This has a beneficial
effect on the prediction accuracy especially in the case of GPS absence (prediction
mode). Furthermore, the IDNN is trained with the New dynamic back propagation
learning algorithm [11, 12].
The training criteria of the network involves New dynamic back propagation
learning algorithm based on input delayed model of IIDDNN. During the time of
GPS absence, the path selector unit is switch over into prediction mode from the
normal mode for training INS values. In this case, IIDDNN model gives the INS
position error value and this value is added with the current samples of INS position
so as to arrive the new target values. The path selector unit allows GPS signal as
target values to the IIDDNN model in normal mode whereas in the prediction mode
658 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

Fig. 1. System configuration of proposed scheme

the GPS signal does not exist. In this case, the adder unit which adds the INS position
error value to the current samples of INS position to determine the new target value
for training the IIDDNN model. At that time, the path selector unit allows the adder
output as target values to the IIDDNN model to carry out the prediction mode training
process.
The IIDDNN system model processes the input delay elements as x(t), x(t -1), x(t -
2), ... and x(t - d), where d is known as the tapped delay line memory length.
The input signal Xi (t) to the input layer of IIDDNN model shown in figure 1 is
mathematically represented as
d
Xi (t) = ∑ Wi(k) x(t - L) + αi
L=0

Wi(k) – Synaptic weight for neuron i at input layer of IIDDNN

αi – bias value between input node to hidden node
d – Tapped delay line memory length

The non linear activation function used in the hidden layer of IIDDNN is given as
d
Hi = S( ∑ Wi(k) x(t - L) + αi)
L=0

Si – Hidden layer output

H(.) – Sigmoidal activation function
S(Xi (t)) = (1 - e - Xi (t) ) / (1+e - Xi (t) )

The linear activation function used in the output layer of IIDDNN is represented as
N
Yj(t) = F( ∑ Wij Hi + βj)
i=1
 Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 659

Yj(t) – neuron output j

βj – bias value between hidden node to output node
Wij – Weight factor between hidden node to output node
L(.) – Transfer function of the output neuron j
N
Cj (t) = C( F( ∑ Wij Hi + βj) )
i=1
C(.) – Comparator function of the trained output neuron j and input neuron i.

B. Training mode
Along each of the East, North and vertical directions, the INS position (PINS) and the
time(t) are the inputs to one of the IIDDNN modules while the error in the
corresponding INS position (dpINS) is the module output. During the availability of the
GPS signal, the IIDDNN module operates in the normal mode. In order to train the
network, the INS position error provided at the output of this IIDDNN module should
be compared to a certain target or a desired response. In this case, the target (or true)
INS position error (dpINS/GPS ) is the difference between the INS original position and
the corresponding GPS position. The difference between IIDDNN module output
(dpINS) and the true position error dpINS/GPS is the estimation error (D (dp)) of the
IIDDNN module. In order to minimize this error, the IIDDNN module is trained to
adjust the IIDDNN parameters that are continuously updated according to the least
square criterion until reaching certain the minimal Mean Square Error (MSE). The
training procedure continues and is repeated for all GPS/INS data windows until a
GPS absence is detected.

C. Prediction mode
When the satellite signal was blocked (during GPS absence), the system was switched
into the prediction mode where the IIDDNN module is used to process the INS
position PINS at the input and predict the corresponding position error (dpINS) using the
latest IIDDNN parameters obtained before losing the satellite signals. The estimated
error for past INS samples along with current INS values is then applied into a
comparator network in order to predict the new GPS position (PGPSC). In order to
utilize the IIDDNN based module in real-time GPS/INS integration, a time varying
pattern with memory size (d) is selected. A number of samples equal to memory size
of INS position component and the corresponding GPS position are collected. The
prediction procedure of the IIDDNN parameters starts after collecting the dth sample
of both INS and GPS position components. The patterns of INS and GPS position
components are used to train the IIDDNN module to reflect the latest vehicle
dynamics and the INS error trend. The IIDDNN module is trained until a certain
number of MSE is reached. This procedure is repeated after collecting completely
new patterns of INS and GPS of a size equal to the memory size. To provide a
complete navigation solution for a aircraft vehicle, the longitude, latitude and altitude
components are considered and applied to the IIDDNN module to estimate the
MSE.
660 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

D. New dynamic back-propagation learning algorithm

 The weight is initialized as small random values

 The Input signal xi is given to each input unit and the input unit transmits the
signal to all hidden layers.
 Each hidden unit sums its weighted input signals as
d
z-inj = ∑ Wi(k) x(t - L) + αi
L=0

and the activation function is applied for training that input signal.
zj =f(z-inj)

 The hidden units output is given as input to output unit. Each output unit
sums it weighted input signals as
N
y-ink = F( ∑ Wjk Hi + βj)
i=1
and the activation function to calculate the output signals
yk=f(y-ink)

 Comparator function of the trained output neuron j and input neuron i is

given as
N
Cj (t) = C( F( ∑ Wjk Hi + βj) )
i=1
 Each output unit receives a GPS target pattern. According to the received
GPS signal corresponding error information is calculated.
δk =(tk - yk)*f(yink)
 Each hidden unit receives the feedback from output unit and sums its delta
inputs from the units in the layer above
m
δ-inj = Σ δj *wjk
k=1
The error information is calculated as
δj =δ-inj* f(z-inj)
 Each output unit updates its bias and weights
 The weight correction term is given by
Δwjk =α δk zj
 Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 661

 And the bias correction term is given by

Δβj =α δk
Therefore
Wjk(new) = wjk(old)+Δwjk

βj(new)= β j(old)+ Δβj

 Each hidden unit updates its bias and weights
The weight correction term
Δ Wi(k) = α δj xi
The bias correction term
Δ αi = α δj
Therefore
Wi(k) (new) = Wi(k) (old)+Δ Wi(k)

αi (new) = αi (old)+Δ αi
 After the errors are minimized, the testing conditions are stopped.

E. Network over-fitting
Network over-fitting is a machine-learning problem that has been investigated by
many researchers. Network over-fitting usually occurs when the network captures the
internal local patterns of the training data set rather than recognizing the global
patterns of the data sets. The knowledge base that is extracted from the training data
set is therefore not general. As a consequence, it is important to realize that the
specification of the training samples is a critical factor in producing a neural network
which is capable of making the correct responses. Two procedures have been
evaluated to overcome the problem of over-fitting namely, early stopping and
regularization.

F. Early stopping
In general, the aim of early stopping is to mimic the prediction of future individuals
form the whole. This will be achieved in case that the training data fully represents the
sample space and each left out individual can lie anywhere in this space. Large
samples and small dimensionality generally satisfy these requirements. During the
IIDDNN training stage, the module performs the function of understanding the input
and output mapping. To ensure the generalization of the IIDDNN model a hold out
kind of early stopping criterion was chosen due to its suitability for real time
implementation. This approach allows IIDDNN to read just its parameters to detect
the dynamics of the data in a transition interval before moving to the prediction stage.
Furthermore, it avoids the possibility of model over-fitting for the training data.

G. Regularization
The second method utilized to avoid the over-fitting problem and to optimize the
IIDDNN model is the regularization technique. This is known to be a very desirable
662 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

procedure when the scaled conjugate gradient descent method is adopted for training.
The regularization technique involves modifying the performance function of mean
square error (MSE) which is normally chosen to be the sum of squares of the error
between the network output and the desired response on the training data set. [9]

4 Experimental Setup
In this experiment, the training mode and the prediction mode of IIDDNN was
performed. The normal mode and prediction mode training were performed using
bipolar sigmoidal activation function (non linear activation function). During the
presence of GPS signal, the proposed system relies on GPS position information to
train the IIDDNN model. During the normal mode, the IIDDNN model is trained to
mimic the latest vehicle dynamic, determine the INS position error, and correct the
corresponding INS position component. The data is processed as follows: first, the INS
and GPS signals are taken as input vector and target vector respectively. The input
vector is send as one of the input signal to input node during the training mode. The
value of the weight is initially considered to be a small value and bias is chosen close
to the first input value. The learning rate of this experiment is also chosen accordingly.
The INS position is then used as the input to IIDDNN, and is trained until it nearly
equals the target GPS position. During the absence of GPS signal, the proposed system
operates in the prediction mode where the IIDDNN model predicts the corresponding
INS position error based on the knowledge stored during the normal mode procedures.
This predicted INS position error is then removed from the corresponding INS position
component to obtain the corrected INS position. [8, 21]

5 Simulation Results and Tables

The trajectory path shown in figure 2 is taken into consideration and the New
Dynamic Back Propagation Learning Algorithm (NDBPLA) for IIDDNN is
implemented using matlab. The position and time value obtained after INS
mechanization are taken as input vectors. The position values obtained from GPS
using GPS receiver is taken as the target vector. Training was performed for latitude,
longitude and altitude components. The actual output i.e., corrected INS value for all
the three components were found using NDBPLA. From the output obtained during
the final epoch of IIDDNN model is shown in Fig. 7,8,9,10,11 and 12 for normal
mode and predicted mode. The mean square error (RMSE) and performance index
(PI) values were calculated for each and every epoch. The formula for the RMSE and
PI are given as

∑ y ay
RMSE
m

∑
∑ | |
 Improved Input-Delaayed Dynamic Neural Network for GPS/INS Integration 663

Where yi is the target outpput, ayi is the computed output, m is the number of ddata
patterns and |yi| is moduluss of target output. The performance of the IIDDNN moodel
is given in table 7. The less RMSE and PI values of prediction mode IIDDNN giives
superior performance than normal
n mode IIDDNN model.

Fig. 2. Trajectory path

Fig. 3. Output of lo
ongitude component in Normal mode when EPOCH=4
664 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

Fig. 4. Output of latitude component in Normal mode when EPOCH=4

Fig. 5. Output of altitude component in Normal mode when EPOCH=4

Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 665

Fig. 6. Output of longitude component in Normal mode when EPOCH=7

Fig. 7. Output of latitude component in Normal mode when EPOCH=7

666 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

Fig. 8. Output of altitude component in Normal mode when EPOCH=7

Fig. 9. Output of longitude component in predicted mode when EPOCH=3

Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 667

Fig. 10. Output of latitude component in predicted mode when EPOCH=3

Fig. 11. Output of altitude component in predicted mode when EPOCH=3

668 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

Fig. 12. Output of longitude component in predicted mode when EPOCH=5

Fig. 13. Output of latitude component in predicted mode when EPOCH=5

 Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 669

Fig. 14. Output of altitude component in predicted mode when EPOCH=5

Table 1. Mean square error of IIDDNN model in Normal mode

Epoch no Longitude Latitude Altitude

(MSE) (MSE) (MSE)
1 120.50 118.46 150.34
2 504.54 367.45 290.89
3 258.99 233.89 180.46
4 69.80 100.48 97.90
5 15.52 40.45 50.67
6 7.65 10.35 15.78
7 1.89 1.76 1.78

Table 2. Weight updates in output layer during each epoch of IIDDNN training in Normal mode

Weight updates in output layer

Epoch
no W1 W2 W3 W4 W5
1 0.5210 0.9053 1.5226 2.4269 3.6583
2 0.4652 0.4676 0.3761 0.4449 1.5453
3 0.4124 0.0731 -0.6298 -1.4743 -1.5273
4 0.3623 -0.2782 -1.4401 -2.7918 -3.8410
5 0.3148 -0.5878 -2.0438 -3.3882 -3.8983
6 0.2700 -0.8581 -2.4651 -3.5756 -3.8813
7 0.2275 -1.0922 -2.7448 -3.6239 -3.8865
670 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

Table 3. Weight updates in hidden layer during each epoch of IIDDNN training in Normal mode

Epoch Weight updates in hidden layer

no
W1 W2 W3 W4 W5
1 -1.5500 -1.5500 -1.5500 -1.5500 -1.5500
2 -1.5500 -1.5500 -1.5500 -1.5499 -1.5496
3 -1.5500 -1.5500 -1.5500 -1.5502 -1.5530
4 -1.5500 -1.5500 -1.5500 -1.5510 -1.5729
5 -1.5500 -1.5500 -1.5502 -1.5616 -1.9903
6 -1.5500 -1.5500 -1.5513 -1.7068 -10.13
7 -1.5500 -1.5500 -1.5593 -3.7926 -168.11

Table 4. Mean square error of IIDDNN model in Predicted Mode

Epoch no Longitude Latitude Altitude

(MSE) (MSE) (MSE)
1 160.40 304.55 267.81
2 128.44 220.78 139.30
3 44.82 110.55 49.87
4 22.75 30.67 28.62
5 1.12 0.98 0.87

Table 5. Weight updates in output layer during each epoch of IIDDNN training in Predicted mode

Weight updates in output layer

Epoch W1 W2 W3 W4 W5
no
1 1.1648 0.2738 -1.0394 -2.0166 -13.1677
2 0.5898 0.3632 0.6924 1.7445 27.6166
3 0.0245 0.4394 1.6642 2.5573 101.7480
4 -0.1676 0.4738 2.0029 2.9893 156.6126
5 -0.3428 0.5328 2.8862 3.4526 182.4536

Table 6. Weight updates in hidden layer during each epoch of IIDDNN training in Predicted mode

Epoch Weight updates in hidden layer

no
W1 W2 W3 W4 W5
1 -1.5500 -1.5500 -1.5500 -1.5500 -1.5502
2 -1.5500 -1.5500 -1.5500 -1.5501 -1.5615
3 -1.5500 -1.5500 -1.5500 -1.5495 0.0412
4 -1.5500 -1.5500 -1.5500 -1.5517 -285.39
5 -1.5500 -1.5500 -1.5500 -1.6243 -328.24
 Improved Input-Delayed Dynamic Neural Network for GPS/INS Integration 671

Table 7. Comparison between Normal mode and Prediction mode of IIDDNN model

Position
components Normal mode training of IIDDNN model Prediction mode training of IIDDNN model
Number RMSE Performance Execution Number RMSE Performance Execution
of index Time(sec) of index Time(sec)
epochs epochs
Longitude 7 0.5428 0.0289 7.3 5 0.4567 0.03456 7.1
Latitude 7 0.5267 0.0278 8.6 5 0.4879 0.00321 8.3
Altitude 7 0.5098 0.0293 8.8 5 0.4673 0.00036 8.4

5 Conclusion
From the results and statistical analysis, the prediction mode training of IIDDNN
model shows superior performance in Root Mean Square Error, Performance index
and Execution time with less number of epochs when compared to the normal mode
training of IIDDNN model. So, IIDDNN is optimized neural network for GPS/INS
Integration in terms of error accuracy and training time even during the signal
blockages in GPS. This model relies on input delay elements at the input layer so that
the output INS position error is modeled based on the present and past samples of the
corresponding INS position. The impact of different delay elements present at the
input layer improves the overall positioning accuracy. So, this novel approach is
proposed to predict the accurate position of the moving vehicle particularly in GPS
absence.

