Deep Learning in

Computer Vision
 Digital Imaging and Computer Vision Series
Series Editor
Rastislav Lukac
Foveon, Inc./Sigma Corporation San Jose, California, U.S.A.
Dermoscopy Image Analysis
by M. Emre Celebi, Teresa Mendonça, and Jorge S. Marques
Semantic Multimedia Analysis and Processing
by Evaggelos Spyrou, Dimitris Iakovidis, and Phivos Mylonas
Microarray Image and Data Analysis: Theory and Practice
by Luis Rueda
Perceptual Digital Imaging: Methods and Applications
by Rastislav Lukac
Image Restoration: Fundamentals and Advances
by Bahadir Kursat Gunturk and Xin Li
Image Processing and Analysis with Graphs: Theory and Practice
by Olivier Lézoray and Leo Grady
Visual Cryptography and Secret Image Sharing
by Stelvio Cimato and Ching-Nung Yang
Digital Imaging for Cultural Heritage Preservation: Analysis,
Restoration, and Reconstruction of Ancient Artworks
by Filippo Stanco, Sebastiano Battiato, and Giovanni Gallo
Computational Photography: Methods and Applications
by Rastislav Lukac
Super-Resolution Imaging
by Peyman Milanfar
 Deep Learning in
Computer Vision
Principles and Applications

Edited by
Mahmoud Hassaballah and Ali Ismail Awad
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2020 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-138-54442-0 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have
been made to publish reliable data and information, but the author and publisher cannot assume responsibility
for the validity of all materials or the consequences of their use. The authors and publishers have attempted to
trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if
permission to publish in this form has not been obtained. If any copyright material has not been acknowledged
please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmit-
ted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented,
including photocopying, microflming, and recording, or in any information storage or retrieval system, with-
out written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive,
Danvers, MA 01923, 978-750-8400. CCC is a not-for-proft organization that provides licenses and registration
for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate
system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used
only for identifcation and explanation without intent to infringe.

Library of Congress Cataloging‑in‑Publication Data

Names: Hassaballah, Mahmoud, editor. | Awad, Ali Ismail, editor.

Title: Deep learning in computer vision : principles and applications /
edited by M. Hassaballah and Ali Ismail Awad.
Description: First edition. | Boca Raton, FL : CRC Press/Taylor and
Francis, 2020. | Series: Digital imaging and computer vision | Includes
bibliographical references and index.
Identifers: LCCN 2019057832 (print) | LCCN 2019057833 (ebook) | ISBN
9781138544420 (hardback ; acid-free paper) | ISBN 9781351003827 (ebook)
Subjects: LCSH: Computer vision. | Machine learning.
Classifcation: LCC TA1634 .D437 2020 (print) | LCC TA1634 (ebook) | DDC
006.3/7--dc23
LC record available at https://lccn.loc.gov/2019057832
LC ebook record available at https://lccn.loc.gov/2019057833

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
Contents
Foreword ..................................................................................................................vii
Preface.......................................................................................................................ix
Editors Bio ............................................................................................................. xiii
Contributors ............................................................................................................. xv

Chapter 1 Accelerating the CNN Inference on FPGAs........................................ 1

Kamel Abdelouahab, Maxime Pelcat, and François Berry

Chapter 2 Object Detection with Convolutional Neural Networks..................... 41

Kaidong Li, Wenchi Ma, Usman Sajid, Yuanwei Wu, and
Guanghui Wang

Chapter 3 Effcient Convolutional Neural Networks for Fire Detection in

Surveillance Applications .................................................................. 63
Khan Muhammad, Salman Khan, and Sung Wook Baik

Chapter 4 A Multi-biometric Face Recognition System Based on

Multimodal Deep Learning Representations..................................... 89
Alaa S. Al-Waisy, Shumoos Al-Fahdawi, and Rami Qahwaji

Chapter 5 Deep LSTM-Based Sequence Learning Approaches for Action

and Activity Recognition.................................................................. 127
Amin Ullah, Khan Muhammad, Tanveer Hussain,
Miyoung Lee, and Sung Wook Baik

Chapter 6 Deep Semantic Segmentation in Autonomous Driving ................... 151

Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab,
Mahmoud Hassaballah, and Mohamed ElHelw

Chapter 7 Aerial Imagery Registration Using Deep Learning for

UAV Geolocalization ....................................................................... 183
Ahmed Nassar, and Mohamed ElHelw

Chapter 8 Applications of Deep Learning in Robot Vision.............................. 211

Javier Ruiz-del-Solar and Patricio Loncomilla

v
vi Contents

Chapter 9 Deep Convolutional Neural Networks: Foundations and

Applications in Medical Imaging..................................................... 233
Mahmoud Khaled Abd-Ellah, Ali Ismail Awad,
Ashraf A. M. Khalaf, and Hesham F. A. Hamed

Chapter 10 Lossless Full-Resolution Deep Learning Convolutional

Networks for Skin Lesion Boundary Segmentation......................... 261
Mohammed A. Al-masni, Mugahed A. Al-antari, and Tae-Seong Kim

Chapter 11 Skin Melanoma Classifcation Using Deep Convolutional

Neural Networks .............................................................................. 291
Khalid M. Hosny, Mohamed A. Kassem, and Mohamed M. Foaud

Index...................................................................................................................... 315
Foreword
Deep learning, while it has multiple defnitions in the literature, can be defned as
“inference of model parameters for decision making in a process mimicking the
understanding process in the human brain”; or, in short: “brain-like model iden-
tifcation”. We can say that deep learning is a way of data inference in machine
learning, and the two together are among the main tools of modern artifcial intel-
ligence. Novel technologies away from traditional academic research have fueled
R&D in convolutional neural networks (CNNs); companies like Google, Microsoft,
and Facebook ignited the “art” of data manipulation, and the term “deep learning”
became almost synonymous with decision making.
Various CNN structures have been introduced and invoked in many computer
vision-related applications, with greatest success in face recognition, autonomous
driving, and text processing. The reality is: deep learning is an art, not a science.
This state of affairs will remain until its developers develop the theory behind its
functionality, which would lead to “cracking its code” and explaining why it works,
and how it can be structured as a function of the information gained with data. In
fact, with deep learning, there is good and bad news. The good news is that the indus-
try—not necessarily academia—has adopted it and is pushing its envelope. The bad
news is that the industry does not share its secrets. Indeed, industries are never inter-
ested in procedural and textbook-style descriptions of knowledge.
This book, Deep Learning in Computer Vision: Principles and Applications—as
a journey in the progress made through deep learning by academia—confnes itself
to deep learning for computer vision, a domain that studies sensory information
used by computers for decision making, and has had its impacts and drawbacks for
nearly 60 years. Computer vision has been and continues to be a system: sensors,
computer, analysis, decision making, and action. This system takes various forms
and the fow of information within its components, not necessarily in tandem. The
linkages between computer vision and machine learning, and between it and arti-
fcial intelligence, are very fuzzy, as is the linkage between computer vision and
deep learning. Computer vision has moved forward, showing amazing progress in
its short history. During the sixties and seventies, computer vision dealt mainly with
capturing and interpreting optical data. In the eighties and nineties, geometric com-
puter vision added science (geometry plus algorithms) to computer vision. During
the frst decade of the new millennium, modern computing contributed to the evolu-
tion of object modeling using multimodality and multiple imaging. By the end of
that decade, a lot of data became available, and so the term “deep learning” crept
into computer vision, as it did into machine learning, artifcial intelligence, and other
domains.
This book shows that traditional applications in computer vision can be solved
through invoking deep learning. The applications addressed and described in the
eleven different chapters have been selected in order to demonstrate the capabilities
of deep learning algorithms to solve various issues in computer vision. The content
of this book has been organized such that each chapter can be read independently

vii
viii Foreword

of the others. Chapters of the book cover the following topics: accelerating the CNN
inference on feld-programmable gate arrays, fre detection in surveillance applica-
tions, face recognition, action and activity recognition, semantic segmentation for
autonomous driving, aerial imagery registration, robot vision, tumor detection, and
skin lesion segmentation as well as skin melanoma classifcation.
From the assortment of approaches and applications in the eleven chapters, the
common thread is that deep learning for identifcation of CNN provides accuracy
over traditional approaches. This accuracy is attributed to the fexibility of CNN
and the availability of large data to enable identifcation through the deep learning
strategy. I would expect that the content of this book to be welcomed worldwide by
graduate and postgraduate students and workers in computer vision, including prac-
titioners in academia and industry. Additionally, professionals who want to explore
the advances in concepts and implementation of deep learning algorithms applied to
computer vision may fnd in this book an excellent guide for such purpose. Finally,
I hope that readers would fnd the presented chapters in the book interesting and
inspiring to future research, from both theoretical and practical viewpoints, to spur
further advances in discovering the secrets of deep learning.

Prof Aly Farag, PhD, Life Fellow, IEEE, Fellow, IAPR

Professor of Electrical and Computer Engineering
University of Louisville, Kentucky
Preface
Simply put, computer vision is an interdisciplinary feld of artifcial intelligence that
aims to guide computers and machines toward understanding the contents of digital
data (i.e., images or video). According to computer vision achievements, the future
generation of computers may understand human actions, behaviors, and languages
similarly to humans, carry out some missions on their behalf, or even communicate
with them in an intelligent manner. One aspect of computer vision that makes it
such an interesting topic of study and active research feld is the amazing diversity
of daily-life applications such as pedestrian protection systems, autonomous driving,
biometric systems, the movie industry, driver assistance systems, video surveillance,
and robotics as well as medical diagnostics and other healthcare applications. For
instance, in healthcare, computer vision algorithms may assist healthcare profession-
als to precisely classify illnesses and cases; this can potentially save patients’ lives
through excluding inaccurate medical diagnoses and avoiding erroneous treatment.
With this wide variety of applications, there is a signifcant overlap between com-
puter vision and other felds such as machine vision and image processing. Scarcely
a month passes where we do not hear from the research and industry communities
with an announcement of some new technological breakthrough in the areas of intel-
ligent systems related to the computer vision feld.
With the recent rapid progress on deep convolutional neural networks, deep learn-
ing has achieved remarkable performance in various felds. In particular, it has brought
a revolution to the computer vision community, introducing non-traditional and eff-
cient solutions to several problems that had long remained unsolved. Due to this prom-
ising performance, it is gaining more and more attention and is being applied widely in
computer vision for several tasks such as object detection and recognition, object seg-
mentation, pedestrian detection, aerial imagery registration, video processing, scene
classifcation, autonomous driving, and robot localization as well as medical image-
related applications. If the phrase “deep learning for computer vision” is searched in
Google, millions of search results will be obtained. Under these circumstances, a book
entitled Deep Learning in Computer Vision that covers recent progress and achieve-
ments in utilizing deep learning for computer vision tasks will be extremely useful.
The purpose of this contributed volume is to fll the existing gap in the literature
for the applications of deep learning in computer vision and to provide a bird’s eye
view of recent state-of-the-art models designed for practical problems in computer
vision. The book presents a collection of eleven high-quality chapters written by
renowned experts in the feld. Each chapter provides the principles and fundamentals
of a specifc topic, introduces reviews of up-to-date techniques, presents outcomes,
and points out challenges and future directions. In each chapter, fgures, tables,
and examples are used to improve the presentation and analysis of covered topics.
Furthermore, bibliographic references are included in each chapter, providing a good
starting point for deeper research and further exploration of the topics considered in
this book. Further, this book is structured such that each chapter can be read inde-
pendently from the others as follows:

ix
x Preface

Chapter 1 presents a state-of-the-art of CNN inference accelerators over FPGAs.

Computational workloads, parallelism opportunities, and the involved memory
accesses are analyzed. At the level of neurons, optimizations of the convolutional and
fully connected layers are explained and the performances of the different methods
compared, while at the network level, approximate computing and data-path optimi-
zation methods are covered and state-of-the-art approaches compared. The methods
and tools investigated in this chapter represent the recent trends in FPGA CNN infer-
ence accelerators and will fuel future advances in effcient hardware deep learning.
Chapter 2 concentrates on object detection problem using deep CNN (DCNN): the
recent developments of several classical CNN-based object detectors are discussed.
These detectors signifcantly improve detection performance either through employ-
ing new architectures or through solving practical issues like degradation, gradi-
ent vanishing, and class imbalance. Detailed background information is provided to
show the progress and improvements of different models. Some evaluation results
and comparisons are reported on three datasets with distinctive characteristics.
Chapter 3 proposes three methods for fre detection using CNNs. The frst method
focuses on early fre detection with an adaptive prioritization mechanism for surveil-
lance cameras. The second CNN-assisted method improves fre detection accuracy with
a main focus on reducing false alarms. The third method uses an effcient deep CNN
for fre detection. For localization of fre regions, a feature map selection algorithm that
intelligently selects appropriate feature maps sensitive to fre areas is proposed.
Chapter 4 presents an accurate and real-time multi-biometric system for identi-
fying a person’s identity using a combination of two discriminative deep learning
approaches to address the problem of unconstrained face recognition: CNN and deep
belief network (DBN). The proposed system is tested on four large-scale challenging
datasets with high diversity in the facial expressions—SDUMLA-HMT, FRGC V
2.0, UFI, and LFW—and new state-of-the-art recognition rates on all the employed
datasets are achieved.
Chapter 5 introduces a study of the concept of sequence learning using RNN,
LSTM, and its variants such as multilayer LSTM and bidirectional LSTM for action
and activity recognition problems. The chapter concludes with major issues of
sequence learning for action and activity recognition and highlights recommenda-
tions for future research.
Chapter 6 discuses semantic segmentation in autonomous driving applications,
where it focuses on constructing effcient and simple architectures to demonstrate
the beneft of fow and depth augmentation to CNN-based semantic segmentation
networks. The impact of both motion and depth information on semantic segmenta-
tion is experimentally studied using four simple network architectures. Results of
experiments on two public datasets—Virtual-KITTI and CityScapes—show reason-
able improvement in overall accuracy.
Chapter 7 presents a method based on deep learning for geolocalizing drones
using only onboard cameras. A pipeline has been implemented that makes use of the
availability of satellite imagery and traditional computer vision feature detectors and
descriptors, along with renowned deep learning methods (semantic segmentation), to
be able to locate the aerial image captured from the drone within the satellite imag-
ery. The method enables the drone to be autonomously aware of its surroundings and
navigate without using GPS.
Preface xi

Chapter 8 is intended to be a guide for the developers of robot vision systems,

focusing on the practical aspects of the use of deep neural networks rather than on
theoretical issues.
The last three chapters are devoted to deep learning in medical applications.
Chapter 9 covers basic information about CNNs in medical applications. CNN
developments are discussed from different perspectives, specifcally, CNN design,
activation function, loss function, regularization, optimization, normalization, and
network depth. Also, a deep convolutional neural network (DCNN) is designed for
brain tumor detection using MRI images. The proposed DCNN architecture is eval-
uated on the RIDER dataset, achieving accurate detection accuracy within a time of
0.24 seconds per MRI image.
Chapter 10 discusses automatic segmentation of skin lesion boundaries from sur-
rounding tissue and presents a novel deep learning segmentation methodology via
full-resolution convolutional network (FrCN). Experimental results show the great
promise of the FrCN method compared to state-of-the-art deep learning segmenta-
tion approaches such as fully convolutional networks (FCN), U-Net, and SegNet
with overall segmentation.
Chapter 11 is about the automatic classifcation of color skin images, where a
highly accurate method is proposed for skin melanoma classifcation utilizing two
modifed deep convolutional neural networks and consisting of three main steps.
The proposed method is tested using the well-known MED-NODE and DermIS &
DermQuest datasets.
It is very necessary to mention here that the book is a small piece in the puzzle
of computer vision and its applications. We hope that our readers fnd the presented
chapters in the book interesting and that the chapters will inspire future research
both from theoretical and practical viewpoints to spur further advances in the com-
puter vision feld.
The editors would like to take this opportunity to express their sincere grati-
tude to the contributors for extending their wholehearted support in sharing some
of their latest results and fndings. Without their signifcant contribution, this
book could not have fulflled its mission. The reviewers deserve our thanks for
their constructive and timely input. Special profound thanks go to Prof Aly Farag,
Professor of Electrical and Computer Engineering, University of Louisville,
Kentucky for writing the Foreword for this book. Finally, the editors acknowledge
the efforts of the CRC Press Taylor & Francis for giving us the opportunity to edit
a book on deep learning for computer vision. In particular, we would like to thank
Dr Rastislav Lukac, the editor of the Digital Imaging and Computer Vision book
series, and Nora Konopka for initiating this project. Really, the editorial staff
at CRC Press has done a meticulous job, and working with them was a pleasant
experience.

Mahmoud Hassaballah
Qena, Egypt

Ali Ismail Awad

Luleå, Sweden
Editors Bio
Mahmoud Hassaballah was born in 1974, Qena, Egypt. He
received his BSc degree in Mathematics in 1997 and his MSc
degree in Computer Science in 2003, both from South Valley
University, Egypt, and his Doctor of Engineering (D Eng) in
computer science from Ehime University, Japan in 2011. He
was a visiting scholar with the department of computer &
communication science, Wakayama University, Japan in
2013 and GREAH laboratory, Le Havre Normandie
University, France in 2019. He is currently an associate professor of computer sci-
ence at the faculty of computers and information, South Valley University, Egypt. He
served as a reviewer for several journals such as IEEE Transactions on Image
Processing, IEEE Transactions on Fuzzy Systems, Pattern Recognition, Pattern
Recognition Letters, IET Image Processing, IET Computer Vision, IET Biometrics,
Journal of Real-Time Image Processing, and Journal of Electronic Imaging. He has
published over 50 research papers in refereed international journals and conferences.
His research interests include feature extraction, object detection/recognition, artif-
cial intelligence, biometrics, image processing, computer vision, machine learning,
and data hiding.

Ali Ismail Awad (SMIEEE, PhD, PhD, MSc, BSc) is cur-

rently an Associate Professor (Docent) with the Department
of Computer Science, Electrical, and Space Engineering,
Luleå University of Technology, Luleå, Sweden, where
he also serves as a Coordinator of the Master Programme
in Information Security. He is a Visiting Researcher with
the University of Plymouth, United Kingdom. He is also
an Associate Professor with the Electrical Engineering
Department, Faculty of Engineering, Al-Azhar University at Qena, Qena, Egypt. His
research interests include information security, Internet-of-Things security, image
analysis with applications in biometrics and medical imaging, and network security.
He has edited or co-edited fve books and authored or co-authored several journal
articles and conference papers in these areas. He is an Editorial Board Member of the
following journals: Future Generation Computer Systems, Computers & Security,
Internet of Things: Engineering Cyber Physical Human Systems, and Health
Information Science and Systems. Dr Awad is currently an IEEE senior member.

xiii
Contributors
Ahmad El Sallab Hesham F.A. Hamed
Valeo Company Egyptian Russian University
Cairo, Egypt Cairo, Egypt
and
Ahmed Nassar Minia University
IRISA Institute Minia, Egypt
Rennes, France
Javier Ruiz-Del-Solar
Alaa S. Al-Waisy University of Chile
University of Bradford Santiago, Chile
Bradford, UK
Kaidong Li
University of Kansas
Ali Ismail Awad
Kansas City, Kansas
Luleå University of Technology
Luleå, Sweden
Kamel Abdelouahab
and
Clermont Auvergne University
Al-Azhar University
Clermont-Ferrand, France
Qena, Egypt
Khalid M. Hosny
Amin Ullah Zagazig University
Sejong University Zagazig, Egypt
Seoul, South Korea
Khan Muhammad
Ashraf A. M. Khalaf Sejong University
Minia University Seoul, South Korea
Minia, Egypt
Mahmoud Hassaballah
François Berry South Valley University
University Clermont Auvergne Qena, Egypt
Clermont-Ferrand, France
Mahmoud Khaled Abd-Ellah
Guanghui Wang Al-Madina Higher Institute for
University of Kansas Engineering and Technology
Kansas City, Kansas Giza, Egypt

Hazem Rashed Maxime Pelcat

Valeo Company University of Rennes
Cairo, Egypt Rennes, France

xv
xvi Contributors

Miyoung Lee Salman Khan

Sejong University Sejong University
Seoul, South Korea Seoul, South Korea

Mohamed A. Kassem Senthil Yogamani

Kafr El Sheikh University Valeo Company
Kafr El Sheikh, Egypt Galway, Ireland

Mohamed Elhelw Shumoos Al-Fahdawi

Nile University University of Bradford
Giza, Egypt Bradford, UK

Sung Wook Baik

Mohamed M. Foaud
Sejong University
Zagazig University
Seoul, South Korea
Zagazig, Egypt
Tae-Seong Kim
Mohammed A. Al-Masni Kyung Hee University
Kyung Hee University Seoul, South Korea
Seoul, South Korea
and Tanveer Hussain
Yonsei University Sejong University
Seoul, South Korea Seoul, South Korea

Mugahed A. Al-Antari Usman Sajid

Kyung Hee University University of Kansas
Seoul, South Korea Kansas City, Kansas
and
Sana’a Community College Wenchi Ma
Sana’a, Republic of Yemen University of Kansas
Kansas City, Kansas
Patricio Loncomilla
University of Chile Yuanwei Wu
Santiago, Chile University of Kansas
Kansas City, Kansas
Rami Qahwaji
University of Bradford
Bradford, UK
 1 Accelerating the CNN
Inference on FPGAs
Kamel Abdelouahab, Maxime Pelcat,
and François Berry

CONTENTS
1.1 Introduction ......................................................................................................2
1.2 Background on CNNs and Their Computational Workload ............................3
1.2.1 General Overview.................................................................................3
1.2.2 Inference versus Training ..................................................................... 3
1.2.3 Inference, Layers, and CNN Models ....................................................3
1.2.4 Workloads and Computations...............................................................6
1.2.4.1 Computational Workload .......................................................6
1.2.4.2 Parallelism in CNNs ..............................................................8
1.2.4.3 Memory Accesses ..................................................................9
1.2.4.4 Hardware, Libraries, and Frameworks ................................ 10
1.3 FPGA-Based Deep Learning.......................................................................... 11
1.4 Computational Transforms ............................................................................. 12
1.4.1 The im2col Transformation ................................................................ 13
1.4.2 Winograd Transform .......................................................................... 14
1.4.3 Fast Fourier Transform ....................................................................... 16
1.5 Data-Path Optimizations ................................................................................ 16
1.5.1 Systolic Arrays.................................................................................... 16
1.5.2 Loop Optimization in Spatial Architectures ...................................... 18
Loop Unrolling ................................................................................... 19
Loop Tiling .........................................................................................20
1.5.3 Design Space Exploration................................................................... 21
1.5.4 FPGA Implementations ...................................................................... 22
1.6 Approximate Computing of CNN Models ..................................................... 23
1.6.1 Approximate Arithmetic for CNNs.................................................... 23
1.6.1.1 Fixed-Point Arithmetic ........................................................ 23
1.6.1.2 Dynamic Fixed Point for CNNs...........................................28
1.6.1.3 FPGA Implementations ....................................................... 29
1.6.1.4 Extreme Quantization and Binary Networks....................... 29
1.6.2 Reduced Computations....................................................................... 30
1.6.2.1 Weight Pruning .................................................................... 31
1.6.2.2 Low Rank Approximation ................................................... 31
1.6.2.3 FPGA Implementations ....................................................... 32

1
2 Deep Learning in Computer Vision

1.7 Conclusions..................................................................................................... 32
Bibliography ............................................................................................................ 33

1.1 INTRODUCTION
The exponential growth of big data during the last decade motivates for innovative
methods to extract high semantic information from raw sensor data such as videos,
images, and speech sequences. Among the proposed methods, convolutional neural
networks (CNNs) [1] have become the de facto standard by delivering near-human
accuracy in many applications related to machine vision (e.g., classifcation [2],
detection [3], segmentation [4]) and speech recognition [5].
This performance comes at the price of a large computational cost as CNNs
require up to 38 GOPs to classify a single frame [6]. As a result, dedicated hard-
ware is required to accelerate their execution. Graphics processing units GPUs
are the most widely used platform to implement CNNs as they offer the best per-
formance in terms of pure computational throughput, reaching up 11 TFLOPs
[7]. Nevertheless, in terms of power consumption, feld-programmable gate array
(FPGA) solutions are known to be more energy effcient (vs. GPU). While GPU
implementations have demonstrated state-of-the-art computational performance,
CNN acceleration will soon be moving towards FPGAs for two reasons. First,
recent improvements in FPGA technology put FPGA performance within striking
distance of GPUs with a reported performance of 9.2 TFLOPs for the latter [8].
Second, recent trends in CNN development increase the sparsity of CNNs and
use extremely compact data types. These trends favor FPGA devices, which are
designed to handle irregular parallelism and custom data types. As a result, next-
generation CNN accelerators are expected to deliver up to 5.4× better computa-
tional throughput than GPUs [7].
As an infection point in the development of CNN accelerators might be near, we
conduct a survey on FPGA-based CNN accelerators. While a similar survey can be
found in [9], we focus in this chapter on the recent techniques that were not covered
in the previous works. In addition to this chapter, we refer the reader to the works
of Venieris et al. [10], which review the toolfows automating the CNN mapping
process, and to the works of Sze et al., which focus on ASICs for deep learning
acceleration.
The amount and diversity of research on the subject of CNN FPGA acceleration
within the last 3 years demonstrate the tremendous industrial and academic interest.
This chapter presents a state-of-the-art review of CNN inference accelerators over
FPGAs. The computational workloads, their parallelism, and the involved memory
accesses are analyzed. At the level of neurons, optimizations of the convolutional
and fully connected (FC) layers are explained and the performances of the differ-
ent methods compared. At the network level, approximate computing and data-path
optimization methods are covered and state-of-the-art approaches compared. The
methods and tools investigated in this survey represent the recent trends in FPGA
CNN inference accelerators and will fuel the future advances on effcient hardware
deep learning.
Accelerating the CNN Inference on FPGAs 3

1.2 BACKGROUND ON CNNS AND THEIR

COMPUTATIONAL WORKLOAD
In this frst section, we overview the main features of CNNs, mainly focusing on the
computations and parallelism patterns involved during their inference.

1.2.1 GENERAL OVERVIEW

Deep* CNNs are feed-forward†, sparsely connected‡ neural networks. A typical
CNN structure consists of a pipeline of layers. Each layer inputs a set of data, known
as a feature map (FM), and produces a new set of FMs with higher-level semantics.

1.2.2 INFERENCE VERSUS TRAINING

As typical machine learning algorithms, CNNs are deployed in two phases. First,
the training stage works on a known set of annotated data samples to create a model
with a modeling power (which semantics extrapolates to natural data outside the
training set). This phase implements the back-propagation algorithm [11], which
iteratively updates CNN parameters such as convolution weights to improve the pre-
dictive power of the model. A special case of CNN training is fne-tuning. When
fne-tuning a model, weights of a previously trained network are used to initialize the
parameters of a new training. These weights are then adjusted for a new constraint,
such as a different dataset or a reduced precision.
The second phase, known as inference, uses the learned model to classify new data
samples (i.e., inputs that were not previously seen by the model). In a typical setup, CNNs
are trained/fne-tuned only once, on large clusters of GPUs. By contrast, the inference
is implemented each time a new data sample has to be classifed. As a consequence,
the literature mostly focuses on accelerating the inference phase. As a result, our dis-
cussion overviews the main methods employed to accelerate the inference. Moreover,
since most of the CNN accelerators benchmark their performance on models trained for
image classifcation, we focus our chapter on this application. Nonetheless, the methods
detailed in this survey can be employed to accelerate CNNs for other applications such
object detection, image segmentation, and speech recognition.

1.2.3 INFERENCE, LAYERS, AND CNN MODELS

CNN inference refers to the feed-forward propagation of B input images across L
layers. This section details the computations involved in the major types of these
layers. A common practice is to manipulate layers, parameters, and FMs as multidi-
mensional arrays, as listed in Table 1.1. Note that when it will be relevant, the type
of the layer will be denoted with superscript, and the position of the layer will be
denoted with subscript.

* Includes a large number of layer, typically above three.

† The information fows from the neurons of a layer ˜ towards the neurons of a layer. ˜ + 1
‡ CNNs implement the weight sharing technique, applying a small number of weights across all the

input pixels (i.e., image convolution).

4 Deep Learning in Computer Vision

TABLE 1.1
Tensors Involved in the Inference of a Given Layer ˜ with Their Dimensions
X Input FMs B×C×H×W B Batch size (Number of input frames)
Y Output FMs B×N×V×U W/H/C Width/Height/Depth of Input FMs
Θ Learned Filters N×C×J×K U/V/N Width/Height/Depth of Output FMs
β Learned biases N K/J Horizontal/Vertical Kernel size

A convolutional layer (conv) carries out the feature extraction process by applying – as
illustrated in Figure 1.1 – a set of three-dimensional convolution flters Θconv to a set
of B input volumes Xconv. Each input volume has a depth C and can be a color image
(in the case of the frst conv layer), or an output generated by previous layers in the
network. Applying a three-dimensional flter to three-dimensional input results in
a 2D (FM). Thus, applying N three-dimensional flters in a layer results in a three-
dimensional output with a depth N.
In some CNN models, a learned offset βconv – called a bias – is added to processed
feature maps. However, this practice has been discarded in recent models [6]. The
computations involved in feed-forward propagation of conv layers are detailed in
Equation 1.1.

˜ {b, n, u, v} ˝[1, B ] × [1, N ] × [1,V ] × [1,U ]

Y conv[b, n, v, u] = b conv[ n]
C J K (1.1)
+ åååX
c=1 j=1 k=1
conv
[b, c, v + j, u + k ] × Qconv[ n, c, j, k]

One may note that applying a depth convolution to a 3D input boils down to applying
a mainstream 2D convolution to each of the 2D channels of the input, then, at each
point, summing the results across all the channels, as shown in Equation 1.2.

FIGURE 1.1 Feed-forward propagation in conv, act, and pool layers (batch size B = 1, bias
β omitted).
Accelerating the CNN Inference on FPGAs 5

˜n °[1, N ]

C
Y[ n] conv
=b conv
[n] + åconv2D ( X[c]
c=1
conv
,Q[ c]conv ) (1.2)

Each conv layer of a CNN is usually followed by an activation layer that applies a
nonlinear function to all the values of FMs. Early CNNs were trained with TanH
or Sigmoid functions, but recent models employ the rectifed linear unit (ReLU)
function, which grants faster training times and less computational complexity, as
highlighted in Krizhevsky et al. [12].

˜ {b, n, u, v} ˝[1, B ] × [1, N ] × [1,V ] × [1,U ]

Y act [b, n, h, w] = act(X act [b, n, h, w]) | act:=TanH, Sigmoid, ReLU… (1.3)

The convolutional and activation parts of a CNN are directly inspired by the
cells of visual cortex in neuroscience [13]. This is also the case with pooling
layers, which are periodically inserted in between successive conv layers. As
shown in Equation 1.4, pooling sub-samples each channel of the input FM by
selecting either the average, or, more commonly, the maximum of a given neigh-
borhood K. As a result, the dimensionality of an FM is reduced, as illustrated
in Figure 1.1.

˜ {b, n, u, v} ˝[1, B ] × [1, N ] × [1,V ] × [1,U ]

(
Y pool [b, n, v, u] = max X pool [b, n, v + p, u + q]
p,q˜[1:K ]
) (1.4)

When deployed for classifcation purposes, the CNN pipeline is often terminated
by FC layers. In contrast with convolutional layers, FC layers do not implement
weight sharing and involve as much weight as input data (i.e., W = K, H = J,U = V = 1).
Moreover, in a similar way as conv layers, a nonlinear function is applied to the
outputs of FC layers.

˜ {b, n} ˝[1, B ] × [1, N ]

C H W
Y [b, n] = b [ n] +
fc fc
åååX [b, c, h, w]× Q [n, c, h, w]
c=1 h=1 w=1
fc fc
(1.5)
6 Deep Learning in Computer Vision

The Softmax function is a generalization of the Sigmoid function, and “squashes”

a N-dimensional vector X to Sigmoid(X) where each output is in the range [0,1].
The Softmax function is used in various multi-class classifcation methods, espe-
cially in CNNs. In this case, the Softmax layer is placed at the end of the net-
work and the dimension of vector it operates on (i.e., N) represents the number of
classes in the considered dataset. Thus, the input of the Softmax is the data gener-
ated by the last fully connected layer, and the output is the probability predicted
for each class.

˜ {b, n} ˝[1, B ] × [1, N ]

exp(X[b, n]) (1.6)

Softmax(X[b, n]) =
°
N
exp(X[b, c])
c=1

Batch normalization was introduced [14] to speed up training by linearly shifting

and scaling the distribution of a given batch of inputs B to have zero mean and unit
variance. These layers fnd also their interest when implementing binary neural net-
works (BNNs) as they reduce the quantization error compared to an arbitrary input
distribution, as highlighted in Hubara et al. [15]. Equation 1.7 details the processing
of batch norm layers, where the mean μ and the variance σ are statistics collected
during the training, α and γ are parameters learned during the training, and ϵ is a
hyper-parameter set empirically for numerical stability purposes (i.e., avoiding divi-
sion by zero).

˜ {b, n, u, v} ˝[1, B ] × [1, N ] × [1,V ] × [1,U ]

X BN [b, n, u, v] − m (1.7)
Y BN [b, n, v, u] = g+a
s2 + ˜

1.2.4 WORKLOADS AND COMPUTATIONS

The accuracy of CNN models has been increasing since their breakthrough in 2012
[12]. However, this accuracy comes at a high computational cost. The main challenge
that faces CNN developers is to improve classifcation accuracy while maintain-
ing a tolerable computational workload. As shown in Table 1.2, this challenge was
successfully addressed by Inception [16] and ResNet models [17], with their use of
bottleneck 1 × 1 convolutions that reduce both model size and computations while
increasing depth and accuracy.

1.2.4.1 Computational Workload

As shown in Equations 1.1 and 1.5, the processing of CNN involves an intensive use
of Multiply Accumulate (MAC) operation. All these MAC operations take place at
conv and FC layers, while the remaining parts of network are element-wise trans-
formations that can be generally implemented with low-complexity computational
requirements.
Accelerating the CNN Inference on FPGAs 7

TABLE 1.2
Popular CNN Models with Their Computational Workload*
AlexNet GoogleNet VGG16 VGG19 ResNet101 ResNet-152
Model [12] [16] [6] [6] [17] [17]

Top1 err (%) 42.9% 31.3% 28.1% 27.3% 23.6% % 23.0%

Top5 err (%) 19.80% 10.07% 9.90% 9.00% 7.1% 6.7%
Lc 5 57 13 16 104 155

˜
Lc
conv 666 M 1.58 G 15.3 G 19.5 G 7.57 G 11.3 G
C˜
˜=1

˜
Lc 2.33 M 5.97 M 14.7 M 20 M 42.4 M 58 M
W˜conv
˜=1

Act ReLU
Pool 3 14 5 5 2 2
Lf 3 1 3 3 1 1

˜
Lf 58.6 M 1.02 M 124 M 124 M 2.05 M 2.05 M
C˜fc
˜=1

˜
Lf 58.6 M 1.02 M 124 M 124 M 2.05 M 2.05 M
W˜ fc
˜=1

C 724 M 1.58 G 15.5 G 19.6 G 7.57 G 11.3 G

W 61 M 6.99 M 138 M 144 M 44.4 M 60 M

* Accuracy Measured on Single-Crops of ImageNet Test-Set

In this chapter, the computational workload C of a given CNN corresponds to the

number of MACs it involves during inference*. The number of these MACs mainly
depends on the topology of the network, and more particularly on the number of
conv and FC layers and their dimensions. Thus, the computational workload can be
expressed as in Equation 1.8, where L c is the number of conv (fully connected) lay-
ers, and C˜conv (C˜fc ) is the number of MACs occurring on a given convolution (fully
connected) layer ˜.

Lc Lf

C= ˜C
˜=1
˜
conv
+ ˜C
˜=1
˜
fc
(1.8)

C˜conv = N ˜ × C˜ × J ˜ × K ˜ × U ˜ × V˜ (1.9)

C˜fc = N ˜ × C˜ × W˜ × H ˜ (1.10)

* Batch size is set to 1 for clarity purposes.

8 Deep Learning in Computer Vision

In a similar way, the number of weights, and consequently the size of a given CNN
model, can be expressed as follows:

Lc Lf

W= ˜W
˜=1
˜
conv
+ ˜W
˜=1
˜
fc
(1.11)

W˜conv = N ˜ × C˜ × J ˜ × K ˜ (1.12)

W˜fc = N ˜ × C˜ × W˜ × H ˜ (1.13)

For state-of-the-art CNN models, L c, N ˜ , and C˜ can be quite large. This makes
CNNs computationally and memory intensive, where for instance, the classifcation
of a single frame using the VGG19 network requires 19.5 billion MAC operations.
It can be observed in the same table that most of the MACs occur on the convolu-
tion parts, and consequently, 90% of the execution time of a typical inference is spent
on conv layers [18]. By contrast, FC layers marginalize most of the weights and thus
the size of a given CNN model.

1.2.4.2 Parallelism in CNNs

The high computational workload of CNNs makes their inference a challenging task,
especially on low-energy embedded devices. The key solution to this challenge is to
leverage on the extensive concurrency they exhibit. These parallelism opportunities
can be formalized as follows:

• Batch Parallelism: CNN implementations can simultaneously classify

multiple frames grouped as a batch B in order to reuse the flters in each
layer, minimizing the number the memory accesses. However, and as
shown in [10], batch parallelism quickly reaches its limits. This is due to the
fact that most of the memory transactions result from storing intermediate
results and not loading CNN parameters. Consequently, reusing the flters
only slightly impacts the overall processing time per image.
• Inter-layer Pipeline Parallelism: CNNs have a feed-forward hierarchical
structure consisting of a succession of data-dependent layers. These layers
can be executed in a pipelined fashion by launching layer (˜) before ending
the execution of layer (˜ −1). This pipelining costs latency but increases
throughput.

Moreover, the execution of the most computationally intensive parts (i.e., conv lay-
ers), exhibits the four following types of concurrency:

• Inter-FM Parallelism: Each two-dimensional plane of an FM can be pro-

cessed separately from the others, meaning that PN elements of Yconv can be
computed in parallel (0 < PN < N).
Accelerating the CNN Inference on FPGAs 9

• Intra-FM Parallelism: In a similar way, pixels of a single output FM plane

are data-independent and can thus be processed concurrently by evaluating
PV × PU values of Yconv[n] (0 < PV × PU < V × U).
• Inter-convolution Parallelism: Depth convolutions occurring in conv lay-
ers can be expressed as a sum of 2D convolutions, as shown in Equation
1.2. These 2D convolutions can be evaluated simultaneously by computing
concurrently Pc elements (0 < Pc < C).
• Intra-convolution Parallelism: The 2D convolutions involved in the pro-
cessing of conv layers can be implemented in a pipelined fashion such as
in [76]. In this case PJ × PK multiplications are implemented concurrently
(0 < PJ × PK < J × K).

1.2.4.3 Memory Accesses

As a consequence of the previous discussion, the inference of a CNN shows large vec-
torization opportunities that can be exploited by allocating multiple computational
resources to concurrently process multiple features. However, this parallelization
can not accelerate the execution of a CNN if no datacaching strategy is implemented.
In fact, memory bandwidth is often the bottleneck when processing CNNs.
In FC parts, the execution can be memory-bounded because of the high number
of weights that these layers contain, and consequently, the high number of memory
reads required.
This is expressed in Equation 1.14, where M˜fc refers to the number of memory
accesses occurring in an FC layer ˜. This number can be written as the sum of
memory accesses reading the inputs X fc˜ , the memory accesses reading the weights
(q ˜fc), and the number of memory accesses writing the results (Y˜fc ).

M˜fc = MemRd(X ˜fc ) + MemRd(q ˜fc ) + MemWr(Y˜fc ) (1.14)

= C˜ H ˜W˜ + N ˜C˜ H ˜W˜ + N ˜ (1.15)

˜ N ˜C˜ H ˜W˜ (1.16)

Note that the fully connected parts of state-of-the-art models involve large values
of N ˜ and C˜ , making the memory reading of weights the most impacting factor,
as formulated in Equation 1.16. In this context, batch parallelism can signifcantly
accelerate the execution of CNNs with a large number of FC layers.
In the conv parts, the high number of MAC operations results in a high number
of memory accesses, as each MAC requires at least 2 memory reads and 1 memory
write*. This number of memory accesses accumulates with the high dimensions of
data manipulated by conv layers, as shown in Equation 1.18. If all these accesses are
towards external memory (for instance, DRAM), throughput and energy consumption

* This is the best-case scenario of a fully pipelined MAC, where intermediate results do not need to be
loaded.
10 Deep Learning in Computer Vision

will be highly impacted, because DRAM access engenders high latency and energy
consumption, even more than the computation itself [21].

M˜conv = MemRd(X ˜conv ) + MemRd(q ˜conv ) + MemWr(Y˜conv ) (1.17)

= C˜ H ˜W˜ + N ˜C˜ J ˜ K ˜ + N ˜U ˜V˜ (1.18)

The number of these DRAM accesses, and thus latency and energy consumption, can
be reduced by implementing a memory-caching hierarchy using on-chip memories.
As discussed in the next sections, state-of-the-art CNN accelerators employ register
fles as well as several levels of caches. The former, being the fastest, is implemented
at the nearest of the computational capabilities. The latency and energy consumption
resulting from these caches is lower by several orders of magnitude than external
memory accesses, as pointed out in Sze et al. [22].

1.2.4.4 Hardware, Libraries, and Frameworks

In order to catch the parallelism of CNNs, dedicated hardware accelerators are
developed. Most of them are based on GPUs, which are known to perform well
on regular parallelism patterns thanks to simd and simt execution models, a dense
collection of foating-point computing elements that peak at 12 TFLOPs, and high
capacity/bandwidth on/off-chip memories [23]. To support these hardware accel-
erators, specialized libraries for deep learning are developed to provide the neces-
sary programming abstraction, such as CudNN on Nvidia GPU [24]. Built upon
these libraries, dedicated frameworks for deep learning are proposed to improve
productivity of conceiving, training, and deploying CNNs, such as Caffe [25] and
TensorFlow [26].
Beside GPU implementations, numerous FPGA accelerators for CNNs have been
proposed. FPGAs are fne-grained programmable devices that can catch the CNN
parallelism patterns with no memory bottleneck, thanks to the following:

1. A high density of hard-wired digital signal processor (DSP) blocks that are
able to achieve up to 20 (8 TFLOPs) TMACs [8].
2. A collection of in situ on-chip memories, located next to DSPs, that can be
exploited to signifcantly reduce the number of external memory accesses.

As a consequence, CNNs can beneft from a signifcant acceleration when running

on reconfgurable hardware. This has caused numerous research efforts to study
FPGA-based CNN acceleration, targeting both high performance computing (HPC)
applications [27] and embedded devices [28].
In the remaining parts of this chapter, we conduct a survey on methods and hard-
ware architectures to accelerate the execution of CNN on FPGA. The next section
lists the evaluation metrics used, then Sections 1.4 and 1.5 respectively study the
computational transforms and the data-path optimization involved in recent CNN
accelerators. Finally, the last section of this chapter details how approximate com-
puting is a key in FPGA-based deep learning, and overviews the main contributions
implementing these techniques.
Accelerating the CNN Inference on FPGAs 11

1.3 FPGA-BASED DEEP LEARNING

Accelerating a CNN on an FPGA-powered platform can be seen as an optimization
effort that focuses on one or several of the following criteria:

• Computational Throughput (T ): A large number of the works studied

in this chapter focus on reducing the CNN execution times on the FPGA
(i.e., the computation latency), by improving the computational throughput
of the accelerator. This throughput is usually expressed as the number of
MACs an accelerator performs per second. While this metric is relevant in
the case of HPC workloads, we prefer to report the throughput as the num-
ber of frames an accelerator processes per second (fps), which better suits
the embedded vision context. The two metrics can be directly related using
Equation 1.19, where C is defned in Equation 1.8, and refers to the number
of computations a CNN involve in order to process a single frame:

T (MACS)
T (FPS) = (1.19)
C(MAC)

• Classifcation/Detection Perf. (A ) : Another way to reduce CNN execution

times is to trade some of their modeling performance in favor of faster exe-
cution timings. For this reason, the classifcation and detection metrics are
reported, especially when dealing with approximate computing methods.
Classifcation performance is usually reported as top-1 and top-5 accura-
cies, and detection performance is reported using the mAP50 and mAP75
metrics.
• Energy and Power Consumption (P ): Numerous FPGA-based accelera-
tion methods can be categorized as either latency-driven or energy-driven.
While the former focus on improving the computational throughput, the
latter considers the power consumption of the accelerator, reported in watts.
Alternatively, numerous latency-driven accelerators can be ported to low-
power-range FPGAs and perform well under strict power consumption
requirements.
• Resource Utilization (R ): When it comes to FPGA acceleration, the utili-
zation of the available resources (lut, DSP blocks, sram blocks) is always
considered. Note that the resource utilization can be correlated to the power
consumption*, but improving the ratio between the two is a technological
problem that clearly exceeds the scope of this chapter. For this reason, both
power consumption and resources utilization metrics will be reported when
available.

An FPGA implementation of a CNN has to satisfy to the former requirements. In this

perspective, the literature provides three main approaches to address the problem

* At a similar number of memory accesses. These accesses typically play the most dominant role in the
power consumption of an accelerator.
12 Deep Learning in Computer Vision

FIGURE 1.2 Main approaches to accelerate CNN inference on FPGAs.

of FPGA-based deep learning. These approaches mainly consists of computational

transforms, data-path optimization, and approximate computing techniques, as illus-
trated in Figure 1.2.

1.4 COMPUTATIONAL TRANSFORMS

In order to accelerate the execution of conv and FC layers, numerous implementa-
tions rely on computational transforms. These transforms, which operate on the FM
and weight arrays, aim at vectorizing the implementations and reducing the number
of operations occurring during inference.
Three main transforms can be distinguished. The im2col method reshapes the
feature and weight arrays in a way to transform depth convolutions into matrix mul-
tiplications. The FFT method operates on the frequency domain, transforming con-
volutions into multiplications. Finally, in Winograd fltering, convolutions boil down
to element-wise matrix multiplications thanks to a tiling and a linear transformation
of data.
These computational transforms mainly appear in temporal architectures and are
implemented by means of variety of linear algebra libraries such OpenBLAS for
CPUs* or cuBLAS for GPUs†. Besides this, various implementations make use of
these transforms to effciently map CNNs on FPGAs.
This section discusses the three former methods, highlighting their use-cases and
computational improvements. For a better understanding, we recall that for each
layer ˜:

• The input feature map is represented as four-dimensional array X, in which

the dimensions B × C × H × W respectively refer to the batch size, the num-
ber of input channels, the height, and the width.

* https://www.openblas.net/
† https://developer.nvidia.com/cublas
Accelerating the CNN Inference on FPGAs 13

• The weights are represented as four-dimensional array Θ, in which the

dimensions N × C × J × K respectively refer to the depth of the output fea-
ture map, the depth of the input feature map, the vertical, and the horizontal
kernel size.

1.4.1 THE IM2COL TRANSFORMATION

In CPUs and GPUs, a common way to process CNNs is to map conv and FC layers
as general matrix multiplications (GEMMs). A number of studies generalize this
approach to FPGA-based implementations.
For FC layers, in which the processing boils down to a matrix-vector multiplica-
tion problem, the GEMM-based implementations fnd their interest when processing
a batch of FMs. As mentioned in Section 1.2.4.1, most of the weights of CNNs are
employed in the FC parts. Instead of loading these weights multiple times to clas-
sify multiple inputs, features extracted from a batch of inputs are concatenated onto
a CHW × B matrix. In this case, the weights are loaded only one time per batch,
as depicted in Figure 1.3a. As a consequence, the former Equation 1.16 – which
expressed the number of memory accesses occurring on FC layers – becomes the
following:

M˜fc = MemRd(q ˜fc ) + MemRd(X ˜fc ) + MemWr(Y˜fc ) (1.20)

= N ˜C˜W˜ H ˜ + BC˜ H ˜W˜ + BN ˜ (1.21)

˜ N ˜C˜ H ˜W˜ (1.22)

As detailed in Section 1.2.4.2, the vectorization of FC layers is often employed in

GPU implementations to increase the computational throughput while maintaining
a constant memory bandwidth utilization. The same concept holds true for FPGA
implementations [31, 48, 49], which batch the FC layers to map them as GEMMs.

FC Weights ˜fc Input FMs Xfc ˜

W
C
1
1 1 1
1 1
1 1 1 1 1 1 1 1 1
1 1
1 1 1 1 1 1 1 1 1

H
1 1 1 1 1 1

N B

Output
N
x N FMs =
= ~
Yfc

CHW B
CHW

(a) B (b)
FIGURE 1.3 GEMM-based processing of FC layers (a) and conv layers (b).
14 Deep Learning in Computer Vision

3D convolutions can also be mapped as GEMMs using the so-called im2col

method introduced in [30]. First, this method fattens all the weights of a given conv
layer onto an N × CKJ matrix ˜ ˜ . Second, it rearranges the input feature maps onto a
CKJ × UV matrix X̃ , squashing each feature map to a column*. With these reshaped
data, the output feature maps Ỹ are computed by multiplying of two former matrices,
as illustrated in Figure 1.3b.
˜ conv = ˜
Y ˜ conv × X
˜ conv (1.23)
Suda et al. [29] and more recently, Zhang et al. [50] and Guan et al. [51] leverage on
im2col to derive OpenCL-based FPGA accelerators for CNN. However, this method
introduces redundant data in the input FM matrix, which can lead to either ineff-
ciency in storage or complex memory access patterns. As a result, and as pointed out
in [22], other strategies to map convolutions have to be considered.

1.4.2 WINOGRAD TRANSFORM

Winograd minimal fltering algorithm, introduced in [52], is a computational trans-
form that can be applied to process convolutions with a stride of 1, which is very
common in CNN topologies.
This algorithm is particularly effcient when processing small convolutions
(where K ≤ 3), as advocated in [53]. In this work, authors outperformed the through-
put of the conventional im2col method by a factor of ×7.2 when executing VGG16
on a TitanX GPU.
In Winograd fltering (Figure 1.4), data is processed by blocks, referred to as tiles,
as follows:

1. An input FM tile x of size (u × u) is pre-processed: x̃ = AT xA

2. In a similar way, θ, the flter tile of size (k × k), is transformed into
q˜ : q˜ = BT xB

Input FM Xconv conv kernel ˜conv Output FM Yconv

W J U
x w ˜ k y u
k K u V
w H

Winograd
~ transform ~
x ˜
~
䎭 = y
EWMM

w+k-1 w+k-1

FIGURE 1.4 Winograd fltering F(u × u, k × k).

* That’s what the im2col name refers to: fattening an image to a column.
Accelerating the CNN Inference on FPGAs 15

3. Winograd fltering algorithm, denoted F(u × u, k × k), outputs a tile y of size

(u × u) that is computed according to Equation 1.24

y = C T °˜q̃ ˜ x̃ ˛˝ C (1.24)

where A, B, C are transformation matrices defned in the Winograd algorithm [52]

and ˜ denotes the Hadamard product also known as EWMM.
While a standard fltering requires u2 × k2 multiplications, Winograd algorithm,
denoted F(u × u, k × k), requires (u + k − 1)2 multiplications [52]. In the case of tiles
of a size u = 2 and kernels of size k = 3, this corresponds to an arithmetic complex-
ity reduction of ×2.25 [53], and in this case, transform matrices can be written as
follows:

°1 0 −1 0˙
˝0 0 ˇˇ
°1 1 1 0˙ ˝ 1 1
AT = ˝ ; BT
=
˛0 1 −1 −1ˇˆ ˝0 −1 1 0ˇ
˝ ˇ
˛0 1 0 −1ˆ

° 1 0 0 ˙
˝1 / 2 1/ 2 1 / 2 ˇˇ
C=˝ (1.25)
˝1 / 2 −1 / 2 1 / 2ˇ
˝ ˇ
˛ 0 0 1 ˆ
Beside this complexity reduction, implementing Winograd fltering in FPGA-based
CNN accelerators has two advantages. First, transformation matrices A, B, C can be
evaluated offine once u and k are determined. As a result, these transforms become
multiplications with the constants that can be implemented by means of lut and shift
registers, as proposed in [54].
Second, Winograd fltering can employ the loop optimization techniques dis-
cussed in Section 1.5.2 to vectorize the implementation. On one hand, the computa-
tional throughput is increased when unrolling the computation of the ewmm parts
over multiple DSP blocks. On the other hand, memory bandwidth is optimized using
loop tiling to determine the size of the FM tiles and flter buffers.
First, utilization of Winograd fltering in FPGA-based CNN accelerators is inves-
tigated in [32] and delivers a computational throughput of 46 GOPs when executing
AlexNet convolutional layers. This performance is signifcantly improved by a factor
of ×42 in [31] when optimizing the data path to support Winograd convolutions (by
employing loop unrolling and tiling strategies), and storing the intermediate FM in
on-chip buffers (cf Section 1.4).
The same method is employed in [54] to derive a CNN accelerator on a Xilinx
ZCU102 device that delivers a throughput of 2.94 TOPs on VGG convolutional lay-
ers. The reported throughput corresponds to half of the performance of a TitanX
device, with 5.7× less power consumption [23]*.

* Implementation in the TitanX GPU employs Winograd algorithm and 32-bit foating point arithmetic.
16 Deep Learning in Computer Vision

1.4.3 FAST FOURIER TRANSFORM

Fast Fourier Transform (FFT) is a well known algorithm to transform the 2D convo-
lutions into ewmm in the frequency domain, as shown in Equation 1.26:

conv2D( X[c],˜[ n, c]) = IFFT(FFT( X[c]) ˜ FFT(˜[n, c])) (1.26)

Using FFT to process 2D convolutions reduces the complexity from O(W2 × K2) to
O(W2log2(W)), which is exploited to derive FPGA-based accelerators and to infer
CNN [34]. When compared to standard fltering and Winograd algorithm, FFT fnds
its interest in convolutions with large kernel size (K > 5), as demonstrated in [53,
55]. The computational complexity of FFT convolutions can be further reduced to
O(Wlog2(K)) using the overlap-and-add method [56], which can be applied when
the signal size is much larger than the flter size, which is typically the case in conv
layers (W >> K ). Works in [33, 57] leverage on the overlap-and-add to implement
frequency domain acceleration for conv layers on FPGA, which results in a compu-
tational throughput of 83 GOPs for AlexNet (Table 1.3).

1.5 DATA-PATH OPTIMIZATIONS

As highlighted in Section 2.4.2, the execution of CNN exhibits numerous sources
of parallelism. However, due to the resource limitations of FPGA devices, it might
be impossible to fully exploit all the concurrency patterns, especially with the
sheer volume of operations involved in deep topologies. In other words, the execu-
tion of recent CNN models cannot fully be unrolled sometimes, not even for a
single conv layer.
To address this problem, the general approach, advocated in state-of-the-art
implementations, is to map a limited number of processing elements (PEs) on the
FPGA. These PEs are then reused by temporally iterating data through them.

1.5.1 SYSTOLIC ARRAYS

Early FPGA-based accelerators for CNN implemented systolic arrays to accelerate
the 2D fltering in convolutions layers [58—61]. As illustrated in Figure 1.5a, systolic
arrays employ a static collection of PE, typically arranged in a 2-dimensional grid.
These PE operate as a co-processor under the control of a central processing unit.
The confguration of systolic arrays is agnostic to the CNN model, making them
ineffcient to process large-scale networks for the following three reasons:
First, the static collection of PE can support convolutions only up to a given flter
size Km, where typical values of Km range from 7 in [59] to 10 in [61]. Therefore, in
convolutional layer (˜) , K ˜ > K m is not supported by the accelerator. Second, systolic
arrays suffer from under-utilization when processing layers in which the kernel size
K ˜ is much smaller then Km. This is for instance the case in [61], where the process-
ing of 3 × 3 convolutions uses only 9% of DSP blocks, while the processing of these
layers can be further parallelized and thus accelerated. Third and fnally, PE in sys-
tolic arrays do not usually include memory caches and have to fetch their inputs from
TABLE 1.3
Accelerators Employing Computational Transforms
Comp Params Freq Through Power LUT Memory
Method Entry Network (GOP) (M) Bit-width Desc. Device (MHz) (GOPs) (W) (K) DSP (MB)

Winograd [33] AlexNet-C 1.3 2.3 Float 32 OpenCL Virtex7 VX690T 200 46 – 505 3683 56.3
[32] AlexNet-C 1.3 2.3 Float16 OpenCL Arria10 GX1150 303 1382 44.3 246 1576 49.7
[55] VGG16-C 30.7 14.7 Fixed 16 HLS Zynq ZU9EG 200 3045 23.6 600 2520 32.8
[55] AlexNet-C 1.3 2.3 Fixed 16 HLS Zynq ZU9EG 200 855 23.6 600 2520 32.8
Accelerating the CNN Inference on FPGAs

FFT [34] AlexNet-C 1.3 2.3 Float 32 – Stratix5 QPI 200 83 13.2 201 224 4.0
[34] VGG19-C 30.6 14.7 Float 32 – Stratix5 QPI 200 123 13.2 201 224 4.0
GEMM [30] AlexNet-C 1.3 2.3 Fixed 16 OpenCL Stratix5 GXA7 194 66 33.9 228 256 37.9
[50] VGG16-F 31.1 138.0 Fixed 16 HLS Kintex KU060 200 365 25.0 150 1058 14.1
[50] VGG16-F 31.1 138.0 Fixed 16 HLS Virtex7 VX960T 150 354 26.0 351 2833 22.5
[51] VGG16-F 31.1 138.0 Fixed 16 OpenCL Arria10 GX1150 370 866 41.7 437 1320 25.0
[51] VGG16-F 31.1 138.0 Float 32 OpenCL Arria10 GX1150 385 1790 37.5 – 2756 29.0
17
18 Deep Learning in Computer Vision

Off-chip
memory Previous partial sum
x1conv
... conv
PE PE PE PE
˜1conv +
...
conv
PE PE PE ... PE
yconv
PC

...

...
...
...
DMA + act
PE PE PE ... PE
xpc
conv +
... Bottleneck

conv
...

...

...

...
...

˜pc
conv
PE PE PE ... PE
...

(a) (b) (c)

FIGURE 1.5 Generic data paths of FPGA-based CNN accelerators: (a) Static systolic array.
(b) Dedicated SIMD accelerator. (c) Dedicated processing element.

an off-chip memory. As a result, the performance of systolic arrays can rapidly be

bounded by memory bandwidth of the device.

1.5.2 LOOP OPTIMIZATION IN SPATIAL ARCHITECTURES

Due to the ineffciency of systolic arrays, fexible and dedicated spatial architectures
for CNN were mapped on FPGA. The general computation fow in these accelerators
is illustrated in Figure 1.5b.
First, FMs and weights are fetched from DRAM to on-chip buffers, and are then
streamed into the PE. At the end of the PE computation, results are transferred back
to on-chip buffers and, if necessary, to the external memory in order to be fetched in
their turn to process the next layers. Each PE – as depicted in Figure 1.5c – is con-
fgurable and has its own computational capabilities by means of DSP blocks, and
its own data caching capabilities by means of on-chip registers. With this paradigm,
the problem of CNN mapping consists of fnding the optimal architectural and tem-
poral confguration of PE: in other words, the best number of DSP blocks per PE, the
optimal temporal scheduling of data that maximizes the computational throughput.
For convolutional layers, in which the processing is described in Listing 1.1, fnd-
ing the optimal PE confguration comes down to a loop optimization problem [28,
29, 39, 40, 62–64].
Listing 1.1: Nested Loops

// Lb : Batch
for (int b =0;b<B,l++) {
// Ll: Layer
for (int l =0;l<L,l++) {
// Ln: Y Depth
for (int n =0;n<N;n++) {
// Lv: Y Columns
for (int v =0;v<V,v++) {
// Lu: Y Raws
Accelerating the CNN Inference on FPGAs 19

for (int u =0;u<U,u++) {

// Lc: X Depth
for (int c =0;n<C;c++) {
// Lj: Theta Columns
for (int j =0;j<J,j++) {
// Lk: Theta Raws
for (int k =0;k<K,k++) {
Y[b,l,n,v,u] +=
X[b,l,c,v+j,u+k] *
Theta [l,n,c,j,k]
}}}}}}}

Listing 1.2: Loop Tiling in conv layers

for (int b =0;b<B,l++){

for (int n =0;n<N;n+= Tn){
for (int v =0;v<V,v+= Tv){
for (int u =0;u<U,u+= Tu){
for (int c =0;n<C;c+= Tc){
// DRAM : Load in on - chip
buffers the tiles :
// X[l,c:c+Tc ,v:v+Tv ,u:u+Tu]
// Theta [l,n:n+Tn ,c:c+Tc ,j,k]
for (int tn =0; tn <Tn;tn ++){
for (int tv =0; tv <Tv ,tv ++){
for (int tu =0; tu <Tu ,tu ++){
for (int tc =0; tn <Tc;tc ++){
for (int j =0;j<J,j++){
for (int k =0;k<K,k++){
Y[l,tn ,tv ,tu] +=
X[l,tc ,tv+j,tu+k] *
Theta [l,tn ,tc ,j,k];
}}}}}} // DRAM : Store output
}}}}

This problem is addressed by applying loop optimization techniques such loop

unrolling, loop tiling, or loop interchange to the 7 nested loops of Listing 1.1. In
this case, the unroll and tiling factors (respectively Pi and Ti) determine the number
of PEs, the computational resources, and the on-chip memory allocated to each PE.