References
1. Noureldin, A., Karamat, T., Eberts, M., El-Shafie, A.: Performance enhancement of
MEMS based INS/GPS integration for low cost navigation applications. IEEE Trans. Veh.
Technol. 58(3), 1077–1096 (2009)
2. El-Sheimy, N., Chiang, K.-W., Noureldin, A.: Developing a low cost MEMS IMU/ GPS
integration scheme using constructive neural networks. IEEE Trans. Aerosp. Electron.
Syst. 44(2), 582–594 (2008)
3. Malleswaran, M., Vaidehi, V., Angel Deborah, S., Manjula, S.: Integration of INS and
GPS using Radial basis function Neural Networks for Vehicular Navigation. IEEE Xplore
(2010), 978-1-4244-7815-6
4. Noureldin, A., El-Shafie, A., Taha, M.R.: Optimizing neuro-fuzzy modules for data fusion
of vehicular navigation systems using temporal cross-validation. Eng. Appl. Artif.
Intell. 20(1), 49–61 (2007)
5. Sharaf, R., Noureldin, A., Osman, A., El-Sheimy, N.: Online INS/GPS integration with a
radial basis function neural network. IEEE Aerosp. Electron. Syst. 20(3), 8–14 (2005)
6. Nassar, S., Noureldin, A., El-Sheimy, N.: Improving positioning accuracy during
kinematic DGPS outage periods using SINS/DGPS integration and sins data de-noising.
Surv. Rev. 37(292), 426–438 (2004)
7. Hiliuta, A., Landry, R., Gagnon, F.: Fuzzy corrections in a GPS/INS hybrid navigation
system. IEEE Trans. Aerosp. Electron. Syst. 2(2), 591–600 (2004)
672 M. Malleswaran, A. Saravanaselvan, and V. Vaidehi

8. Noureldin, A., Osman, A., El-Sheimy, N.: A neuro-wavelet method for multi-sensor
system integration for vehicular navigation. J. Meas. Sci. Technol., Lond., UK 15(2), 404–
412 (2004)
9. Noureldin, A., Irvine-Halliday, D., Mintchev, M.P.: Accuracy limitations of FOG-based
continuous measurement-while-drilling surveying instruments for horizontal wells. IEEE
Trans. Instrum. Meas. 51(6), 1177–1191 (2002)
10. Hirokawa, R., Nakakuki, K., Sato, K., Ishihara, R.: Threading the maze GPS/INS
landmark sensing and obstacle avoidance. GPS World, 20–26 (2004)
11. Malleswaran, M., Vaidehi, V., Angel Deborah, S., Manjula, S.: Comparison of RBF and
BPN Neural Networks applied in INS and GPS Integration for Vehicular Navigation.
International Journal of Electronics Engineering research 2(5), 753–763 (2010)
12. Malleswaran, M., Vaidehi, V., Angel Deborah, S.: Data fusion using MLFFNN for land
vehicle application. International Journal of Engineering Science and Technology 2(12),
7676–7690 (2010)
13. Dissanayake, G., Sukkarieh, S.: The aiding of a low-cost strapdown inertial measurement
unit using vehicle model constraints for land vehicle applications. IEEE Trans. Robot.
Automat. 17(5), 731–747 (2001)
14. Sharaf, R., Tarbouchi, M., El-Shafie, A., Noureldin, A.: Real-time implementation of
INS/GPS data fusion utilizing adaptive neuro-fuzzy inference system. In: Proceedings of
the Institute of Navigation National Technical Meeting – ION NTM 2005, San Diego, CA,
January 24-26 (2005)
15. Reda Taha, M.M., Noureldin, A., El-Sheimy, N.: Improving INS/GPS positioning
accuracy during GPS outages using fuzzy Logic. In: Proceedings of the 16th International
Technical Meeting of the Satellite Division of the Institute of Navigation ION GPS/GNSS
2003, September 9-12 Oregon Convention Center, Portland (2003)
16. Lobo, J., Lucas, P., Dias, J., Traca de Almeida, A.: Inertial navigation system for mobile
land vehicles. In: Proceedings of the IEEE International Symposium on Industrial
Electronics, ISIE 1995, vol. 2, pp. 843–848 (July 1995)
17. Mynbaev, D.K.: Errors of an inertial navigation unit caused by ring laser gyro errors. In:
Proceedings of the IEEE Position Location and Navigation Symposium, pp. 833–838
(1994)
18. El-Sheimy, N., Schwarz, K.P.: Integrating differential GPS receivers with an inertial
navigation system (INS) and CCD cameras for a mobile GIS data collection system. In:
Proceedings of the International Society for Photogrammetry and Remote Sensing
Conference, Ottawa, Canada, pp. 241–248 (1994)
19. Wong, R.V.C., Schwarz, K.P., Cannon, M.E.: High-accuracy kinematic positioning by
GPS-INS. Navig.: J. Inst. Navig. 35, 275–287 (Summer 1988)
20. Sivanandam, S.N., Sumathi, S., Deepa, S.N.: Introduction to Neural Networks using
Matlab 6.0. Tata McGraw-Hill Publication, New York (2002)
21. Bishop, C.M.: Neural Networks for Pattern Recognition. Oxford University Press, New
York (1995)
 Application of Artificial Neural Network for the
Prediction of Groundwater Level in Hard Rock Region

M. Kavitha Mayilvaganan1 and K.B. Naidu2

1
Research Scholar
2
Professor
Department of Mathematics,
Sathyabama University, Chennai
[email protected]

Abstract. Artificial neural network has been shown to be an efficient tool for
non-parametric modeling of data in a variety of different contexts where the
output is the nonlinear function of inputs. Neural networks are the preferred tool
for many predictive data mining applications because of their power, flexibility
and ease of use. In the present study Feed-Forward Network based Artificial
Neural Network (ANN) model is used as a method to predict the groundwater
levels. The ANN model was trained using back propagated algorithm with two
hidden layer, and with logsig activation function. The models are evaluated
using three statistical performance criteria namely Mean Average Error (MAE),
Root Mean Squared Error (RMSE) and Regression coefficient (R²). The results
of the study clearly showed that ANN can be used to predict groundwater level
in a hard rock region with reasonably good accuracy even in case of limited
data situation.

Keywords: Artificial Neural Networks Model, Back Propagation Algorithm,

MATLAB, Observation Well, Ground Water level.

1 Introduction

In this paper, groundwater level has been estimated for the hard rock aquifer system
of a small region of Tamilnadu named Thurinjapuram watershed using the method of
ANN. It is an information processing paradigm that is inspired by the way biological
nervous systems, such as the brain processes information. It determines the
relationship between inputs and outputs of physical systems by a network of
interconnecting nodes adjusted by connecting weights based on the training samples,
and extracts patterns and detects trends that are too complex to be noticed by either
humans or other computational techniques. Neural network takes a different approach
to problem solving than that of conventional computers. It has remarkable ability to
derive meanings from complicated and imprecise data. It has an ability to learn and
apply the knowledge based on the data given for training or initial experience
[1].Moreover, it validates the learning model with a chosen validation set from the
input data to achieve a desired level of accuracy. The accuracy could be maximized
by repeating the training process with different topological model [2]. Recently, ANN

D. Nagamalai, E. Renault, and M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 673–682, 2011.
© Springer-Verlag Berlin Heidelberg 2011
674 M. Kavitha Mayilvaganan and K.B. Naidu

has been used for forecasting groundwater level successfully [3-5]. Thus proper
design and implementation of ANN models could provide an efficient tool in water
resource management and forecasting. Moreover, ANN could be employed to predict
various aspects in hydrology such as rainfall-runoff model, precipitation forecasting
and water quality model [6].

2 Artificial Neural Network (ANN)

ANN based methods are data analysis methods and algorithms loosely based on
nervous systems of humans and animals. In this study, feed forward network based
neural network is used for the input data set to make the future predictions. In this
network, the information moves in only forward direction from the input nodes to the
output nodes, through the hidden nodes. There are no cycles or loops in the network.
Input nodes are the resultant where identified weight functions (depending on the
input data set) are applied to the inputs. Based on the topology it traverses through
multiple intermediate weight functions before concluding to a resultant output
(represented as output node). All these intermediate weighted resultants are known as
hidden nodes and the cumulative application of weight function decides the number of
hidden layers (depending on the topology)

Fig. 1. Structure of feed-forward ANN

In this paper, Levenberg–Marquardt algorithm (LMA) has been used to train the
ANN and it has surfaced as the best model resulting high level of accuracy and
efficiency with the given input data. [7]. Next is Training Parameters Selection. In this
phase, training algorithm’s parameters like quick propagation coefficient, learning
rate, momentum and a condition to stop the network training process, mean square
error and minimum improvement has been decided. Error is the kind of conditions
based on which training process are said to be reached at the permissible level for
forecasting. In this phase the ANN validates its’ authenticity forecasting from the
training subset and comparing it with the validation subset. Once it achieves this, it
certifies itself for forecasting.
 Application of Artificial Neural Network for the Prediction of Groundwater Level 675

3 Study Area and Data

The Thurinjapuram watershed covers geographical area of 151.38 sq. km and is
located in between 12o12’58” and 12o 21’11” North latitudes and 78o59’45” and
79o9’28” East longitudes (Fig. 2.) It is mainly situated in Thiruvannamalai district of
Tamilnadu, India. It is mainly located in Thurinjapuram block and partially falls into
two other blocks (Chengam and Thiruvannamalai). Thurinjalar is one of the major
tributaries of Ponnaiyar Major River originating from Kavuttimalai reserve forest in
Chengam Taluk of Tiruvannamalai district. Thurinjalar River, which is the major
stream draining the area, exhibits only sporadic flow during the rainy season. The
drainage characteristics are very good. Bedrock is peninsular gneiss of Archean age.
The Thurinjapuram area can be classified as ‘‘hard rock terrain’’. The predominant
soil types in this river basin are Entiso, Inceptisols, Vertisol and Alfisols. The soil in
this minor basin is observed to have good infiltration characteristics. Hence
groundwater recharge is possible in this area. The economy of the Thurinjapuram sub
watershed depends mainly on agriculture.

Fig. 2. Study area

The input data used for water level prediction are monthly Rainfall and Ground
water (level in the observation well) data of Thurinjapuram watershed in Tamilnadu,
India, and one month ahead groundwater level as output. For the present study
monthly water level data for three observation wells (23112, 23142, and 23143)
during 1985 to 2008 has been collected from Thiruvannamalai Groundwater
subdivision. In the same period monthly Rainfall data were collected from
Kilnatchipattu Raingauge station.
676 M. Kavitha Mayilvaganan and K.B. Naidu

4 Methodology
There is no fixed theory for the selection of a particular network type or various
internal parameters within it. The various parameters, which affect the performance of
network, are number of hidden nodes, number of hidden layers, learning rate and
momentum factor. The ANN architecture consist of an input layer with two nodes, a
hidden layer with varying nodes and one output layer with single node, thus leading to
a multi input single output network. The input data are the rainfall the ground water
levels during time ‘t’, ‘t-1’ and ‘t-2’ etc and output data is ground water level during
time ‘t+1’. The data length of 276 data sets has been divided into 70% for training and
30% for testing. Thus the ANN model was trained using 192 randomly selected data
sets while the testing set belonged to remaining data sets. The closeness of predictions
with the actual observations was judged quantitatively by drawing a scatter plot and a
time series comparison plot and quantitatively by evaluating three error measures,
namely, Regression coefficient, mean average error and root mean square error.

5 ANN Model
Neural network with single hidden layer was trained with different number of
neurons. Its regression percentage and performance values were given in Table1.

5.1 The Effect of Different Number of Neurons Per Layer

At 32 neurons per layer the data had a good enough fit, with an increase of neurons
above this level having no real effect on the prediction anymore, see Table 1. In order
to get the best fit and resolution a range of 30 to a 35 were found to be optimal
number of neurons in two layer neural networks.

Table 1. Training results for two layer neural networks with different number of neurons.

Number of Linear regression Performance value

neurons percentage
10 48 0.01796
15 63 0.01094
20 77 0.00701
25 82 0.00660
30 86 0.00581
32 88 0.00508
35 85 0.00683
40 79 0.00695
50 73 0.00723
60 68 0.00755

5.2 Number of Layers in the Neural Network

The influence of the number of layers was also investigated. These results indicate
that the 3 layer network was better than the 2 layer network, while 4 and more layers
 Application of Artificial Neural Network for the Prediction of Groundwater Level 677

gave no results. This allows for neural networks to be constructed for 3 layer neural
networks, since 2 layer networks have less resolution than 3 layer neural networks.
Table 2.presents the results of the number of layers.

Table 2. Result of the ANN for different number of layers

Number of Layers Linear regression Performance value

percentage (Reg %)
2 88 0.00508
3 94 0.00581
4 No Fit No Fit

Fig. 3. Feed forward neural network with two hidden layers

5.3 The Effect of Transfer Function on Data Estimation

Three layer networks was constructed with different transfer functions contained
within it to determine the stability of the neural network. The results are given below
in Table 3.
Table 3. Result of ANN for different activation function

No of neurons Activation fun RMSE Reg % performance

12-20-1 Tansig 0.5832 94 0.00581
12-20-1 Logsig 0.4665 96 0.00463

The difference can be observed from the average fit value in Table 3, in which the
three layer log sigmoid neural network performed the best (Fig3).
Different neural network architectures were developed in order to establish a
relationship between the input and output. All the networks were of the feed forward
type. The network architectures were trained by varying the number of hidden layer
and then by varying number of neurons in each hidden layer.
The momentum factor and learning rate were also fixed by trial and error process.
Momentum factor of 0.9 with learning rate of 0.6 were found better for all networks.
Also for the given data log sigmoid as an activation function performed better than the
other activation functions. Number of training epochs given were 2000 to 3000 early
stopping methods was applied, and error goal was set to be 0.001( Fig 4(a)(b)).
678 M. Kavitha Mayilvaganan and K.B. Naidu

Fig. 4(a). Neural network training tool Fig. 4(b). Performance plot

6 ANN vs. Linear Regression (LR)

As a benchmark for ANN performance, LR was also conducted on the same data sets
with the identical input and output variables. Pearson’s correlation coefficient test (2-
tailed) was used to test correlations between rainfall and groundwater level . Water
level was estimated by a linear curve estimation regression model of SPSS 13.0. All
the statistical analyses were conducted with SPSS 13.0 for Windows. A comparison
in predictive accuracy between the two methods is shown in Fig 5.

2.5

2
1.5

1
0.5
0
ANN LR ANN LR ANN LR
23112 23141 23143

M AE 0.5 1.6 0.17 1.9 0.53 1.68

RM SE 0.84 1.91 0.47 2.38 0.82 2.11
R² 0.86 0.27 0.96 0.11 0.91 0.39

Fig. 5. Comparison of ANN and LR error for all the three wells
 Application of Artificial Neural Network for the Prediction of Groundwater Level 679

7 Results and Discussion

A quantitative evaluation of the results was made by deriving the correlation
coefficient for various network architectures. Large numbers of ANN models has
been developed based upon type of network, network architecture, training algorithm,
transfer function, momentum rate, training and testing data set etc. among these
models a model of two hidden layer with 12 and 20 nodes performed better .
High R² value, small MAE and RMSE values of ANN shows that the model fits the
data well. Analysis of data in randomized sets ( Fig 5) clearly showed that ANN
model is best fit for predicting the groundwater level than the LR model. Since
statistical regression is problematic if there is vagueness in the relationship between
the independent (input) and the dependent variables (output).

Well No.23112
Actual
ANN Predicted
Groundwater level (m)

12.00
10.00
8.00
6.00
4.00
2.00
0.00
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86
Number of Data (Months)

(a)

Well No.23112
Estimated Ground Water

12
10
Level (m)

8
2
6 R = 0.8596
4
2
0
0.00 2.00 4.00 6.00 8.00 10.00 12.00
Observed Ground Water Level (m)

(b)

Fig. 6(a), (c), (e) Comparison of predicted and actual values of groundwater level.6 (b), (d), (f)
Scatter Plot between Observed and Predicted groundwater level of the wells
680 M. Kavitha Mayilvaganan and K.B. Naidu

Well No. 23141 Actual

Groundwater level (m) ANN Predicted
1.2
1
0.8
0.6
0.4
0.2
0
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81
Number of Data (Months)

(c)

Well No. 23141

Estimated groundwater

12
10
level (m)

8
2
6 R = 0.9661
4
2
0
0 2 4 6 8 10 12
Observed groundwater le vel (m)

(d)

Well No.23143 Actual

ANN Predicted
Groundwater level (m)

15.00

10.00

5.00

0.00
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81
Number of Data (Months)

(e)

Fig. 6. (Continued)
 Application of Artificial Neural Network for the Prediction of Groundwater Level 681

Well No. 23143

Estimated Ground Water 16

14
12
Level (m)
10
8 2
6 R = 0.9122
4
2
0
0.00 2.00 4.00 6.00 8.00 10.00 12.00 14.00 16.00
Observed Ground Water Level (m)

(f)

Fig. 6. (Continued)

Further, the data were analyzed separately for each independent well point to have
a clear comparison of the mean observed and estimated water levels. Here also, the
results exhibited similar trend as in the case of set-wise data analysis, i.e. ANN model
was found to be the best fit for predicting groundwater levels at the Thurinjapuram
watershed. The performance trend of the model, individually for each of the three
wells along with observed and simulated values indicated in Fig 6 showing that the
difference between observed and simulated values of groundwater levels with more
accuracy and stability over conventional methods. Scatter plot between observed and
predicted groundwater level is also given for the goodness of fit.

8 Conclusion
In this study, a better forecasting model using ANN has been developed for predicting
monthly groundwater level fluctuations in the Thurinjapuram watershed. As evident
from Fig 6 there have been few deviations of the predicted groundwater level value
from the actual. In conventional regression analysis, deviations between observed and
estimated values are due to indefiniteness of structure of a system or imprecise
observations. As rainfall and groundwater level predication involves tremendous
amount of imprecision and uncertainty, soft computing technique might warrant the
ideal prediction model. On the other hand, the proposed prediction models based on
soft computing are easy to implement and produces desirable mapping function by
training on the given dataset. The model requires information only on the input
variables for generating forecasts.
ANN configurations were tested in terms of optimum results for a prediction
horizon of 84 months. The most suitable configuration for this task proved to be a 12-
20-1 feed forward network trained with the Levenberg–Marquardt method as it
showed the most accurate predictions of the decreasing groundwater levels. From the
results of the study it can also be inferred that the Levenberg–Marquardt algorithm is
682 M. Kavitha Mayilvaganan and K.B. Naidu

more appropriate for this problem. In general, the results of the case study are
satisfactory and demonstrate that neural networks can be a useful prediction tool in
the area of groundwater hydrology.

Acknowledgement. The First author gratefully acknowledges the support given by

DST Govt. of India, under women scientist scheme.

References
1. Caulibaly, P., Anctil, F., Bobee, B.: Multivariate reservoir inflow forecasting using temporal
neural networks. J. Hydrol. Eng. (9-10), 367–376 (2001)
2. Prochazka, A.: Seasonal time neural networks and –series prediction by artificial neural
networks. In: Fifth International Conference, Digital Object Identifier, vol. 440(7-9), pp.
36–41 (1997)
3. Coppola, E.J., Anthony, J.R., Poulton, M., Szidarovszky, V.W.: A neural network model for
predicting aquifer level elevations. Groundwater 43(2), 231–241 (2005)
4. Nayak, P.C., Satyaji, U.R., Sudheer, K.P.: Groundwater level forecasting in shallow aquifer
using artificial neural network approach. Water Resour. Manage. 20, 77–90 (2006)
5. Daliakopoulos, I.N., Coulibaly, P., Tsanis, K.I.: Groundwater level forecasting using
artificial neural network. J. Hydrol. 229, 229–240 (2005)
6. Govindaraju, R.S., Ramachandra, R.A.: Artificial neural networks in hydrology. Kluwer,
Netherlands (2000)
7. Banerjee, P., Prasad, R.K., Singh, V.S.: Forecasting of groundwater level in hard rock
region using artificial neural network. Environ. Geol. 58, 1239–1246 (2009)
 Energy Efficient and Fault Tolerant GPSR
in Ad Hoc Wireless Network

Jyotsana Jaiswal and Pabitra Mohan Khilar

Department of Computer Science and Engineering,

National Institute of Technology Rourkela, India
[email protected], [email protected]

Abstract. Routing in wireless network is a key research area which es-

tablishes path between source and destination node pairs. In this paper,
we have designed and evaluated an energy-efficient and fault tolerant
Greedy perimeter Stateless routing (EFGPSR) protocol for wireless Ad
hoc network. The proposed protocol is divided into four phases: Fault
testing phase, Planarization phase, Energy efficient greedy forwarding
phase and Energy efficient perimeter forwarding phase. In fault testing
phase, all nodes come to know about their fault free neighbours. Next is
planarization phase which is subset of perimeter forwarding phase and
can be done reactively or proactively. Next are energy efficient greedy
forwarding and perimeter forwarding phases. Both these phases try to
maintain balance between the metrics to choose the next hop (i.e. dis-
tance from destination in greedy forwarding phase and minimum angle
w.r.t the line connecting the forwarding node and destination in perime-
ter forwarding phase) and selection of node having highest energy among
the neighbouring node to extend network lifetime. Evaluation and com-
parison of GPSR and EFGPSR is done through NS-2 simulator. Simu-
lation shows that EFGPSR performs better in terms of increasing the
network lifetime, successful packet delivery ratio with insignificant in-
crease in number of hop count.

Keywords: Wireless network, geographic routing, fault tolerance, GPSR,

planarization.

1 Introduction
In recent years, wireless and mobile communication technologies have grown
rapidly. As wireless network evolves there is a trend towards decentralized, de-
ployable and self organizing networks.
Routing plays a vital role in establishing a communication links between var-
ious nodes. Since last few years, geographic routing (GR) protocols, also known
as georouting or position based routing for wireless networks has gained a sig-
nificant attention. The idea behind GR protocol is that the source node sends a
packet to destination node using the geographic location of destination instead
of using network address. The essential requirements of GR are a) each node
should be capable of determining its own location (geographic coordinates) and

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 683–692, 2011.

c Springer-Verlag Berlin Heidelberg 2011
684 J. Jaiswal and P.M. Khilar

b) source should be aware of the location of destination. The main advantage

of GR over traditional routing protocol is that each node require to maintain
only location of itself and its neighbours for its functioning. However, in tradi-
tional routing protocols for wireless networks (e.g. AODV [1], DSDV [2]), nodes
usually have to keep significant amount of routing information. Greedy Perime-
ter Stateless Routing (GPSR) [3] is a well known and most commonly used
position-based routing protocol for wireless networks. In GPSR, source includes
the location information of destination in the header of every packet. If the des-
tination is not directly reachable, the source starts with greedy forwarding, i.e.,
the source node forwards the data packet to the neighbour that is closest to
the destination in the coordinate space. Such greedy forwarding is repeated at
the intermediate node, until the destination is reached. However, GPSR itself
suffers from few drawbacks. Firstly, greedy forwarding over geographic coordi-
nates may not be optimal due to unawareness of connectivity information of
the network. Secondly, packet may get stuck in local minimum condition (a for-
warding node could not find a neighbour that lies closer to the destination than
itself) for sparse networks. To deal with above stated local minimum problem,
the nodes switch to perimeter forwarding. In perimeter forwarding, the packet
is forwarded to that neighbouring node which comes first in a planar sub graph
of the network , when the line connecting the forwarding node and destination
is rotated in the counter clockwise direction. The location of forwarding node
where perimeter forwarding starts is recorded in the header of the data packet.
Greedy forwarding is resumed when the data packet reaches a forwarding node
which can find a neighbour node whose distance is smaller than the distance
between the destination node and the node at which perimeter forwarding be-
gun. The probability of finding a route between source destination node pairs is
very high.
Most existing designs of wireless network routing protocols are based on the
assumption that every node in the network is fault free. However, such assump-
tion usually does not hold in realistic environments. Various kinds of faults such
as crash fault, transient fault, Byzantine faults can occur in a node. The occur-
rence of faults affects the routing process therefore fault tolerance is of increasing
importance in applications where it is essential to maintain efficient routing. In
routing, fault tolerance helps in controlling the overhead which is there due to
the faulty node and thus helps in considering the reliable routes. Many fault
tolerance routing protocols have been proposed in wireless network. In Wire-
less sensor network(WSN), fault tolerance mechanism in routing protocol has
been classified into two schemes, retransmission and replication. In MANET,
works that focus on fault tolerant routing problem are very less. Wireless net-
work comprises of large number of energy constraint node, some node may run
out of energy and die and shorten the lifetime of network. Therefore, many
routing protocols in MANET and WSN have been proposed with the aim to
reduce energy consumption. GPER protocol was proposed to provide power effi-
cient geographic routing in WNS. Traditionally, greedy forwarding and perimeter
 Energy Efficient and Fault Tolerant GPSR in Ad Hoc Wireless Network 685

forwarding consider only single metrics i.e. minimum distance from destination
and minimum angle in counter clockwise direction with respect to the line con-
necting the forwarding node and destination. These parameters does not take
into account the energy conservation for evaluating the routing performance.
In this paper, an energy efficient and fault tolerant routing algorithm has been
proposed. The rest of the paper is organized as follows: The system models are
described in Section 2. The energy-efficient and fault tolerant GPSR routing
protocol is presented in Section 3. Section 4 presents the simulation result and
comparison with other existing protocol. Section 5 conclude the paper and lists
the future works.

2 Preliminaries
2.1 System, Fault and Energy Model
System Model. The system is composed of homogeneous n nodes, each having
a unique id. All nodes have similar computing, storage resources and identical
communication range. A set of nodes with circular radio range r, can be seen as
a graph: each node is a vertex, and edge (n, m) exists between nodes n and m if
the distance between n and m, d(n, m) ≤ r. Source node knows the location of
final destination.

Fault Model. Wireless network consists of many nodes which may be either
fault free or faulty. The following assumptions have been made about faulty
nodes:
1. No new faults occur during the execution of the routing protocol i.e. faults are
permanent (a faulty node remains faulty until it is repaired and/or replaced).
2. Faulty nodes can be either hard faulty or soft faulty.
Nodes become hard faulty due to two main reasons. First, a node may be dam-
aged during deployment or immediately after that. Second, depletion of battery
power. When a node becomes hard faulty it does not participate in any further
communication. Soft faulty nodes are more subtle than hard faulty nodes, since
a soft-faulted node continues to communicate with the other node in the sys-
tem although with altered specifications i.e., the faulty nodes may produce some
random results instead of expected results. In this work, both hard-faulted and
soft-faulted nodes have been considered.
The proposed routing protocol apply the following testing model given in Ta-
ble 1 to check whether the node is faulty or fault free. The testing model uses
the test task to check the validity of the node. The test task is to find out two’s
complement of five bit number. Tester node knows the result of the test task.
Tester node broadcast the test task along with its id and location information.
On receiving the test task, all the neighbouring node unicast the result of the
686 J. Jaiswal and P.M. Khilar

test task along with their id and location information. As shown in the table of
test model, five cases are observed. First case, the status of the tester node and
tested node are both fault free then test result is zero. This means that there
is match between the expected result of the tester node and the actual result
returned by the tested node. Second case, the tester node is fault free and the
tested node is soft faulty then test result is 1. This means there is a mismatch
between the expected result and actual result. Third case, the tester node is fault
free or soft faulty and the tested node is hard faulty then test result is NULL
because a hard faulty node can receive a beacon test message but cannot send
reply to it. Fourth case, the tester node is soft faulty and tested node is soft
faulty or fault free then test result is 1. Here also there is mismatch between the
expected result and actual result. Fifth case, the tester node is hard faulty and
tested node is also faulty (soft or hard) or fault free then test result is NULL
because a hard faulty node cannot send a beacon test message.
When the test result is 0, the node is considered as fault free, otherwise
faulty. The test model given in Table 1 is used to select fault free nodes while
establishing the paths between the nodes pair.

Table 1. Test model

Status of Status of Test

Tester node Tested node Result
Fault free Fault free 0
Fault free Soft faulty 1
Fault free Hard faulty NULL
Soft faulty Hard faulty NULL
Soft faulty Soft faulty 1
Soft faulty Fault free 1
Hard faulty Soft faulty NULL
Hard faulty Hard faulty NULL
Hard faulty Fault free NULL

2.2 Energy Model

In wireless network energy consumption at each node can be due to transmission
and reception of message. The energy model proposed by Heinzelman et al [6]
to transmit an n-bit message over a distance d or receive an n-bit message are
as follows:

ET X (m, d) = ET X−elec (m) + ET X−amp (m, d) (1)

= m ∗ Ee lec + m ∗ εamp ∗ d 2
(2)
ERX (m) = ERX−elec (m) = m ∗ Eelec (3)

Table 2 gives related parameters and their definations.

 Energy Efficient and Fault Tolerant GPSR in Ad Hoc Wireless Network 687

Table 2. Radio parameters

Parameter Definition Unit

Eelec Energy dissipation 50nJ/bit
rate to run the radio
εamp Energy dissipation 100pJ/bit/m2
rate to run transmit amplifier

3 The Proposed Algorithm

The proposed algorithm consists of four different but dependent phases which
are as given below:

3.1 Fault Testing Phase

In this phase, all nodes come to know about the location information and residual
energy information of their fault free neighbours at regular interval. Each tester
node broadcast a beacon-test message containing its own id (e.g. IP address),
location information and test task. Information regarding the test task in given
in 2.1. Each tester node knows the result of the test task. On receiving the
beacon-test message, all the neighbouring node execute the test task and unicast
the reply message. The reply message contains the result of the test task, their
unique Id, location information and residual energy information of the node. In
this phase, tester node checks whether the neighbouring nodes present are fault
free or faulty (i.e., soft or hard) by comparing the result of test task using the
test model given in Table 1. Each node maintains the location information and
residual energy information of all fault free neighbouring nodes in their range.

3.2 Planarization Phase

After fault testing phase, planarization phase starts. This phase can be done
reactively or proactively. In proactive version, each node begins the planarization
phase immediately after the fault testing phase. In reactive version, each node
begins the planarization phase when local minimum condition occur. The main
reason for using the planarization phase is that the right-hand rule does not
work properly on full connected graphs with crossing edge. The right hand rule
is used by perimeter forwarding phase. Each node should run the planarization
algorithm in a distributed fashion. The Relative Neighbourhood Graph (RNG)
and Gabriel Graph (GG) are two well known planar graph [7]. RNG planarization
has been considered in this paper. During this phase one important property to
be taken care is that, while removing edges from the graph to reduce it to RNG
must not disconnect the graph.
688 J. Jaiswal and P.M. Khilar

3.3 Energy Efficient Greedy Forwarding Phase (EEGF)

In this phase, a forwarding node F uses full graph in choosing next hop. The
optimal choice of next hop considers two metrics first minimum distance from
destination(D) and second the energy. Algorithm 1 explains the pseudo code for
the proposed energy efficient greedy forwarding used at a forwarding node. In
traditional GPSR, the distance between forwarding node as well as its neigh-
bour with respect to destination is calculated. The forwarding node selects that
neighbour as the next hop that lies closest to the destination. If the forward-
ing node F could not find a neighbour node that lies closer to the destination
than itself, then node switches to perimeter forwarding. However in EEGF, for-
warding node first form a set of Selected-Neighbour(F ) which is a subset of
the Neighbour-list(F ). Each node I ε Selected-Neighbour(F ), if the distance be-
tween the neighbour node I and the destination node D is less than the distance
between the forwarding node F and D. After this, Cost(I) of every node I ε
Selected-Neighbour(F) is calculated. Cost(I) is defined as
ResidualEnergy(I)
Cost(I) = 2 (4)
Distance(F, I)
where, ResidualEnergy(I) is the AvailableEnergy(I) divided by InitialEnergy
(I) and Distance(F, I) is the difference in the distance between F And D and
the distance between I and D divided by the distance between F and D. The
reason for calculating this cost is that relation between energy and distance are
inversely proportional to each other because from energy balance point of view,
the node with more residual energy should be selected as the next hop and
from distance point of view, the node which is closer to the destination should
be selected as the next hop. Thus, Cost(I) balances the two metrics energy
and distance and helps in finding the optimal next hop. The forwarding node
selects that neighbour I ε Selected-Neighbour(F ) as next hop, whose Cost(I) is
maximum. If the forwarding node F could not find a neighbour node that lies
closer to the destination than itself, that is Selected-Neighbour(F ) is empty then
node switches to perimeter forwarding.