Loop Unrolling
Unrolling a loop Li with an unrolling factor Pi (Pi ˜ i, i °{L, V , U , N , C , J , K}) acceler-
ates its execution by allocating multiple computational resources. Each of the par-
allelism patterns listed in Section 1.2.4.2 can be implemented by unrolling one of
the loops of Listing 1.1, as summarized in Table 1.4. For the confguration given in
Figure 1.5c, the unrolling factor PN sets the number of PEs. The remaining factors
– PC, PK, PJ – determine the number of multipliers, as well as the size of buffer con-
tained in each PE (Figure 1.6).
20 Deep Learning in Computer Vision

TABLE 1.4
Loop Optimization Parameters Pi and Ti
Parallelism Intra layer Inter FM Intra FM Inter conv. Intra conv.

Loop LL LN LV LU Lc LJ LK
Unroll Factor PL PN PV PU Pc PJ PK
Tiling Factor TL TN TU TU TC TJ TK

Input FMs Xconv Ms Yconv

Output FMs conv
W U xi TW

TC TN
V
H
TH
TV +
TW × × × +
TU PK + +
C N conv × × ×
θi

...
× × × +
TC J PJ × PK Adder
conv Weights PJ
conv T ... mult. tree
ϴ J
C
TK K

FIGURE 1.6 Loop tiling and unrolling in convolutional layers.

Loop Tiling
In general, the capacity of on-chip memory in current FPGA is not large enough to
store the weights and intermediate FM of all CNN layers*. For example, AlexNet’s
convolutional layers resort to 18.6 Mbits of weights, and generate a total 70.7 Mbits
of intermediate feature maps†. In contrast, the highest-end Stratix V FPGA provides
a maximum of 52 Mbits of on-chip ram.
As a consequence, FPGA-based accelerators resort to external DRAM to store
these data. As mentioned in Section 1.2.4.3, DRAM accesses are costly in terms of
energy and latency, and data caches must be implemented by means of on-chip buf-
fers and local registers. The challenge is thus to build a data path in a way that every
data transferred from DRAM is reused as much as possible.
For conv layers, this challenge can be addressed by tiling the nested loops of
Listing 1.1. Loop tiling [66] divides the FM and weights of each layer into multiple
groups that can fit into the on-chip buffers. For the configuration given in Figure 1.5c,
the size of the buffers containing input FM, weights, and output FM is set according
to the tiling factors listed in Table 1.4.

* Exception can be made for [6666], where a large cluster of FPGAs is interconnected and resorts only
to on-chip memory to store CNN weights and intermediate data.
† Estimated by summing the number of outputs for each convolution layer.
Accelerating the CNN Inference on FPGAs 21

B˜conv = TN × TC × TJ × TK (1.28)

BYconv = TN × TV × TU (1.29)

With these buffers, the memory accesses occurring in the conv layer (cf Equation
1.18) are respectively divided by BXconv , B˜conv and BYconv , as expressed in Equation 1.30.

C˜ H ˜W˜ N ˜C˜ J ˜ K ˜ N ˜U ˜V˜

M˜conv = + + (1.30)
TCTH TW TN TCTJ TK TN TV TU

Since the same hardware is reused to accelerate the execution of multiple conv layers
with different workloads, the tiling factors are agnostic to the workload of a specifc
layer, as can be noticed in the denominator of Equation 1.30. As a result, the value
of the tiling factors is generally set to optimize the overall performance of a CNN
execution.

1.5.3 DESIGN SPACE EXPLORATION

Finding the optimal unrolling and tiling factors for a specifc device is a complex
problem that is generally solved using brute-force design space exploration [29,
39, 40, 48, 67, 68]. This exploration is driven by an analytical model, in which the
inputs are loop factors Pi, Ti and outputs are theoretical predictions of the computa-
tional throughput (T ) , the size of buffers (B ), and the number of external memory
accesses (M) . This model is parametrized by the available resources of a given
FPGA platform and the workload of the considered CNN. To select feasible solu-
tions for this optimization problem, most literature approaches rely on the Roofine
method [69] to accept or reject design solutions that do not match with the maxi-
mum computational throughput or the maximum memory bandwidth of a given
device (Figures 1.7).
A typical design space exploration driven by the roofine model is illustrated in
Figure 1.8. In this graph, each point represents the performance of an explored solu-
tion (Pi, Ti). For a given FPGA platform, the attainable bandwidth and computational
throughput are respectively reported by the diagonal and horizontal lines. Point A is
an invalid solution, as it is above the bandwidth roof, while point A′ is feasible but
delivers mediocre computational throughput. Acceptable solutions are represented
by points C and D, the latter being better than the former since it has lower band-
width requirements.

FIGURE 1.7 Design space exploration methodology.

22 Deep Learning in Computer Vision

FIGURE 1.8 Example of a design selection driven by the roofine model.

1.5.4 FPGA IMPLEMENTATIONS

Employing loop optimizations to derive FPGA-based CNN accelerator was frst
investigated in [39]. In this work, Zhang et al. report a computational throughput of
61.62 GOPs in the execution of AlexNet convolutional layers by unrolling loops LC
and L N. This accelerator, described with Vivado HLS tools, relies on 32-bit foating-
point arithmetic. Works in [68] follow the same unrolling scheme and feature a 16-bit
fxed-point arithmetic, resulting in a ×2.2 improvement in terms of computational
throughput. Finally, the same unrolling and tiling scheme is employed in recent work
[48], where authors report a ×13.4 improvement, thanks to a deeply pipelined FPGA
cluster of four Virtex7-XV960t devices.
In all these implementations, loops L J and LK are not unrolled because J and K are
usually small, especially in recent topologies. Works of Motamedi et al. [40] study
the impact of unrolling these loops in AlexNet, where the frst convolutional layers
use large 11 × 11 and 5 × 5 flters. Expanding loop unrolling and tiling to loops L J
and LK results in a 1.36× improvement in computational throughput vs. [39] on the
same VX485T device when using 32-bit foating-point arithmetic. Nevertheless, and
as pointed out in [63], unrolling these loops is ineffective for recent CNN models that
employ small convolution kernels.
The values of U, V, N can be very large in CNN models. Consequently, unroll-
ing and tiling loops LU, LV, L N can be effcient only for devices with high computa-
tional capabilities (i.e., DSP blocks). This is demonstrated in works of Rahman et al.
[67] that report an improvement of ×1.22 over [39] when enlarging the design space
exploration to loops LU, LV, L N, which comes at the price of very long exploration tim-
ing. In order to keep data in on-chip buffer after the execution of a given layer, works
of Alwani et al. [62] advocate the use of fused-layer accelerators by tiling across
layer L L. As a result, authors are able to remove 95% of DRAM accesses at the cost
of 362 KB of extra on-chip memory.
In all these approaches, loops L N, LC, L J, LK are unrolled in a similar way they
are tiled (i.e., Ti = Pi). By contrast, the works of Ma et al. [63, 70] fully explore all
Accelerating the CNN Inference on FPGAs 23

the design variables searching for optimal loop unroll and tiling factors. More
particularly, the authors demonstrate that the input FM and weights are optimally
reused when unrolling only computations within a single input FM (i.e., when
PC = PJ = PK = 1). Tiling factors are set in such a way that all the data required to
compute an element of Y are fully buffered (i.e., TC = C, TK = K, TJ = J). The remain-
ing design parameters are derived after a brute-force design exploration. The same
authors leverage on these loop optimizations to build an RTL compiler for CNNs
in [71]. To the best of our knowledge, this accelerator outperforms all the previ-
ous implementations that are based on loop optimization in terms of computational
throughput (Tables 1.5 through 1.7).

1.6 APPROXIMATE COMPUTING OF CNN MODELS

Besides the computational transforms and data-path optimization, the CNN execu-
tion can be accelerated when employing approximate computing, which is known to
perform effciently on FPGAs [73].
In the methods detailed in this section, a minimal amount of the CNN accu-
racy is traded to improve the computational throughput or energy effciency of the
accelerator. Two main strategies are employed. The frst implements approximate
arithmetic to process the CNN layers with a reduced precision. The second aims
at reducing the number of operations occurring in CNN models without criti-
cally affecting the modeling performance. Note that both approaches can resort
to fne-tuning in order to compensate the accuracy loss introduced by approximate
computing.

1.6.1 APPROXIMATE ARITHMETIC FOR CNNS

Several studies have demonstrated that the precision of both operations and operands
in CNN, and more generally in neural networks, can be reduced without critically
affecting their predictive performance. This reduction can be achieved by quantizing
either or both of the CNN inputs, weights, and/or FM using a fxed-point numerical
representation.

1.6.1.1 Fixed-Point Arithmetic

In a general way, CNN models are deployed in CPU and GPU using the same numer-
ical precision they were trained with, relying on the single-precision foating-point
representation. This format employs 32 bits, arranged according to the IEEE754
standard. As current FPGAs support foating operations, various implementations
[39, 62, 67] employ such data representation.
Nonetheless, numerous studies such [74–76] demonstrate that the inference of
CNNs can be achieved with a reduced precision of operands. More particularly,
works in [77, 78] demonstrate the applicability of fxed-point (F × P) arithmetic to
train and infer CNNs. The F × P representation encodes numbers with a given bit-
width b, using i bits for the integer part, and f bits for the fractional part (b = i + f).
Note that the value of i is selected according the desired numerical range, and the
value of f is selected according to the desired numerical precision.
TABLE 1.5
24

Accelerators Employing Loop Optimization

Comp Params Freq Through Power LUT Memory
Entry Network (GOP) (M) Bit-width Desc. Device (MHz) (GOPs) (W) (K) DSP (MB)

[40] AlexNet-C 1.3 2.3 Float 32 HLS Virtex7 VX485T 100 61.62 18.61 186 2240 18.4
[29] VGG16SVD-F 30.8 50.2 Fixed 16 RTL Zynq Z7045 150 136.97 9.63 183 780 17.5
[30] AlexNet-C 1.3 2.3 Fixed 16 OpenCL Stratix5 GSD8 120 187.24 33.93 138 635 18.2
[30] AlexNet-F 1.4 61.0 Fixed 16 OpenCL Stratix5 GSD8 120 71.64 33.93 272 752 30.1
[30] VGG16-F 31.1 138.0 Fixed 16 OpenCL Stratix5 GSD8 120 117.9 33.93 524 1963 51.4
[68] AlexNet-C 1.3 2.3 Float 32 HLS Virtex7 VX485T 100 75.16 33.93 28 2695 19.5
[49] AlexNet-F 1.4 61.0 Fixed 16 HLS Virtex7 VX690T 150 825.6 126.00 N.R 14400 N.R
[49] VGG16-F 31.1 138.0 Fixed 16 HLS Virtex7 VX690T 150 1280.3 160.00 N.R 21600 N.R
[69] NIN-F 2.2 61.0 Fixed 16 RTL Stratix5 GXA7 100 114.5 19.50 224 256 46.6
[69] AlexNet-F 1.5 7.6 Fixed 16 RTL Stratix5 GXA7 100 134.1 19.10 242 256 31.0
[38] AlexNet-F 1.4 61.0 Fixed 16 RTL Virtex7 VX690T 156 565.94 30.20 274 2144 34.8
[63] AlexNet-C 1.3 2.3 Float 32 HLS Virtex7 VX690T 100 61.62 30.20 273 2401 20.2
[64] VGG16-F 31.1 138.0 Fixed 16 RTL Arria10 GX1150 150 645.25 50.00 322 1518 38.0
[42] AlexNet-C 1.3 2.3 Fixed 16 RTL Cyclone5 SEM 100 12.11 N.R 22 28 0.2
[42] AlexNet-C 1.3 2.3 Fixed 16 RTL Virtex7 VX485T 100 445 N.R 22 2800 N.R
[72] NiN 20.2 7.6 Fixed 16 RTL Stratix5 GXA7 150 282.67 N.R 453 256 30.2
[72] VGG16-F 31.1 138.0 Fixed 16 RTL Stratix5 GXA7 150 352.24 N.R 424 256 44.0
[72] ResNet-50 7.8 25.5 Fixed 16 RTL Stratix5 GXA7 150 250.75 N.R 347 256 39.3
[72] NiN 20.2 7.6 Fixed 16 RTL Arria10 GX1150 200 587.63 N.R 320 1518 30.4
[72] VGG16-F 31.1 138.0 Fixed 16 RTL Arria10 GX1150 200 720.15 N.R 263 1518 44.5
[72] ResNet-50 7.8 25.5 Fixed 16 RTL Arria10 GX1150 200 619.13 N.R 437 1518 38.5
[73] AlexNet-F 1.5 7.6 Float 32 N.R Virtex7 VX690T 100 445.6 24.80 207 2872 37
[73] VGG16SVD-F 30.8 50.2 Float 32 N.R Virtex7 VX690T 100 473.4 25.60 224 2950 47
Deep Learning in Computer Vision
TABLE 1.6
Accelerators Employing Approximate Arithmetic
Bit-width
Comp Params Acc Freq Through. Power LUT Memory
A×C Entry Dataset (GOP) (M) In/Out FMs θconv θFC (%) Device (MHz) (GOPs) (W) (K) DSP (MB)

FP32 [51] ImageNet 30.8 138.0 32 32 32 32 90.1 Arria10 370 866 41.7 437 1320 25.0
GX1150
FP16 [32] ImageNet 30.8 61.0 16 16 16 16 79.2 Arria10 303 1382 44.3 246 1576 49.7
GX1150
[64] ImageNet 30.8 138.0 16 16 8 8 88.1 Arria10 150 645 N.R 322 1518 38.0
GX1150
DFP [72] ImageNet 30.8 138.0 16 16 16 16 N.R Arria10 200 720 N.R 132 1518 44.5
GX1150
[51] ImageNet 30.8 138.0 16 16 16 16 N.R Arria10 370 1790 N.R 437 2756 29.0
Accelerating the CNN Inference on FPGAs

GX1150
[91] Cifar10 1.2 13.4 20 2 1 1 87.7 Zynq Z7020 143 208 4.7 47 3 N.R
[46] Cifar10 0.3 5.6 20/16 2 1 1 80.1 Zynq Z7045 200 2465 11.7 83 N.R 7.1
BNN [93] MNIST 0.0 9.6 8 2 1 1 98.2 Stratix5 GSD8 150 5905 26.2 364 20 44.2
[93] Cifar10 1.2 13.4 8 8 1 1 86.3 Stratix5 GSD8 150 9396 26.2 438 20 44.2
[93] ImageNet 2.3 87.1 8 32 1a 1 66.8 Stratix5 GSD8 150 1964 26.2 462 384 44.2
[94] Cifar10 1.2 13.4 8 2 2 2 89.4 Xilinx7 250 10962 13.6 275 N.R 39.4
VX690T
TNN [94] SVHN 0.3 5.6 8 2 2 2 97.6 Xilinx7 250 86124 7.1 155 N.R 12.2
VX690T
[94] GTSRB 0.3 5.6 8 2 2 2 99.0 Xilinx7 250 86124 6.6 155 N.R 12.2
VX690T
25
 26

TABLE 1.7
Accelerators Employing Pruning and Low Rank Approximation
Comp Params Removed Acc Freq Through. Power LUT Memory
Reduc. Entry Dataset (GOP) (M) Param. (%) Bit-width (%) Device (MHz) (GOPs) (W) (K) DSP (MB)

SVD [29] ImageNet 30.8 50.2 63.6 16 Fixed 88.0 Zynq 7Z045 150 137.0 9.6 183 780 17.50
Pruning [44] Cifar10 0.3 13.9 89.3 8 Fixed 91.5 Kintex 7K325T 100 8620.7 7.0 17 145 15.12
[7] ImageNet 1.5 9.2 85.0 32 Float 79.7 Stratix 10 500 12000.0 141.2 N.R N.R N.R
Deep Learning in Computer Vision
Accelerating the CNN Inference on FPGAs 27

In the simplest version of fxed-point arithmetic, all the numbers are encoded
with the same fractional and integer bit-widths. This means that the position of the
radix point is similar for all the represented numbers. In this chapter, we refer to this
representation as static F × P.
When compared to foating point, F × P is known to be more effcient in terms of
hardware utilization and power consumption. This is especially true in FPGAs [79],
where – for instance – a single DSP block in Intel devices can either implement one
32-bit foating-point multiplication or three concurrent F × P multiplications of 9 bits
[8]. This motivated early FPGA implementations such as [61, 80] to employ fxed-
point arithmetic in deriving CNN accelerators. These implementations mainly use
a 16-bit Q8.8 format, where 8 bits are allocated to the integer parts, and 8 bits to the
fractional part. Note that the same Q8.8 format is used for representing the features
and the weights of all the layers.
In order to prevent overfow, the former implementations also expand the bit-
width when computing weighted sums of convolutions. Equation 1.31 explains how
the bit-width is expanded; if bX bits are used to quantize the input FM and bΘ bits are
used to quantize the weights, an accumulator of bacc bits is required to represent a
weighted sum of C˜ K ˜2 elements, where:

bacc = bx + b˘ + max ˝ log2 C˜ K ˜2 ˆ

˜ ˙ ˇ ( ) (1.31)

In practice, most FPGA accelerators use 48-bit accumulators, such as in [59, 60]
(Figure 1.9).

bx bits bq + bx bits
conv
X

X + Accumulate

conv bq bits bq + bx +
Q log2(CK²) bits

(a) bit-width of accumulators

Integer part Fractional part

sign mantissa
{
{

0 0 1 1 0 0 1 0

0 0 1 1 0 0 1 0

(b) Dynamic FxP

FIGURE 1.9 Fixed-point arithmetic for CNN accelerators.

28 Deep Learning in Computer Vision

1.6.1.2 Dynamic Fixed Point for CNNs

In deep topologies, it can be observed that distinct parts of a network can have sig-
nifcantly different ranges of data. In particular, the features of the deep layers tend
to have a much larger numerical range when compared to the features of the frst
CNN layers.
The histograms of Figure 1.10a depict this phenomenon for AlexNet convolu-
tional layers*. While the CNN inputs (data column) are normalized take their values
between 0 and 1, the outputs of the frst convolutional layer (conv1 column) have a
wider numerical range, between 2−7 and 22. This is even more salient for the ffth
convolutional layer, where most of the outputs take their values between 2−1 and 26.
The same problem appears when comparing the numericals of the CNN weights,
and CNN activations. In this case, the weights are numerically much smaller when
compared to the activations, as illustrated in Figure 1.10b†.
As a consequence, large bit-widths have to be allocated to the integer fractional
parts in order to keep a uniform precision across the network while preventing over-
fow. This expansion badly increases the resource requirements of a given FPGA
mapping. As a result, static F × P, with its unique shared fxed exponent, is ill-suited
to deep learning, as pointed out in. [81]
To address this problem, works in [77, 81, 82] advocate the use of dynamic F × P
[83]‡. In dynamic F × P, different scaling factors are used to process different parts
of the network. In other words, the position of the radix point varies from one layer

(a) Histogram of the layer outputs

(b) Histogram of weights and activations. Inputs and weights encoded in 8 bits

FIGURE 1.10 Distribution of AlexNet activations and weights.

* Code made available at github.com/KamelAbdelouahab/CNN-Data-Distribution.

† This fgure deliberately multiplies the weights and activations by a scale factor of 27 − 1 to emulate an

8-bit quantization.
‡ Another approach to address this problem is to use custom foating point representations, as detailed

in [31].
Accelerating the CNN Inference on FPGAs 29

to another. More particularly, weights, weighted sums, and outputs of each layer are
assigned distinct integer and fractional bit-widths.
The optimal values of these bit-widths (i.e., the ones that deliver the best trade-off
between accuracy loss and computational load) for each layer can be derived after
a profling process performed by dedicated frameworks that support F × P. Among
these frameworks, Ristretto [81] and FixCaffe [84] are compatible with Caffe, while
TensorFlow natively supports 8-bit computations. Most of these tools can fne-tune a
given CNN model to improve the accuracy of the quantized network.
In particular, the works in [85] demonstrate the effciency dynamic of F × P, point-
ing out how the inference of AlexNet is possible using 6 bits in dynamic F × P instead
of 16 bits with a conventional fxed-point format.

1.6.1.3 FPGA Implementations

The FPGA-based CNN accelerator proposed in [29] is built upon this quantization
scheme and employs different precisions to represent the FM, convolution kernels,
and FC weights with 16, 8, and 10 bits, respectively. Without fne-tuning, the authors
report a drop of 1% in the classifcation accuracy of AlexNet. In a similar way, Qiu
et al. employ F × P to quantize the VGG network with respectively 8 bits for the
weights, 8 bits for activations, and 4 bits for FC layers, resulting in an accuracy drop
of 2%. In all these accelerators, dynamic quantization is supported by means of data
shift modules [28, 82]. Finally, the accelerator in [41] relies on the Ristretto frame-
work [81] to derive an AlexNet model wherein the data is quantized in 16 bits with
distinct integer bit-widths per layer*.

1.6.1.4 Extreme Quantization and Binary Networks

Training and inferring CNNs with extremely compact data representations is an
area that has recently gained a lot of research interest. Early works of Courbariaux
et al. in BinaryConnect [86] demonstrate the feasibility of training neural networks
using binary weights, i.e., weights with either a value of −θ or θ encoded in 1 bit.
BinaryConnect lowers the bandwidth requirements of a network by a factor of ×32 at
the price of an accuracy loss, evaluated at 19.2% on ImageNet†. The same authors go
further in their investigations in [15] and propose BNNs that represent both feature
maps and weights with only 1 bit. In these networks, negative values are represented
as 0, while positive values are represented as 1. BNNs greatly simplify the process-
ing of convolutions, boiling down the computations of MACs into bitwise XNOR
operations followed by a pop-count (see Figure 1.11b). Moreover, the authors use the
sign function as activation and apply batch normalization before applying the activa-
tion, which reduces the information lost during binarization (see Figure 1.11a). In
turn, a higher drop in classifcation accuracy occurs when using BNNs, evaluated at
29.8% for ImageNet. This accuracy drop is then lowered to 11% by Rastegari et al.,
using different scale factors for binary weights (i.e., −θ1 or +θ2).

* Since the same PEs are reused to process different layers, the same bit width is used with a variable
radix point for each layer.
† When compared to an exact 32-bit implementation of AlexNet.
30 Deep Learning in Computer Vision

1 0 0 XNOR 0 0 1 Pop count

1 0 0 0 0 1
conv act 0 1 0 0 1 0
5
Xconv
pool BN 0 1 0 = 0 1 0
Z (bin) (sign)
Xconv
Count number
1 1 0 1 1 0
1 1 0 1 1 0 of ones over C,J,K

˜conv ˜ conv
0 1 0
0 1 0
conv
Xconv X
conv 1 1 0
Xconv X xconv 1 1 0
Z Z
Xconv Yconv
1 0 0
1 0 0

(a) Processing Pipeline (b) Binary Convolutions

FIGURE 1.11 Binary neural networks.

Beside BNNs, pseudo-binary networks, such as DoReFa-Net [87] and quantized

neural networks (QNNs) [88], reduce the accuracy drop to 6.5% when employing a
slightly expanded bit-width (2 bits) to represent the intermediate FM. Similarly, in
TTQ [89], weights are constrained to three values (2 bits) − θ1, 0, −θ2, but FMs are
represented in a 32-bit foat scheme. As a consequence, the effciency gain of TTQ is
not as high as in BNNs. In turn, TTQ achieves comparable accuracy on ImageNet,
within 0.7% of full-precision.
In FPGAs, BNNs beneft from a signifcant acceleration, as the processing of
“binary” convolutions can be mapped on XNOR gates followed by a pop-count oper-
ation, as depicted in Figure 1.11b. Furthermore, and as suggested in [7], a pop-count
operation can be implemented using lookup tables in a way that convolutions are
processed only with logical elements. Thus, the DSP blocks can be used to process the
batch norm calculation (Equation 1.7, which can be formulated as a linear transform
in order to reduce the number of operations). This approach is followed in the imple-
mentation of [90] to derive an FPGA-based accelerator for BNNs that achieves 207.8
GOPs while only consuming 4.7 W and 3 DSP blocks to classify the Cifar10 dataset.
For the same task, works in [45, 91] use a smaller network confguration* and
reach a throughput of 2.4 TOPs when using a larger Zynq 7Z045 device with 11W
power consumption.
For ImageNet classifcation, binary net implementation of [92] delivers an overall
throughput of 1.9 TOPs on a Stratix V GSD device. In all these works, the frst layer
is not binarized to achieve better classifcation accuracy. As pointed out in [92], the
performance in this layer can be improved when using a higher number of DSP
blocks. Finally, an accelerator for TTQ proposed in [93] achieves a peak perfor-
mance of 8.36 TMACs when classifying the Cifar10 dataset with a 2-bit precision.

1.6.2 REDUCED COMPUTATIONS

In addition to approximate arithmetic, several studies attempt to reduce the number
of operations involved in CNNs. For FPGA-based implementations, two main strate-
gies are investigated: weight pruning, which increases the sparsity of the model,
and low-rank approximation of flters, which reduces the number of multiplications
occurring in the inference.

* The network topology used in this work involves 90% fewer computations and achieves 7% less clas-
sifcation accuracy on Cifar10.
Accelerating the CNN Inference on FPGAs 31

1.6.2.1 Weight Pruning

As highlighted in [94], CNNs as overparametrized networks and a large amount of
the weights can be removed – or pruned – without critically affecting the classifca-
tion accuracy. In its simplest form, pruning is performed according to the magnitude,
such that the lowest values of the weights are truncated to zero [95]. In a more recent
approach, weight removal is driven by energy consumption of a given node of the
graph, which is 1.74× more effcient than magnitude-based approaches [96]. In both
cases, pruning is followed by a fne-tuning of the remaining weights in order to
improve the classifcation accuracy. This is for instance the case in [97], where prun-
ing removes respectively 53% and 85% of the weights in AlexNet conv and FC layers
for less then 0.5% accuracy loss (Figure 1.12).

1.6.2.2 Low Rank Approximation

Another way to reduce the computations occurring in CNNs is to maximize the
number of separable flters. A 2D-separable flter, denoted θsep, has a unitary rank*†,
and can be expressed as two successive 1D flters (q J´1 then q1´K ). Filter decomposi-
sep sep

tion reduces the number of multiplications from J × K to J + K. This is illustrated in

Figure 1.13, where the 3 × 3 averaging flter is separable, and can thus be decom-
posed into two successive one-dimensional convolutions.

FIGURE 1.12 Histogram of conv weights in a compressed AlexNet model†.

FIGURE 1.13 Example of a separable flter.

* Meaning that rank(θsep) = 1.

† Pruned flters treated as zero-valued weights.
32 Deep Learning in Computer Vision

The same concept expands to depth convolutions, where a separable flter requires
C + J + K multiplications instead of C × J × K multiplications.
Nonetheless, only a small proportion of CNN flters are separable. To increase
this proportion, a frst approach is to force the convolution kernels to be separable by
penalizing high-rank flters when training the network [98]. Alternatively, and after
the training, the weights Θ of a given layer can be approximated into a small set of r
low-rank flters. In this case, r × (C + J + K) multiplications are required to process a
single depth convolution.
Finally, CNN computations can be reduced further by decomposing the weight
matrix ˜ ˜ through single-value decomposition (SVD). As demonstrated in the early
works of [99], SVD greatly reduces the resource utilization of a given 2D-flter
implantation. Moreover, SVD also fnds its interest when processing FC layers and
convolutions that employ the im2col method (cf Section 1.4.1). In a similar way to
pruning, low rank approximation or SVD is followed by a fne-tuning in order to
counterbalance the drop in classifcation accuracy.

1.6.2.3 FPGA Implementations

In FPGA implementations, SVD is applied on FC layer to signifcantly reduce the
number of weights, such as in [28], where the authors derive a VGG16-SVD model
that achieves 87.96% accuracy on ImageNet with 63% fewer parameters.
Alternatively, one can take advantage of the numerous research efforts given to
accelerate Sparse GEMM on FPGA [100]. In this case, the challenge is to deter-
mine the optimal format of matrices that maximizes the chance to detect and skip
zero computations, such compressed sparse column (CSC) or compressed sparse row
(CSR) formats*. Based on this, Sze et al. [22] advocate the use of the CRC to pro-
cess CNN. Indeed, this format requires lower memory bandwidths when the output
matrix is smaller then the input, which is typically the case in CNNs where N < CJK,
as in Figure 1.3b.
However, this effciency of CRC format is valid only for extremely sparse matri-
ces (typically with ≤1% of non-zeros), while in practice, pruned CNN matrices are
not that sparse (typically, ≤4% – 80% of non-zeros). Therefore, works in [7] propose
a zero skip scheduler that identifes zero elements and skips them in the scheduling
of the MAC processing. As a consequence, the number of cycles required to compute
the sparse GEMM is reduced. For AlexNet layers, the zero skip scheduler results in
a 4× speedup. The same authors project a throughput of 12 TOPs for pruned CNN in
the next Intel Stratix10 FPGAs, which outperforms the computational throughput of
state-of-the-art GPU implementations by 10%.

1.7 CONCLUSIONS
In this chapter, a number of methods and tools have been compared that aim at
porting convolutional neural networks onto FPGAs. At the network level, approxi-
mate computing and data-path optimization methods have been covered, while at

* This format represents a matrix by three one-dimensional arrays, that respectively contain nonzero
values, row indices, and column indices.
Accelerating the CNN Inference on FPGAs 33

the neuron level, the optimizations of convolutional and fully connected layers have
been detailed and compared. All the different degrees of freedom offered by FPGAs
(custom data types, local data streams, dedicated processors, etc.) are exploited by
the presented methods. Moreover, algorithmic and data-path optimizations can and
should be jointly implemented, resulting in additive hardware performance gains.
CNNs are by nature overparameterized and support particularly well approxi-
mate computing techniques such as weight pruning and fxed-point computation.
Approximate computing already constitutes a key to CNN acceleration over hard-
ware and will certainly continue driving the performance gains in the years to come.

BIBLIOGRAPHY
1. Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp.
436–444, 2015.
2. O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy,
A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei, “ImageNet large scale visual
recognition challenge,” International Journal of Computer Vision, vol. 115, no. 3, pp.
211–252, September 2014.
3. R. Girshick, “Fast R-CNN,” in Proceedings of the IEEE International Conference on
Computer Vision - ICCV ’15, 2015, pp. 1440–1448.
4. J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic seg-
mentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition - CVPR ’15, 2015, pp. 3431–3440.
5. Y. Zhang, M. Pezeshki, P. Brakel, S. Zhang, C. L. Y. Bengio, and A. Courville,
“Towards end-to-end speech recognition with deep convolutional neural networks,”
arXiv preprint, vol. arXiv:1701, 2017.
6. K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale
image recognition,” arXiv preprint, vol. arXiv:1409, pp. 1–14, 2014.
7. E. Nurvitadhi, S. Subhaschandra, G. Boudoukh, G. Venkatesh, J. Sim, D. Marr, R.
Huang, J. OngGeeHock, Y. T. Liew, K. Srivatsan, and D. Moss, “Can FPGAs beat GPUs
in accelerating next-generation deep neural networks?” in Proceedings of the ACM/
SIGDA International Symposium on Field-Programmable Gate Arrays - FPGA ’17,
2017, pp. 5–14.
8. Intel FPGA, “Intel Stratix 10 variable precision DSP blocks user guide,” pp. 4–5, 2017.
9. G. Lacey, G. Taylor, and S. Areibi, “Deep learning on FPGAs: Past, present, and
future,” arXiv e-print, 2016.
10. S. I. Venieris, A. Kouris, and C.-S. Bouganis, “Toolfows for mapping convolutional
neural networks on FPGAs,” ACM Computing Surveys, vol. 51, no. 3, pp. 1–39, June
2018.
11. Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient based learning applied to
document recognition,” in Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
12. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classifcation with deep con-
volutional neural networks,” in Advances in Neural Information Processing Systems -
NIPS’12, 2012, p. 19.
13. D. H. Hubel and T. N. Wiesel, “Receptive felds, binocular interaction and functional
architecture in the cat’s visual cortex,” The Journal of Physiology, vol. 160, no. 1, pp.
106–154, 1962.
14. S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by
reducing internal covariate shift,” in Proceedings of the International Conference on
Machine Learning - ICML ’15, F. Bach and D. Blei, Eds., vol. 37, 2015, pp. 448–456.
34 Deep Learning in Computer Vision

15. I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio, “Binarized neu-

ral networks,” in Advances in Neural Information Processing Systems – NIPS ’16,
February 2016, pp. 4107–4115.
16. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke,
and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition - CVPR ’15, pp. 1–9, 2015.
17. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recogni-
tion,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition - CVPR ’16, 2016, pp. 770–778.
18. J. Cong and B. Xiao, “Minimizing computation in convolutional neural networks,”
in Proceedings of the International Conference on Artifcial Neural Networks -
ICANN ’14. Springer, 2014, pp. 281–290.
19. A. Canziani, A. Paszke, and E. Culurciello, “An analysis of deep neural network mod-
els for practical applications,” arXiv e-print, May 2016.
20. R. G. Shoup, “Parameterized convolution fltering in a feld programmable gate array,”
in Proceedings of the International Workshop on Field Programmable Logic and
Applications on More FPGAs, 1994, pp. 274–280.
21. M. Horowitz, “Computing’s energy problem (and what we can do about it),” in
Proceedings of the IEEE International Solid-State Circuits - ISSCC ’14. IEEE,
February 2014, pp. 10–14.
22. V. Sze, Y.-H. Chen, T.-J. Yang, and J. Emer, “Effcient processing of deep neural net-
works: A tutorial and survey,” Proceedings of the IEEE, vol. 105, no. 12, pp. 2295–
2329, December 2017.
23. Nvidia, “GPU-based deep learning inference: A performance and power analysis,”
White Paper, 2015. https://www.nvidia.com/
24. S. Chetlur, C. Woolley, P. Vandermersch, J. Cohen, J. Tran, B. Catanzaro, and E.
Shelhamer, “cuDNN: Effcient primitives for deep learning,” arXiv e-print, 2014.
25. Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama,
and T. Darrell, “Caffe: Convolutional architecture for fast feature embedding,” in
Proceedings of the ACM International Conference on Multimedia – MM ’14, 2014, pp.
675–678.
26. M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat,
G. Irving, M. Isard, M. Kudlur, J. Levenberg, R. Monga, S. Moore, D. G. Murray,
B. Steiner, P. Tucker, V. Vasudevan, P. Warden, M. Wicke, Y. Yu, X. Zheng, and G.
Brain, “TensorFlow: A system for large-scale machine learning,” in Proceedings of the
USENIX Symposium on Operating Systems Design and Implementation - OSDI ’16,
2016, pp. 265–284.
27. K. Ovtcharov, O. Ruwase, J.-y. Kim, J. Fowers, K. Strauss, and E. Chung, “Accelerating
deep convolutional neural networks using specialized hardware,” White Paper, pp. 3–6,
February 2015.
28. J. Qiu, J. Wang, S. Yao, K. Guo, B. Li, E. Zhou, J. Yu, T. Tang, N. Xu, S. Song, Y. Wang,
and H. Yang, “Going deeper with embedded FPGA platform for convolutional neu-
ral network,” in Proceedings of the ACM/SIGDA International Symposium on Field-
Programmable Gate Arrays - FPGA ’16. ACM, 2016, pp. 26–35.
29. N. Suda, V. Chandra, G. Dasika, A. Mohanty, Y. Ma, S. Vrudhula, J.-s. Seo, and Y. Cao,
“Throughput-optimized openCL-based FPGA accelerator for large-scale convolutional
neural networks,” in Proceedings of the ACM/SIGDA International Symposium on
Field-Programmable Gate Arrays - FPGA ’16, 2016, pp. 16–25.
30. K. Chellapilla, S. Puri, and P. Simard, “High performance convolutional neural net-
works for document processing,” in Proceedings of the International Workshop on
Frontiers in Handwriting Recognition – FHR ’06. Suvisoft, October 2006.
Accelerating the CNN Inference on FPGAs 35

31. U. Aydonat, S. O’Connell, D. Capalija, A. C. Ling, and G. R. Chiu, “An openCL(TM)

deep learning accelerator on arria 10,” in Proceedings of the ACM/SIGDA International
Symposium on Field-Programmable Gate Arrays - FPGA ’17, ACM, Ed. ACM, 2017,
pp. 55–64.
32. R. DiCecco, G. Lacey, J. Vasiljevic, P. Chow, G. Taylor, and S. Areibi, “Caffeinated
FPGAs: FPGA framework for convolutional neural networks,” in Proceedings of the
International Conference on Field- Programmable Technology - FPT ’16, pp. 265–
268, 2016.
33. C. Zhang and V. Prasanna, “Frequency domain acceleration of convolutional neural
networks on CPU-FPGA shared memory system,” in Proceedings of the ACM/SIGDA
International Symposium on Field- Programmable Gate Arrays - FPGA ’17, 2017, pp.
35–44.
34. J. H. Ko, B. A. Mudassar, T. Na, and S. Mukhopadhyay, “Design of an energy-effcient
accelerator for training of convolutional neural networks using frequency-domain compu-
tation,” in Proceedings of the Annual Conference on Design Automation - DAC ’17, 2017.
35. S. Venieris and C. Bouganis, “FpgaConvNet: A framework for mapping convolu-
tional neural networks on FPGAs,” in Proceedings of the IEEE Annual International
Symposium on Field-Programmable Custom Computing Machines - FCCM ’16, 2016,
pp. 40–47.
36. H. Sharma, J. Park, D. Mahajan, E. Amaro, J. K. Kim, C. Shao, A. Mishra, and H.
Esmaeilzadeh, “From high-level deep neural models to FPGAs,” in Proceedings of the
International Symposium on Microarchitecture - MICRO ’16, 2016, pp. 1–12.
37. H. Li, X. Fan, L. Jiao, W. Cao, X. Zhou, and L. Wang, “A high performance FPGA-
based accelerator for large-scale convolutional neural networks,” in Proceedings of the
International Conference on Field Programmable Logic and Applications - FPL ’16.
IEEE, August 2016, pp. 1–9.
38. K. Abdelouahab, M. Pelcat, J. Serot, C. Bourrasset, and F. Berry, “Tactics to directly
map CNN graphs on embedded FPGAs,” IEEE Embedded Systems Letters, vol. 9, no.
4, pp. 113–116, December 2017.
39. C. Zhang, P. Li, G. Sun, Y. Guan, B. Xiao, and J. Cong, “Optimizing FPGA-based
accelerator design for deep convolutional neural networks,” in Proceedings of the
ACM/SIGDA International Symposium on Field-Programmable Gate Arrays - FPGA
’15, ser. FPGA, 2015, pp. 161–170.
40. M. Motamedi, P. Gysel, V. Akella, and S. Ghiasi, “Design space exploration of FPGA-
based deep convolutional neural networks,” in Proceedings of the Asia and South
Pacifc Design Automation Conference – ASPDAC ’16, January 2016, pp. 575–580.
41. M. Motamedi, P. Gysel, and S. Ghiasi, “PLACID: A platform for FPGA-based
accelerator creation for DCNNs,” ACM Transactions on Multimedia Computing,
Communications, and Applications, vol. 13, no. 4, pp. 62:1–62:21, September 2017.
42. P. Molchanov, S. Tyree, T. Karras, T. Aila, and J. Kautz, “Pruning convolutional neural
networks for resource effcient learning,” arXiv preprint, 2017.
43. T. Fujii, S. Sato, H. Nakahara, and M. Motomura, “An FPGA realization of a deep
convolutional neural network using a threshold neuron pruning,” in Proceedings of the
International Symposium on Applied Reconfgurable Computing – ARC ’16, vol. 9625,
2017, pp. 268–280.
44. S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and Y. Zou, “DoReFaNet: Training low bit-
width convolutional neural networks with low bitwidth gradients,” arXiv e-print, 2016.
45. Y. Umuroglu, N. J. Fraser, G. Gambardella, M. Blott, P. Leong, M. Jahre, and K.
Vissers, “FINN: A framework for fast, scalable binarized neural network inference,”
in Proceedings of the ACM/SIGDA International Symposium on Field-Programmable
Gate Arrays - FPGA ’17, 2017, pp. 65–74.
46. R. Andri, L. Cavigelli, D. Rossi, and L. Benini, “YodaNN: An ultralow power convo-
lutional neural network accelerator based on binary weights,” in 2016 IEEE Computer
Society Annual Symposium on VLSI (ISVLSI), pp. 236–241, July 2016.
36 Deep Learning in Computer Vision

47. R. Zhao, W. Ouyang, H. Li, and X. Wang, “Saliency detection by multicontext deep
learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition - CVPR ’15, 2015, pp. 1265–1274.
48. C. Zhang, D. Wu, J. Sun, G. Sun, G. Luo, and J. Cong, “Energy-effcient CNN imple-
mentation on a deeply pipelined FPGA cluster,” in Proceedings of the International
Symposium on Low Power Electronics and Design - ISLPED ’16, 2016, pp. 326–331.
49. C. Zhang, Z. Fang, P. Zhou, P. Pan, and J. Cong, “Caffeine: Towards uniformed repre-
sentation and acceleration for deep convolutional neural networks,” in Proceedings of the
International Conference on Computer-Aided Design - ICCAD ’16. ACM, 2016, pp. 1–8.
50. J. Zhang and J. Li, “Improving the performance of openCL-based FPGA accelerator
for convolutional neural network,” in Proceedings of the ACM/SIGDA International
Symposium on Field-Programmable Gate Arrays - FPGA ’17, 2017, pp. 25–34.
51. Y. Guan, H. Liang, N. Xu, W. Wang, S. Shi, X. Chen, G. Sun, W. Zhang, and J. Cong,
“FP-DNN: An automated framework for mapping deep neural networks onto FPGAs
with RTL-HLS hybrid templates,” in Proceedings of the IEEE Annual International
Symposium on Field- Programmable Custom Computing Machines - FCCM ’17. IEEE,
2017, pp. 152–159.
52. S. Winograd, Arithmetic Complexity of Computations. SIAM, 1980, vol. 33.
53. A. Lavin and S. Gray, “Fast algorithms for convolutional neural networks,” arXiv
e-print, vol. arXiv: 150, September 2015.
54. L. Lu, Y. Liang, Q. Xiao, and S. Yan, “Evaluating fast algorithms for convolutional neu-
ral networks on FPGAs,” in Proceedings of the IEEE Annual International Symposium
on Field-Programmable Custom Computing Machines - FCCM ’17, 2017, pp. 101–108.
55. J. Bottleson, S. Kim, J. Andrews, P. Bindu, D. N. Murthy, and J. Jin, “ClCaffe: OpenCL accel-
erated caffe for convolutional neural networks,” in Proceedings of the IEEE International
Parallel and Distributed Processing Symposium – IPDPS ’16, 2016, pp. 50–57.
56. T. Highlander and A. Rodriguez, “Very effcient training of convolutional neural net-
works using fast fourier transform and overlap-and- add,” arXiv preprint, pp. 1–9, 2016.
57. H. Zeng, R. Chen, C. Zhang, and V. Prasanna, “A framework for generating high
throughput CNN implementations on FPGAs,” in Proceedings of the ACM/SIGDA
International Symposium on Field- Programmable Gate Arrays - FPGA ’18. ACM
Press, 2018, pp. 117–126.
58. M. Sankaradas, V. Jakkula, S. Cadambi, S. Chakradhar, I. Durdanovic, E. Cosatto, and
H. P. Graf, “A massively parallel coprocessor for convolutional neural networks,” in
Proceedings of the IEEE International Conference on Acoustics, Speech and Signal
Processing - ICASSP ’17. IEEE, July 2009, pp. 53–60.
59. C. Farabet, C. Poulet, J. Y. Han, Y. LeCun, D. R. Tobergte, and S. Curtis, “CNP: An
FPGA-based processor for convolutional networks,” in Proceedings of the International
Conference on Field Programmable Logic and Applications - FPL ’09, pp. 32–37, 2009.
60. S. Chakradhar, M. Sankaradas, V. Jakkula, and S. Cadambi, “A dynamically con-
fgurable coprocessor for convolutional neural networks,” ACM SIGARCH Computer
Architecture News, vol. 38, no. 3, pp. 247–257, June 2010.
61. V. Gokhale, J. Jin, A. Dundar, B. Martini, and E. Culurciello, “A 240 G-ops/s mobile
coprocessor for deep neural networks,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition - CVPR ’14, June 2014, pp. 696–701.
62. M. Alwani, H. Chen, M. Ferdman, and P. Milder, “Fused-layer CNN accelerators,” in
Proceedings of the Annual International Symposium on Microarchitecture - MICRO
’16, vol. 2016, December 2016.
63. Y. Ma, Y. Cao, S. Vrudhula, and J.-s. Seo, “Optimizing loop operation and datafow
in FPGA acceleration of deep convolutional neural networks,” in Proceedings of the
ACM/SIGDA International Symposium on Field-Programmable Gate Arrays - FPGA
’17, 2017, pp. 45–54.
Accelerating the CNN Inference on FPGAs 37

64. V. Gokhale, A. Zaidy, A. Chang, and E. Culurciello, “Snowfake: An effcient hardware

accelerator for convolutional neural networks,” in Proceedings of the IEEE International
Symposium on Circuits and Systems - ISCAS ’17. IEEE, May 2017, pp. 1–4.
65. Microsoft, “Microsoft unveils Project Brainwave for real-time AI,” 2017. https://www.
microsoft.com/en-us/research/blog/microsoft-unveils-project-brainwave/
66. S. Derrien and S. Rajopadhye, “Loop tiling for reconfgurable accelerators,” in
Proceedings of the International Conference on Field Programmable Logic and
Applications - FPL ’01. Springer, 2001, pp. 398–408.
67. R. Atul, L. Jongeun, and C. Kiyoung, “Effcient FPGA acceleration of convolutional
neural networks using logical-3D compute array,” in Proceedings of the Design,
Automation & Test in Europe Conference & Exhibition - DATE ’16. IEEE, 2016,
pp. 1393–1398.
68. Y. Ma, N. Suda, Y. Cao, J. S. Seo, and S. Vrudhula, “Scalable and modularized RTL
compilation of Convolutional Neural Networks onto FPGA,” in Proceedings of the
26th International Conference on Field Programmable Logic and Applications (FPL),
pp. 1–8. IEEE, 2016.
69. S. Williams, A. Waterman, and D. Patterson, “Roofine: An insightful visual perfor-
mance model for multicore architectures,” Communications of the ACM, vol. 52, no. 4,
p. 65, April 2009.
70. Y. Ma, M. Kim, Y. Cao, S. Vrudhula, and J.-s. Seo, “End-to-end scalable FPGA acceler-
ator for deep residual networks,” in Proceedings of the IEEE International Symposium
on Circuits and Systems - ISCAS ’17. IEEE, May 2017, pp. 1–4.
71. Y. Ma, Y. Cao, S. Vrudhula, and J.-s. Seo, “An automatic RTL compiler for high-
throughput FPGA implementation of diverse deep convolutional neural networks,”
in Proceedings of the International Conference on Field Programmable Logic and
Applications - FPL ’17. IEEE, September 2017, pp. 1–8.
72. Z. Liu, Y. Dou, J. Jiang, J. Xu, S. Li, Y. Zhou, and Y. Xu, “Throughput-optimized
FPGA accelerator for deep convolutional neural networks,” ACM Transactions on
Reconfgurable Technology and Systems, vol. 10, no. 3, pp. 1–23, 2017.
73. S. Mittal, “A survey of techniques for approximate computing,” ACM Computing
Surveys, vol. 48, no. 4, pp. 1–33, March 2016.
74. S. Anwar, K. Hwang, and W. Sung, “Fixed point optimization of deep convolutional
neural networks for object recognition,” in 2015 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP). IEEE, April 2015.
75. S. Gupta, A. Agrawal, P. Narayanan, K. Gopalakrishnan, and P. Narayanan, “Deep
learning with limited numerical precision,” in Proceedings of the International
Conference on Machine Learning - ICML ’15, 2015, pp. 1737–1746.
76. D. Lin, S. Talathi, and V. Annapureddy, “Fixed point quantization of deep convolu-
tional networks,” in Proceedings of the International Conference on Machine Learning
- ICML ’16, 2016, pp. 2849–2858.
77. M. Courbariaux, Y. Bengio, and J.-P. David, “Training deep neural networks with low
precision multiplications,” arXiv e-print, December 2014.
78. S. Zhou, Y. Wang, H. Wen, Q. He, and Y. Zou, “Balanced quantization: An effective
and effcient approach to quantized neural networks,” Journal of Computer Science
and Technology, vol. 32, pp. 667–682, 2017.
79. J.-P. David, K. Kalach, and N. Tittley, “Hardware complexity of modular multiplication
and exponentiation,” IEEE Transactions on Computers, vol. 56, no. 10, pp. 1308–1319,
October 2007.
80. C. Farabet, B. Martini, B. Corda, P. Akselrod, E. Culurciello, and Y. LeCun, “NeuFlow:
A runtime reconfgurable datafow processor for vision,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition - CVPR ’11. IEEE, June
2011, pp. 109–116.
38 Deep Learning in Computer Vision

81. P. Gysel, M. Motamedi, and S. Ghiasi, “Hardware-oriented approximation of convolu-

tional neural networks,” arXiv preprint, 2016, p. 8.
82. A. Kouris, S. I. Venieris, and C.-S. Bouganis, “CascadeCNN: Pushing the perfor-
mance limits of quantisation in convolutional neural networks,” in Proceedings of the
International Conference on Field Programmable Logic and Applications - FPL ’18,
pp. 155–1557, July 2018.
83. D. Williamson, “Dynamically scaled fxed point arithmetic,” in Proceedings of the
IEEE Pacifc Rim Conference on Communications, Computers and Signal Processing
Conference. IEEE, 1991, pp. 315–318.
84. S. Guo, L. Wang, B. Chen, Q. Dou, Y. Tang, and Z. Li, “FixCaffe: Training CNN
with low precision arithmetic operations by fxed point caffe,” in Proceedings of the
International Workshop on Advanced Parallel Processing Technologies - APPT ’17.
Springe, August 2017, pp. 38–50.
85. P. Gysel, J. Pimentel, M. Motamedi, and S. Ghiasi, “Ristretto: A framework for empiri-
cal study of resource-effcient inference in convolutional neural networks,” IEEE
Transactions on Neural Networks and Learning Systems, vol. 29, pp. 1–6, 2018.
86. M. Courbariaux, Y. Bengio, and J.-P. David, “BinaryConnect: Training deep neural
networks with binary weights during propagations,” in Advances in Neural Information
Processing Systems – NIPS ’15, 2015, pp. 3123–3131.
87. H. Nakahara, T. Fujii, and S. Sato, “A fully connected layer elimination for a bina-
rizec convolutional neural network on an FPGA,” in Proceedings of the International
Conference on Field Programmable Logic and Applications - FPL ’17. IEEE,
September 2017, pp. 1–4.
88. I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio, “Quantized neu-
ral networks: Training neural networks with low precision weights and activations,”
Journal of Machine Learning Research, vol. 18, pp. 187:1–187:30, September 2018.
89. C. Zhu, S. Han, H. Mao, and W. J. Dally, “Trained ternary quantization,” in Proceedings
of the International Conference on Learning Representations – ICLR ’17, December
2017.
90. R. Zhao, W. Song, W. Zhang, T. Xing, J.-H. Lin, M. Srivastava, R. Gupta, and Z. Zhang,
“Accelerating binarized convolutional neural networks with software-programmable
FPGAs,” in Proceedings of the ACM/SIGDA International Symposium on Field-
Programmable Gate Arrays - FPGA ’17, 2017.
91. N. J. Fraser, Y. Umuroglu, G. Gambardella, M. Blott, P. Leong, M. Jahre, and K.
Vissers, “Scaling binarized neural networks on reconfgurable logic,” in Proceedings
of the Workshop on Parallel Programming and Run-Time Management Techniques for
Many-core Architectures and Design Tools and Architectures for Multicore Embedded
Computing Platforms - PARMA-DITAM ’17. ACM, 2017, pp. 25–30.
92. S. Liang, S. Yin, L. Liu, W. Luk, and S. Wei, “FP-BNN: Binarized neural network on
FPGA,” Neurocomputing, vol. 275, pp. 1072–1086, January 2018.
93. A. ProstBoucle, A. Bourge, F. Ptrot, H. Alemdar, N. Caldwell, and V. Leroy, “Scalable
high-performance architecture for convolutional ternary neural networks on FPGA,”
in Proceedings of the International Conference on Field Programmable Logic and
Applications - FPL ’17, pp. 1–7, July 2017.
94. B. Liu, M. Wang, H. Foroosh, M. Tappen, and M. Pensky, “Sparse convolutional neural
networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition - CVPR ’15, 2015, pp. 806–814.
95. S. Han, J. Pool, J. Tran, and W. J. Dally, “Learning both weights and connections for
effcient neural network,” in Advances in Neural Information Processing Systems –
NIPS ’15, 2015, pp. 1135–1143.
Accelerating the CNN Inference on FPGAs 39

96. T.-J. Yang, Y.-H. Chen, and V. Sze, “Designing energy-effcient convolutional neural
networks using energy-aware pruning,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition - CVPR ’17, pp. 5687–5695, 2017.
97. S. Han, H. Mao, and W. J. Dally, “Deep compression - compressing deep neural net-
works with pruning, trained quantization and huffman coding,” in Proceedings of the
International Conference on Learning Representations – ICLR ’16, 2016, pp. 1–13.
98. A. Sironi, B. Tekin, R. Rigamonti, V. Lepetit, and P. Fua, “Learning separable flters,”
IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 37, no. 1, pp.
94–106, 2015.
99. C. Bouganis, G. Constantinides, and P. Cheung, “A novel 2D flter design methodology
for heterogeneous devices,” in Proceedings of the Annual IEEE Symposium on Field-
Programmable Custom Computing Machines - FCCM ’05. IEEE, 2005, pp. 13–22.
100. R. Dorrance, F. Ren, and D. Markovi, “A scalable sparse matrix-vector multiplication
kernel for energy-effcient sparse-blas on FPGAs,” in Proceedings of the ACM/SIGDA
International Symposium on Field- Programmable Gate Arrays - FPGA ’14. ACM,
2014, pp. 161–170.
 2 Object Detection
with Convolutional
Neural Networks
Kaidong Li, Wenchi Ma, Usman Sajid,
Yuanwei Wu, and Guanghui Wang

CONTENTS
2.1 Introduction .................................................................................................... 41
2.1.1 Major Evolution .................................................................................. 42
2.1.2 Other Development............................................................................. 43
2.2 Two-Stage Model............................................................................................44
2.2.1 Regions with CNN Features (R-CNN)...............................................44
2.2.2 Fast R-CNN ........................................................................................ 45
2.2.3 Faster R-CNN .....................................................................................46
2.3 One-Stage Model ............................................................................................ 48
2.3.1 You Only Look Once (YOLO) ........................................................... 48
2.3.2 YOLOv2 and YOLO9000 .................................................................. 49
2.3.3 YOLOv3 ............................................................................................. 50
2.3.4 Single-Shot Detector (SSD) ................................................................ 50
2.4 Other Detector Architectures ......................................................................... 51
2.4.1 Deep Residual Learning (ResNet)...................................................... 52
2.4.2 RetinaNet ............................................................................................ 53
2.5 Performance Comparison on Different Datasets............................................ 54
2.5.1 PASCAL VOC .................................................................................... 54
2.5.2 MS COCO .......................................................................................... 55
2.5.3 VisDrone-DET2018 ............................................................................ 55
2.6 Conclusion ...................................................................................................... 58
Bibliography ............................................................................................................ 58

2.1 INTRODUCTION
Deep learning was frst proposed in 2006 [1]; however, it did not attract much atten-
tion until 2012. With the development of computational power and a large number
of labeled datasets [2, 3], deep learning has proven to be very effective in extract-
ing intrinsic structural and high-level features. Signifcant progress has been made
in solving various problems in the artifcial intelligence community [4], espe-
cially in areas where the data are multidimensional and the features are diffcult to

41
42 Deep Learning in Computer Vision

hand-engineer, including speech recognition [5, 6, 7], natural language processing

[8, 9, 10], and computer vision [11, 12, 13, 14]. Deep learning also shows dominance
on areas like business analysis [15], medical diagnostic [16, 17], art creation [18], and
image translation [19, 20]. Since little human input is required, deep learning will
continue to make more and more impact in the future with the increase of computing
power and exploding data growth.
Among all these areas, computer vision has witnessed remarkably successful
applications of deep learning [21, 22, 23]. Vision, as the most important sense in
terms of navigation and recognition for humans, provides the most information about
the surroundings. Thanks to the rapid development of digital cameras in the last 10
years, image sensors have become one of the most accessible types of hardware.
Researchers tend to explore the possibility of using computer vision in applica-
tions like autonomous driving, visual surveillance, facial recognition [24], etc. The
demand in these areas, in return, draws tremendous attention and provides huge
resources to the computer vision feld. As a result, it is developing at a speed that
has been rarely seen in other research felds. Every month, sometimes even within
a week, we have seen record performance achieved on datasets like ImageNet [11],
PASCAL VOC [25], and COCO [26].
Object detection is a fundamental step for many computer vision applications [23,
27, 28, 29]. The architecture of convolutional neural networks (CNN) designed for
object detection is usually used as the frst step to extract objects’ spatial and clas-
sifcation information. Then more modules are added for specifc applications. For
example, tracking shares a very similar task to object detection with the addition of a
temporal module. Therefore, the performance of object detection will affect almost
all other computer vision research. A huge amount of effort has been put into its
development [30, 31, 32].
In this chapter, we present a brief overview of the recent developments in object
detection using convolutional neural networks (CNNs). Several classical CNN-based
detectors are presented. Some developments are based on the detector architectures
[33, 34], while others are focused on solving certain problems, like model degrada-
tion and small-scale object detection [35, 36, 37]. The chapter also presents some
performance comparison results of different models on several benchmark datasets.
Through a discussion of these models, we hope to give readers a general idea about
the developments of CNN-based object detection.

2.1.1 MAJOR EVOLUTION

To talk about object detection, we have to mention image classifcation. Image clas-
sifcation is the task of assigning a label to an input image from a fxed set of cat-
egories. The assigned label usually corresponds to the most salient object in the
image. It works best when the object is centered and dominating in the image frame.
However, most images contain multiple objects, scattering in the frame with dif-
ferent scales. One label is far from enough to describe the contextual meaning of
images. Therefore, object detection is introduced not only to output multiple labels
corresponding to an image but also to generate the spatial region associated with
each label.
Object Detection with Convolutional Neural Networks 43

Two-stage models. Inspired by image classifcation, it is natural for researchers

to detect objects by exploring a two-stage approach. In most two-stage models, the
frst stage is for region proposal, followed by image classifcation on the proposed
regions. With this modifcation, additional work is done on localization, with few
changes to previous classifcation models. For the region proposal part, some sys-
tems employ a sliding window technique, like Deformable Parts Models (DPM) [38]
and OverFeat [39]. With certain strategies, these methods usually apply classifers
on windows at different locations with different scales. Another region proposal is
selective search [40], which is adopted by regions with CNN features (R-CNN) [41].
It selectively extracts around 2,000 bottom-up region proposals [41], which greatly
reduces the regions needed in sliding window methods.
The two-stage models, generally speaking, yield higher accuracy. With each stage
doing its specifc task, these models perform better on objects with various sizes.
However, with the demands for real-time detection, two-stage models show its weak-
ness in processing speed.
One-stage models. With more and more powerful computing ability available,
CNN layers are becoming deeper and deeper. Researchers are able to utilize one-
stage methods, like YOLO (You Only Look Once) [34] and SSD (single-shot multi-
box detector) [42], with faster detection speed and similar or sometimes even higher
accuracy. One-stage models usually divide the image into N × N grids. Each grid cell
is responsible for the object whose center falls into that grid. Thus, the output is an
N × N × S tensor. Each of the S feature maps is an N × N matrix, with each element
describing the feature on its corresponding grid cell. A common design of these fea-
ture maps for one object consists of the following S = (5 + C) values:

• Four values for the bounding-box dimensions (x coordinate, y coordinate,

height, and width);
• One value for the possibility of this grid cell containing the object;
• C values indicating which class this object belongs to.