(a) (b)

Fig. 1. (a) Greedy forwarding and (b) Local minima condition

 Energy Efficient and Fault Tolerant GPSR in Ad Hoc Wireless Network 689

Require: Destination D, Forwarding node F , Neighbour-list(F )

Ensure: Next-Node // if Energy efficient greedy forwarding is successful
Switch to perimeter forwarding // if Energy efficient greedy forwarding is not
successful
Intialization: Next-Node=NULL; Maximum-Cost=0.0;
Selected-Neighbour(F)=0

1. Dist(F-D) = (xF − xD )2 + (yF − yD )2
2. while neighbour-node
 I ε Neighbour-list(N ) do
3. Dist(I-D) = (xI − xD )2 + (yI − yD )2
4. if Distance(I-D) < Dist(F-D) then
5. Selected-Neighbour(F ) → Selected-Neighbour(F )∪(I)
6. end if
7. end while
8. while neighbour-node I ε Selected-Neighbour-list(N ) do
9. ResidualEnergy(I ) = AvailableEnergy(I
InitialEnergy(I )
)

Dist(F −D)−Dist(I−D)
10. Distance(F, I) = Dist(F −D)
ResidualEnergy
Cost(I) = Distance(F,I)2

11. if Maximum-Cost < Cost(I) then

12. M aximum − Cost = Cost(I)
13. N ext − N ode ← I
14. end if
15. end while

Algorithm 1. Energy-Efficient Greedy Forwarding Algorithm

3.4 Energy Efficient Perimeter Forwarding

Energy efficient perimeter forwarding uses right hand rule/left hand rule as given
below:

Right Hand Rule. The right-hand rule to traverse a graph is shown in

Figure2 (a). This rule states that when arriving at node b from node a, the
next edge traversed is the next one sequentially counterclockwise about b from
edge (a, b). On graphs with edges that cross, the right-hand rule may take a
degenerate tour of edges that does not trace the correct path as shown in figure
2 (b). Now on removing the crossing edges right hand rule traces the correct
path (Figure 2 (c)). In this context planarization comes into picture. The left
hand rule also works in a similar fashion the only difference is that the next edge
traversed is in clockwise direction

Energy Efficient Perimeter Forwarding. In this phase, complete EFGPSR

is described which combines fault testing phase, energy efficient greedy forward-
ing phase and energy efficient perimeter forwarding phase. Greedy forwarding
phase is executed on the full network graph whereas energy efficient perime-
ter forwarding phase on the planarized network graph where greedy forwarding
690 J. Jaiswal and P.M. Khilar

(a) (b) (c)

Fig. 2. (a) Right Hand Rule (RHR), (b) RHR with crossing edge, (c) RHR without
crossing edge

Fig. 3. Perimeter forwarding example

fails. A flag is used in the packet header of EFGPSR which indicate whether
the packet is in greedy mode or perimeter mode. Initially, source node marks all
data packets as greedy-mode. On receiving greedy-mode packet, a forwarding
node uses the energy efficient greedy forwarding algorithm to find the optimal
next hop. If no optimal node is found i.e. greedy forwarding fails, the node marks
the packet into perimeter mode. In traditional perimeter forwarding as shown
in figure 3 the packet is forwarded to that neighbour node that comes first in a
planar sub graph of the network , when the line connecting the forwarding node
and destination(D) is rotated in the counter clockwise direction. The location of
forwarding node where perimeter forwarding starts is recorded in the header of
the data packet. Greedy forwarding is resumed when the data packet reaches a
forwarding node or a node which can find a neighbour node whose distance is
smaller than the distance between the destination node and the node at which
perimeter forwarding begun. However, in energy efficient perimeter forwarding,
the forwarding node select the next hop using the following algorithm:

4 Simulations and Results

In this paper, the performance of GPSR and EFGPSR is evaluated based on NS-2
network simulator. The network dimension used for simulation is 1000×1000meter
square. The transmission range of each node is assumed to be 250m. For traffic
source 20 traffic flow originated by 20 sending nodes has been considered. The
MAC layer protocol used is IEEE 802.11. Simulation results are shown in figure 4
and figure 5.
 Energy Efficient and Fault Tolerant GPSR in Ad Hoc Wireless Network 691

Require: Forwarding node F , Neighbouring-node-list containing index, angle and

energy of each node
Ensure: Next-Node
Intialization: Next-Node=NULL; Temp=0; Max=Max-energy;
Min=Min-energy; Mid=Mid-energy
1. Sort the Neighbouring-node-list in increasing order of angle
2. for all Neighbouring-node
3. Temp=First nodes after sorting
4. if Temp[energy] < Mid and Temp[energy] >= Max then
5. Select node index
6. break
7. end if
8. End for

Algorithm 2. Energy-Efficient Perimeter Forwarding Algorithm

0.99
GPSR
0.988 EFGPSR
Fraction of packet delivered

0.986

0.984

0.982

0.98

0.978

0.976
20 40 60 80 100 120
Pause time (ms)

Fig. 4. Packet Delivery Success Rate. GPSR and EFGPSR with becon interval 1. 50
nodes
100
GPSR
FGPSR
90
Percentage of paths established

80

70

60

50

40

30

20
0 5 10 15 20 25 30 35 40
Pecentage of faulty nodes

Fig. 5. GPSR vs EFGPSR

692 J. Jaiswal and P.M. Khilar

5 Conclusions

In this paper, fault tolerant GPSR routing algorithm has been proposed. EFG-
PSR and GPSR have been simulated for the 1000×1000-sized static networks
of 50 node. 20 traffic flow originated by 20 sending nodes has been considered.
The result carried out shows that EFGPSR provides higher no. of path between
source and destination out of the total no. of paths (i.e., 20) than GPSR due to
the fact that GPSR does not check the faulty node which causes long detouring
paths and may not establish path to the destination shown in figure 5. And
also EFGPSR try to maintain balance between the metrics to choose the next
hop (i.e. distance from destination in greedy forwarding phase and minimum
angle with respect to the line connecting the forwarding node and destination in
perimeter forwarding phase) and selection of node having highest energy among
the neighbouring node to extend network lifetime. Evaluation and comparison of
GPSR and EFGPSR shows that EFGPSR performs better in terms of increasing
the network lifetime, successful packet delivery ratio with insignificant increase
in number of hop count.

References
1. Perkins, C., Belding-Royer, E., Das, S.: Ad hoc On-Demand Distance Vector
(AODV) Routing, RFC Editor (2003)
2. Charles, E., Perkins, B.P.: Highly dynamic Destination-Sequenced Distance-Vector
routing (DSDV) for mobile computers. SIGCOMM Comput. Commun. Rev. 24,
234–244 (1994)
3. Karp, B., Kung, H.: GPSR: greedy perimeter stateless routing for wireless networks.
In: Proceedings of the 6th Annual International Conference on Mobile Computing
and Networking, Boston, MA, USA, pp. 243–254 (August 2000)
4. Wu, S., Candan, K.: GPER: Geographic Power Efficient Routing in Sensor Network.
In: Proceeding of Int’l Conference Network Protocols (ICNP), pp. 161–172. IEEE,
Los Alamitos (2004)
5. Melodia, T., Pompili, D., Akyildiz, I.: Optimal Local Topology Knowledge for En-
ergy Efficient Geographical Routing in Sensor Network. In: Proceeding of INFO-
COM. IEEE, Los Alamitos (2004)
6. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: Energy-Efficient Communica-
tion Protocol for Wireless Microsensor Networks. In: Proceedings of the 33rd Hawaii
International Conference on System Sciences, vol. 8. IEEE Computer Society, Hawai
(2000)
7. Agarwal, P., Matausek, J.: Relative neighborhood graphs in three dimensions. In:
Proceedings of the Third Annual ACM-SIAM Symposium on Discrete Algorithms,
pp. 58–65. Society for Industrial and Applied Mathematics, Philadelphia (1992)
 An IDE for Developing Mobile Applications for
Smart Phones

Sakila Banu Ali and Nisar Hundewale

College of Computers and Information Technology,

Taif University,
Saudi Arabia
[email protected],
[email protected]
http://cit.tu.edu.sa

Abstract. Mobile Devices have been gaining increasing acceptance for

developing Multimedia-rich applications. Everyday new Technologies in
Mobile Devices have been developed by developers and are compet-
ing with the Previous in Market. Nowadays developing applications for
Smart phones is common today and nearly thousands of new applications
are coming to the Market every day. Software firms have also developed
many programming platforms and tools for the developers to write pro-
grams for Mobile Devices. But so far there is no Specific IDE developed
to create mobile application easily by just Drag and Drop method to
make even the non-programmers to develop application for the smart
phones.
This paper presents an IDE for developing Mobile Application without
Programming for Smart phones. The IDE allows non-programmer to
develop a mobile application for smart phones efficiently. The IDE has
navigation control tools in iconic format that can be used to incorporate
within the application by drag and drop method and directly installed to
the smart phones as applications. With the help of the GUI Builders and
ADT plug-in the IDE is created and tested against the Android Mobile
Phones.

Keywords: IDE-Integrated Development Environment, GUI-Graphical

User Interface, ADT-Android Development Tool.

1 Introduction

The Mobile Applications are classified into the following types. They are Com-
munications, Games, Multimedia and productivity. Almost all developed appli-
cations come into this one category. Mobile devices are gaining acceptance for
multimedia applications more as they are attracting the mobile users mostly.
Based on the users Area of Interests Developers are providing applications for
smart phones.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 693–703, 2011.

c Springer-Verlag Berlin Heidelberg 2011
694 S. Banu Ali and N. Hundewale

For example, if the user is more interested in learning Islamic topics, then the
developers has to collect the useful pages from various Islamic sources based on
the users interests and the pages have to be wrapped in one unit to present it
as a Mobile Application for the smart phones.
This paper provides an IDE that allows non-programmers to develop such ap-
plications without the knowledge of Programming language. The IDE has tools in
Iconic format needed to wrap the pages. The UI presented in IDE are designed for
Drag and Drop method. Thus an application can be developed within minutes.
This IDE is mainly designed for the non-programmer who will develop mobile
application without the knowledge of API and programming language. This
IDE is first aimed to develop applications for android mobile phone developers
as android is open source and hence applications created can be easily installed
on android mobile devices as well as the developers has complete access to use
android tools. In addition, android phones are competing today with apples
iPhones in market, day by day android mobile users are increasing.
The paper is structured as follows. The proposed IDE is developed using
Eclipse with Android plug-in. Eclipse IDE is used here because android can be
used as a plug-in to the eclipse IDE hence the created applications using the
proposed IDE can be tested right there on android emulator in eclipse. This
paper first presents the design of the proposed IDE. The IDE is designed using
Jigloo GUI designer as Jigloo can also be used as a plug-in to the eclipse. Using
Jigloo GUI designer the IDE is designed first with all necessary Icons with drag
and drop functions. Programs are written in java for each icon in the IDE to
perform the actions. The created IDE is tested with various applications and the
results are shown below. The application thus created with this IDE is tested
again with android emulator in eclipse. A table is provided at last to show the
time consuming for different applications using this IDE.

1.1 PROPOSED IDE

The IDE is divided into the following. They are Widgets, Main View and Slider
View. The Main View is the Heart of this IDE where the Dragged Icons and
Slides are placed to this Main View. The slider view is shown in the Right Side
of the IDE. The slides will be distributed in the Slider View once the file has
been chosen from the File option. The Slides are dragged to the Main View by
selecting the Slides from the Slider View and dragging it to the Main View.
The widget is shown in the left side of the IDE. The widget has the icons
arranged vertically. The icons are dragged to the main view and are placed in
the selected slides.
Once the icons are dropped into the slide, the input box is shown to enter
the slide number based on the icon selected. Finally, the slides are consolidated
and converted to .apk files to install into the android mobile phones. The IDE
is divided into three views.
 An IDE for Developing Mobile Applications for Smart Phones 695

I. Icon View
II. Main View
III.Slider View

I.Icon View: The Icons like Previous, Next, Home and Scroll Bar are placed in
the Icon View. These icons are programmed with Drag and Drop functionalities.
The Icon is dragged to the main view to place on the selected slide. Once the
icon is dropped on the slide in the Main View, input box will be shown to enter
the Next or Previous Slide Number based on the Icon selected.

II.Main View: The Main View is the heart of the IDE where the slides and
Icons are Dragged and Dropped to this Main View. The slides are dragged to
the Main View to place the icon on the Slide. The slides with the icons are then
consolidated by the Slide Number which the user enters while placing the icons
on the slide. The consolidated slides are then saved with the .apk file type and
are send to the Android Emulator to test the developed application and installed
to the Android Device.

III.Slider View: The Slides are placed vertically one by one in the Slider View.
Once the file is selected from the File Menu, the entire slides from the selected
file are distributed to the Slider View. The slides are arranged vertically with
the slide number.

Fig. 1. Architecture of the Proposed IDE:

696 S. Banu Ali and N. Hundewale

2 Tools Used to Develop IDE

The tools used for this IDE are the following;

I. Eclipse
II. Jigloo GUI Builder
III. Android Development Tools
IV. Android Emulator to test the developed Application

2.1 Eclipse

Nowadays Eclipse IDE is used by all software developers to develop applications.

Eclipse is easy to use and any number of plug-ins can be added to the Eclipse
IDE. Moreover, eclipse is an open source IDE and includes a number of unique
features such as code refactoring, automatic code updates. In its entirety, the
Eclipse Platform contains the functionality required to build an IDE. However,
the Eclipse Platform is itself a composition of components; by using a subset
of these components, it is possible to build arbitrary applications. The Eclipse
Rich Client Platform (RCP) is one such subset of components. Figure 2 shows a
representation of some of the components in the Eclipse Platform and highlights
the subset that makes up the RCP (in reality there are a great many more
components).

Fig. 2. Eclipse

2.2 Jigloo GUI Builder

Jigloo is a visual Java GUI builder. Jigloo is an Eclipse plug-in that enables to
build sophisticated graphical user interfaces (GUIs) rapidly that run on the Java
platform. It can be used to build Swing-based applications and Standard Widget
Toolkit (SWT)-based applications[11]. Jigloo is an easy-to-use visual editor, and
is easy to create the UI for desktop application quickly.
Desktop Applications can be easily developed by Jigloo GUI Builder. The
main advantage of Jigloo is that it can be used as a Plug in for Eclipse to design
the UI for the Application. With Jigloo, it is easy to build UIs targeting Swing
or SWT.
 An IDE for Developing Mobile Applications for Smart Phones 697

Things Possible with Jigloo: Jigloo is a classic WYSIWYG editor for build-
ing Swing and SWT GUIs. To develop a graphical, desktop application in Java,
Jigloo is a great option. Jigloo supports round-tripping. That means not only to
make visual changes that cause the underlying code to change but also directly
edit the code and see the changes that causes in the GUI. This makes Jigloo a
great choice. Jigloo was built with Swing/SWT compatibility in mind. It sup-
ports many layout options supported by either technology. Not only to use either
technology for building a new GUI but it can even convert between Swing and
SWT using Jigloo.

2.3 Android Development Tools

Android Development Tools (ADT) is a plug-in for the Eclipse IDE that is de-
signed to give you a powerful, integrated environment to build Android applica-
tions. ADT extends the capabilities of Eclipse to let quickly set up new Android
projects, create an application UI, add components based on the Android Frame-
work API, and debugs the applications using the Android SDK tools, and even
export signed (or unsigned) .apk files in order to distribute the application to
the smart phones.
Developing in Eclipse with ADT is highly recommended and is the fastest
way to get started. With the guided project setup, as well as tools integration,
custom XML editors, and debug output pane, ADT gives an incredible boost in
developing Android applications.

2.4 Android Emulator

Android is a software stack for mobile devices that includes an operating system,
middleware and key Applications. The Android SDK provides the tools and
APIs necessary to begin developing applications on the Android platform using
the Java programming language. Android emulator is a device that appears
like Android phones, where the application developed can be tested right there
without the need of external android device.