One-stage models, thanks to their simpler architecture, usually require less com-
putational resource. Some recent networks could achieve more than 150 frames
per second (fps) [43]. However, the trade-off is accuracy, especially for small-scale
objects as shown in Table 2.2 in Section 2.5.1. Another drawback is the class imbal-
ance during training. In order to detect B objects in a grid, the output tensor we
mentioned above includes information for N × N × B number of anchor boxes. But
among them, only a few contain objects. The ratio most of the time is about 1,000:1
[36]. This results in low effciency during training.

2.1.2 OTHER DEVELOPMENT

With the general architecture established, new detector development mainly focuses
on certain aspects to improve the performance.
Studies [44, 12] show strong evidence that network depth is crucial in CNN model
performance. However, when we expect the model to converge for a loss function
during training, the problem of vanishing/exploding gradients prevents it from
44 Deep Learning in Computer Vision

behaving this way. The study from Bradley [65] back in 2010 found that the mag-
nitude of gradient update decreases exponentially from output layer to input layer.
Therefore, training becomes ineffcient towards the input layers. This shows the dif-
fculty of training very deep models. On the other hand, research shows that simply
adding additional layers will not necessarily result in better detection performance.
Monitoring the training process shows that added layers are trained to become iden-
tity maps. Therefore, it can only generate models whose performance is equal to a
shallower network at most after a certain amount of layers. To address this issue,
skip connections are introduced into the networks [35, 46, 47] to pass information
between nonadjacent layers.
Feature pyramid networks (FPN) [48], which output feature maps at different
scales, can detect objects at very different scales. This idea could also be found when
the features were hand-engineered [49]. Another problem is the trailing accuracy
in one-stage detector. According to Lin et al. [36], the lower accuracy is caused by
extreme foreground-background class imbalance during training. To address this
issue, Lin et al. introduced RetinaNet [36] by defning focal loss to reduce the weight
of background loss.

2.2 TWO-STAGE MODEL

A two-stage model usually consists of regional proposals’ extraction and classifca-
tion. It is an intuitive idea to do this after the success of image classifcation. This
type of model could use the proven image classifcation network after the region pro-
posal stage. In addition, to some extent, the two steps resemble how humans receive
visual information and arrange attentions on regions of interest (RoI). In this section,
we will introduce the R-CNN, fast R-CNN, and Faster R-CNN models [41, 50, 33],
and discuss the improvements of each model.

2.2.1 REGIONS WITH CNN FEATURES (R-CNN)

R-CNN was developed using a multi-stage process, which is shown in Figure 2.1. It
generally can be divided into two stages.

SVM Classifers

CONV
NET
Bounding Box
Regressors

Selec˙ve Search Warped region Compute

~2K region proposals for each proposal feature map

FIGURE 2.1 An overview of the R-CNN detection system. The system generates 2k
proposals. Each warped proposal produces feature maps. Then the system classifes each
proposal and refnes the bounding-box prediction.
Object Detection with Convolutional Neural Networks 45

The frst stage is the region proposal extraction. The R-CNN model utilizes selec-
tive search [40], which takes the entire image as input and generates around 2,000
class-independent region proposals. In theory, R-CNN is able to work with any
region proposal methods. Selective search is chosen because it performs well and has
been employed by other detection models. The second stage starts with a CNN that
takes a fxed-dimension image as input and generates a fxed-length feature vector as
output. The input comes from the regional proposals. Each proposal is wrapped into
the required size regardless of its original size and aspect ratio. Then using the CNN
feature vector, pre-trained class-specifc linear support vector machines (SVMs) are
applied to calculate the class scores. Girshick et al. [41] conducted an error analysis,
and based on the analysis result, a bounding-box regression is added to reduce the
localization errors using the CNN feature vector. The regressor, which is class spe-
cifc, can refne bounding-box prediction.

2.2.2 FAST R-CNN

R-CNN [50] has achieved great improvement compared to the previous best meth-
ods. Let us take the results on PASCAL VOC 2012 as an example. R-CNN has a 30%
relative performance increase over the previous best algorithm. Apart from all the
progress, R-CNN suffers from two major limitations. The frst one is the space and
time cost in training. In the second stage, the features for the 2,000 region proposals
are all extracted separately using CNN and stored to the disk. With VGG16, VOC07
trainval dataset of 5,000 images requires 2.5 GPU-days*, and the memory expands
to hundreds of GB. The second limitation is the slow detection speed. For each test
image, it takes a GPU 47 seconds with VGG16. This weakness prevents the R-CNN
model from real-time applications.
From the analysis above, both of the problems are caused by individually cal-
culating a huge number of feature vectors for region proposals. Fast R-CNN is
designed to overcome this ineffciency. The architecture of fast R-CNN is illustrated
in Figure 2.2. The frst stage remains unchanged, and generates region proposals

CONV Net
Somax
Classifca˙on
FC

ROI FC
Bounding
Pooling
Box
FC Regressor
Input Image Selec˙ve Search: Feature ROI Feature
~2K region proposals Maps Vector

FIGURE 2.2 An overview of Fast R-CNN structure. Feature maps of the entire image are
extracted. RoI is projected to the feature maps and pooled into a fxed size. Then, RoI fea-
ture vector is computed using fully connected layers. Using the vector, softmax classifcation
probability and bounding-box regression are calculated as outputs for each proposal.

* A unit of computational complexity equivalent to what a single GPU can complete in a day.
46 Deep Learning in Computer Vision

using selective search. For the second stage, instead of generating feature vectors
for each region proposal separately, fast R-CNN calculates one feature map from
the entire image and uses a RoI pooling layer to extract a fxed-length feature vec-
tor for each region proposal. Each RoI feature vector is processed by a sequence of
fully connected layers and forked into two branches. The frst branch calculates class
scores with a softmax layer, and the other branch is again a bounding-box regressor
that refnes the bounding-box estimation.
In fast R-CNN, the RoI pooling layer is the key part where the region proposals
are translated into fxed-dimension feature vectors. This layer takes the feature map
of the entire image and RoI as input. A four-element tuple (r,c,h,w) represents a RoI,
with (r,c) and (h,w) specifying its top-left corner and its height and width respectively
[50]. Then, the (h,w) dimension of RoI is divided into an H × W grid, where H and W
are layer hyper-parameters. Standard max pooling is conducted to each grid cell in
each feature map channel.
With similar mAP (mean average precision) levels, fast R-CNN is very successful
in improving the training/testing speed. For large-scale objects defned in He et al.
[35], fast R-CNN is 8.8 times faster in training and 146 times faster in detection [50].
It achieves 0.32s/image performance, which is a signifcant leap towards real-time
detection.

2.2.3 FASTER R-CNN

To further increase the detection speed, researchers discover that the region proposal
computation is the bottleneck of performance improvement. Before fast R-CNN, the
time taken at the second stage is signifcantly more than that at the region proposal
stage. Considering that selective search [40] works robustly and is a very popular
method, improving region proposal seems unnecessary. However, with the improve-
ment in fast R-CNN, selective search now is an order slower compared to classifca-
tion stage. Even with some proposal methods that are balanced between speed and
quality, region proposal stage still costs as much as the second stage. To solve this
issue, Ren et al. [33] introduced a Region proposal network (RPN), which shares the
feature map of the entire image with stage two.
The frst step in Faster R-CNN is a shared CNN network. The RPN then
takes the shared feature maps as input and generates region proposals with
scores indicating how confdent the network is that there are objects in them. In
the RPN, a small network will slide over the input feature map as indicated in
Figure 2.3. Each sliding window will generate a lower-dimensional vector. This
lower-dimensional vector will be the input of two sibling fully connected layers,
a bounding-box regression layer (reg layer), and a bounding-box classifcation
layer (cls layer). To increase the nonlinearity, Rectifed Linear Units (ReLUs) are
applied to the output of the CNN layer. For each sliding window, k region pro-
posals will be generated. Correspondingly, the reg layer will have 4k outputs for
bounding-box coordinates, and the cls layer will have 2k probabilities of object/
not-object for each bounding-box proposals. The k proposals are parameterized
relative to k reference boxes, called anchors. The parameters include the scales
and aspect ratio.
Object Detection with Convolutional Neural Networks 47

So°max
FC Classifca˜on
ROI
Pooling
Input Image FC
Bounding Box
FC Regressor

ROI Feature
Vector
For each
anchor box
cls layer
2k scores
CONV Bbox
regressor
Net 4k output

k anchor boxes
for each loaca˜on

RPN
FIGURE 2.3 An overview of the Faster R-CNN framework. Region proposal network
(RPN) shares feature maps with a fast R-CNN network; a small convolutional network
slides over the shared feature maps to produce region proposal; and each location has k
anchor boxes.

To train RPN, binary class labels are assigned to each anchor. Positive is assigned
to two situations.

• Anchors with highest Intersection-over-Union (IoU) with ground-truth

boxes;
• Anchors with IoU over 0.7 with ground-truth boxes.

Non-positive is given to anchors with IoU lower than 0.3 for all ground-truth boxes.
During training, only positive or non-positive anchors contribute to loss functions.
With the introduction of RPN, the total time to complete object detection on GPU
is 198 ms using VGG as the CNN layer. Compared to the selective search, it is
almost 10 times faster. The proposal stage is improved from 1,510 ms to only 10 ms.
Combined, the new Faster R-CNN achieves 5 fps.
Two-stage detectors usually have higher accuracy in comparison to one-stage
detectors. Most of the models at the top detection dataset leader board are two-
stage models. Region proposals with regressor to refne localization means that
they inherently produce much better results for bounding-box predictions. In terms
of speed, these detectors offered far from real-time performance when they were
frst introduced. With recent development, they are getting closer by simplifying
48 Deep Learning in Computer Vision

the architecture. It evolves from models with multiple computationally heavy steps
to models with a single shared CNN. After the Faster R-CNN [33] was published,
the networks are mainly based on shared feature maps. For example, R-FCN
(region-based fully convolutional networks) [51] and Cascade R-CNN [52] focus on
localization accuracy, and at the same time, they can both achieve around 10 fps.
Region-based two-stage detectors are becoming not only more accurate but also
faster. Therefore, it becomes one of the main branches for object detection.

2.3 ONE-STAGE MODEL

From the development of R-CNN, we can observe the success of sharing CNN in
almost all the steps of a detection network. Region proposal, classifcation, and
bounding-box refnement all take the output of the last shared CNN. This indicates
that, if designed properly, the feature maps contain information about all of the
above tasks. It is natural to explore the possibility of implementing detection as a
single regression problem.

2.3.1 YOU ONLY LOOK ONCE (YOLO)

YOLO [34] is a simple model that takes the entire image as input and simultaneously
generates multiple bounding boxes and class probabilities. Since YOLO takes input
from the entire image, it reasons globally for detection, which means each detection
takes the background and other objects into consideration.
YOLO divides input images into S × S grids. If an object center falls into a
grid cell, this grid cell is responsible for that object’s detection. Each grid cell
will predict B bounding boxes, and the bounding boxes are described by fve
parameters: x, y, w, h, and the confdence, where (x,y) represents the center of
the box relative to the bounds of the grid cell, and (w,h) represents the width and
height relative to the dimensions of the whole image, respectively. The conf-
dence is defned as:

Pr(Object)* IOU truth

pred

which refects how confdent the prediction is that an object exists in this bounding
box and how accurate the bounding box is relative to the ground-truth box. The fnal
parameter is C, conditional class probabilities. It represents the conditional prob-
abilities of the grid cell containing an object, regardless of how many bounding
boxes a grid cell has. Therefore, the fnal predictions are an S × S ×(B × 5 + C) tensor.
The network design is shown in Table 2.1. It is inspired by GoogLeNet [12], fol-
lowed by 2 fully connected layers: FC1 of 4,096 dimensional vector, and the fnal
output FC2 of 7 × 7 × 30 tensor. In the paper, the authors also introduced a fast ver-
sion with only 79 convolutional layers.
During training, the loss function is shown in Equation 2.1, where 1ijobj denotes
whether the jth prediction in cell i should make a prediction, and 1iobj denotes whether
there’s an object in cell i [34]. From the function, we can see that it penalizes the
Object Detection with Convolutional Neural Networks 49

TABLE 2.1
The Architecture of the Convolutional Layers of YOLO [34]
Stage Image conv1 conv2 conv3 conv4 conv5 conv6

Filter/ 7×7 3×3 1×1 1×1 1×1 3×3

üï üï
Size 64/2 192 128 256 512 1024
ý´ 4 ý´ 2
Stride 3×3 3×3 ïþ 3×3 ïþ 3×3
256 512 1024 1024
1×1 1×1 3×3
256 512 1024
3×3 3×3 3×3
512 1024 1024/2

Maxpool 2×2 2×2 2×2 2×2

Stride 2 2 2 2
Output 448 × 448 112 × 112 56 × 56 28 × 28 14 × 14 7×7 7×7
3 192 256 512 1024 1024 1024

error only when an object exists and when a prediction is actually responsible for the
ground truth.

Loss = Lossbounding box + Lossconfidence + Lossclassification (2.1)

where

S2

˝ ˝
B
Lossbounding box = lcoord 1ijobj ˙ˆ(xi − x̂i )2 + (yi − ŷi )2 ˇ˘
i=0 j =0

S2

˝ ˝ ˙
( ) +( ) ˇ
B 2 2
+ lcoord 1ijobj ˆ wi − ŵi yi − ŷi ˘ ,
i=0 j =0

S2
å å 1ijobj ( Ci - Cˆ i )
B 2
Lossconfidence =
i=0 j=0

S2
å å 1ijobj ( Ci - Cˆ i ) ,
B 2
+ l noodj
i=0 j=0

S2
Lossclassification = ˙ i=0
1iobj ˙ cˆclasses
( pi (c) − pˆ i (c))2 .

2.3.2 YOLOV2 AND YOLO9000

YOLOv2 [53] is an improved model based on YOLO. The original YOLO back-
bone network is replaced by a simpler Darknet-19, and the fully connected layers at
the end are removed. Redmon et al. also tested other different design changes and
applied modifcations only with accuracy increases. YOLO9000 [53], as the name
50 Deep Learning in Computer Vision

suggests, can detect 9,000 object categories. Based on a slightly modifed version of
YOLOv2, this is achieved with a joint training strategy on classifcation and detec-
tion datasets. A WorldTree hierarchy [34] is used to merge the ground-truth classes
from different datasets.
In this section, we will discuss some of the most effective modifcations. Batch
normalization can help to reduce the effect of internal covariate shifts [54], thus
accelerating convergence during training. By adding batch normalization to all CNN
layers, the accuracy is improved by 2.4%.
In YOLO, the classifer is trained on the resolution of 224 × 224. At the stage of
detection, the resolution is increased to 448 × 448. In YOLOv2, during the last 10
epochs of classifer training, the image is changed to a full 448 resolution, so the
detection training can focus on object detection rather than adapting to the new reso-
lution. This gives a 4% mAP increase. While trying anchor boxes with YOLO, the
issue of instability is exposed. YOLO predicts the box by generating offsets to each
anchor. The most effcient training is when all objects are predicted by the closest
anchor with a minimal number of offsets. However, without offset constrains, an
original anchor could predict an object at any location. Therefore in YOLOv2 [53],
a logistic activation constraint on offset is introduced to limit the predicted bound-
ing box near the original anchor. This makes the network more stable and increases
mAP by 4.8%.

2.3.3 YOLOV3
YOLOv3 [43] is another improved detector based on YOLOv2. The major improve-
ment is on the convolutional network. Based on the idea of Feature Pyramid Networks
[48], YOLOv3 predicts the boxes at 3 different scales, which helps to detect small
objects. The idea of skip connection, as discussed in the following section, is also
added into the design. As a result, the network becomes larger, with 53 layers com-
pared to its original 19 layers. Thus, it is called Darknet-53 [43]. Another achieve-
ment is that Darknet-53 runs at the highest measured foating point operation speed,
which is an indication that the network is better at utilizing GPU resources.

2.3.4 SINGLE-SHOT DETECTOR (SSD)

SSD [42] was introduced to make one-stage detector run in real time with com-
parative accuracy to the region proposal detectors. It is faster than YOLO [34] and
generates competitive accuracy compared to the latest two-stage models like Faster
R-CNN [33]. The architecture of SSD, which is shown in Figure 2.4, has three dis-
tinctive features. First, it uses multi-scale feature maps. From Figure 2.4, we can
observe that the output layers decrease in size progressively after a truncated base
network. The layers are chosen to be output and perform detection at different scales.
It allows prediction to be made at different scales. The shallower layers with more
details yield better results for smaller objects, while deeper layers with information
about the background are suited for larger objects. Second, the network in SSD is
fully convolutional, unlike YOLO [34], which employs fully connected layers at the
end. The third is the default bounding boxes and aspect ratios. In SSD, the image
Object Detection with Convolutional Neural Networks 51

Basee
Bas CONV
C ONV CONV
CO NV
Input
In put Image
Image
Netwoork
or
Network

Feature Feature Detec˜on Non Max

Non-Maximum
Map Map Layer Suppression

FIGURE 2.4 SSD architecture. The frst CONV layer, which is called the base network, is
truncated from a standard network. The detection layer computes confdence scores for each
class and offsets to default boxes.

is divided into grid cells. Each cell in the feature maps associates a set of default
bounding boxes and aspect ratios. SSD then computes category confdence scores
and bounding-box offsets to those default bounding boxes for each set. During pre-
diction, SSD performs detection for objects with different sizes on the feature maps
with various scales (Figure 2.4).
One-stage models inherently have the advantage of speed. In recent years, models
with more than 150 fps have been published, while a fast two-stage model [33] only
achieves around 20 fps. To make the prediction more accurate, networks combine
ideas from both one-stage and two-stage models to fnd the balance between speed
and accuracy. For example, Faster R-CNN [33] resembles the one-stage model in
sharing one major CNN network. One-stage models normally produce prediction
from the entire feature map. So they are good at taking context information into con-
sideration. However, this also means they are not very sensitive to smaller objects.
Methods like deconvolutional single-shot detector (DSSD) [55] and RetinaNet [36]
have been proposed to fx this problem from different viewpoints. DSSD learned
from FPN [48] modifes its CNN to predict objects at different scales. RetinaNet
develops a unique loss function to focus on hard objects. A detailed comparison of
results from different models on the PASCAL VOC dataset is given in Table 2.2.

2.4 OTHER DETECTOR ARCHITECTURES

As we discussed at the beginning of this chapter, emerging detectors published after
2016 mainly focus on the performance improvement of detection. Some focus on opti-
mizing the feature extraction ability of backbone networks [35, 56, 46, 48]. Some meth-
ods improve performance by modifying certain metrics, such as the loss function in
RetinaNet [36] and a new IoU defnition given in GIoU [57]. With new measuring met-
rics introduced in Lin et al. [26] and a comprehensive analysis performed after 2017,
detection of objects with different scales has become a major focus [48, 58, 59, 37].
In this section, we are going to discuss two examples, ResNet [35] and RetinaNet
[36]. ResNet [35] addresses some problems caused by increasing network layers,
while RetinaNet [36] tries to optimize the training effciency.
52 Deep Learning in Computer Vision

TABLE 2.2
Detection Results on PASCAL VOC 2007 Test Set
Detector mAP (%) fps

R-CNN [41] 58.5 0.02

Fast R-CNN [50] 70.0 0.5
Faster R-CNN (with VGG16) [33] 73.2 7
Faster R-CNN (with ZF) [33] 62.1 18
YOLO [34] 63.4 45
Fast YOLO [34] 52.7 155
YOLO (with VGG16) [34] 66.4 21
YOLOv2 288 × 288 [53] 69.0 91
YOLOv2 352 × 352 [53] 73.1 81
YOLOv2 416 × 416 [53] 76.8 67
YOLOv2 480 × 480 [53] 77.8 59
YOLOv2 544 × 544 [53] 78.6 40

2.4.1 DEEP RESIDUAL LEARNING (RESNET)

Degradation is the detection accuracy drop when the network depth increases.
Theoretically, this should not happen. If we assume the optimal network for a prob-
lem domain consists of a certain number of layers, an even deeper network should
be as good as the optimal one at least. The reason is that the added layers can be
trained as identity mappings, which simply take the input and pass it to the next
layer. The fact that degradation exists in practice implies some systems are more
diffcult to learn [35]. ResNets was published in December 2015. The effort evolves
around assisting the network to fnd the desired mappings. The key block diagram
for ResNets is shown in Figure 2.5.
The desired mapping without a skip connection is denoted as H( x ) = F ( x ) + x .
By introducing the skip connection, which is an identity connection, the layers are
now forced to ft the new mapping, F ( x ) = H( x ) − x . F ( x ) is defned as the residual
mapping. In addition, the paper hypothesizes that optimizing the residual mapping
is easier than the original mapping. For example, when a layer’s original optimal

Iden˜ty Map

ReLu ReLu
Input:

CONV CONV
Layer Layer

FIGURE 2.5 ResNet building block. Convert the underlying mapping from H( x ) to
F ( x ), where H( x ) = F ( x ) + x .
Object Detection with Convolutional Neural Networks 53

solution is an identity mapping, the modifed layer is going to ft zero by adding the
skip connection. The hypothesis is proven in the test results, as discussed in the next
session.
Another attractive property of ResNets is that the design does not introduce
any extra parameters or computational complexity. This is desirable in two
ways. First, it helps the network to achieve better performance without extra
cost. Second, it facilitates fair comparison between plain networks and residual
networks. During the test, we can run same networks with and without the skip
connection. Therefore, the performance difference is caused just by the skip
connection.
Based on a 34-layer plain network, the skip connection is added as indicated
in Figure 2.5. When the input and output dimensions are the same, the skip con-
nection can be added directly. And when the skip connection goes across different
dimensions, the paper proposes two options: (i) identity mapping with extra zeros
paddings; and (ii) using a projection shortcut, which is the projection in the plain
network block. With either option, a stride of 2 will be performed.
The signifcance of this network design is to reduce the diffculties in fnding
the optimal mappings of each layer. Without this design, researchers have to design
architectures with different depths, train on the dataset, and then compare the detec-
tion accuracy to narrow down the optimal depth. The process is very time-consum-
ing, and different problem domains may have various optimal depths. The common
training strategy, which is introduced in R-CNN [41], is supervised pre-training on a
large auxiliary dataset, followed by domain-specifc fne-tuning on a small dataset.
On the domain-specifc dataset, there is a high chance that the network architecture
is not optimal. Plus, it is practically impossible to fnd the optimal depth for each
specifc problem. However, with ResNets, we can safely add more layers and expect
better performance. We can rely on the networks and training to fnd the best per-
formance model as long as the optimal model is shallower. Another advantage of
ResNets is its effciency, which achieves performance improvement without adding
any computational complexity.

2.4.2 RETINANET
With all the advancements made in recent years, both two-stage methods and one-
stage methods have their own pros and cons. One-stage methods are usually faster in
detection, but trailing in detection accuracy. According to Lin et al. [36], the lower
accuracy is mainly caused by the extreme foreground-background class imbalance
during training. For the one-stage detector, it must sample much more candidate
locations in an image. In practice, the number of candidate locations normally goes
up to around 100k covering different positions, scales, and aspect ratios. Candidate
locations are dominated by background examples, which makes the training very
ineffcient. In two-stage detectors, however, most background regions are fltered out
by region proposals.
To address this issue, Lin et al. proposed RetinaNet [36]. The problem of a nor-
mal one-stage network is the overwhelming number of easy background candidates.
During training, it contributes little to the improvement of accuracy. Candidates can
54 Deep Learning in Computer Vision

be easily classifed as background, which means little information will be added to

the model. Since the number of them is dominating, it will overpower other hard
foreground candidates, resulting in degenerating models. A regular cross-entropy
(CE) loss for binary classifcation is shown in Equation 2.2.

° p, if y = 1
CE( p, y) = − log( pt ), where pt = ˛ (2.2)
˝1 − p, otherwise

To reduce the imbalance between the easy background and hard foreground samples,
Lin et al. [36] introduced the following focal loss.

FL( pt ) = −(1 − pt )g log( pt ) (2.3)

When a sample is misclassifed with a high score, pt will have small value. Therefore,
the loss is not affected that much. But when a sample is correctly classifed with a
high confdence, pt is close to 1, which will signifcantly down-weigh its effect. In
addition, a weighting factor α is employed to balance the importance of positive/
negative examples.

FL ( pt ) = −at (1 − pt )g log( pt ) (2.4)

Both models in this section address the problems encountered during experiments.
ResNet [35] has become one of the most widely used backbone networks. Its model
is provided in most of the popular deep learning frameworks, such as Caffe, PyTorch,
Keras, MXNet, etc. RetinaNet [36] achieves remarkable performance improvement
with minimal addition in computational complexity by designing a new loss function.
More importantly, it shows a new direction in optimizing the detector. Modifying
evaluation metrics could result in signifcant accuracy increase. Rezatofghi et al.
[57] in 2019 proposed Generalized Intersection-over-Union (GIoU) to replace the
current IoU and used GIoU as the loss function. Before the study of Rezatofghi et al.
[57], all CNN detectors calculate IoU during the test to evaluate the result, while
employing other metrics as loss function during training to optimize. Other recent
detectors proposed after 2017 mainly focus on detecting objects at various scales,
especially at small scale [52, 58, 60]. Among them, the trend in design is refnement
based on previous successful detectors [59, 37].

2.5 PERFORMANCE COMPARISON ON DIFFERENT DATASETS

In this section, we will present some comparative evaluation results for different
models on three benchmark datasets: PASCAL VOC [25], MS COCO [26], and
VisDrone-DET2018 [61].

2.5.1 PASCAL VOC

The PASCAL VOC [25] dataset is one of the pioneering works in generic object
detection, which is designed to provide a standardized testbed for object detection,
Object Detection with Convolutional Neural Networks 55

image classifcation, object segmentation, person layout, and action classifcation

[62]. The latest version is PASCAL VOC 2012. It has 20 classes and 11,530 images
in the trainval dataset with 27,450 RoI annotated objects. The initial 2007 version
was the frst dataset with 20 classes of objects.
To compare the performance of different networks on this dataset, we choose
the PASCAL VOC 2007 testing results from models trained on PASCAL VOC
2007 + 2012. The comparison results of different models are shown in Table 2.2.
Link between one-stage and two-stage models. From Table 2.2, we can tell that
the speed advantage of the one-stage detector is obvious. The frst version of YOLO
is almost three times faster than the fastest R-CNN detector. Looking back at the
evolution of these detectors, we can see that every modifcation from R-CNN leading
to one-stage detectors actually fts in a trend. R-CNN has two stages, with each stage
having its own separate computations. The improvements are made by eliminating
duplicate calculations and combining shared features. The fnal version of R-CNN
could be viewed as one main CNN network doing most of the work and some small
regressors and CNN layers generating predictions. This architecture has become
very similar to the one-stage model.
Speed and accuracy. For the same model, if we want to achieve higher accuracy,
speed usually needs to be compromised. So it is important to fnd the balance for
specifc applications. The test results show two ways to modify. A more complex
classifer or a higher-resolution image usually yields higher accuracy. At the same
time, more time will be needed to detect the objects.

2.5.2 MS COCO
MS COCO [26] is a large-scale object detection, segmentation, and captioning data-
set with 330,000 images. It aims at addressing three core problems at scene under-
standing: detection of non-iconic views (or non-canonical perspectives) of objects,
contextual reasoning between objects, and the precise 2D localization of objects
[26]. COCO defnes 12 metrics to evaluate the performance of a detector, which
gives a more detailed and insightful look. Since MS COCO is a relatively new data-
set, not all detectors have offcial results. We only compare the results of YOLOv3
and RetinaNet in Table 2.3. Although the two models were tested under different
GPUs, M40 and Titan X, their performances are almost identical [43].
Performance difference with higher IoU threshold. From Table 2.3, we can see
that YOLOv3 performs better than RetinaNet when the IoU threshold is 50%, while
at 75% IoU, RetinaNet has better performance, which indicates that RetinaNet has
higher localization accuracy. This aligns with the effort by Lin et al. [36]. The focal
loss is designed to put more weight on learning hard examples.

2.5.3 VISDRONE-DET2018
VisDrone-DET2018 dataset [61] is a special dataset that consists of images from
different drone platforms. It has 8,599 images in total, including 6,471 for train-
ing, 548 for validation, and 1,580 for testing. The images feature diverse real-
world scenes, collected in various scenarios (across 14 different cities in different
56 Deep Learning in Computer Vision

TABLE 2.3
Detection Results on MS COCO*
Detector Backbone Scale (pixels) AP AP50 AP75 fps

RetinaNet [36] ResNet-101-FPN 600 36.0 55.2 38.7 8.2

YOLOv3 [43] Darknet-53 608 33.0 57.9 34.4 19.6

* AP50 and AP75 Are Average Precision with IOU Thresholds of 0.5 and 0.75, Respectively. AP (Average
Precision) Is the Average over 10 Thresholds ([0.5:0.05:0.95]). RetinaNet Is Measured on an Nvidia
M40 GPU [36], while YOLOv3 Is Tested on an Nvidia Titan X [43].

regions), and under various environmental and lighting conditions [62]. This data-
set is challenging since most of the objects are small and densely populated as
shown in Figure 2.6.
To check the performance of different models on this dataset, we implemented
the latest version of each detector and show the result in Tables 2.4 and 2.5. The
results are calculated on the VisDrone validation set since the test set is not publicly
available.
From Table 2.4, it is evident that Faster R-CNN [33] performs signifcantly better
than YOLOv3 [43]. At IoU = 0.5, Faster R-CNN has 10% higher accuracy. YOLO
detectors inherently would struggle in datasets like this. The fully connected lay-
ers at the end take the entire feature map as input. It enables YOLO to have enough
contextual information. However, it lacks local details. In addition, each grid cell can
predict only a certain number of objects, which is determined before training starts.
So they have a hard time in the small and densely populated VisDrone-DET2018

FIGURE 2.6 Six sample images from the VisDrone-DET2018 dataset. They are pictures
taken from drones. The rectangular boxes on each image are ground truth. It has a large
number of small objects at distance.
Object Detection with Convolutional Neural Networks 57

TABLE 2.4
Performance Metrics on VisDrone-DET2018*
Detector Iterations AP Score

*YOLOv3 832 × 832 40 k AP @ [IoU=0.50:0.05:0.95 | maxDets=500] 18.74

AP @ [IoU=0.50 | maxDets=500] 29.77
AP @ [IoU=0.75 | maxDets=500] 19.53
AR @ [IoU=0.50:0.05:0.95 | maxDets=1] 0.91
AR @ [IoU=0.50:0.05:0.95 | maxDets=10] 5.63
AR @ [IoU=0.50:0.05:0.95 | maxDets=100] 27.16
AR @ [IoU=0.50:0.05:0.95 | maxDets=500] 27.42
*Faster R-CNN 80 k AP @ [IoU=0.50:0.05:0.95 | maxDets=500] 24.33
AP @ [IoU=0.50 | maxDets=500] 39.74
AP @ [IoU=0.75 | maxDets=500] 24.78
AR @ [IoU=0.50:0.05:0.95 | maxDets=1] 0.73
AR @ [IoU=0.50:0.05:0.95 | maxDets=10] 6.11
AR @ [IoU=0.50:0.05:0.95 | maxDets=100] 34.26
AR @ [IoU=0.50:0.05:0.95 | maxDets=500] 44.62

* AR @ [maxDets=1] and [maxDets=10] mean the maximum recall given 1 detection per
image and 10 detections per image, respectively.