Android Application Framework: By providing an open development plat-

form, Android offers developers the ability to build extremely rich and innova-
tive applications. Developers are free to take advantage of the device hardware,
access location information, run background services, set alarms, add notifica-
tions to the status bar, and much, much more. Developers have full access to
the same framework APIs used by the core applications. The application ar-
chitecture is designed to simplify the reuse of components; any application can
publish its capabilities and any other application may then make use of those
capabilities. This same mechanism allows components to be replaced by the
user.
698 S. Banu Ali and N. Hundewale

Android Runtime: Android includes a set of core libraries that provides most
of the functionality available in the core libraries of the Java programming lan-
guage. Every Android application runs in its own process, with its own instance
of the Dalvik virtual machine. Dalvik has been written so that a device can run
multiple VMs efficiently. The Dalvik VM executes files in the Dalvik Executable
(.dex) format which is optimized for minimal memory footprint. The VM is
register-based, and runs classes compiled by a Java language compiler that have
been transformed into the .dex format by the included ”dx” tool. The Dalvik
VM relies on the Linux kernel for underlying functionality such as threading and
low-level.

3 Image of IDE
The left side panel shows the Icon View, the Middle panel shows the Main View
and the Right Panel shows the Slider View. The Previous, Next, Scroll view
and the Home Icons are arranged vertically in the Icon View. The Main View is
the place where the selected slide and the Icons reside. The Slider View has the
Slides that are arranged one by one vertically.

Fig. 3. Image Of IDE

3.1 Program Structure

The IDE is developed using Jigloo GUI Builder. Jigloo is an Eclipse Plugin.
Jigloo can be installed using the eclipse powerful automatic updates.
The Structure of the Program is shown from the following figure
 An IDE for Developing Mobile Applications for Smart Phones 699

Fig. 4. Program Structure

Eclipse with Jigloo Plugin to Create the Application IDE. Steps Fol-
lowed in Creating the IDE:

1. The IDE is developed as a package using Jigloo as a plugin in eclipse. Package

Name: org.developerworks.jiglooIDE
2. A Swing Application is created in the Package Swing Application Name:
NewIDE.java
3. The NewIDE.java contains the main source code required to build the IDE
Sample Source Code for NewIDE.java:

NewIDE.java\\
{
jButton1 = new JButton();
Sliderpanel.add (jButton1);
jButton1.setText ("slide1");
jButton1.setBounds(31, 26, 73, 92);
---------
---------
}
4. The Xml Code of the IDE is generated using Jigloo Code Parser. JAXB
is used to read and write the XML data for the created IDE[12]. Example
NewIDE Schema
<dw:workflow xmlns:dw="org:developerworks:newIde"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="org:developerworks:newIde
workflow.xsd ">
<complexType name="AppIDE">
<sequence>
<element name="slide1" type="dw:priority"/>
700 S. Banu Ali and N. Hundewale

Fig. 5. Result:Design of NewIde.java

<element name="slide2" type=" dw:priority "/>

<element name="previous" type=" dw:priority " minOccurs="0"/>
<element name="next" type="string"/>
</sequence>
</complexType>
------
------
------
>
5. The XML code for the design of the IDE is copied and saved in a new file
called AppIDE.xml

3.2 Eclipse with ADT Plugin(Android) to Test the Application

Steps Followed to Run the Developed AppIDE in Android Project.
1. ADT Plug in is used with Eclipse to test the Application.
2. A New Android Test Project is Created.
3. The created XML file AppIDE.xml is imported to the Android -¿res-¿Layout
Folder
4. Applications are created using this ADT plugin in Eclipse using the new
Developed IDE AppIDE.xml file.
5. Once the Application are created with ADT plugin it will create the .APK
files automatically.
6. The Application .APK files are installed to the Android Mobile Phones.
Thus Website Application for Android Mobile Devices is developed in few min-
utes and the Experimental Results are shown below.
 An IDE for Developing Mobile Applications for Smart Phones 701

3.3 Image of the Created IDE in Jigloo

Fig. 6. Image of Created IDE in Jigloo

4 Snapshot of IDE Using Java

4.1 Experimental Setup and Results

The IDE is first experimented for Islamic Sources. Various Islamic Sites are col-
lected for Dhua, Historic Places and Prophets Sayings etc. Selected Pages from

Fig. 7. Snapshot of IDE

702 S. Banu Ali and N. Hundewale

various sites are collected for different types of application like audio files, Pow-
erPoint and PDF. Audio files are tested by collecting Dhuas from various sources
and developed into mobile application as .apk files using the IDE .Applications
to install into the Android Phones are experimented with the Created IDE and
the Results are Shown Below.

4.2 Use Case 1: Development Using PowerPoint

The below figure shows three different kinds of applications that fall into the
same category. For Example, the PDF, the web pages and the word documents
are treated here as a slides. These applications will disperse the contents as
the slides and the slides are consolidated using this IDE. The slides are then
converted to .APK files to install into the smart phones. The Android OS will
accept only .APK files to install into the smart phones. For this the file type
is converted to .APK extensions. This is done by the Android plug-in that is
plugged into the eclipse IDE. The xml source code is then placed into the Android
Plug-in into the Eclipse IDE to obtain the Target file Extension .APK. The
applications that use similar file types are shown below. Some Applications are
tested with the developed IDE and the time required to create the application
using this IDE is described below in the Table.

Fig. 8. Common Files

Table 1. TEST CASES

Application Tested
SerialNo TypeOfApplication TypeOfFiles NoOfPages TimeTaken
1 Dhua AudioFiles 10 5 Minutes
2 Historic Places PowerPoint 10 4 Minutes
3 Quran PDF 30 10 Minutes
4 Prophet sayings AudioFiles 15 6 Minutes
5 Prophets Life PowerPoint 20 6 Minutes
 An IDE for Developing Mobile Applications for Smart Phones 703

4.3 Conclusion and Future Work

This IDE is a useful tool for Developing Mobile applications easily and the main
advantage of this IDE is even a non-programmer who is programming for first
time can use this IDE to develop applications for mobile devices. Currently the
IDE has limited controls and in future more controls will be added to this IDE
to make more efficient for mobile application developers. Using this IDE Mobile
Applications can be created more easily with Drag and Drop functionalities.

Future Work: The IDE currently has limited icons that are necessary to de-
velop application, and in future more icons will be added to develop advance
applications for android smart phones. More Functionalities will be added in
future in order to save the created application in different files. Currently the
IDE is designed for Android Smart Phones and in future it will be extended to
develop applications for all types of smart phones.

References
1. Bloom, S., Book, M., Gruhn, V., Hrushchak, R., Kohler, A.: Write Once Run
Anywhere. A survey of Mobile Runtime Environments. In: Wu, S., Yang, L.T.,
Xu, T.L. (eds.) GPC 2008. LNCS, vol. 5036, pp. 132–137. Springer, Heidelberg
(2008)
2. Kenteris, M., Gavalas, D., Economou, D.: An Innovative Mobile EI (2009)
3. Kelly, M.D.: Getting started with Jigloo, a GUI builder for Eclipse
4. Holzer, A., Ondrus, J.: Trends in Mobile Application Development. In: Proceed-
ings of the 2nd International Conference Mobile Wireless Middleware, Operating
Systems and Applications (Mobileware 2009), pp. 55–64 (2009)
5. Android-An Open Handset Alliance Project, http://code.google.com/android
6. J2ME Polish, http://www.j2mepolish.org (retrieved December 13, 2009)
7. JavaFX Mobile, http://javafx.com/ (retrieved December 13, 2009)
8. Java ME, http://java.sun.com/javame (retrieved December 13, 2009)
9. JForm Designer, http://www.formdev.com/jformdesigner
10. Android- An Open Handset Alliance Project, http://code.google.com/android
11. Jigloo GUI Builder,
http://www.ibm.com/developerworks/opensource/tutorials
12. Jigloo SWT/Swing GUI Builder For Eclipse and Webshphere,
http://www.cloudgarden.com/jigloo/
13. A Guide to Install APK apps On Google Android phones,
http://www.talkandroid.com/guides/install-apk-files-on-android/
Function Approximation Using SVM with FCM
and Slope Based Partition

Amera Almas and Nisar Hundewale

College of Computers and Information Technology,

Taif University, Saudi Arabia
[email protected], [email protected]
http://cit.tu.edu.sa

Abstract. For a large training data set, the memory space required to
store the kernel matrix is a difficult task for the solution of QP prob-
lems for SVR. Support Vector Regression (SVR) is used to approximate
the function. So we are proposing the slope based partition algorithm,
which automatically evolves the partitions based on change in slope. In
this paper we are improving the performance of function approximation
with SVM by preprocessing the dataset with Fuzzy c-Mean clustering
(FCM) and slope based partition methods. Using FCM and slope, we
are portioning the data, and for every partitioned dataset we are find-
ing the function approximation, and aggregating the result. The results
indicate that Root Mean Square Error (RMSE) has been reduced with
the partitioned data, compare to the processing entire dataset, with the
increase in performance approximately 40%

Keywords: Support Vector Regression (SVR), Function approxima-

tion, FCM, slope based partition, partition of data sets.

1 Introduction
During the last decade a most of the researchers are using Support Vector
Machines (SVM) to solve pattern classification and function approximation prob-
lems. Function approximation or regression problems, unlike classification prob-
lems, have continuous output variable. SVM is a kernel based tool, can be
applied to both classification and Regression problems. In classification prob-
lems ( [Burges, 1998], [Osuna et. al, 1997], [Pontil and Verri, 1997]) SVM finds
a maximal margin hyper plane that separates two classes and increases the per-
formance in accuracy. SVR aims to find a linear regressor through a given set
of data points, where the points may be in the linear space or in a higher di-
mensional feature space. So the main objective is to find the hyper plane with
small norm which minimizes the distances from data points to the hyperplane.
The SVR method was proposed for estimating function approximation for the
regression data sets, constructing multidimensional splines, and solving linear
operator equations [Vapnik,1995]. The need for an approximating function often
arises in scientific and economic problems. One possibility is that one has a finite
set of data points and wants to determine the underlying functional form.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 704–714, 2011.

c Springer-Verlag Berlin Heidelberg 2011
 Function Approximation Using SVM with FCM and Slope Based Partition 705

Recently Fuzzy Functions and SVMs have been proposed to solve the regres-
sion estimation problems Fuzzy logic system and SVM is used to be powerful
tool for modeling nonlinear functions. Fuzzy logic system design is mainly neural
fuzzy networks. One major disadvantage of using neural fuzzy networks is that
network learning is based empirical risk minimization which does not account
for small test error. Whereas SVR learning formulation is based on the princi-
ple of structural risk minimization and attempts to minimize a bound on the
generalization error. The SVR learning approach also exhibits desirable proper-
ties such as generalization and over-fitting prevention capabilities. Many studies
have used SVRs to deal with regression estimation problem due to their high
generalization ability. SVRis a novel method for tackling the problems of func-
tion approximation and regression estimation based on the statistical learning
theory. [Osuna et. al, 1997] suggested optimizing only a fixed-size subset (work-
ing set) of the training data each time, while the variables corresponding to the
other patterns are frozen. Going to the extreme, the sequential minimal opti-
mization (SMO) algorithm [Platt, 1999] breaks the original QP into a series of
smallest possible QPs, each involving only two variables.
As proved by [Chia et. al, 2010] Fuzzy function along with SVR gives better
performance . In this paper we are to deal mainly with dataset which is parti-
tioned to according to the logic used in slope based and FCM. First the given
system domain is fuzzy partitioned into c clusters using fuzzy c-means cluster-
ing (FCM) algorithm. Then, one regression function is calculated to model the
behavior of each partition. In the second approach, the dataset is partitioned
based on maximum change of slope and for each partition, regression function
is calculated.

2 Literature Survey
The T-FR methods represent the fuzzy regression approaches proposed by Tanaka
et al. [Tanaka et. al, 1982], where the regression coefficients of input variables
are fuzzy numbers. Fuzzy Functions and SVMs have been proposed to solve the
regression estimation problems [Asli et. al, 2007]. Fuzzy clustering based Neural
Fuzzy Network with SVR has been proposed to differentiate the performance in
noisy and clean training data [Chia et. al, 2010].
The organization of this paper is as follows. Section 3 introduces Support Vec-
tor Regression, Section 4 introduces partition methods, FCM, and slope based
partition, Section 5 Experiments Section and Section 6 Results and Discussion.
In the Results section, we display results obtained from the application of the
proposed slope based partition and FCM on operating system dataset.

3 Support Vector Regression

The Support Vector method can also be applied to the case of regression, main-
taining all the main features that characterise the maximal margin algorithm: a
non-linear function is learned by a linear learning machine in a kernel-induced
706 A. Almas and N. Hundewale

feature space while the capacity of the system is controlled by a parameter that
does not depend on the dimensionality of the space. As in the classification
case the learning algorithm minimises a convex functional and its solution is
sparse. We are given a training set {(xi , yi )}ni=1 ⊂ Rd × R, where xi and yi
are the input and output, respectively, of the ith pair. Support vector regres-
sion ( [Smola and Scholkopf,2004] is a kernel method that performs nonlinear
regression based on the kernel trick. Essentially, each input xi ∈ Rd is mapped
implicitly via a nonlinear feature map ϕ(x) to some kernel-induced feature space
F where linear regression is performed. In the case of linear functions f taking
the form
n
f (x) = wi xi + b with w ∈ R, b ∈ R. (1)
i=1

n wayT to ensure minimum ε deviation is to minimize the norm, i,e ||w|| =

2
One
i=1 Wi Wi The problem can be written as a convex optimization problem:
minimize 12 ||w||2 subject to
n
 n

subject to yi − Wi xi − b ≤ ε, Wi xi + b − yi ≤ ε (2)
i=1 i=1

The assumption in (2) is that such a function f actually exists that approximates
all pairs (xi , yi ) with ε precision. In the case where constraints are infeasible,
slack variables , ξi , ξi∗ were introduced. This case is called soft margin formulation
[Bennet and Mangasarian, 1992] and is described by the following problem;

 l
1
M inimize ||w||2 + C (ξi + ξi∗ ) (3)
2
i=1

subject to
n

yi − Wi xi − b ≤ ε + ξi (4)
i=1
n

Wi xi + b − yi ≤ ε + ξi∗ (5)
i=1

ξi , ξi∗ ≥ 0 (6)
The constant C ¿ 0 determines the amount up to which deviations larger than
ε are tolerated. This is called -insensitive loss function |ξ|ε and is described by

|ξ|ε = 0 if |ξ| ≤ ε, (7)

|ξ|ε = |ξ| − ε otherwise, (8)

Figure 1 depicts this ε -insensitive loss function graphically. Only the points
inside the shaded region contribute to the prediction accuracy, whereas the points
outside the shaded region contribute to the error, as the deviations are penalized
 Function Approximation Using SVM with FCM and Slope Based Partition 707

Fig. 1. The soft margin loss setting for a linear SVR

in a linear fashion. It is observed that in most cases the optimization problem

can be solved more easily in its dual formulation. Hence, Lagrange multipliers
are used to get the dual formulation as described in [Fletcher, 1989], which is as
follows;

l l
1
L= ||w||2 + C (ξi + ξi∗ ) − (ηi ξi + ηi∗ ξi∗ )
2 i=1 i=1
l
 n

− αi (ε + ξi − yi + wi xi + b)
i=1 i=1
l
 n
 l

− α∗i (ε + ξi − yi + wi xi + b) − (ηi ξi + ηi∗ ξi∗ ) (9)
i=1 i=1 i=1

Here L is the Lagrangian and ηi , ηi∗ , αi , α∗i are Lagrange multipliers. Hence the
dual variables in (4) have to satisfy positivity constraints i.e.