TABLE 2.5
RetinaNet on VisDrone-DET2018 without Tuning
Pedestrian Person Bicycle Car Van

12.6 3.6 4.6 50.2 21.0

Truck Tricycle Awning-tricycle Bus Motor
17.8 10.7 4.9 32.1 11.7

dataset. It can be noted that the gap is reduced to 5% at IoU = 0.75, which means
YOLOv3 is catching up in terms of localization precision. We suggest this is the
result of adding prediction at 3 different scales in YOLOv3 and detecting on high-
resolution (832 × 832) images.
Table 2.5 shows the mAPs for each class from RetinaNet [36]. It shows that recent
detectors still need to be improved for small and morphologically similar objects.
The VisDrone dataset is a unique dataset, which has many potential real-life applica-
tions. Actually, most practical applications have to face different challenging situa-
tions, like bad exposure, lack of lighting, and a saturated number of objects. More
investigation has to be done to develop more effective models to handle these com-
plex real-life applications.
The evolution of object detection is partially linked to the availability of labeled
large-scale datasets. The existence of various datasets helps train the neural network
to extract more effective features. The different characteristics of datasets motivate
58 Deep Learning in Computer Vision

researchers to focus on different problems. In addition, new metrics proposed by

the dataset help us to better evaluate the detector performance. For example, MS
COCO style mAP [26] helps us to understand the performance related to localiza-
tion accuracy.

2.6 CONCLUSION
In this chapter, we have made a brief review of CNN-based object detection by pre-
senting some of the most typical detectors and network architectures. The detector
starts with a two-stage multi-step architecture and evolves to a simpler one-stage
model. In the latest models, even two-stage ones employ an architecture that shares
a single CNN feature map so as to reduce the computational load [52]. Some models
gain accuracy increase by fusing different ideas into one detector [47, 59]. In addi-
tion to the direct effort on detectors, training strategy is also an important factor to
produce high-quality results [41, 63]. With recent developments, most detectors have
decent performance in both accuracy and effciency. In practical applications, we
need to make a trade-off between the accuracy and the speed by choosing a proper
set of parameters and network structures. Although great progress has been made in
object detection in past years, some challenges still need to be addressed, like occlu-
sion [64] and truncation. In addition, better-designed datasets, like VisDrone, need
to be developed for specifc practical applications.

BIBLIOGRAPHY
1. G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for deep belief
nets,” Neural Computation, vol. 18, no. 7, pp. 1527–1554, 2006.
2. Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new
perspectives,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
35, no. 8, pp. 1798–1828, 2013.
3. I. Arel, D. C. Rose, T. P. Karnowski, et al., “Deep machine learning-a new frontier in
artifcial intelligence research,” IEEE Computational Intelligence Magazine, vol. 5, no.
4, pp. 13–18, 2010.
4. Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, p.
436, 2015.
5. G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V.
Vanhoucke, P. Nguyen, T. N. Sainath, et al., “Deep neural networks for acoustic mod-
eling in speech recognition: The shared views of four research groups,” IEEE Signal
Processing Magazine, vol. 29, no. 6, pp. 82–97, 2012.
6. T. N. Sainath, A.-r. Mohamed, B. Kingsbury, and B. Ramabhadran, “Deep convo-
lutional neural networks for lvcsr,” in IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pp. 8614–8618, IEEE, 2013.
7. G. E. Dahl, D. Yu, L. Deng, and A. Acero, “Context-dependent pre-trained deep neu-
ral networks for large-vocabulary speech recognition,” IEEE Transactions on Audio,
Speech, and Language Processing, vol. 20, no. 1, pp. 30–42, 2012.
8. R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, and P. Kuksa, “Natural
language processing (almost) from scratch,” Journal of Machine Learning Research,
vol. 12, pp. 2493–2537, 2011.
9. I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence learning with neural net-
works,” in Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.
Object Detection with Convolutional Neural Networks 59

10. R. Socher, C. C. Lin, C. Manning, and A. Y. Ng, “Parsing natural scenes and natural
language with recursive neural networks,” in Proceedings of the 28th International
Conference on Machine Learning (ICML-11), pp. 129–136, 2011.
11. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifcation with deep con-
volutional neural networks,” in Advances in Neural Information Processing Systems,
pp. 1097–1105, 2012.
12. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke,
and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp. 1–9, 2015.
13. L. He, G. Wang, and Z. Hu, “Learning depth from single images with deep neural net-
work embedding focal length,” IEEE Transactions on Image Processing, vol. 27, no. 9,
pp. 4676–4689, 2018.
14. J. Gao, J. Yang, G. Wang, and M. Li, “A novel feature extraction method for scene
recognition based on centered convolutional restricted boltzmann machines,”
Neurocomputing, vol. 214, pp. 708–717, 2016.
15. M. M. Najafabadi, F. Villanustre, T. M. Khoshgoftaar, N. Seliya, R. Wald, and E.
Muharemagic, “Deep learning applications and challenges in big data analytics,”
Journal of Big Data, vol. 2, no. 1, p. 1, 2015.
16. G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. Van
Der Laak, B. Van Ginneken, and C. I. Sánchez, “A survey on deep learning in medical
image analysis,” Medical Image Analysis, vol. 42, pp. 60–88, 2017.
17. X. Mo, K. Tao, Q. Wang, and G. Wang, “An effcient approach for polyps detection in
endoscopic videos based on faster R-CNN,” in 2018 24th International Conference on
Pattern Recognition (ICPR), pp. 3929–3934, IEEE, 2018.
18. A. Elgammal, B. Liu, M. Elhoseiny, and M. Mazzone, “CAN: Creative adversarial
networks, generating “art” by learning about styles and deviating from style norms,”
arXiv preprint arXiv:1706.07068, 2017.
19. W. Xu, S. Keshmiri, and G. Wang, “Adversarially approximated autoencoder for
image generation and manipulation,” IEEE Transactions on Multimedia, doi:10.1109/
TMM.2019.2898777, 2019.
20. W. Xu, S. Keshmiri, and G. Wang, “Toward learning a unifed many-to-many mapping
for diverse image translation,” Pattern Recognition, doi:10.1016/j.patcog.2019.05.017,
2019.
21. W. Ma, Y. Wu, Z. Wang, and G. Wang, “MDCN: Multi-scale, deep inception con-
volutional neural networks for effcient object detection,” in 2018 24th International
Conference on Pattern Recognition (ICPR), pp. 2510–2515, IEEE, 2018.
22. Z. Zhang, Y. Wu, and G. Wang, “Bpgrad: Towards global optimality in deep learning
via branch and pruning,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pp. 3301–3309, 2018.
23. L. Liu, W. Ouyang, X. Wang, P. Fieguth, J. Chen, X. Liu, and M. Pietikäinen, “Deep
learning for generic object detection: A survey,” arXiv preprint arXiv:1809.02165, 2018.
24. F. Cen and G. Wang, “Dictionary representation of deep features for occlusion-robust
face recognition,” IEEE Access, vol. 7, pp. 26595–26605, 2019.
25. M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal
visual object classes (voc) challenge,” International Journal of Computer Vision, vol.
88, no. 2, pp. 303–338, 2010.
26. T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L.
Zitnick, “Microsoft coco: Common objects in context,” in European Conference on
Computer Vision, pp. 740–755, Springer, 2014.
27. S. P. Bharati, S. Nandi, Y. Wu, Y. Sui, and G. Wang, “Fast and robust object tracking
with adaptive detection,” in 2016 IEEE 28th International Conference on Tools with
Artifcial Intelligence (ICTAI), pp. 706–713, IEEE, 2016.
60 Deep Learning in Computer Vision

28. Y. Wu, Y. Sui, and G. Wang, “Vision-based real-time aerial object localization and
tracking for UAV sensing system,” IEEE Access, vol. 5, pp. 23969–23978, 2017.
29. S. P. Bharati, Y. Wu, Y. Sui, C. Padgett, and G. Wang, “Real-time obstacle detection and
tracking for sense-and-avoid mechanism in UAVs,” IEEE Transactions on Intelligent
Vehicles, vol. 3, no. 2, pp. 185–197, 2018.
30. Y. Wei, X. Pan, H. Qin, W. Ouyang, and J. Yan, “Quantization mimic: Towards very tiny
CNN for object detection,” in Proceedings of the European Conference on Computer
Vision (ECCV), pp. 267–283, 2018.
31. K. Kang, H. Li, J. Yan, X. Zeng, B. Yang, T. Xiao, C. Zhang, Z. Wang, R. Wang, X.
Wang, et al., “T-CNN: Tubelets with convolutional neural networks for object detection
from videos,” IEEE Transactions on Circuits and Systems for Video Technology, vol.
28, no. 10, pp. 2896–2907, 2018.
32. W. Chu and D. Cai, “Deep feature based contextual model for object detection,”
Neurocomputing, vol. 275, pp. 1035–1042, 2018.
33. S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-time object detec-
tion with region proposal networks,” in Advances in Neural Information Processing
Systems, pp. 91–99, 2015.
34. J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unifed, real-
time object detection,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pp. 779–788, 2016.
35. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,”
in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 770–778, 2016.
36. T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detec-
tion,” in IEEE Transactions on Pattern Analysis and Machine Intelligence, p. 1, 2018.
37. Q. Zhao, T. Sheng, Y. Wang, Z. Tang, Y. Chen, L. Cai, and H. Ling, “M2det: A sin-
gle-shot object detector based on multi-level feature pyramid network,” CoRR, vol.
abs/1811.04533, 2019.
38. P. Felzenszwalb, D. McAllester, and D. Ramanan, “A discriminatively trained, multi-
scale, deformable part model,” in IEEE Conference on Computer Vision and Pattern
Recognition, pp. 1–8, IEEE, 2008.
39. P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, “Overfeat:
Integrated recognition, localization and detection using convolutional networks,” arXiv
preprint arXiv:1312.6229, 2013.
40. J. R. Uijlings, K. E. Van De Sande, T. Gevers, and A. W. Smeulders, “Selective search
for object recognition,” International Journal of Computer Vision, vol. 104, no. 2, pp.
154–171, 2013.
41. R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for accurate
object detection and semantic segmentation,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 580–587, 2014.
42. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg, “Ssd:
Single shot multibox detector,” in European Conference on Computer Vision, pp. 21–37,
Springer, 2016.
43. J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv preprint
arXiv:1804.02767, 2018.
44. K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale
image recognition,” arXiv preprint arXiv:1409.1556, 2014.
45. X. Glorot and Y. Bengio, “Understanding the diffculty of training deep feedforward
neural networks,” in Proceedings of the Thirteenth International Conference on
Artifcial Intelligence and Statistics, pp. 249–256, 2010.
46. G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected con-
volutional networks,” in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 4700–4708, 2017.
Object Detection with Convolutional Neural Networks 61

47. Y. Chen, J. Li, H. Xiao, X. Jin, S. Yan, and J. Feng, “Dual path networks,” in Advances
in Neural Information Processing Systems, pp. 4467–4475, 2017.
48. T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature
pyramid networks for object detection,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 2117–2125, 2017.
49. D. G. Lowe, “Object recognition from local scale-invariant features,” in The
Proceedings of the Seventh IEEE International Conference on Computer Vision, vol.
99, no. 2, pp. 1150–1157, IEEE, 1999.
50. R. Girshick, “Fast R-CNN,” in Proceedings of the IEEE International Conference on
Computer Vision, pp. 1440–1448, 2015.
51. J. Dai, Y. Li, K. He, and J. Sun, “R-FCN: Object detection via region-based fully con-
volutional networks,” in Advances in Neural Information Processing Systems, pp.
379–387, 2016.
52. Z. Cai and N. Vasconcelos, “Cascade R-CNN: Delving into high quality object detec-
tion,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 6154–6162, 2018.
53. J. Redmon and A. Farhadi, “Yolo9000: Better, faster, stronger,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pp. 7263–7271, 2017.
54. S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by
reducing internal covariate shift,” in Proceedings of the 32nd International Conference
on Machine Learning, vol. 37, pp. 448–456, 2015.
55. C.-Y. Fu, W. Liu, A. Ranga, A. Tyagi, and A. C. Berg, “DSSD: Deconvolutional single
shot detector,” arXiv preprint arXiv:1701.06659, 2017.
56. C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception
architecture for computer vision,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 2818–2826, 2016.
57. H. Rezatofghi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized
intersection over union: A metric and a loss for bounding box regression,” arXiv pre-
print arXiv:1902.09630, 2019.
58. P. Zhou, B. Ni, C. Geng, J. Hu, and Y. Xu, “Scale-transferrable object detection,” in
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.
528–537, 2018.
59. S. Zhang, L. Wen, X. Bian, Z. Lei, and S. Z. Li, “Single-shot refnement neural network
for object detection,” in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 4203–4212, 2018.
60. B. Singh and L. S. Davis, “An analysis of scale invariance in object detection snip,” in
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.
3578–3587, 2018.
61. P. Zhu, L. Wen, X. Bian, L. Haibin, and Q. Hu, “Vision meets drones: A challenge,”
arXiv preprint arXiv:1804.07437, 2018.
62. P. Zhu, L. Wen, D. Du, X. Bian, H. Ling, Q. Hu, Q. Nie, H. Cheng, C. Liu, X. Liu, et
al., “VisDrone-DET2018: The vision meets drone object detection in image challenge
results,” In Proceedings of the European Conference on Computer Vision (ECCV),
pp. 437–468, 2018.
63. J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, “Algorithms for hyper-parameter
optimization,” in Advances in Neural Information Processing Systems, pp. 2546–2554,
2011.
64. B. Wu and R. Nevatia, “Detection of multiple, partially occluded humans in a single
image by bayesian combination of edgelet part detectors,” in Tenth IEEE International
Conference on Computer Vision, vol. 1, pp. 90–97, 2005.
65. Bradley, David M. “Learning in modular systems,” No. CMU-RI-TR-09-26.
CARNEGIE-MELLON UNIV PITTSBURGH PA ROBOTICS INST, 2010.
 3 Effcient Convolutional
Neural Networks
for Fire Detection in
Surveillance Applications
Khan Muhammad, Salman Khan,
and Sung Wook Baik

CONTENTS
3.1 Introduction ....................................................................................................64
3.2 Related Work .................................................................................................. 65
3.2.1 Traditional Sensor-Based Fire Detection............................................ 65
3.2.2 Vision-Based Fire Detection...............................................................66
3.2.2.1 Color- and Motion-Based Fire Detection Methods..............66
3.2.2.2 Deep Learning-Based Fire Detection Methods ...................66
3.3 CNN-Based Fire Detection Methods ............................................................. 67
3.3.1 EFD: Early Fire Detection for Effcient Disaster Management ......... 67
3.3.1.1 EFD Architecture................................................................. 68
3.3.1.2 EFD Training and Fine-Tuning............................................ 68
3.3.2 BEA: Balancing Effciency and Accuracy ......................................... 69
3.3.2.1 BEA Framework .................................................................. 70
3.3.3 FLCIE: Fire Localization and Contextual Information Extraction........ 72
3.3.3.1 FLCIE Architecture............................................................. 72
3.3.3.2 Feature Map Selection for Localization .............................. 72
3.3.3.3 Contextual Information Extraction Mechanism .................. 74
3.4 Experimental Results and Discussion ............................................................ 75
3.4.1 Datasets............................................................................................... 75
3.4.2 Experimental Setup ............................................................................ 77
3.4.3 Experimental Results.......................................................................... 78
3.4.3.1 Results and Comparison ...................................................... 78
3.4.3.2 FLCIE Results .....................................................................80
3.4.3.3 Experiments for Contextual Information Extraction........... 81
3.5 Conclusions..................................................................................................... 83
Bibliography ............................................................................................................84

63
64 Deep Learning in Computer Vision

3.1 INTRODUCTION
Recently, a variety of sensors have been introduced for different applications such as
setting off a fre alarm [1], detecting vehicle obstacles [2], visualizing the interior of
the human body for diagnosis [3–5], monitoring animals and ships, and surveillance.
Of these applications, surveillance has primarily attracted the attention of research-
ers due to the enhanced embedded processing capabilities of cameras. Using smart
surveillance systems, various abnormal events such as road accidents, fres, medical
emergencies, etc. can be detected at early stages, and the appropriate authority can
be autonomously informed [6, 7]. A fre is an abnormal event that can cause signif-
cant damage to lives and property within a very short time. Such disasters, the causes
of which include human error and system failures, result in severe loss of human life
and other damage [8]. In June 2013, fre disasters killed 19 frefghters and ruined
100 houses in Arizona. Similarly, another forest fre in August 2013 in California
ruined an area of land the size of 1,042 km2, causing a loss of US$127.35 million.
According to an annual disaster report, fre disasters alone affected 494,000 people
and resulted in a loss of US$3.1 billion in 2015. According to [9], at least 37 people
were killed and about 130 injured in a fre at Sejong Hospital, Miryang, South Korea,
in December 2017, and a fre in Greece killed around 80 people, injured hundreds,
and destroyed about 1,000 homes in July 2018. It is important to detect fres at early
stages utilizing smart surveillance cameras, to avoid such disasters.
Two broad categories of approach can be identifed for fre detection: traditional
fre alarms and vision sensor-assisted fre detection. Traditional fre alarm systems
are based on sensors that require close proximity for activation, such as infrared
and optical sensors. To overcome the limitations of existing approaches, numerous
vision sensor-based methods have been explored by researchers in this feld; these
have many other advantages, e.g., less human interference, faster response, afford-
able cost, and larger surveillance coverage. Despite these advantages, there are still
some issues with these systems, e.g., the complexity of the scenes under observation,
irregular lighting, and low-quality frames; researchers have made several efforts to
address these aspects, taking into consideration both color and motion features. For
instance, [8, 10–17] used color features for fre detection by exploring different color
models including HIS [8], YUV [13], YCbCr [14], RGB [15], and YUC [10]. The
major issue with these methods is their high rate of false alarms. Several attempts
have been made to solve this issue by combing the color information with motion
and analyses of fre’s shape and other characteristics. However, maintaining a well-
balanced trade-off between accuracy, false alarms, and computational effciency still
have remained a challenge. In addition, the above-mentioned methods fail to detect
small fres or fres at larger distances.
In this chapter, two major issues related to fre detection have been investigated:
image representation with feature vectors and feature map selection for fre local-
ization and contextual information extraction. The proposed solutions for image
representation used highly discriminative deep convolutional features to construct
a robust representation. The extensive fne-tuning process enabled the feature
extraction pipeline to focus on fre regions in images, thereby effectively captur-
ing the essential features needed for fre and ignoring irrelevant and trivial or often
Effcient CNNs for Fire Detection 65

misleading features. Further details about feature extraction, feature map selection,
and contextual information extraction are provided in the respective sections of the
chapter.
Section 1.2 presents an overview of image representation techniques for fre detec-
tion based on traditional hand-crafted features and recent deep learning approaches.
Section 1.3 introduces the main contributions of this chapter to fre detection lit-
erature for both indoor and outdoor surveillance. Section 1.4 contains experimental
results on benchmark datasets and evaluations from different perspectives of the pro-
posed methods included in this chapter. Finally, Section 1.5 concludes this chapter
by focusing on the key fndings, strengths, and weaknesses of the proposed methods
along with some future research directions.

3.2 RELATED WORK

The available literature dictates that fre can be detected either by traditional sensors
or by using visible light camera. A broader classifcation of different fre detection
methods is shown in Figure 3.1. Details about these methods are described in sub-
sequent sections.

3.2.1 TRADITIONAL SENSOR-BASED FIRE DETECTION

The recent advancements in technology have resulted in a variety of sensors for
different applications such as wireless capsule sensors for visualization of the inte-
rior of a human body [3], vehicle sensors for obstacle detection [18], and fre alarm
sensors [1]. The current fre alarm sensors such as infrared, ion, and optical sensors
need close proximity of the heat, fre, radiation, or smoke for activation; hence, such
sensors are not considered as good candidates for environments of a critical nature.
Among fre deaths, 78% occurred only due to home fres [19]. One of the main rea-
sons is the delayed escape for disabled people, as traditional fre alarm systems need
strong fres or close proximity, failing to generate an alarm on time for such people.

FIGURE 3.1 General classifcation of fre detection methods.

66 Deep Learning in Computer Vision

This necessitates the existence of effective fre alarm systems for surveillance. The
majority of the research is conducted for fre detection using cameras and visual sen-
sors, whose details are given in the next section.

3.2.2 VISION-BASED FIRE DETECTION

The conventional fre alerting systems are based on optical sensors, needing close
proximity to the fre, thus failing to facilitate in providing additional information,
e.g., burning degree, fre location, and size. In addition to this, such systems involve
much human intervention such as visiting the fre location for confrming the fre
in case of any fre alarm. To cope with these limitations, many visual sensor-based
fre detection systems have been presented. Visual sensor-based fre detection sys-
tems are motivated by several encouraging advantages, including the following: 1)
low cost due to existing setup of installed cameras for surveillance, 2) monitoring
of larger regions, 3) comparatively fast response time due to elimination of waiting
time for heat diffusion, 4) fre confrmation without visiting the fre location, 5) fex-
ibility for detection of smoke and fames by adjustment of certain parameters, and
6) availability of fre details such as size, location, and burning degree. Due to these
characteristics, such systems have attracted the attention of many researchers and
as a result, many fre detection methods have been investigated based on numerous
visual features, achieving reasonable performance with specifc constraints.

3.2.2.1 Color- and Motion-Based Fire Detection Methods

Color is one of the important clues for determining fre or fame in a video frame.
Researchers have presented several methods for detection of fre by exploiting color
features in numerous color models. Chen et al. [8] investigated the dynamic behavior
and irregularity of fames in both RGB and HSI color spaces for fre detection. Since
their method considers the frame difference during prediction, it fails to differenti-
ate real fre from fre-like moving outliers and objects. Another method is presented
by Celik et al. [14], who used YCbCr with specifc rules for separating the chromi-
nance component from luminance. The method has potential to detect fames with
good accuracy but only at short distances and larger fre sizes. Summarizing the
color-based methods, it can be noted that such methods are sensitive to brightness
and shadows. As a result, the number of false warnings produced by these methods
is high. Because of this, such methods may interpret red vehicles or people wear-
ing red clothes as fre due to the predominance of red. Later on, possible solutions
were introduced based on the fact that fre changes its shape continuously, which
can differentiate it from moving rigid objects. Kolesov et al. [18] used optimal mass
transport optical fow as a low-dimensional descriptor along with RGB color chan-
nels and neural networks for differentiating between smoke and a similarly colored
white wall, as well as fre from a similarly colored background.

3.2.2.2 Deep Learning-Based Fire Detection Methods

Deep learning has shown encouraging performance for solving different problems in
the areas of computer vision [20–24], image classifcation [25–28], action recogni-
tion [29, 30], object tracking and detection, medical image analysis [31], and image
Effcient CNNs for Fire Detection 67

indexing and retrieval [32, 33]. Recently, some preliminary results for fre detec-
tion using CNNs have also been reported. It is therefore important to highlight and
analyze their performance in comparison with our proposed works. Frizzi et al. [34]
presented a CNN-based method for fre and smoke detection. This work is based
on a limited number of images, having no comparison with existing methods that
could prove its performance. Sharma et al. [35] explored VGG16 and Resnet50 for
fre detection. The dataset is very small (651 images only) and the reported test-
ing accuracy is less than 93%. This work is compared with that of Frizzi et al. [34]
with testing accuracy of 50%. Another recent work has been presented in Zhong
et al. [36], where a fame detection algorithm based on CNN processed the video
data generated by an ordinary camera monitoring a scene. Firstly, a candidate target
area extraction algorithm was used for dealing with the suspected fame area and
for improving the effciency of recognition. Secondly, the extracted feature maps of
candidate areas were classifed by the designed deep neural network model based
on CNN. Finally, the corresponding alarm signal was obtained by the classifcation
results. Experiments conducted on a homemade database show that this method can
identify fre and achieve reasonable accuracy but with a higher false alarm rate. The
method can process 30 fps on a system with the following specifcations: standard
desktop PC equipped with OctaCore, CPU 3.6 GHz, and 8 GB RAM.
It can be observed that some of the methods are too naïve: their execution time
is fast, but they compromise on accuracy, producing a large number of false alarms.
Conversely, some methods have achieved good fre detection accuracies but their
execution time is too high; hence, they cannot be applied in real-world environments,
especially in critical areas where minor delay can lead to huge disasters. Therefore,
for more accurate and early detection of fre, we need a robust mechanism that can
detect fre during varying conditions and can send the important keyframes and
alerts immediately to disaster management systems.

3.3 CNN-BASED FIRE DETECTION METHODS

The maturity of sensors has resulted in several useful applications for e-health, smarter
surveillance, law enforcement, and disaster management. Among these applications,
disaster management is of a critical nature and has attracted much attention from both
academia and industry. Disaster management is mainly based on smoke/fre detection
systems [37–39], for which researchers have presented both traditional and learned
representation-based methods as described in Section 1.2. In order to avoid fre disas-
ters, it is important to detect fres at early stages utilizing smart surveillance cameras
and analyze the fre scene for collecting further necessary information. For this pur-
pose, three methods [7, 40, 41] are described in this section.

3.3.1 EFD: EARLY FIRE DETECTION FOR EFFICIENT DISASTER MANAGEMENT

Early fre detection in the context of disaster management systems during surveil-
lance of public areas, forests, and nuclear power plants can result in the prevention of
ecological, economic, and social damage. However, early detection is a challenging
problem due to varying lighting conditions, shadows, and movement of fre-colored
68 Deep Learning in Computer Vision

FIGURE 3.2 Early fre detection using CNN with reliable communication for effective
disaster management [7].

objects. Thus, there is a need for an algorithm that can achieve better accuracy in the
aforementioned scenarios while minimizing the number of false alarms. To achieve
this goal, we explored deep CNNs and devised a fne-tuned architecture for early
fre detection during surveillance for effective disaster management. Our system is
overviewed in Figure 3.2.

3.3.1.1 EFD Architecture

The successful use of CNNs for various tasks motivated us to investigate them for
fre detection and fre scene analysis. Thus, a model with similar architecture to the
AlexNet [42] with modifcations according to the problem of fre detection is used
as shown in Figure 3.3. The output of the last fully connected layer is fed into the
Softmax classifer, which computes probabilities for the two classes.

3.3.1.2 EFD Training and Fine-Tuning

As previously discussed, a typical CNN consists of three different neural layers,
i.e., convolutional, pooling, and fully connected layers. All of these layers play dif-
ferent roles in the overall modeling process. The loss cost between the predicted
output computed bypassing input image and labels via network layers is calculated;
this is designed to simplify the fnal output. To select the best model, the data are
trained over various models and the performance of the transfer learning approach
was evaluated; a 4–5% improvement was predicted. In the context of fre detec-
tion, a fne-tuned AlexNet model was used on our fre dataset by modifying the last
fully connected layer and keeping a slower learning rate. Model fne-tuning was
performed for 10 epochs, achieving an improvement of about 5% in classifcation
accuracy, compared to the freshly trained model. After training and the fne-tuning
Effcient CNNs for Fire Detection 69

FIGURE 3.3 Architecture of the proposed CNN [7].

FIGURE 3.4 Sample query images along with their probabilities for CNN-based fre
detection [7].

process, a target model that can be used for prediction of fre at early stages was
achieved. For testing, the query image is passed through the proposed model, which
results in probabilities for both classes: fre and non-fre (normal). Based on the
higher probability, the image is assigned to its appropriate class. Examples of query
images along with their probabilities are shown in Figure 3.4.

3.3.2 BEA: BALANCING EFFICIENCY AND ACCURACY

The research since the last decade has been focused on traditional feature extraction meth-
ods for fame detection. The issues with such methods are their time-consuming process
of feature engineering and their low performance for fame detection. Such methods also
generate high numbers of false alarms, especially in surveillance with shadows, varying
lighting, and fre-colored objects. To cope with such issues, deep learning architectures
have been extensively studied and explored for early fame detection. Motivated by the
70 Deep Learning in Computer Vision

FIGURE 3.5 Early fame detection in surveillance videos using deep CNN [40].

recent improvements in embedded processing capabilities and the potential of deep fea-
tures, numerous CNNs have been investigated to improve fame detection accuracy and
minimize the false warning rate. In this section, we propose a cost-effective fre detec-
tion CNN architecture for surveillance videos. The model is inspired from GoogleNet
architecture, considering its reasonable computational complexity and suitability for
the intended problem compared to other computationally expensive networks such as
“AlexNet.” To balance the effciency and accuracy, the model is fne-tuned considering
the nature of the target problem and fre data. An overview of our framework for fame
detection in CCTV surveillance networks is given in Figure 3.5.

3.3.2.1 BEA Framework

In this section, the architectural details of this method have been described. For the
intended classifcation of images into fre and non-fre, a model similar to GoogleNet
[43] with amendments as per our problem is used. The inspirational reasons for using
GoogleNet rather than other models such as AlexNet include its better classifcation
accuracy, small model size, and suitability of implementation on Field Programmable
Gate Arrays (FPGAs) and other low-memory hardware architectures. The intended
architecture consists of 100 layers with 2 main convolutions, 4 max pooling, one
average pooling, and 7 inception modules as given in Figure 3.6.
Effcient CNNs for Fire Detection 71

FIGURE 3.6 Architectural overview of the proposed deep CNN [40].