ηi , ηi∗ , αi , α∗i ≥ 0 (10)

According to the saddle point condition the partial derivatives of L with respect
to the primal variables (w, b, ξi , ξi∗ ) have to vanish for optimality. Therefore, we
get
l
∂L  ∗
= (αi − αi ) = 0, (11)
∂b i=0

l
∂L
=w− (αi − α∗i )xi = 0, (12)
∂w i=1

∂L
= C − αi − ηi = 0, (13)
∂ξi
∂L
= C − α∗i − ηi∗ = 0, (14)
∂ξi∗
708 A. Almas and N. Hundewale

Substituting (11), (12), (13) and (14) into (9) yields the dual optimization
problem;
l n l
1  ∗ ∗
 
maximize − (αi − αi )(αj − αj ) xi xj − ε (αi − α∗i )
2 i,j=1 i,j=1 i=1


l
+ yi (αi − α∗i ) (15)
i=1

subject to
l

(αi − α∗i ) = 0 and αi , α∗i ∈ [0, C]. (16)
i=1

To make the SV algorithm nonlinear all training patterns xi are mapped φ :

x → f into some feature space f as described in ( [Aizerman et al, 1964],
[Nilsson, 1965]) and then using the standard SV regression a linear hyperplane
is constructed the feature space. Using the trick of kernel functions
[Cortes and Vapnik, 1995] following QP problem is formulated.
l
1 
M aximize − (αi − α∗i )(αj − α∗j )K(xi , xj ), (17)
2
i,j=1

l
 l

−ε (αi − α∗i ) + yi (αi − α∗i ) (18)
i=1 i=1

subject to
l

(αi − α∗i ) = 0 and αi , α∗i ∈ [0, C]. (19)
i=1

The optimal solution obtained

l
 l

w= (αi − α∗i )φ(xi ) and f (x) = (αi − α∗i )K(xi , x) + b, (20)
i=1 i=1

where K(.,.) is a kernel function. Usually more than one kernel used in the liter-
ature to map the input space into feature space [Christianini and Shawe, 2000].

4 Data Partition
4.1 Based on Fuzzy Clustering Means(FCM)

In Fuzzy clustering, each point belongs to one or more cluster, whereas in fuzzy
logic, the point belongs to only one cluster. Thus, points which are outlier of
cluster, has less membership value than the points which lies near the center of
cluster. For each point x we have a coefficient giving the membership value of
 Function Approximation Using SVM with FCM and Slope Based Partition 709

the kth cluster is μk (x). Usually, the sum of those coefficients for any given x is
defined to be 1:
num.clusters 

∀x μk (x) = 1 (21)
k=1

The centroid of a cluster is the mean of all points, weighted by their membership
value of belonging to the cluster:

μk (x)m x
centerk = x m
(22)
x μk (x)

The membership value is related to the inverse of the distance to the cluster
center:
1
μk (x) = , (23)
d(centerk , x)
then the coefficients are normalized and fuzzyfied with a real parameter m > 1
so that their sum is 1. So
For m equal to 2, this is equivalent to normalizing the coefficient linearly to
make their sum 1. When m is close to 1, then cluster center closest to the point
is given much more weight than the others, and the algorithm is similar to k-
means. [WIKI]. Clustering is then obtained by considering the points which has
more membership value than the other clusters. The mean shift algorithm does
not have prior knowledge about the shape of the clusters, and the number of
classes can be determined based on the number of clusters.

4.2 Data Partition Based on Slope

For a large training data set, the memory space required to store the kernel
matrix becomes very challenging for the solution of QP problems for SVR. Thus,
the traditional optimization algorithms, such as Newton, cannot be directly used.
In this section, we propose an unsupervised partition algorithm, through which
fixed no. of partitions are automatically generated based on the slope of data
set, based upon the random change of slope. The resultant partitions of the
algorithm are then used for SVR training. Preprocessing of data includes taking

preprocess the data p=(xi , yi )N

i=1
Find Z = Δ(y)
sort z in ascending order
select top m values
select Minimum length of the partition
For j = 1tom
Fork = z[1]toz[m]
pj = (xi , yi )ki=k−1
For each partition pi apply SVR and find RMSE
Calculate the aggregate RMSE
710 A. Almas and N. Hundewale

care of missing values, noise, etc. Also, fetching the set of target values from the
data into the program is also a part of that pre-processing. The main idea is
to partition the data into multiple small sections. The section limits are defined
by any great change in the data. The secondary condition is to make partitions
of greater than a or equal to a minimum length in order not to end up in very
badly sized subsets.
To find a greatest amount of change in the data, first the difference of the
data is found. This difference data is also called as slope (of data). Now, the
maximum value in this slope is actually the greatest change in the original data.
There might be multiple points where the data has same highest amount of
change. The program is set to find all those positions in the data where the slope
is the highest. From these found positions, the program checks the points that
are so apart from each other (and boundaries) that they can make a partition of
the minimum required size or more. The positions that cannot make partitions
large enough are ignored. If the required number of partitions is not found, this
process has to be redone. In order to make next iteration possible, the highest
value is made to a lowest or Not-a-Number value in the same iteration. That
makes the next highest value as the highest value. Then, this process is repeated
until either the data is divided into a set number of partitions (or more) or to a
point where there are no more divisions possible.
A dataset may contain a slope which has more common highest values than
the number of required partitions. Also, in a certain iteration the number of
partitions made might be much more than the remaining number of partitions
required. Hence, it is certainly possible that the above process ends up giving
more number of partitions than required. But it should also be noted that itera-
tion will not take place unless there is a need of partition. So, any extra partition
generated has to be from the last iteration. Now, all these new partitions has
same peak value. At this juncture, the program is set to skip the partitions from
the last iterations which had minimum size. Or in other words, the partitions
which give larger size of partitions, if removed, are skipped. This way, the data
is divided into the exact number of partitions. . In FCM, every point belongs to
all the clusters, but we have considered only those points in a cluster which has
the maximum value of membership compare to other clusters. After finding the
partitions, each partition is used to train SVR separately and the performance
is evaluated in RMSE. And all the partitions RMSE values are aggregated.

5 Experiment Section

For the experiments we have considered data set from the website [DACS] re-
liability datasets, SS1A, SS1B, SS1C, SS2, SS3,SS4. The Software Reliability
Dataset was compiled by John Musa of Bell Telephone Laboratories. These
datasets contains t failure interval data to assist software managers in moni-
toring test status and predicting schedules and to assist software researchers in
 Function Approximation Using SVM with FCM and Slope Based Partition 711

validating software reliability models. These models are applied in the discipline
of Software Reliability Engineering. The dataset consists of software failure data
for operating system. This data set consists of 3 attributes, called Failure Num-
ber, Failure Interval, Day of Failure. For these datasets we have applied the FCM
and slope based algorithms and portioned the dataset in to five datasets. Later
each subset of dataset is processed through LibSVM to get the performance in
RMSE. While training the SVR we have chosen the RBF kernel as it gives the
better performance than the other kernels. In this we are doing experiments
with the values Cost (c) value is 6.4 and epsilon (p) is 0.026, as it is proved by
Shawon et. al., that the performance of trined SVM model is better than the
SVM trained with other combination of values [Shaowu et.al, 2007].
The figure 2 shows the results of this process in a graphical form. This graph
has multiple scale for the original data and its slope are completely different in
magnitude. The left vertical axis has the the scale for original data magnitude.
The right one is for the difference or slope of the data. And partitions are shown
with vertical bars across the data. Also shown are the skipped partitions, if extra
partitions were found. It also shows the graph representing the max peaks of
SS1A data set, in which the lines representing with pink color has been selected
for the partition and the blue color represents the skipped partition, as in this the
partition length color represents the skipped partition, as in this the partition
length is less than 4.

Fig. 2. Graph Based on Slope Partition

Fig 3 explains the proposed approach. First the dataset is processed through
the FCM and slope based partition then each sub dataset is passed to train SVR
through LibSVM measuring the performance in RMSE.
712 A. Almas and N. Hundewale

Fig. 3. Data Flow Diagram

6 Results and Discussion

The results are shown in the Tables 1 and Table2. In Table 1, it is representing
the results of the datasets before and after applying slope based partition. From
the table it is observed that SVR error rate is reduced when it is trained with
partitioned data. Considering the SS1A dataset we made 5 partitions, based up
on the max peaks, which is the change in slope. All the 5 partitions are used to
train data individually, passed through LibSVM using RBF kernel, by setting
the parameter. And the SVR model is tested by other partitions. The perfor-
mance is measured in RMSE, and tabulated the RMSE rate on the original data
and the aggregated RMSE of partitioned data. This table shows that the results
are better than the original data. Similarly, Table 2 represents the results of
the datasets for training SVR with the original data and datasets after parti-
tioning. From the table we can conclude that in all datasets we are getting the
better performance, after partitioning the data and aggregating the RMSE, than
considering complete data.

Table 1. Shows the results of slope based partition

Dataset RMSE(original Data) RMSE(after partitioning) % of performance improved

SS1A 17.85 7.3766 58.67
SS1B 90.1762 51.8399 42.51
SS1C 60.9547 29.7321 51.22
SS2 44.8908 24.4178 45.61
SS3 63.2017 48.4788 23.3
SS4 38.1437 30.3019 20.56
Over all average performance= 40.31

Table 2. Shows the results based on FCM

Dataset RMSE(original Data) RMSE(after partitioning) % of performance improved
SS1A 17.85 10.2024 42.84
SS1B 90.1762 52.7324 41.52
SS1C 60.9547 26.5313 56.47
SS2 44.8908 19.6897 56.14
SS3 63.2017 30.2127 52.20
SS4 38.1437 23.0987 39.44
Over all average performance= 48.1017
 Function Approximation Using SVM with FCM and Slope Based Partition 713

7 Conclusions and Future Work

In this paper we have proposed new approach called Slope based partition, and
compared the results with the FCM partitions, for training SVR. The proposed
method reduces model complexity by first grouping similar components with
same slope into a partition and then training SVR and getting local function
approximation. Which is the simple slope logic based partition, which reduces
the complexity, and also SVR model simplification. We have considered the lo-
cal maxima, and restricted the no. of partitions for all data sets. This is the
limitation of our algorithm. We could not handle the stopping criteria. In FCM
algorithm we need to mention the no. of clusters has to be initialized. While
training data set should be known in prior, which is used to train SVR in both
the algorithms, which is the limitation of both the algorithms. Considering origi-
nal data training SVR we get good functional approximation. But preprocessing
the data with FCM and slope based partition and training SVR, overall finding
the aggregating the RMSE gives reduced error rate. So from our algorithmic
approach, we can conclude that both FCM and slope based partitions are com-
parable. The study about joining the results of the subsets is crucial. The effect
of partitioning on a data which has periodical slope should also be studied.
Also, the effectiveness of this technique on a data with non-increasing values
should also be critically studied. This research also calls for the study of simpler
selection of peaks based on thresholds rather than analysis.

References

[Aizerman et al, 1964] Aizerman, M.A., Braverman, E.M., Rozonoer, L.I.: Theoreti-
cal foundations of the potential function method in pattern recognition learning.
Automation and Remote Control 25, 821–837 (1964)
[Asli et. al, 2007] Celikyilmaz, A., Turksen, B.I.: Fuzzy functions with support vector
machines. Science Direct, Information Sciences 177, 5163–5177 (2007)
[Bennet and Mangasarian, 1992] Bennet, K.P., Mangasarian, O.L.: Robust linear pro-
gramming discrimination of two linearly inseparable sets. Optimization Methods
and Software 1, 23–34 (1992)
[Burges, 1998] Burges, C.J.C.: A tutorial on support vector machines for pattern recog-
nition. Data Mining and Knowledge Discovery 2(2), 121–167 (1998)
[Chia et. al, 2010] Juang, C.-F., Hsieh, C.-D., Hong, J.-L.: Fuzzy Clustering-Based
Neural Fuzzy Network, with Support Vector Regression. IEEE, Los Alamitos (2010)
[Cortes and Vapnik, 1995] Cortes, C., Vapnik, V.: Support vector networks. Machine
Learning 20, 273–297 (1995)
[Christianini and Shawe, 2000] Cristianini, N., Shawe-Taylor, J.: An introduction to
support vector machines and other kernel-based learning methods. Cambridge Uni-
versity Press, New York (2000)
[Fletcher, 1989] Fletcher, R.: Practical Methods of Optimization. John Wiley and
Sons, New York (1989)
[Nilsson, 1965] Nilsson, N.J.: Learning machines: Foundations of trainable pattern clas-
sifying systems. McGraw-Hill, New York (1965)
714 A. Almas and N. Hundewale

[Osuna et. al, 1997] Osuna, E.E., Freund, R., Girosi, F.: Support vector machines:
Training and applications. Technical report, Massachusetts Institute of Technol-
ogy, Artificial Intelligence Laboratory, AI Memo No. 1602 (1997)
[Pontil and Verri, 1997] Pontil, M., Verri, A.: Properties of support vector machines.
Technical Report, Massachusetts Institute of Technology, Boston MA, USA (1997)
[Platt, 1999] Platt, J.C.: Fast training of support vector machines using sequential
minimal optimization. In: Schölkopf, B., Burges, C., Smola, A. (eds.) Advances in
Kernel Methods - Support Vector Learning, pp. 185–208. MIT Press, Cambridge
(1999)
[Shaowu et.al, 2007] Zhou, S., Wu, L., Yuan, X., Tan, W.: Parameter Selection of SVM
for Function Approximation Based on Differential Evolution (2007)
[Smola and Scholkopf,2004] Smola, A.J., Schölkopf, B.: A tutorial on support vector
regression. Statistics and Computing 14(3), 199–222 (2004)
[Tanaka et. al, 1982] Tanaka, H., Uejima, S., Asai, K.: Linear regression analysis with
fuzzy model. IEEE Transactions on Systems, Man, Cybernetics SMC-2, 903–907
(1982)
[Vapnik,1995] Vapnik, V.N.: The nature of statistical learning theory. Springer-Verlag,
New York Inc., New York (1995)
[DACS] www.thedacs.com
[WIKI] http://en.wikipedia.org/wiki/Clusteranalysis
A PIC-Controller Based Doubly Fed Induction Generator
(DFIG) for Wind Energy Conversion Systems

K. Sureshkumar1, P. Vijaya kumar2, R. Dhandayuthabani3, and A. Sakthivel4

1
Department of EEE, Velammal Engineering College,
Chennai-600066
[email protected]
2
Department of EEE, Karpagam College of Engineering,
Coimbatore-641 032
[email protected]
3
Product Development Engineer, Automotive Infotronics, Chennai
[email protected]
4
Department of EEE, Dr. Mahalingam College of Engg & Tech,
Pollachi-641003
[email protected]

Abstract. The design details of a PIC-controller based Doubly Fed Induction

Generator (DFIG) for wind power generation is described. The proposed PIC-
DFIG not only provides a low cost Solution to today’s electric power needs, but
is also aimed at delivering better power quality, making it ideally suited for
wind energy conversion systems operating at variable wind speeds and
connected to variable and often, dynamic utility loads. Power quality
improvement is achieved by cancellation of the largest and the most annoying
harmonics in the supply grid. Additionally, power factor correction and
adaptive response to variable reactive loading are also addressed by this design.
The proposed design involves two current-controlled power converters
connected through a slip ring assembly. The rotor control is achieved using a
commercially available PIC controller (model 16F877A) employing a position
sensorless, field oriented control algorithm. Some preliminary simulation
efforts and modeling scenarios are also presented.

Keywords: PIC-controller, doubly-fed induction generator, wind energy,

harmonic compensation, power factor improvement.