The size of the input image is 224×224×3 pixels passed through the proposed
CNN architecture visualized in Figure 3.6. The motivation for using the proposed
architecture is to avoid an uncontrollable increase in the computational complexity
and network fexibility and signifcant increases in the number of units at each stage.
To achieve this, a dimensionality reduction mechanism is applied before computa-
tion-hungry convolutions of patches with larger sizes. Further, the architecture is
modifed according to the problem of fre classifcation by keeping the number of
output classes to 2, i.e., fre and non-fre. The better results for using this architec-
ture over benchmark datasets with prediction scores for fre and non-fre classes are
shown in Figure 3.7.

FIGURE 3.7 Probability scores and predicted labels produced by the proposed deep CNN
framework for different images from benchmark datasets [40].
72 Deep Learning in Computer Vision

3.3.3 FLCIE: FIRE LOCALIZATION AND CONTEXTUAL INFORMATION EXTRACTION

In this section, an energy-friendly and computationally effcient CNN architec-
ture is proposed, inspired by the SqueezeNet [44] architecture for fre detection,
localization, and semantic understanding of the scene of the fre. It uses smaller
convolutional kernels and contains no fully connected layers, which helps keep the
computational requirements to a minimum. Despite its low computational needs, the
proposed solution achieves accuracies that are comparable to other, more complex
models, mainly due to its increased depth. Moreover, this work shows how a trade-
off can be reached between fre detection accuracy and effciency, by considering the
specifc characteristics of the problem of interest and the variety of fre data.

3.3.3.1 FLCIE Architecture

Details of the proposed architecture are described in this section. A model with an
architecture similar to that of SqueezeNet is modifed in accordance with our target
problem of fre detection. The original model was trained on the ImageNet dataset
and is capable of classifying 1,000 different objects. In the fre detection case, how-
ever, this architecture is used to detect fre and non-fre images. This was achieved
by reducing the number of neurons in the fnal layer from 1,000 to 2. Keeping the
rest of the architecture similar to the original was done in an attempt to reuse the
parameters to solve the fre detection problem more effectively. During the experi-
ments, a transfer learning strategy was also explored in an attempt to further improve
the accuracy. Interestingly, an improvement in classifcation accuracy of approxi-
mately 5% for the test data after fne-tuning was achieved. A set of sample images
with their predicted class labels and probability scores is given in Figure 3.8.

3.3.3.2 Feature Map Selection for Localization

In this section, the process of localizing fre regions in video frames is described. To
localize a fre in a sample image, the framework is employed as given in Figure 3.10.

FIGURE 3.8 Prediction scores for a set of query images using the proposed deep CNN [45].
Effcient CNNs for Fire Detection 73

FIGURE 3.9 Sample images and the corresponding localized fre regions using our
approach. The frst row shows the original images, while the second row shows the localized
fre regions [45].

First, a prediction is obtained from our deep CNN. In non-fre cases, no further
action is performed; in the case of fre, further processing of its localization is per-
formed, as given in Algorithms 1 and 2.
After analyzing all the feature maps of the different layers of our proposed CNN
using Algorithm 1, feature maps 8, 26, and 32 of the “Fire2/Concat” layer were found
to be sensitive to fre regions and to be appropriate for fre localization. Therefore,
these three feature maps are fused and binarization is applied to segment the fre
[46]. A set of sample fre images with their segmented regions is given in Figure 3.9.

Algorithm 1. Feature Map Selection Algorithm for Localization

Input: Training samples (TS), ground truth (GT), and the proposed deep CNN model (CNN-M)
1. Forward propagate TS through CNN-M
2. Select the feature maps FN from layer L of CNN-M
3. Resize GT and FN to 256×256 pixels
4. Compute mean activations map FMA(i) for FN
5. Binarize each feature map Fi as follows:

˛1, F ( x, y)i > FMA(i )

F ( x, y)bin(i ) = ˝
˙0, Otherwise

6. Calculate the hamming distance HDi between GT and each feature map Fbin(i) as follows:

HDi = Fbin(i ) − GT

This results in TS × FN hamming distances

7. Calculate the sum of all resultant hamming distances, and shortlist the minimum hamming
distances using threshold T
8. Select appropriate feature maps according to the shortlisted hamming distances
Output: Feature maps sensitive to fre
74 Deep Learning in Computer Vision

3.3.3.3 Contextual Information Extraction Mechanism

The fre detection methods presented so far in the literature can either detect fre or
localize fre regions with reasonable accuracy. To the best of our knowledge, there
is no published work that can extract contextual information from the fre scenes.
“Contextual information” refers to anything that is of interest to fre brigades and
disaster management systems. Whenever a person reports to a fre brigade, the fol-
lowing information is collected by them:

1. What is on fre?
2. Are there any people in the building?
3. What type of building is on fre?
4. Are there any hazards in or around the building?
5. Is evacuation in progress?
6. What is the address of the premises?
7. What is your phone number?
8. Other information as relevant.

Algorithm 2. Fire Localization Algorithm

Input: Image I of the video sequence and the proposed deep CNN model (CNN-M)
1.  Select a frame from the video sequence and forward propagate it through CNN-M
2.  IF predicted label = non-fre THEN
No action
ELSE
a)  Extract feature maps 8, 26, and 32 (F8, F26, F32) from the “Fire2/Concat” layer of CNN-M
b)  Calculate mean activations map (FMA) for F8, F26, and F32
c)  Apply binarization on FMA through threshold T as follows:

˛1, FMA > T

FLocalize = ˝
˙0, Otherwise
d)  Segment fre regions from FMA
END
Output: Binary image with segmented fre ILocalize

The proposed method can extract similar contextual information as well as more
useful information (such as size and expansion of fre) automatically by processing
the video stream of surveillance cameras. To achieve such information, the seg-
mented fre regions of Algorithm 2 are further processed to determine the severity
level/burning degree of the scene under observation and fnd the zone of infuence
(ZOI) from the input fre image. The burning degree can be determined from the
number of pixels in the segmented fre.
The zone of infuence can be calculated by subtracting the segmented fre regions
from the original input image. The resultant zone of infuence image is then passed
from the original SqueezeNet model [44], which predicts its label from 1,000 objects.
The object information can be used to determine the situation in the scene, such as a
fre in a house, a forest, or a vehicle. This information along with the severity of the
fre can be reported to the fre brigade to take appropriate action.
Effcient CNNs for Fire Detection 75

FIGURE 3.10 Fire localization and contextual information extraction using the proposed
deep CNN [45].

3.4 EXPERIMENTAL RESULTS AND DISCUSSION

In this section, the proposed methods (EFD, BEA, and FLCIE) are evaluated both
qualitatively and quantitatively on challenging benchmark datasets. Two sets of eval-
uation metrics are used for the evaluation of each method. Performance and various
other aspects of the presented methods are highlighted in the relevant sections.

3.4.1 DATASETS
This section presents the benchmark datasets used for experimental results and eval-
uation. The main focus during experiments is on two famous benchmark dataset:
Foggia’s video dataset [10] and Chino’s dataset [47]. The frst dataset contains a total
of 31 videos consisting of indoor as well outdoor scenarios, where 14 videos belong
76 Deep Learning in Computer Vision

to the fre class and the remaining 17 videos belong to the non-fre class. The main
motivation for selecting this dataset is that it contains videos recorded in both indoor
and outdoor environments. Furthermore, this dataset is very challenging because
of the fre-like objects in the non-fre videos, which can be predicated as fre, mak-
ing the classifcation more diffcult. To this end, color-based methods may fail to
differentiate between real fre and scenes with red objects. Similarly, motion-based
techniques may wrongly classify a scene with mountains containing smoke, cloud,
or fog. These compositions made the dataset more challenging, enabling us to stress
our framework and investigate its performance in various situations of the real envi-
ronment. Representative frames of these videos are shown in Figure 3.11.
The second dataset (DS2) [47] is comparatively small but very challenging.
There are 226 images in this dataset, out of which 119 images contain fre while

FIGURE 3.11 Sample images from DS1 videos, containing fre and without fres [7].
Effcient CNNs for Fire Detection 77

FIGURE 3.12 Sample images from dataset DS2.

the remaining 107 are fre-like images containing sunsets, fre-like lights, sunlight
coming through windows, etc. A set of selected images from this dataset is shown
in Figure 3.12.

3.4.2 EXPERIMENTAL SETUP

The models used in the proposed methods were trained using a system with specif-
cations as follows: Intel Core i5 CPU equipped with 64 GB RAM with Ubuntu OS,
NVidia GeForce GTX TITAN X (Pascal) having 12 GB onboard memory, and Caffe
78 Deep Learning in Computer Vision

deep learning framework [48]. The rest of the experiments were conducted using
MATLAB R2015a with a Core i5 system containing 8 GB RAM.
Furthermore, two different sets of evaluation metrics were employed to evalu-
ate the performance of each method from all perspectives. The frst set of metrics
contains true positives (accuracy), false positives, and false negatives as used in
Muhammad et al. [49]. Dividing the number of images identifed as fre by the sys-
tem by the number of actual fre images provides the “true positive” (TP) rate. “False
negatives” (FN) are calculated by dividing the number of fre images identifed as
non-fre by the system by the number of actual fre images. The “false positives”
(FP) or false alarm rate is achieved by dividing the number of non-fre images pre-
dicted as fre by the system by the total number of non-fre images. For better evalu-
ation of the performance, another set of metrics including precision [4], recall, and
F-measure [50] is also employed. These metrics are computed as follows:

TP
Precision = (3.1)
(TP + FP)

TP
Recall = (3.2)
(TP + FN)

(Precision * Recall)
F − Measure = 2 * (3.3)
(Precision + Recall)

3.4.3 EXPERIMENTAL RESULTS

This section explains in detail the experiments conducted for the performance evalu-
ation of our proposed methods, i.e., EFD, BEA, and FLCIE. First, experiments per-
formed on various datasets of literature are provided and comparisons are made
of all the proposed works with state-of-the-art methods. Next, the results of fre
localization are compared with the existing methods. Finally, the experiments for
contextual information extraction from the object that is on fre are presented.

3.4.3.1 Results and Comparison

For the training and testing phases of the experiments, the experimental strategy of
Foggia et al. [10] are followed, where 20% and 80% of the data are used for training
and testing, respectively. Using this strategy, we trained our proposed three methods,
i.e., EFD, BEA, and FLCIE, as discussed in Section 1.3. The collected experimental
results using Dataset1 are given in Table 3.1.
For the selection of existing literature for assessment, a criterion was employed
based on the features used in the related works, years of publication, and the datasets
under consideration. Then, the proposed work was compared with the selected fre
detection algorithms, as shown in Table 3.1. The results show that Celik et al. [14]
and Foggia et al. [10] are the best algorithms in terms of false negatives. However,
the results of these methods are not impressive in terms of the other metrics of false
positives and accuracy. From the perspective of false positives, the algorithm of
Effcient CNNs for Fire Detection 79

TABLE 3.1
Comparison of Various Fire Detection Methods for Dataset1
Technique False Positives (%) False Negatives (%) Accuracy (%)

FLCIE after FT 8.87 2.12 94.50

FLCIE before FT 9.99 10.39 89.8
BEA after FT 0.054 1.5 94.43
BEA before FT 0.11 5.5 88.41
EFD after FT 9.07 2.13 94.39
EFD before FT 9.22 10.65 90.06
Foggia et al. [10] 11.67 0 93.55
De Lascio et al. [51] 13.33 0 92.86
Habibuglu et al. [16] 5.88 14.29 90.32
Rafee et al. (RGB) [15] 41.18 7.14 74.20
Rafee et al. (YUV) [15] 17.65 7.14 87.10
Celik et al. [14] 29.41 0 83.87
Chen et al. [8] 11.76 14.29 87.10

Habibuglu et al. [16] performs best and dominates the other methods. However, its
false-negative rate is 14.29%, the worst result of all the methods examined. The
accuracy of the other four methods is also better than this method, with the most
recent method [10] being the best. However, the false positive score of 11.67% is
still high, and the accuracy could be further improved. To achieve a high accuracy
and a low false positive rate, the use of deep features for fre detection is explored.
For this reason, three different methods, i.e., EFD, BEA, and FLCIE, are proposed.
In EFD, frst we used the AlexNet architecture without fne-tuning, which resulted
in an accuracy of 90.06% and reduced false positives from 11.67% to 9.22%. In
the baseline AlexNet architecture, the weights of kernels are initialized randomly
and are modifed during the training process considering the error rate and accu-
racy. The strategy of transfer learning [52] is also applied, whereby the weights from
a pre-trained AlexNet model are initialized with a low learning rate of 0.001 and
modifed the last fully connected layer according to the problem of fre detection.
Interestingly, an improvement in accuracy of 4.33% and reductions in false negatives
and false positives of up to 8.52% and 0.15%, respectively, are obtained.
The false alarm score is still high and needs further improvement. Therefore,
deep learning architecture GoogleNet is explored for this purpose in BEA. Initially,
a GoogleNet model is trained with its default kernel weights, which resulted in an
accuracy of 88.41% and a false positive score of 0.11%. The baseline GoogleNet
architecture randomly initializes the kernel weights, which are tuned according to
the accuracy and error rate during the training process. In an attempt to improve
the accuracy, transfer learning [52] by initializing the weights from pre-trained
GoogleNet model and keeping the learning rate threshold to 0.001 is explored.
Further, the last fully connected layer as per the nature of the intended problem is
changed. With this fne-tuning process, the false alarm rate is reduced from 0.11% to
0.054% and the false-negative score from 5.5% to 1.5%.
80 Deep Learning in Computer Vision

TABLE 3.2
Comparison of Different Fire Detection Methods for Dataset2
Technique Precision Recall F-Measure

FLCIE After FT 0.86 0.97 0.91

Before FT 0.84 0.87 0.85
BEA After FT 0.80 0.93 0.86
Before FT 0.86 0.89 0.88
EFD After FT 0.82 0.98 0.89
Before FT 0.85 0.92 0.88
Chino et al. (BoWFire) [47] 0.51 0.65 0.57
Rudz et al. [53] 0.63 0.45 0.52
Rossi et al. [54] 0.39 0.22 0.28
Celik et al. [14] 0.55 0.54 0.54
Chen et al. [8] 0.75 0.15 0.25

Although the results of EFD and BEA are good compared to other existing meth-
ods, there are still certain limitations. Firstly, the size of this model is comparatively
large, thereby restricting its implementation in CCTV networks. Secondly, the accu-
racy is still low and would be problematic for fre brigades and disaster management
teams.
With these strong motivations, SqueezeNet, a lightweight architecture, was
explored for this problem in FLCIE. The experiments for this new architecture were
repeated, and an improvement of 0.11% in accuracy was achieved. Furthermore, the
rate of false alarms was reduced from 9.07% to 8.87%. The rate of false negatives
remained almost the same. Finally, the major achievement of the proposed frame-
work was the reduction of the model size from 238 MB to 3 MB, thus saving an extra
235 MB, which can greatly minimize the cost of CCTV surveillance systems.
For further experimentation, Dataset2 was considered and compared with the
results from four other fre detection algorithms in terms of their relevancy, datasets,
and years of publication as shown in Table 3.2.
To ensure a fair evaluation and a full overview of the performance of our
approach, we considered another set of metrics (precision, recall, and F-measure)
as used by Chino et al. [47]. In a similar way to the experiments on Dataset1, we
tested Dataset2 using EFD, BEA, and FLCIE. For the fne-tuned models in EFD
and BEA, F-measure scores of 0.89 and 0.86, respectively, were achieved. Further
improvement was achieved using FLCIE, increasing the F-measure score to 0.91 and
the precision to 0.86. It is evident from Table 3.2 that FLCIE achieved better results
than the state-of-the-art methods, confrming the effectiveness of the proposed deep
CNN framework.

3.4.3.2 FLCIE Results

In this section, the proposed approach (FLCIE) is assessed with respect to fre local-
ization. The true positive and false positive rates are computed in order to have a
better and unbiased evaluation of our proposed fre localization module. The feature
Effcient CNNs for Fire Detection 81

FIGURE 3.13 Comparison of the CNNFire approach with other methods [45].

maps from different convolutional and pooling layers of our proposed CNN archi-
tecture are used to localize fre in an image precisely. Next, the number of over-
lapping fre pixels in the detection maps and ground truth images are used as true
positives. Similarly, the number of non-overlapping fre pixels in the detection maps
are also determined and then interpreted as false positives. The localization results
are compared with those of several state-of-the-art methods such as Chen et al. [8],
Celik et al. [14], Rossi et al. [54], Rudz et al. [53], and Chino et al. (BoWFire) [47],
as shown in Figure 3.13. Figures 3.14 and 3.15 show the results of all methods for
a sample image from Dataset2. The results of BoWFire, color classifcation, Celik
et al., and Rudz et al. are almost the same. Rossi et al. gives the worst results in this
case, and Chen et al. is better than Rossi et al. The results from CNNFire are similar
to the ground truth.
3.4.3.3 Experiments for Contextual Information Extraction
Along with fre detection and localization, the proposed approach is able to determine
the intensity of the fre detected in an input image accurately. To achieve this target,
the zone of infuence (ZOI) inside the input image along with the segmented fre
regions is extracted. The ZOI image is then advanced to the pre-trained SqueezeNet
model on the ImageNet dataset comprising 1,000 classes. The label assigned by the
SqueezeNet model to the ZOI image is then combined with the severity of the fre
for reporting to the fre brigade. A set of sample cases from this experiment is given
in Figure 3.16.
In Figure 3.16, the frst column shows input images with labels predicted by the
CNN model and their probabilities, with the highest probability taken as the fnal
class label; the second column shows three feature maps (F8, F26, and F32) selected
by EFD; the third column highlights the results for each image using BEA; the
fourth column shows the severity of the fre and ZOI images with a label assigned
82 Deep Learning in Computer Vision

FIGURE 3.14 Visual fre localization results of our CNNFire approach and other fre local-
ization methods [45].

FIGURE 3.15 Fire localization results from our CNNFire and other schemes with false
positives [45].
Effcient CNNs for Fire Detection 83

FIGURE 3.16 Sample outputs from the overall system [45].

by the SqueezeNet model; and the fnal column shows the alert that should be sent to
emergency services, such as the fre brigade.

3.5 CONCLUSIONS
Recently, convolutional neural networks (CNNs) have shown substantial progress in
surveillance applications; however, their key roles in early fre detection, localiza-
tion, and fre scene analysis have not been investigated. With this motivation, CNNs
were investigated for addressing the aforementioned problems in this chapter for fre
detection and localization, thereby achieving better detection accuracy with a mini-
mum false alarm rate compared to existing methods. Several highly effcient and
intelligent methods that selected discriminative deep features for fre scene analysis
in both indoor and outdoor surveillance have been presented.
In this chapter, a total of three CNN-based frameworks have been proposed. The
frst framework was proposed for early fre detection using CNNs in surveillance
84 Deep Learning in Computer Vision

applications. A cost-effective fre detection CNN architecture for surveillance videos

was described in the second framework. Finally, an effcient deep CNN-based fre
detection system was proposed for fre detection and localization. In the future, we
plan to study the feature maps of other recent and lightweight architectures (e.g.,
ShuffeNet) for localizing fre regions in video frames, helping determining fre
severity and burning degree. Further, we will also explore its possibility for extract-
ing semantic information about the fre scene as well as its integration with smoke
detection systems for detailed surveillance of sensitive areas.

BIBLIOGRAPHY
1. B. C. Ko, K.-H. Cheong, and J.-Y. Nam, “Fire detection based on vision sensor and sup-
port vector machines,” Fire Safety Journal, vol. 44, pp. 322–329, 2009.
2. V. D. Nguyen, H. Van Nguyen, D. T. Tran, S. J. Lee, and J. W. Jeon, “Learning frame-
work for robust obstacle detection, recognition, and tracking,” IEEE Transactions on
Intelligent Transportation Systems, vol. 18, pp. 1633–1646, 2017.
3. I. Mehmood, M. Sajjad, and S. W. Baik, “Mobile-cloud assisted video summarization
framework for effcient management of remote sensing data generated by wireless cap-
sule sensors,” Sensors, vol. 14, pp. 17112–17145, 2014.
4. K. Muhammad, M. Sajjad, M. Y. Lee, and S. W. Baik, “Effcient visual attention driven
framework for key frames extraction from hysteroscopy videos,” Biomedical Signal
Processing and Control, vol. 33, pp. 161–168, 2017.
5. R. Hamza, K. Muhammad, Z. Lv, and F. Titouna, “Secure video summarization frame-
work for personalized wireless capsule endoscopy,” Pervasive and Mobile Computing,
vol. 41, pp. 436–450, 2017/10/01/ 2017.
6. K. Muhammad, R. Hamza, J. Ahmad, J. Lloret, H. H. G. Wang, and S. W. Baik, “Secure
surveillance framework for IoT systems using probabilistic image encryption,” IEEE
Transactions on Industrial Informatics, vol. 14(8), pp. 3679–3689, 2018.
7. K. Muhammad, J. Ahmad, and S. W. Baik, “Early fre detection using convolutional neu-
ral networks during surveillance for effective disaster management,” Neurocomputing,
vol. 288, pp. 30–42, 2018.
8. C. Thou-Ho, W. Ping-Hsueh, and C. Yung-Chuen, “An early fre-detection method
based on image processing,” in 2004 International Conference on Image Processing,
2004. ICIP ‘04, vol. 3, 2004, pp. 1707–1710.
9. http://www.bbc.com/news/world-asia-42828023 (Visited 31 January, 2018, 9 AM).
10. P. Foggia, A. Saggese, and M. Vento, “Real-time fre detection for video-surveillance applica-
tions using a combination of experts based on color, shape, and motion,” IEEE Transactions
on Circuits and Systems for Video Technology, vol. 25(9), pp. 1545–1556, 2015.
11. B. U. Töreyin, Y. Dedeoğlu, U. Güdükbay, and A. E. Cetin, “Computer vision based
method for real-time fre and fame detection,” Pattern Recognition Letters, vol. 27(1),
pp. 49–58, 2006.
12. D. Han, and B. Lee, “Development of early tunnel fre detection algorithm using
the image processing,” in International Symposium on Visual Computing, 2006, pp.
39–48.
13. G. Marbach, M. Loepfe, and T. Brupbacher, “An image processing technique for fre
detection in video images,” Fire Safety Journal, vol. 41(4), pp. 285–289, 2006.
14. T. Celik, and H. Demirel, “Fire detection in video sequences using a generic color
model,” Fire Safety Journal, vol. 44(2), pp. 147–158, 2009.
15. A. Rafee, R. Dianat, M. Jamshidi, R. Tavakoli, and S. Abbaspour, “Fire and smoke
detection using wavelet analysis and disorder characteristics,” in 2011 3rd International
Conference on Computer Research and Development, 2011, pp. 262–265.
Effcient CNNs for Fire Detection 85

16. Y. H. Habiboğlu, O. Günay, and A. E. Çetin, “Covariance matrix-based fre and fame
detection method in video,” Machine Vision and Applications, vol. 23(6), pp. 1103–
1113, 2012.
17. A. Sorbara, E. Zereik, M. Bibuli, G. Bruzzone, and M. Caccia, “Low cost optronic
obstacle detection sensor for unmanned surface vehicles,” in 2015 IEEE Sensors
Applications Symposium (SAS), 2015, pp. 1–6.
18. I. Kolesov, P. Karasev, A. Tannenbaum, and E. Haber, “Fire and smoke detection in
video with optimal mass transport based optical fow and neural networks,” in 2010
IEEE International Conference on Image Processing, 2010, pp. 761–764.
19. H. J. G. Haynes, “Fire loss in the United States during 2015,” http://www.nfpa.org/, 2016.
20. K. Muhammad, T. Hussain, and S. W. Baik, “Effcient CNN based summarization of
surveillance videos for resource-constrained devices,” Pattern Recognition Letters, 2018.
21. M. Sajjad, S. Khan, T. Hussain, K. Muhammad, A. K. Sangaiah, A. Castiglione, et
al., “CNN-based anti-spoofng two-tier multi-factor authentication system,” Pattern
Recognition Letters, 126(2019): 123–131.
22. M. Hassaballah, A. A. Abdelmgeid, and H. A. Alshazly, “Image features detection,
description and matching,” in Image Feature Detectors and Descriptors, Awad, A. I.,
Hassaballah, M., (Eds.): Springer, 2016, pp. 11–45.
23. A. I. Awad, and M. Hassaballah, Image Feature Detectors and Descriptors. Studies in
Computational Intelligence. Springer International Publishing, Cham, 2016.
24. F. U. M. Ullah, A. Ullah, K. Muhammad, I. U. Haq, and S. W. Baik, “Violence detec-
tion using spatiotemporal features with 3D Convolutional Neural Network,” Sensors,
vol. 19(11), p. 2472, 2019.
25. M. Sajjad, S. Khan, Z. Jan, K. Muhammad, H. Moon, J. T. Kwak, et al., “Leukocytes
classifcation and segmentation in microscopic blood smear: A resource-aware health-
care service in smart cities,” IEEE Access, vol. 5, pp. 3475–3489, 2017.
26. I. U. Haq, K. Muhammad, A. Ullah, and S. W. Baik, “DeepStar: Detecting starring
characters in movies,” IEEE Access, 7(2019): 9265–9272.
27. M. Hassaballah, H. A. Alshazly, and A. A. Ali, “Ear recognition using local binary pat-
terns: A comparative experimental study,” Expert Systems with Applications, vol. 118,
pp. 182–200, 2019.
28. A. I. Awad, and K. Baba, “Singular point detection for effcient fngerprint classifca-
tion,” International Journal on New Computer Architectures and Their Applications
(IJNCAA), vol. 2, pp. 1–7, 2012.
29. A. Ullah, K. Muhammad, J. D. Ser, S. W. Baik, and V. Albuquerque, “Activity recogni-
tion using temporal optical fow convolutional features and multi-layer LSTM,” IEEE
Transactions on Industrial Electronics, vol. 66(12), pp. 9692–9702, 2019.
30. A. Ullah, J. Ahmad, K. Muhammad, M. Sajjad, and S. W. Baik, “Action recognition in
video sequences using deep Bi-directional LSTM with CNN features,” IEEE Access,
vol. 6, pp. 1155–1166, 2018.
31. M. Sajjad, S. Khan, K. Muhammad, W. Wu, A. Ullah, and S. W. Baik, “Multi-grade
brain tumor classifcation using deep CNN with extensive data augmentation,” Journal
of Computational Science, vol. 30, pp. 174–182, 2019.
32. J. Ahmad, K. Muhammad, J. Lloret, and S. W. Baik, “Effcient conversion of deep fea-
tures to compact binary codes using Fourier decomposition for multimedia big data,”
IEEE Transactions on Industrial Informatics, vol. 14(7), pp. 3205–3215, 2018.
33. J. Ahmad, K. Muhammad, S. Bakshi, and S. W. Baik, “Object-oriented convolu-
tional features for fne-grained image retrieval in large surveillance datasets,” Future
Generation Computer Systems, vol. 81, pp. 314–330, 2018.
86 Deep Learning in Computer Vision

34. S. Frizzi, R. Kaabi, M. Bouchouicha, J. Ginoux, E. Moreau, and F. Fnaiech,

“Convolutional neural network for video fre and smoke detection,” in IECON 2016 -
42nd Annual Conference of the IEEE Industrial Electronics Society, 2016, pp. 877–882.
35. J. Sharma, O.-C. Granmo, M. Goodwin, and J. T. Fidje, “Deep convolutional neural
networks for fre detection in images,” in International Conference on Engineering
Applications of Neural Networks, 2017, pp. 183–193.
36. Z. Zhong, M. Wang, Y. Shi, and W. Gao, “A convolutional neural network-based fame
detection method in video sequence,” Signal, Image and Video Processing, vol. 12(8), pp.
1619–1627, 2018.
37. K. Muhammad, S. Khan, M. Elhoseny, S. H. Ahmed, and S. W. Baik, “Effcient fre
detection for uncertain surveillance environment,” IEEE Transactions on Industrial
Informatics, vol. 15(5), pp. 3113–3122, 2019.
38. S. Khan, K. Muhammad, S. Mumtaz, S. W. Baik, and V. H. C. d. Albuquerque, “Energy-
effcient deep CNN for smoke detection in foggy IoT environment,” IEEE Internet of
Things Journal, vol. 6(6), pp. 9237–9245, 2019.
39. K. Muhammad, S. Khan, V. Palade, I. Mehmood, and V. H. C. D. Albuquerque, “Edge
intelligence-assisted smoke detection in foggy surveillance environments,” IEEE
Transactions on Industrial Informatics, pp. 1–1, 2019.
40. K. Muhammad, J. Ahmad, I. Mehmood, S. Rho, and S. W. Baik, “Convolutional neural
networks based fre detection in surveillance videos,” IEEE Access, vol. 6, pp. 18174–
18183, 2018.
41. K. Muhammad, J. Ahmad, Z. Lv, P. Bellavista, P. Yang, and S. W. Baik, “Effcient deep
CNN-based fre detection and localization in video surveillance applications,” IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 49(7), pp. 1419–1434,
2018.
42. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifcation with deep con-
volutional neural networks,” Advances in Neural Information Processing Systems, pp.
1097–1105, 2012.
43. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, et al., “Going deeper
with convolutions,” in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2015, pp. 1–9.
44. F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer,
“SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 1MB model
size,” arXiv preprint arXiv:1602.07360, 2016.
45. K. Muhammad, J. Ahmad, Z. Lv, P. Bellavista, P. Yang, and S. W. Baik, “Effcient deep
CNN-based fre detection and localization in video surveillance applications,” IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 49(7), pp. 1419–1434, 2018.
46. K. Wattanachote, and T. K. Shih, “Automatic dynamic texture transformation based on
a new motion coherence metric,” IEEE Transactions on Circuits and Systems for Video
Technology, vol. 26(10), pp. 1805–1820, 2015.
47. D. Y. Chino, L. P. Avalhais, J. F. Rodrigues, and A. J. Traina, “BoWFire: Detection
of fre in still images by integrating pixel color and texture analysis,” in 2015 28th
SIBGRAPI Conference on Graphics, Patterns and Images, 2015, pp. 95–102.
48. Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, et al., “Caffe:
Convolutional architecture for fast feature embedding,” in Proceedings of the 22nd
ACM International Conference on Multimedia, 2014, pp. 675–678.
49. K. Muhammad, J. Ahmad, Z. Lv, P. Bellavista, P. Yang, and S. W. Baik, “Effcient deep
CNN-based fre detection and localization in video surveillance applications,” IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 49(7), pp. 1419–1434,
2018.
Effcient CNNs for Fire Detection 87