1 Introduction
In view of the rapid depletion of fossil fuels and the increasing global concern for
clean environment, considerable efforts are being spent on developing green,
renewable energy sources. Of such resources, wind energy offers a sustainable, low-
cost and environmentally friendly solution to energy needs. However, the intermittent
and variable nature of wind currents offers many challenges to tapping the wind
power effectively. Sometimes a Variable Speed Power Generation (VSPG) system is
used to harness wind energy, where a generator parameter, such as the wind turbine
speed, is dynamically adjusted to optimize the output power as the ambient wind

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 715–723, 2011.
© Springer-Verlag Berlin Heidelberg 2011
716 K. Sureshkumar et al.

speed changes. Fixed speed generating systems are also used widely in wind power
generation and owe their popularity to a few advantages such as low cost, simplicity
of design, the ability to control both active and reactive currents independently and so
on. The system we propose, the Doubly Fed Induction Generator (DFIG) is one such
fixed speed system. The optimization of output power in this system is achieved
through regulating the power flow in the two generator windings, stator and rotor.
Since the rotor side winding handles only the slip power, which is easier to control,
often it is enough to control only the rotor side flow in order to regulate the overall
flow and obtain an improved power rating. We propose the use of PIC controllers to
regulate such rotor side converter flows because they offer the following advantages
among others: (1) Low cost (2) Off the shelf units easily available (3) The Assembly
Language coding results in fast response and therefore more dynamic control (4)
multi-functionality and flexibility, leading to the development of customizable
optimization algorithms. Once the rotor parameters’ ranges are determined and
control algorithm is fine-tuned, we may be able to implement the design using
commercially available PLC units. in place of PIC controllers for easier end-user
programming, if needed. We will discuss some concepts on which we base such
control algorithms and propose how best to generate and optimize wind power.
First, in addition to wind power conversion, it is also important to improve the
power quality of the generated power, especially given the wide variety of electrical
appliances with stringent supply quality requirements that are being used by today’s
utility consumers. Power quality, in terms of harmonic current distortion in the
distributed power and power factor, may also be subject to government regulations
and compliance can be quite challenging. The DFIG design described in this paper
addresses these concerns [1-6] as well and is capable of compensating for the most
prevalent harmonic current components in the grid, in addition to its ability to
generate effective wind power. The proposed design also corrects power factor and
offers reactive power control. . The current sensing in the rotor is done using a position
sensorless, field oriented control [10] method and is used in regulating the rotor
circuit dynamically via PIC controllers.
As a prelude to actual implementation of the design, we offer a simulation of the
design and some scenarios using MATLAB. In particular, we provide some sampling
of PIC programming to give a flavor of the controller programming.

2 Architecture of DFIG Power Grid System

The proposed DFIG design and a section of the generator-grid interface are shown
schematically in Fig. 1. The exact circuitry of the controllers, the power converters
and associated electronics, details of the load etc. are not shown for brevity.
It must be noted that this design affords independent control of active (wind) power
generation, reactive (load) power compensation and grid harmonics distortion
cancellation. The two power converters used, one for the stator and one for the rotor,
are the current-controlled voltage source type [10]. The stator flux field orientation is
sensed in a position-less manner and is used to regulate flow in rotor side windings.
This way, both the fundamental as well as the harmonics are sensed. The exact level
of harmonic currents to be used in command for the rotor side is obtained either by
estimating the harmonics current theoretically based on the externally connected load
or by actually sensing the current in the load. In general there are two major power
 A PIC-Controller Based Doubly Fed Induction Generator (DFIG) 717

Fig. 1. Block diagram of DFIG-POWER grid system

components in this rotor side regulation algorithm, one due to the active process of
wind energy generation and associated variability due to wind speed and other
characteristics of wind turbine and another harmonics component due to the utility
load and grid harmonics. By suitable controller commands, these components are
dynamically translated into the appropriate rotor side excitation. This in turn results in
high quality power output with little harmonic distortion and adjusted for reactive
load. i.e., better power factor.

3 Design Details

3.1 Rotor Side Design

In our design, the non linear load currents and phase currents of the line are sampled
and converted into d-q co-ordinate values by using Parks transformation. Based on the
quadrature and direct rotor current, the gate signal for the back to back ac/dc/ac
converter is given.
In order to have a common frame of reference for both the stator and the rotor
equations and also to reduce the complexity of the voltage equations, one can use a
rotating set of d-q axes attached to the stator flux[9] which rotates at an angular
frequency of θ. The flux linkages of stator and rotor in the Excitation reference
frame will be:

λeqs = Ls ieqs +Lm ieqr, (1)

λeds = Ls ieds +Lm iedr (2)

λeqr = Lr ieqr +Lm ieqs, (3)

λedr = Lr iedr +Lm ieds (4)

718 K. Sureshkumar et al.

Rotor voltages in the d–q excitation reference frame can be shown as

veqr = rr ieqr + (ωe −ωr ) λedr +P λeqr (5)

vedr = rr iedr− (ωe −ωr ) λeqr +P λedr (6)

If the stator flux vector is aligned with the d-axis in the synchronously rotating
reference frame, then λeqs = 0, and the rotor currents components can be shown as

ieqr = -(Ls/Lm) ieqs (7)

iedr = -(Ls/Lm) ieds + (1/Lm) λeds (8)

The rotor voltage equations in the excitation reference frame

veqr = rrieqr+ωslipσLriedr+ωslip(Lm/Ls)λeds+σLr(d/dt)ieqr (9)

vedr=rriedr−ωslipσLrieqr+σLr(d/dt)iedr+(Lm/Ls) (d/dt)λeds (10)

where,
ωslip = ωe- ωr (11)

The quadrature and direct rotor current demand can be derived from the active and
reactive power references P*, Q*, and harmonic currents needed for compensation of
nonlinear load current. The variables ie∗qr and ie∗dr can be derived as follows:

ie∗qr = ie∗qr−active + ie∗qrh (12)

ie∗dr = ie∗dr−reactive + ie∗drh. (13)

where,

ie∗qrh = − (Ls/Lm). ieqLh (14)

ie∗drh = − (Ls/Lm)iedLh +(1/Lm)λeds (15)

3.2 Direct and Quadrature Rotor Current Command Generation

The block diagram of quadrature and direct rotor currents commands generation is
shown in Fig. 3. In this method, wind velocity is measured, and if it is higher than the
cut-in value, the system goes to current mode control. Active and reactive power
references are being generated in respect to the wind speed and turbine characteristics.
This will generate the active power command I*qr-active. If the wind turbine power
 A PIC-Controller Based Doubly Fed Induction Generator (DFIG) 719

falls below Pmin, the speed control mode is activated, and using speed regulator, the
active power command is obtained.
In this topology, while direct axis is used to control the reactive power and the
quadrature axis current component is used to regulate the active power, ac signal
portion of the quadrature and the direct rotor current components are set to control
harmonics compensation of the grid. In the proposed algorithm, the decoupled control
of active power, reactive power, and harmonic compensation are achieved.

Fig. 2. Direct and quadrature rotor current command generation

The active and reactive powers of the stator are:

P∗s=-3/2(vm(Lm/Ls)i∗eqr−active) (16)

Q∗s=-3/2vm(λeds−Lmi∗edr−reactive/Ls) (17)

3.3 Position – Sensorless Rotor Position Detection

Proper rotor current orientation [7] with respect to the stator flux requires
transformations between the stator and the rotor coordinates. It means that
information of the instantaneous rotor position is necessary to implement correct field
orientation. The following method is based on simple trigonometric computations3
Therefore the estimation is possible instantaneously, and the control on the fly is
enforceable without any appreciable transient. The stator flux depends on the stator
voltage and frequency of excitation.
The d–q components of the rotor current in the stationary reference frame can be
obtained from

isdr = (1/Lm)(∫(vsds− rs isds)d-(Lm + Lls )isds) (18)

720 K. Sureshkumar et al.

isqr = (1/Lm)(∫(vsqs− rs isqs)dt-(Lm+Lls)isqs) (19)

If we project these currents into two-phase orthogonal reference frame fixed on the
rotor we can write,

ir =√(isdr2 +isqr2) (20)

ir =√(irdr2 +irqr2) (21)

θ1 = tan−1 (isqr/isdr) (23)

θ2 = tan−1 (irqr/irdr) (24)

The rotor current vectors in the orthogonal d–q stationary and rotary reference frames
are illustrated in Fig. 4. It can be inferred that the rotor current vector in the stationary
reference frame rotates with synchronous speed.
However, in the rotor reference frame, this vector rotates with slip frequency. It is
obvious that the rotor position angle can be calculated from

θr = θ1 − θ2 (25)

Fig. 3. Sensorless rotor position calculator

4 Simulation Results
We have simulated some scenarios using our model.
 A PIC-Controller Based Doubly Fed Induction Generator (DFIG) 721

Case Study 1. Grid status without non linear load

Fig. 4. Case study 1. Grid voltage and current.

Case Study 2. Grid status with non linear load

Nonlinear load current is generated by a diode rectifier feeding a variable resistor. The
magnitude and phase of the non-linear load current are varied to evaluate the
performance of the proposed Method. After connecting the non linear load across the
grid, the Total harmonic distortion of current is 29.73%.

Fig. 5. Case study 2. Grid voltage and Grid current with connecting non linear load

Fig. 6. Case study 2. (a)Direct and Quadrature axis current of stator. (b)Direct and Quadrature
axis current of Rotor without controller.
722 K. Sureshkumar et al.

Case Study 3. Grid status with controller

The controller has been developed to cancel harmonics needed for non linear load by
varying the stator flux based on the command gate signal. After connecting the
controller, the total harmonic distortion is reduced (THD = 3.07%).

Fig. 7. Case study 3 .Grid voltage and current with connecting non linear load with controller.

Fig. 8. Case study 3. (a)Direct and Quadrature axis current of stator. (b)Direct and Quadrature
axis current of Rotor with controller.

5 Conclusion
The design and modeling of a PIC controller-based DFIG-Wind power system is
described. The power generation is not only optimized for maximum conversion of
wind energy, but is also capable of producing high quality power with dynamic
cancellation of grid harmonics. The design algorithm is based upon controlling rotor
current flows and in the control of reactive current components due to external utility
load. The use of PIC controller in the rotor side control affords a simple, low cost way
of implementing the proposed algorithm and consequent generation of high quality
wind power. Such a system is well suited for low to medium scale wind power
generation units.

References
1. Abolhassani, T., Niazi, P., Toliyat, H.A., Enjeti, P.: Integrated Doubly Fed Electric
Alternator/ActiveFilter (IDEA), a Viable Power Quality Solution, for Wind Energy
Conversion System. Proc. IEEE Trans. Energy Conversion 23(2), 642–650 (2008)
 A PIC-Controller Based Doubly Fed Induction Generator (DFIG) 723

2. Abolhassani, M.T., Niazi, P., Toliyat, H.A., Enjeti, P.: A sensoreless Integrated doubly-fed
electric alternator/active filter (IDEA) for variable speed wind energy system. In: Proc.
IEEE Ind. Appl. Soc., vol. 1, pp. 507–514 (2003)
3. Datta, R., Ranganathan, V.T.: A Simple Position-Sensorless Algorithm for Rotor-Side
Field-Oriented Control of Wound-Rotor Induction Machine. IEEE Transaction on
Industrial Electronics 48(4) (August 2001)
4. Xu, L., Cheng, W.: Torque and reactive power control of a doubly fed induction machine
by position sensorless scheme. IEEE Trans. Ind. Applicat. 31, 636–642 (1995)
5. Akagi, H.: New Trends in Active Filters for Power Conditioning. IEEE Transactions on
Industry Applications 32(6), 1312–1322 (1996)
6. Pena, R., Clare, J.C., Asher, G.M.: Doubly fed induction generator using back-to-back
PWM converters and its application to vari-able-speed wind- energy generation. Proc. Inst.
Elect. Eng., pt. B 143(3), 231–241 (1996)
7. Datta, R.: Rotor side control of grid-connected wound rotor Induction machine and its
application to wind power generation. Ph.D. dissertation, Dept. Elect. Eng., Indian Inst.
Science, Bangalore, India (February 2000)
8. Akagi, H., Sato, H.: Control and Performance of a Doubly-Fed Induction Machine
Intended For a flywheel Energy Storage System. IEE Transaction on Power
Electronics 17(1) (January 2002)
9. Hopfensperger, B., et al.: Stator-flux-oriented control of a doubly-fed induction machine
with and without position encoder. IEE Proc. Electr., Power Appl. 147(4) (July 2000)
10. Morel, L., Godfroid, H., Mirzaian, A., Kauffmann, J.M.: Double-Fed induction machine:
converter optimization and field oriented control without position sensor. Proc. Inst. Elect.
Eng., pt. B 145(4), 360–368 (1998)
11. Akagi, H., Sato, H.: Control and performance of a doubly-fed Induction Machine intended
for a flywheel energy storage system. IEE Trans. Power Electron. 17, 109–116 (2002)
 Packet Scheduling Scheme with Enhanced
Quality of Service for Mobile WiMAX

V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

M.V.G.R.College of Engineering, Vizianagaram

[email protected], [email protected], priya [email protected]

Abstract. A new packet scheduling scheme for Mobile WiMAX net-

work is proposed. The network architecture includes one BS and several
MSs which are in different locations. The uplink traffic from Mobile Sta-
tions (MS’s) that belong to various classes to the Base Station (BS) is
scheduled.It is assumed that the MS’s request the BS for bandwidth by a
Bandwidth request and the BS station grants the bandwidth. The avail-
able slots for scheduling of uplink traffic is done based on the predefined
priorities of the traffic classes. The BS then allocates the bandwidth to
each MS based on the class it belongs and distance from the BS. All the
mobile WiMAX parameters like transmit power, receive gain, transmit
gain, mobility have been considered for simulation. Results indicate a
clear reduction in delay and increase in throughput for different classes
of flows compared to the existing Quality of Service based 2-Tier Ad-Hoc
Scheduling Scheme (2T-AHSS).

1 Introduction

The WiMAX family of standards (802.16) concentrates on two types of usage

models a fixed WiMAX usage model and a mobile WiMAX usage model. The
basic element that differentiates these systems is the ground speed at which
the systems are designed to manage. Based on mobility, wireless access systems
are designed to operate on the move without any disruption of service; wireless
access can be divided into three classes; stationary, pedestrian and vehicular.
A mobile WiMAX network access system is one that can address the vehicular
class, whereas the fixed WiMAX serves the stationary and pedestrian classes.
This raises a question about the nomadic wireless access system, which is referred
to as a system that works as a fixed WiMAX network access system but can
change its location.
Mobile WiMAX systems offer scalability in both radio access technology and
network architecture, thus providing a great deal of flexibility in network de-
ployment options and service offerings. Some of the salient features supported
by Mobile WiMAX are high data rates, Quality of Service, Scalability, Security
and Mobility.
Scheduling is the main component of the MAC layer that helps assure QoS
to various service classes. Scheduling can be done for uplink and downlink.

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 724–734, 2011.

c Springer-Verlag Berlin Heidelberg 2011
 Packet Scheduling Scheme with Enhanced Quality of Service 725

The scheduling algorithms developed in WiMAX Networks are vendor depen-

dent. Every algorithm has its own advantages and disadvantages depending on
what scenario it is being used. Hence selecting the appropriate algorithm for
scheduling the packets for the downlink channel plays a key role [1]. Example is
the Round Robin Scheduler [2] that fairly assigns allocation one by one to all
connections. With packet-based allocation, stations with larger packets have an
unfair advantage. Moreover, RR may be non-work conserving in the sense that
the allocation is still made for connections that may have nothing to transmit;
since RR cannot assure QoS for different service classes. The Weighted Round
Robin algorithm has been applied for WiMAX scheduling [3]. A WRR algorithm
assigns weight to each SS and the bandwidth is then allocated according to the
weights. The weights can be used to adjust for the throughput and delay re-
quirements. The WRR is an extension of the Round Robin algorithm. It is a
work-conserving algorithm in that it will continue allocating bandwidth to the
SSs as long as they have backlogged packets. Though the above algorithms are
effective for fixed WiMAX where the data rates at the subscriber stations change
less frequently, In Mobile WiMAX the rates at which the mobile stations trans-
mit the data are changed at every instant. We describe a new efficient scheme
for mobile station downlink scheduling that meets the challenges faced by a Mo-
bile WiMAX network. A study of an optimized scheduling scheme in WiMAX
networks has been done which includes allocation of resources to the mobile
stations and scheduling the traffic flows according to their QoS parameters[4],
channel aware scheduling algorithms. The goal of the work is to achieve the key
performance objectives such as complete QoS requirements, fair to all services
and using resources efficiently.
In this study, we propose a new packet scheduling scheme for Mobile WiMAX
network. The network architecture includes one BS and several MSs which are in
different locations. We schedule the uplink traffic from Mobile Stations (MS’s)
that belong to various classes to the Base Station(BS). Here we assume that
the MS’s request the BS for bandwidth by a Bandwidth request and the BS
station grants the bandwidth. The available slots for scheduling of uplink traffic
is done based on the predefined priorities of the traffic classes. The BS then
allocates the bandwidth to each MS based on the class it belongs and distance
from the BS.
In our work, we have considered all the mobile wimax parameters like transmit
power, receive gain, transmit gain, mobility, restricting the no of users to a limit
of 20 mobile station each requesting single type of service class at a time.

2 Previous Work

Our algorithm is an extension to the 2-Tier Ad-Hoc Scheduling Scheme (2T-

AHSS)[5] that makes the bandwidth allocation for the TDD mode of operation
of the IEEE 802.16 standard, and supports all types of service flows as indicated
by the standard. 2T-AHSS is fair, efficient, and granular; and it follows the IEEE
802.16 variable transmission rate and signaling mechanism rules and regulations
726 V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

through dynamic adjustments of the bandwidth allocation. 2T-AHSS operates

in two stages - Ad-Hoc STAGE and Dynamic Bandwidth Allocating Priority
Queue (DBPAQ) for scheduling different classes of Traffic.
The first stage involves the process of identifying the traffic class of the in-
coming requests by the MSs and applying corresponding packet scheduling al-
gorithms for each class. Table below gives the different traffic classes and their
corresponding scheduling algorithms.

Table 1. Traffic classes and their corresponding algorithms

Service Type Scheduling Algorithm Used

UGS traffic EDF Scheduling
ertPS traffic Weighted Fair Queuing
rtPS traffic Weighted Fair Queuing
nrtPS traffic RoundRobin Scheduling
BE traffic FIFO Scheduling

In stage 2, As all the classes are given a predefined share of total bandwidth
, previously a strict priority algorithm is used for allocating the bandwidth for
the packets in AQL (Active Queue List). As for the entire scheduling process,
we first perform the bandwidth allocation for various classes of traffic flows in a
priority manner, and then perform the admission control for each type of traffic
flows, respectively. The relevant traffic scheduling algorithm is described in the
following algorithm.

1. Traffic Scheduling at each MSi.At the beginning of each frame ft Perform

the bandwidth allocation in order for UGS(ertPS), rtPS, nrtPS and BE,
respectively
2. Get a transmission rate Ri(t)for each MSi;
3. Perform Admission Control for each type as follows
a For a UGS or ertPS session ii with request of rii
If (rii ¿ Ri(t ))
Accept flow ii; Ri (t ) ? Ri (t ) ?rii ;
Else Reject flow ii

b For an rtPS flow ij with request of rij

If (rij ¿ Ri(t ))
Accept flow ij; Ri (t ) ? Ri (t ) ?rij ;
Else If (flow ij is within its delay deadline d max(ij))
Put flow ij into the rtPS waiting queue;

c For a nrtPS flow ik with a minimal request of min (ik)

If ( min (ik) ¿ Ri (t ))
Accept flow ij; Ri (t ) ? Ri (t ) ? min (ik);
Else Put ik into the nrtPS waiting queue;
 Packet Scheduling Scheme with Enhanced Quality of Service 727

d For a BE flow il
If ( Ri(t) ¿ 0)
Allocate a fair share of Ri (t ) to il;
Else Put il into the BE waiting queue;

4. End Process

3 Proposed Work - Enhanced DBAPQ for

Improved QoS
In the above algorithm we deal with all the traffic flows in a strict priority manner
according to their service urgency status, and then schedule them according to
the available resource (bandwidth) Ri (t ).
However, it has a disadvantage of starvation as least priority classes get least
chance for bandwidth allocation.
In the Enhanced DBAPQ scheme, the bandwidth to each service class is
allocated in a dynamic way instead of allocating in static percentages [6]. Here
we introduce a new value Bandwidth Remaining (BWR). At start of scheduling,
this value is same as bandwidth of base station ie 100.
As the mobility in the mobile WiMAX is most vital factor, we need to allocate
the bandwidths dynamically based on the distance of the MS node. If the MS
node is far away from BS, then its actual data rates are lesser than sending
data rates since as distance increases, the path losses of the signal will also
increase during the transmission. So while considering mobility, we also consider
the data rates of the mobile station at different distances. But because of this
, the MS nodes far away suffer with low bandwidths and nearer MSs will have
higher bandwidths which causes the bandwidth allocation useless and may also
exceeds the actual request than actually it needed as per the requesting service
class. So we calculate the distance of the MS node and find out the bandwidth
it can actually receive at that distance. If that bandwidth is more than what
it actually needed, then remaining bandwidth is re-allocated to the MS nodes
which are suffering from lower bandwidths.
The bandwidth based on the distance from the base station that is calculated
from the formula[7]:
  2  n
P ∗ Gt ∗ Gr λ 1
SN R =  (1)
KT Bη 4 d
Where Pt=transmit power
Gt=antenna transmit
Gr=antenna gain
K= Boltzmann constant
T=temperature
B= Bandwidth
d = distance of MS
SNR=Signal Noise Ratio
728 V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

Thus our proposed algorithm helps in allocating the bandwidth dynamically

not only based on class but also the distance of the MS. As the nodes far from the
BS suffer from improper service they are allocated an equal share of bandwidth
same as those of nearer MSs. Thus it results in higher throughput of MSs which
are far from the base station same as those nearer to BS. The proposed algorithm
also makes use of unutilized bandwidth of a particular class and distributes it to
the remaining classes which are requesting for bandwidth in that particular frame.
For the purpose of Dynamic Bandwidth allocation, we introduce a new value
Bandwidth Remaining (BWR). At start of scheduling, this value is same as
bandwidth of base station ie 100% of BS bandwidth. Based on the type of service
class, we allocate the dynamic % of BWR to each service class.
The proposed scheduling scheme is able to send the data packets to the base
station at higher data rates with lesser delay compared to the existing scheme.
By using the bandwidth re-allocation technique it is possible to grant bandwidth
to all users preventing improper service and starvation.
Table below indicates the allocation of bandwidths considered for different
service classes.

Table 2. Allocated bandwidth on the frame, per service class

Service Type Default Bandwidth(%)

UGS traffic 45% of BWR
ertPS traffic 36.5% of BWR
rtPS traffic 57.2% of BWR
nrtPS traffic 75% of BWR
BE traffic 5%

4 Simulation Results

To simulate our proposed algorithm, we have developed a simulator. The sim-

ulator provides options to the end users to create as many Mobile stations and
set their properties and global properties accordingly.
As the distance varies the data rates to that mobile station varies. So as
the station moves far from the base station, the data rates decreases. Since the
resources in wireless networks are limited, another important consideration is
how to maximize the total system throughput. One of the best ways to increase
throughput is to reduce delay between the success transmissions, which is the
actual transmitted data not including the overhead and lost packets. The band-
width request is indicated in number of bytes. If we allocate the bandwidths
dynamically as per the request of the service classes, appropriate bandwidths
are allocated and the services flows are scheduled in lesser time reducing the
delays between the rounds. Since the time taken is reduced, the throughput is
increased. Our scheduling scheme using Dynamic Bandwidth allocation clearly
shows reduced delay compared to the existing scheme within a given Frame time.
Packet Scheduling Scheme with Enhanced Quality of Service 729

Fig. 1. Setting Global Parameters

Fig. 2. Setting Properties for Mobile Stations (MS)

730 V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

Fig. 3. Mobile Station statistics of the packet generation

Fig. 4. Time Frame Vs Rounds taken to schedule

Also the comparison of Data Rates for various classes of traffic with increase
in distance has been measured.
Packet Scheduling Scheme with Enhanced Quality of Service 731

Fig. 5. Distance Vs Data rates for UGS Packets

Fig. 6. Distance Vs Data rates for ERTPS Packets

732 V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

Fig. 7. Distance Vs Data rates for RTPS Packets

Fig. 8. Distance Vs Data rates for NRTPS Packets

 Packet Scheduling Scheme with Enhanced Quality of Service 733

Fig. 9. Distance Vs Data rates for BE Packets

5 Conclusion
In a mobile WiMAX scenario, The same mobile subscriber station may have dif-
ferent service requirements at different times in the WiMAX network.Scheduling
strategies that can improve network throughput, maintain relative fairness, and
lower delay over the existing scheduling schemes such as the round robin and
weighted round robin algorithms are desirable.In this project, a packet scheduling
scheme considering the QOS in mobile WiMAX network is proposed. In Mobile
WiMAX the rates at which the mobile stations transmit the data are changed
at every instant. Even in this condition the proposed scheduling scheme is able
to send the data packets to the base station with better bandwidth utilization
and lesser delay compared to the existing scheme.

References
1. Lu, S., Bharghavan, V.: Fair Scheduling in Wireless Packet Networks. IEEE/ACM
Trans. Networking 7 (1999)
2. So-In, C., Jain, R., Al-Tamim, I.A.: Generalized Weighted Fairness and its sup-
port in Deficit Round Robin with Fragmentation in IEEE 802.16 WiMAX. In: The
2nd International Conference on Computer and Automation Engineering (ICCAE),
Singapore, vol. 1, pp. 784–788 (2010)
734 V. Sampath Kumar, C. Kalyana Chakravarthy, and V.R.L. Priyanka

3. Katevenis, M., Sidiropoulos, S., Courcoubetis, C.: Weighted Round-Robin Cell

Multiplexing in a General-Purpose ATM Switch Chip. IEEE J. Selected Areas in
Comm. 9(8), 1265–1279 (1991)
4. Mobile WiMAX QoS Classes,
http://www.wimaxforum.org/technology/downloads/
Mobile WiMAX Part1 Overview and Performance.pdf
5. Esmailpour, A., Nasser, N.: Packet Scheduling Scheme with Quality of Service Sup-
port for Mobile WiMAX Networks. In: The 9th IEEE International Workshop on
Wireless Local Networks (WLN 2009), Zürich, Switzerland, October 20-23 (2009)
6. Vaidya, N.H., Bahl, P., Gupta, S.: Distributed fair scheduling in a Wireless LAN.
IEEE Trans. Mobile Comput. 4, 616–629 (2005)
7. Tran, M., Zaggoulos, G., Nix, A., Doufexi, A.: Mobile wimax: Performance Analysis
and Comparision with Experimental Results. IEEE, Los Alamitos (2008)
 Author Index

Abhilash, Kaligotla 186 Gomathi, M. 42, 55

Abirami, S. 275 Gursaran, 463
Agarwal, Vinti 476
Almas, Amera 704 Hemalatha, G. 516
Anisha, R. 42, 55 Hota, Chittaranjan 609
Arockiam, L. 619 Hundewale, Nisar 693, 704
Arul Kumar, V. 619
Asim, Mohammad 32 Indira, K. 171
Asokan, Gayathri 571
Jacob, Japhynth 179
Balakrishnan, Arunkumar 65 Jaiswal, Jyotsana 683
Balakrishnan, Kannan 534 Jayalakhsmi, Vaithiyanathan 142
Balasubramanian, K. 244 Jayamalar, T. 130
Banu Ali, Sakila 693 Jayashree, D. 638
Basak, Swati 542 Jesudasan, Isaac Balasingh 179
Baskar, S. 344 J.K., Kishore 486
Bharadwaj, Kamal K. 476 Joe Prathap, P.M. 355
Bhushan Petlu, Paul Bharath 112 John, Anita 564
Bhutkar, Ganesh 628 John Prakash, A. 233
Brindha, G.R. 355 Joy, Jestin 564

Chand, Padmabati 336 Kalyana Chakravarthy, C. 724

Charles, S. 619 Kanmani, S. 171
Chel, Haradhan 196 Kannan, S. 344
Chithralekha, T. 392 Kannapiran, B. 404
Chitra, K. 275 Karmakar, Anupam 297
Chittineni, Suresh 186 Kavitha Mayilvaganan, M. 673
Cigith, Pillai 94 Kerji, Vijay K. 225
Cijo, P. 619 Khilar, Pabitra Mohan 683
Krishnaveni, M. 211, 254, 414
D., Minnie 121 Kumar, B. Vijay 581
Das, Asha 507 Kumaran, S. 171
Das, Sanjoy 326 Kumar Anantapalli, N.L. 112
Datta, Biswajit 297 Kumari, Sunita 589
Dhandayuthabani, R. 715 Kurian, Cini 534
Dhannaseth, Nilesh 628
Lecroq, Thierry 599
El Falah, Tarek 599 Lobiyal, D.K. 326
Eliahim Jeevaraj, P.S. 376 Loganayagi, B. 442
Elloumi, Mourad 599
Maheswari, D. 265
Ganapathy, Gopinath 307 Majhi, Banshidhar 589
Gaur, Vibha 101 Majumder, Aurpan 196
Gaurav Sethia, D. 171 Malleswaran, M. 654
Geethalakshmi, S.N. 316 Mandal, Indrajit 384
736 Author Index

Modanwal, Neelam 542 Saravanaselvan, A. 654

Mohanavalli, S. 571 Satapathy, Suresh Chandra 186
Mohanty, Hrushikesha 65 Satsangi, Dharna 463
Mohanty, J.R. 336 Senthamarai Kannan, K. 275
Mohanty, Vakul 609 SenthilRajaManokar, B. 142
Mounica, Vabbilisetty 186 Seshasaayee, Ananthi 162
Mundra, Punit 496 Shanmugalakshmi, R. 367
Muralidhara Rao, M. 112 Shanmuganathan, C. 523
Murugesan, Keerthiram 11 Shanmugavadivu, P. 244, 287, 376
Shantha Kumari, K. 392
Naidu, K.B. 673 Sharma, Madhavi 496
Nandi, Debashis 196 Shukla, Shobhit 496
Naskar, Mrinal K. 297 Singh, Sanjay 486, 496
Singh, Vivek Kumar 542
Padma, S. 162 Singhal, Mukesh 11
Panimalar, S. 42, 55, 75 Sivapriya, T.R. 553
Pendlimarri, Devaki 112 Somasundaram, K. 244
Philip, Mintu 507 Soni, Anuja 101
Poovammal, E. 94 Srivastava, Kamal 463
Pothiraj, Sivakumar 432 Subashini, P. 211
Potti, Subbaraj 432 Subbaraj, P. 404
Prabavathy, K. 647 Sujatha, S. 442
Prabhakar, J. 171 Sumathi, P. 647
Prabu, M. 367 Sumathy, P. 287
Prasad Reddy, P.V.G.D. 186 Sureshkumar, K. 715
Priyanka, V.R.L. 724 Suresh Kumar, S. 19
Surianarayanan, Chellammal 307
Radha, V. 130, 254, 265, 414, 581
Rajagopal, Karthikeyan 84 Thanushkodi, K. 516
Rajsingh, Elijah Blessing 179 Thiyagarajan, M. 151
Ramesh, S. 344 T.T., Sampath Kumar 423
Ramesh Gorrila, N.V.K.D. 581
Ranjit Jeba Thangaiah, P. 553 Uma Selvi, M. 19
Raviraj, P. 523 Upadhyay, Nitin 453
Reddy, Sateesh 486 Upender, Surabhi 581
R., Gowri 423
Vadivel, A. 287
S., Kanmani 423 Vaidehi, V. 654
S., Srinivasan 121 Vaidyanathan, Sundarapandian 1, 84
Sairam, N. 384 Valliammal, N. 316
Sakthivel, A. 715 Venkateswaran, H. 233
Sampath Kumar, V. 724 Victor, Monisha 453
Samundeeswari, S. 151 Vijaya kumar, P. 715
Sapna, P.G. 65 Vimala, C. 254
Saralaya, Vikram 486 Vinil Dani, W. 355
Saravana Kumar, S. 75
Saravanan, V. 553 Yogesh, P. 638