50. K. Muhammad, J. Ahmad, M. Sajjad, and S. W. Baik, “Visual saliency models for sum-
marization of diagnostic hysteroscopy videos in healthcare systems,” SpringerPlus,
vol. 5(1), p. 1495, 2016.
51. R. Di Lascio, A. Greco, A. Saggese, and M. Vento, “Improving fre detection reliability
by a combination of videoanalytics,” in International Conference Image Analysis and
Recognition, 2014, pp. 477–484.
52. S. J. Pan, and Q. Yang, “A survey on transfer learning,” IEEE Transactions on
Knowledge and Data Engineering, vol. 22(10), pp. 1345–1359, 2010.
53. S. Rudz, K. Chetehouna, A. Hafane, H. Laurent, and O. Séro-Guillaume, “Investigation
of a novel image segmentation method dedicated to forest fre applicationsThis paper is
dedicated to the memory of Dr Olivier Séro-Guillaume (1950–2013), CNRS Research
Director,” Measurement Science and Technology, vol. 24(7), p. 075403, 2013.
54. L. Rossi, M. Akhlouf, and Y. Tison, “On the use of stereovision to develop a novel
instrumentation system to extract geometric fre fronts characteristics,” Fire Safety
Journal, vol. 46(1–2), pp. 9–20, 2011.
 4 A Multi-biometric Face
Recognition System
Based on Multimodal
Deep Learning
Representations
Alaa S. Al-Waisy, Shumoos Al-Fahdawi,
and Rami Qahwaji

CONTENTS
4.1 Introduction ....................................................................................................90
4.2 Related Work ..................................................................................................92
4.2.1 Handcrafted-Descriptor Approaches.................................................. 93
4.2.2 Learning-Based Approaches ..............................................................94
4.3 The Proposed Approaches.............................................................................. 95
4.3.1 Fractal Dimension............................................................................... 95
4.3.2 Restricted Boltzmann Machine ..........................................................96
4.3.3 Deep Belief Networks......................................................................... 98
4.3.4 Convolutional Neural Network...........................................................99
4.4 The Proposed Methodology.......................................................................... 100
4.4.1 The Proposed FDT-DRBM Approach.............................................. 101
4.4.1.1 Fractal Dimension Transformation.................................... 102
4.4.1.2 Discriminative Restricted Boltzmann Machines............... 104
4.4.2 The Proposed FaceDeepNet System................................................. 105
4.5 Experimental Results.................................................................................... 106
4.5.1 Description of Face Datasets ............................................................ 107
4.5.2 Face Identifcation Experiments ....................................................... 108
4.5.2.1 Parameter Settings of the FDT-DRBM Approach............. 109
4.5.2.2 Testing the Generalization Ability of the
Proposed Approaches ........................................................ 114
4.5.3 Face Verifcation Experiments.......................................................... 117
4.6 Conclusions................................................................................................... 119
Bibliography .......................................................................................................... 121

89
90 Deep Learning in Computer Vision

4.1 INTRODUCTION
A reliable and effcient identity administration system is a vital component in many
governmental and civilian high-security applications that provide services to only
legitimately registered users [1]. Many pattern recognition systems have been used
for establishing the identity of an individual based on different biometric traits (e.g.,
physiological, behavioral, and/or soft biometric traits) [2–3]. In a biometric system,
these traits establish a substantial and strong connection between the users and
their identities [4]. Face recognition has received considerably more attention in the
research community than other biometric traits, due to its wide range of commercial
and governmental applications, its low cost, and the ease of capturing the face image
in a non-intrusive manner. This is unlike other biometric systems (e.g., fngerprint
recognition), which can cause some health risks by transferring viruses and/or some
infectious diseases from one user to the other by using the same sensor device to
capture the biometric trait from multiple users [5]. Despite these advantages, design-
ing and implementing a face recognition system is considered a challenging task in
the image processing, computer vision, and pattern recognition felds. Therefore,
different factors and problems need to be addressed when developing any biometric
system based on the face image, especially when face images are taken in uncon-
strained environments. Such environments are challenging due to the large intraper-
sonal variations, such as changes in facial expression, illumination, multiple views,
aging, occlusion from wearing glasses and hats, and the small interpersonal differ-
ences [6]. Basically, the performance of the face recognition system depends on two
fundamental stages: feature extraction and classifcation. The correct identifcation
of a person’s identity in the latter stage is dependent on the discriminative power of
the extracted facial features in the former stage. Thus, extracting and learning pow-
erful and highly discriminating facial features to minimize intrapersonal variations
and interpersonal similarities remains a challenging task [7].
Recently, considerable attention has been paid to employing multi-biometric
systems in many governmental and private sectors due to their ability to sig-
nifcantly improve the recognition performance of biometric systems. Moreover,
multimodal systems have the following advantages [8–11]: (i) improving popula-
tion coverage; (ii) improving the biometric system’s throughput; (iii) deterring
spoofng attacks; (iv) maximizing the interpersonal similarities and minimizing
the intrapersonal variations; and (v) providing a high degree of fexibility allow-
ing people to choose to provide either a subset or all of their biometric traits
depending on the nature of the implemented application and the user’s conve-
nience. As illustrated in Figure 4.1, to satisfy the multi-biometric system concept,
the face biometric trait can be engaged in four out of fve different types of multi-
biometric systems.
In this chapter, a novel multi-biometric system for identifying a person’s iden-
tity using two discriminative deep learning approaches is proposed based on the
combination of a convolutional neural network (CNN) and deep belief network
(DBN) to address the problem of unconstrained face recognition. CNN is one of
the most powerful supervised deep neural networks (DNNs), which is widely used
to resolve many tasks in image processing, computer vision, and pattern recognition
with high ability to automatically extract discriminative features from input images.
A Multi-biometric Face Recognition System 91

FIGURE 4.1 The fve different types of a multimodal biometric system, adapted from Ross
and Jain [10].

Furthermore, CNN is considered to be invariant to image deformations, such as

translation, rotation, and scaling [12]. On the other hand, DBN, which is one of the
most popular unsupervised DNNs, has been effectively applied to learn hierarchi-
cal feature representations from unlabeled data in a wide range of felds, including
face recognition [13], speech recognition [14], audio classifcation [15], and natu-
ral language understanding [16]. Finally, both CNNs and DBNs have a number of
advantages compared to other DNNs, for instance simpler architecture, adaptability,
fast convergence, and fewer free parameters. The main contributions of the work
presented here can be summarized as follows:

1. A novel local facial feature-based approach based on a combination of the

fractal dimension transformation (FDT) and the discriminative restricted
Boltzmann machines (DRBM) as a nonlinear classifer to effciently model
the joint distribution of inputs and target classes is proposed and termed
the FDT-DRBM approach. The proposed FDT-DRBM approach has man-
aged to produce an illumination-invariant representation of the face image.
It also managed to effciently extract the local facial features (e.g., mouth,
nose, eyes, and eyebrows) along with the face texture roughness and fuc-
tuations in the surface effciently by exploiting the fractal dimension prop-
erties, such as self-similarity.
92 Deep Learning in Computer Vision

2. An effcient and real-time multi-biometric face recognition system based

on building deep learning representations and fusing the results obtained
from two discriminative deep learning approaches (i.e., CNN and DBN) is
proposed and termed the FaceDeepNet system. To the best of our knowl-
edge, this is the frst work that investigates the potential use of the CNN and
the DBN together for fusing the face-matching scores in a unifed multi-
biometric system. In this chapter, a parallel architecture is considered,
which allows users a high degree of fexibility to use different approaches
for encoding the facial features depending on the required security level
and the users’ convenience.
3. Discriminative training schemes equipped with a number of training strate-
gies (e.g., Dropout method, AdaGrad algorithm, Data augmentation, etc.)
are also employed in order to prevent overftting during the learning process
and increase the generalization ability of the proposed DNNs.
4. Another contribution of this work is to further examine and validate the
current fndings for improving the generalization ability and reducing the
computational complexity of the DNNs (e.g., CNN and DBN) by assign-
ing the handcrafted features (e.g., feature representations obtained from
the FDT approach) to the visible units of the DNNs instead of the pixel
intensity representations to learn additional and complementary representa-
tions. We demonstrate that the proposed FaceDeepNet system can represent
large-scale face datasets, with less time required to obtain the fnal trained
model compared to the direct use of the raw data. In addition, the proposed
FaceDeepNet system is able to effciently handle nonlinear variations of
face images and is unlikely to overft the training data due to the high non-
linearity of the proposed DNNs.
5. The performance of the proposed approaches has been tested and evaluated
in both face identifcation and verifcation tasks by conducting a number
of extensive experiments on four large-scale unconstrained face datasets:
SDUMLA-HMT, FRGC V 2.0, UFI, and LFW. Recognition rates compa-
rable to those of the state-of-the-art methods have been achieved using only
the proposed FDT-DRBM approach. New state-of-the-art recognition rates
were achieved using the FaceDeepNet system on all the employed datasets,
in particular when it is trained on top of the output feature representations
of the FDT approach.

4.2 RELATED WORK

Generally, a number of approaches have been proposed and developed to overcome
all/some of the aforementioned drawbacks and problems in the face recognition sys-
tem. However, very few of them were capable of working in fully unconstrained
environments. These approaches can be broadly divided into two categories: hand-
crafted-descriptor approaches, and learning-based approaches [17]. The next subsec-
tions are devoted to outlining these approaches and their usage to address the face
recognition problem, before discussing their respective strengths and limitations.
A Multi-biometric Face Recognition System 93

4.2.1 HANDCRAFTED-DESCRIPTOR APPROACHES

The majority of previous face recognition systems depend on feature represen-
tations produced by local/global handcrafted-descriptor approaches, such as
scale-invariant feature transform (SIFT), local binary patterns (LBP), 2D Gabor
Wavelet, etc. Handcrafted-descriptor approaches can be further divided into three
groups: feature-based, holistic-based, and hybrid-based approaches. In the frst
category, a vector of geometrical facial features is extracted by measuring and
calculating the locations and geometric relationships among the facial features
(e.g., the mouth, eyes, and nose). For instance, Biswas et al. [18] developed a face
recognition system based on feature representations obtained using SIFT-based
descriptors [19] at fducial locations on the face image, and then the SIFT features
of all facial landmarks were formed into a single vector representing the face. A
best Rank-1 identifcation rate of 91% was achieved on the CMU Multi-PIE face
database containing a total of 337 subjects. Unlike the feature-based approaches,
the holistic-based approaches (e.g., Eigenface methods) operate on the whole face
image instead of measuring the local geometrical features. The major limitation
of these approaches is caused by the need to perform the matching process in very
high dimensionality space. Thus, a number of face recognition algorithms have
been developed that employ different dimensionality reduction methods to handle
the issue of high dimensionality before the matching process can take place. The
frst well-known face recognition system based on the Eigenface methods was pro-
posed by Sirovich and Kirby [20].
The hybrid-based approaches use a combination of both feature and holis-
tic approaches to address the face recognition problems. For example, one can
argue that hybrid approaches that use both local and global features (e.g., obtained
from both feature and holistic approaches, respectively) can be an effcient way to
make use of complementary information, reduce the complexity of classifers, and
increase their generalization ability. However, the main challenging factors that can
infuence the performance of such approaches lie in the identifcation of the fea-
tures to be fused, and the fusion method [21]. For example, local features are very
sensitive to the illumination changes, while expression changes have more infuence
on global features. Some of the recent examples belonging to this category were
proposed by Fischer et al. [22], Berg and Belhumeur [23], Annan et al. [24], Zhu et
al. [25], and Al-Waisy [26, 27], and other examples can be found in Meethongjan
and Mohamad [21]. Although previous research has demonstrated the effciency of
handcrafted-descriptor approaches used as robust and discriminative feature detec-
tors to solve the face recognition problem, their performance declines dramatically
in unconstrained environments. The decline in performance is caused by the fact
that constructed facial feature representations are very sensitive to highly nonlinear
intrapersonal variations such as illumination, expression, pose, and occlusion [28].
In addition, this approach usually requires signifcantly more effort and domain
knowledge from the researcher to fnd the best compact and discriminative feature
set for a given problem. Therefore, they are considered to be challenging and time
demanding.
94 Deep Learning in Computer Vision

4.2.2 LEARNING-BASED APPROACHES

Inspired by their ability to learn more discriminative feature representations, consid-
erable attention has been paid recently to deep learning approaches, such as CNN.
The main advantage of the deep learning approaches over handcrafted-descriptor
approaches is their ability to learn from experience and their robustness in handling
nonlinear variations of face images caused by pose variations, self-occlusion, etc.
[29]. Moreover, in contrast to handcrafted-descriptor approaches, the applications
making use of deep learning approaches can generalize well to other new felds.
Some recent examples belonging to this category were proposed by Sun et al. A
face recognition system, called Deep Identifcation-verifcation features (DeepID2),
is presented in [28]. In this work, two supervisory signals (e.g., the face identifcation
and verifcation signals) were combined simultaneously to boost the discriminative
power of extracted features using CNNs. A verifcation rate of 99.15% was achieved
using the “Unrestricted, Labeled Outside Data” evaluation protocol. The same group
of researchers developed the deep ConvNet-RBM model based on a combination of a
set of CNNs and a restricted Boltzmann machine (RBM) model for face verifcation
in the wild environment [30]. Verifcation rates of 91.75% and 92.52% were achieved
on the LFW and CelebFaces databases, respectively.
Taigman et al. [31] proposed the DeepFace framework for face verifcation using
the frontal face images generated from a 3D shape model of a large-scale face data-
set. They proposed to use the Softmax function on the top of CNN as the supervi-
sory signal to train the proposed framework. Experiments on the LFW database
using the “Unrestricted, Labeled Outside Data” setting yielded a verifcation rate
of 97.35%. In the literature, several metric learning approaches have been pro-
posed to address the face verifcation problem, including Multimodal Deep Face
Representation (MM-DFR) [32], Information-Theoretic Metric Learning (ITML)
[33], Discriminative Deep Metric Learning (DDML) [34], Pairwise-constrained
Multiple Metric Learning (PMML) [35], Logistic Discriminant-based Metric
Learning (LDML) [36], Large Scale Metric Learning (LSML) [37], and Cosine
Similarity Metric Learning (CSML) [38]. The key limitations of the deep learning
approaches include requiring a large amount of training data to avoid the overftting
problem and increase the generalization ability of the neural network, and the num-
ber of hyper-parameters that need to be set. The DNNs mainly depend on a tedious
iterative optimization procedure that can be computationally expensive for large-
scale databases. However, these issues can be effectively alleviated using high-per-
formance computing systems, equipped with graphics processing units (GPUs) [29].
Finally, a number of approaches have been proposed to address the problem of the
multi-biometric face recognition system. For instance, Peng et al. [39] have proposed
a novel graphical representation-based HFR method (G-HFR) to address the hetero-
geneous face recognition (HFR) problem. The HFR problem refers to identifying
the person’s identity using face images captured from different sources (e.g., differ-
ent sensors or different wavelengths). The same research group has also addressed
another face recognition task, known as cross-modality face recognition. This task
aims to match face images across different modalities, include matching sketches
with real photos, low-resolution face images with high-resolution images, and
A Multi-biometric Face Recognition System 95

near-infrared images with visual lighting images [40, 41]. Gao et al. [42] have also
addressed the problem of the insuffcient labeled samples by proposing the Semi-
Supervised Sparse Representation-based Classifcation method. This method has
been tested on four different datasets, including AR, Multi-PIE, CAS-PEAL, and
LFW. For more information on the most signifcant challenges, achievements, and
future directions related to the face recognition task, the reader is referred to [43].

4.3 THE PROPOSED APPROACHES

This section provides a brief description of the proposed approaches, including the
FDT and DRBM approaches, which are used in the proposed local facial feature-
based extraction approach. In addition, a brief description of the proposed deep
learning approaches is given, including the CNN and DBN. The main goal here is
to review and identify their strengths and shortcomings to empower the proposal of
a novel multi-biometric face recognition system that consolidates the strengths of
these approaches.

4.3.1 FRACTAL DIMENSION

The term “fractal dimension” was frst introduced by the mathematician Benoit
Mandelbrot as a geometrical quantity to describe the complexity of objects that
show self-similarity at different scales [44]. The most important characteristic of
the fractal dimension is self-similarity, which means that an object has a similar
representation to the original under different magnifcations. This property has been
effciently used in refecting the roughness and fuctuation of the surface of an image,
where increasing the scale of magnifcation provides more and more details of the
imaged surface. Moreover, the non-integer value of the fractal dimension provides
a quantitative measure of objects that have complex geometry, and cannot be well
described by an integral dimension (e.g., the length of a coastline) [45–46]. The
fractal dimension has been widely employed to resolve many problems in image
processing and computer vision (e.g., texture segmentation, texture classifcation,
medical imaging, etc.), due to its simplicity and robustness in refecting the rough-
ness and fuctuations of the imaged surface. However, not much work has been done
to explore and address the potential of using the fractal dimension to resolve pattern
recognition problems. Lin et al. [47] proposed an algorithm for human eye detec-
tion by exploiting the fractal dimension as an effcient approach for representing the
texture of facial features. Farhan et al. [48] developed a personal identifcation sys-
tem based on fngerprint images using the fractal dimension as a feature extraction
method. Therefore, it appears that the texture of the facial image can be effciently
described by using the fractal dimension. Various methods have been proposed in
the literature to estimate fractal dimension [44], e.g., box counting (BC), differential
box counting (DBC), fractional Brownian motion (FBM), and others. In this work,
the fractal dimension for each pixel over the whole face image is calculated using
an improved differential box counting (IDBC) algorithm, which could be used as a
powerful edge enhancement technique to enhance the edge representations for the
face image without increasing the noise level.
96 Deep Learning in Computer Vision

4.3.2 RESTRICTED BOLTZMANN MACHINE

A restricted Boltzmann machine (RBM) is an energy-based bipartite graphical
model composed of two fully connected layers via symmetric undirected edges, but
there are no connections between units of the same layer. The frst layer consists of
m visible units v = (v1 , v2 ,…, vm), which represent observed data, while the second
layer consists of n hidden units h = (h1 , h2 ,…, hn), which can be viewed as nonlin-
ear feature detectors to capture higher-order correlations in the observed data. In
addition, W = {w11 , w12 ,…, w1n ,…, wmn} is the connecting weights matrix between
the visible and hidden units. A typical RBM structure is shown in Figure 4.2 (a).
The standard RBM was designed to be used only with binary stochastic visible
units, and is termed Bernoulli RBM (BRBM). However, using binary units is not
suitable for real-valued data (e.g., pixel intensity values in images). Therefore, a
new model has been developed called the Gaussian RBM (GRBM) to address this
limitation of the standard RBM [49]. The energy function of the GRBM is defned
as follows:

m n m
( vi − bi )2 − n

˙˙ ˙ ˙c h
v
E ( v, h) = − wi , j h j i − i j (4.1)
i=1 j =1
s i
i=1
s 2i
2s j =1

Here, σi is the standard deviation of the Gaussian noise for the visible unit vi, wij
represents the weights for the visible unit vi and the hidden unit hj, and bi and cj are
biases for the visible and hidden units, respectively. The conditional probabilities for

FIGURE 4.2 (a) A typical RBM structure. (b) A discriminative RBM modeling the joint
˜
distribution of input variables v and target classes y (represented as a one-hot vector by y ).
(c) A greedy layer-wise training algorithm for the DBN composed of three stacked RBMs.
(d) Three layers of the DBN as a generative model, where the top-down generative path is
represented by the P distributions (solid arcs), and bottom-up inference and the training path
are represented by the Q distributions (dashed arcs).
A Multi-biometric Face Recognition System 97

the visible units, given hidden units, and vice versa for the hidden units are defned
as follows:

˙ ˘
p ( vi = 1 | h) = N ˇ v | bi +
ˇˆ wj
i, j h j , s 2i 

(4.2)

˙ vi ˘
p ( h j = 1 | v) = f ˇ c j +
ˆ
wi
i, j 
s 2i 
(4.3)

Here, N(·| µ,σ2) refers to the Gaussian probability density function with mean µ and
standard deviation σ. f(x) is a sigmoid function. During the training process, the log-
likelihood of the training data is maximized using stochastic gradient descent, and
the update rules for the parameters are defned as follows:

˛ 1 1 ˆ
Dwi , j = ˜ ˙ 2 vi h j − vi h j ˘ (4.4)
˝ si data
s 2i mod el ˇ

˛ 1 1 ˆ
D
Dbi = ˜ ˙ 2 vi − vi ˘ (4.5)
˝ si data
s 2i model ˇ

D (
Dci = ˜ h j data
− hj model ) (4.6)

Here, ˜ is the learning rate and ˜ data and ˜ model represent the expectations under the
distribution specifed by the input data (Positive phase) and the internal representa-
tions of the RBM model (Negative phase), respectively. Finally, bi and ci are bias
terms for visible and hidden units, respectively. More details on the GRBM model
can be found in [49]. Typically, RBMs can be used in two different ways: either as
generative models or as discriminative models, as shown in Figure 4.2 (a, b). The
generative models use a layer of hidden units to model a distribution over the visible
units, as described above. Such models are usually trained in an unsupervised way
and used as feature extractors to model only the inputs for another learning algo-
rithm. On the other hand, discriminative models can also model the joint distribution
of the input data and associated target classes. The discriminative models aim to
train a joint density model using an RBM that has two layers of visible units. The frst
represents the input data and the second is the Softmax label layer that represents
the target classes. A discriminative RBM with n hidden units is a parametric model
of the joint distribution between a layer of hidden units (referred to as features) h =
(h1, h2 ,…, hn) and visible units made of the input data v = (v1, v2 ,…, vm) and targets
y ∈ {1,2,…C}, which can be defned as follows:

p ( y, v, h) ˛ e ( )
− E y, v , h
(4.7)
98 Deep Learning in Computer Vision

where
˜ ˜
E ( y, v, h) = −hTWv − bT v − cT h − d T y − hTUy (4.8)

˜
Here, (W, b, c, d, U) refer to the model parameters and y = (1 y = i )i =1 to the C classes.
C

More details about the discriminative RBM model can be found in [50].

4.3.3 DEEP BELIEF NETWORKS

A deep belief network (DBN) is a generative probabilistic model that differs from
conventional discriminative neural networks. DBNs are composed of one visible
layer (observed data) and many hidden layers that have the ability to learn the statis-
tical relationships between the units in the previous layer. They model the joint prob-
ability distribution over the input data (observations) and labels, which facilitates
the estimation of both P(Observations |Labels) and P(Labels |Observations), while
conventional neural networks are limited only to the latter [51]. DBN has been pro-
posed to address issues encountered when applying the Back-propagation algorithm
to very deep neural networks, including:

1. The constraint of having a labeled dataset in the training phase.

2. The long time required to converge (slow learning process).
3. The increased number of free parameters that get trapped in poor local
optima.

As depicted in Figure 4.2 (c), a DBN can be viewed as a composition of bipartite

undirected graphical models, each of which is an RBM. Therefore, DBNs can be
effciently trained using an unsupervised greedy layer-wise algorithm, in which the
stacked RBMs are trained one at a time in a bottom-to-top manner. For instance,
consider training a DBN composed of three hidden layers, as shown in Figure 4.2(c).
According to the greedy layer-wise training algorithm proposed by Hinton et al.
[52], the frst RBM is trained using the CD algorithm to learn a layer (h1) of feature
representations from the visible units, as described in Section 4.1.2. Then, the hidden
layer units (h1) of the frst RBM are used as visible units to train the second RBM.
The whole DBN is trained when the learning of the fnal hidden layer is completed.
A DBN with l layers can model the joint distribution between the observed data vec-
tor v and l hidden layers hk as follows:

˙ l −2
˘
( 1
P v, h , …, h = ˇ
l
) P ( h k
) (
| h k+1  P hl −1 , hl ) (4.9)
ˆ k=0 

( )
Here, ν = h0, P h k | h k+1 is the conditional distribution for the visible units given
hidden units of the RBM associated with level k of the DBN, and P hl −1 , hl is ( )
the visible-hidden joint distribution in the top-level RBM. An example of a three-
layer DBN as a generative model is shown in Figure 4.2 (d), where the symbol Q
is introduced for exact or approximate posteriors of that model, which are used
A Multi-biometric Face Recognition System 99

for bottom-up inference. During the bottom-up inference, the Q posteriors are all
approximate except for the top-level P ( hl | hl −1 ), which is formed as an RBM and
then the exact inference is possible. Although the DBN is considered as one of
the most popular unsupervised deep learning approaches, and has been success-
fully applied in a wide range of applications (e.g., face recognition [13], speech
recognition [14], audio classifcation [15], etc.), one of the main issues of the DBN
is that the feature representations it learns are very sensitive to the local transla-
tions of the input image, in particular when the pixel intensity values (raw data)
are assigned directly to the visible units. This can cause the local facial features
of the input image, which are known to be important for face recognition, to be
disregarded.

4.3.4 CONVOLUTIONAL NEURAL NETWORK

A convolutional neural network (CNN) is a feed-forward multi-layer neural net-
work, which differs from traditional fully connected neural networks by combin-
ing a number of locally connected layers aimed at automated feature recognition,
followed by a number of fully connected layers aimed at classifcation [51]. As
shown in Figure 4.3, the CNN architecture is composed of several distinct lay-
ers, including sets of locally connected convolutional layers (each layer contains a
number of learnable kernels), sub-sampling pooling layers, and one or more fully
connected layers. The internal structure of the CNN is associated with three archi-
tectural concepts, which make the CNN successful in resolving different problems
in image processing, computer vision, and pattern recognition. The frst concept
is applied in both convolutional and pooling layers, in which each neuron receives
input from a small region of the previous layer called the local receptive feld,
which is equal in size to a convolution kernel [53]. This local connectivity scheme
ensures that the trained CNN produces strong responses to capture local depen-
dencies and extracts elementary features in the input image (e.g., edges, curves,
etc.), which can play a signifcant role in minimizing the intra-class variations
and inter-class similarities and hence increasing the discriminative power of the
learned facial feature representations. Secondly, the convolutional layer applies the
sharing parameters (weights) scheme in order to control the model capacity and

FIGURE 4.3 An illustration of the CNN architecture, where the gray and green squares
refer to the activation maps and the learnable convolution kernels, respectively. The crossed
lines between the last two layers show the fully connected neurons [55].
100 Deep Learning in Computer Vision

reduce its complexity. At this point, a form of translational invariance is obtained

using the same convolution kernel to detect a specifc feature at different locations
in the face image [54]. Finally, the nonlinear down-sampling applied in the pool-
ing layers reduces the spatial size of the convolutional layer’s output and reduces
the model’s number of free parameters. Together, these characteristics make the
CNN very robust and effcient at handling image deformations and other geomet-
ric transformations, such as translation, rotation, and scaling [51]. More details on
the CNN can be found in [55].

4.4 THE PROPOSED METHODOLOGY

As depicted in Figure 4.4, a novel multi-biometric face recognition system named
FaceDeepNet system is proposed to learn high-level facial feature representations
by training the proposed deep learning approaches (e.g., CNN and DBN) on top of
local facial feature representations obtained from the FDT approach instead of the
pixel intensity representations. Firstly, the main stage of the proposed FDT-DRBM
approach is described in detail. This is followed by describing how to learn addi-
tional and complementary representations by applying a FaceDeepNet system on top
of existing local feature representations.

FIGURE 4.4 Illustration of the proposed multi-biometric face recognition system.

A Multi-biometric Face Recognition System 101

4.4.1 THE PROPOSED FDT-DRBM APPROACH

The proposed face recognition algorithm starts by detecting the face region using
a Viola-Jones face detector [56]. Then, after rescaling the input image f(x, y) to the
range of [0, 1], a simple pre-processing algorithm using a sigmoid function is applied
as follows:

1
g( x, y) = (4.10)
1+ e
(c*(Th − f ( x, y)))

Here, g(x, y) is the enhanced image. The contrast factor (c) and the threshold value
(Th) are empirically set to be 5 and 0.3, respectively. As shown in Figure 4.5, the
advantage of the sigmoid function is to reduce the effect of lighting changes by
expanding and compressing the range of values of the dark and bright pixels in the
face image, respectively. This is followed by estimating the fractal dimension value
of each pixel in the enhanced face image using the proposed FDT approach. As men-
tioned above, the fractal dimension has many important characteristics, for instance
its ability to refect the roughness and fuctuations of a face image’s surface, and to
represent the facial features under different environmental conditions (e.g., illumina-
tion changes).
However, fractal estimation approaches are very time consuming and cannot
meet real-time requirements [7]. Thus, the face image is rescaled to (64×64) pixel
after detecting the face region in order to speed up the experiments and meet the
real-time system demands. The result is then reshaped into a row feature vector
FDTVector, which is used as an input to a nonlinear classifer using DRBM to eff-
ciently model the joint distribution of inputs and target classes. In this chapter, the

FIGURE 4.5 Output of the image contrast enhancement procedure. (a) The original face
image. (b) The enhanced face image. pmax and pmin refer to maximum and minimum pixel
values in the enhanced face image.
102 Deep Learning in Computer Vision

number of the produced matching scores and predicted classes using the DRBM
classifer is varied according to the applied recognition task. For instance, the goal
of the identifcation system is to determine the true identity (I) of the query trait
based on these N matching scores, where I ∈ {I1, I2 , I3 , ··· , IN , IN+1}. Here, (I1, I2 ,
I3 , ··· , IN) correspond to the identities of the N persons enrolled in the dataset, while
(IN+1) point to the “reject” option, which is generated when no associated identity
can be determined for the given query trait; this is known as open set identifcation.
However, when the biometric system assumes that the given query trait is among the
templates enrolled in the dataset, this is referred to as closed set identifcation [57].
In this proposed system, the open set identifcation protocol is adopted, because of
its effciency with large-scale, real-world applications (e.g., surveillance and watch
list scenarios) [58]. On the other hand, in the verifcation task the DRBM classifer is
used as a binary classifer to determine whether a pair of face images belong to the
same person or not. The next two subsections describe in more detail the FDT and
the DRBM methods mentioned above.

4.4.1.1 Fractal Dimension Transformation

Using the proposed FDT approach, the face image of size I(M×N) is transformed
to its fractal dimension form using the differential box-counting (DBC) algorithm
where a kernel function f(p, q) of size (n×n) pixel is applied on the entire face image
to produce a 3D matrix that represents the number of boxes necessary to overlay the
whole face image at each pixel as follows:
a b 2

N d ( x, y, d ) = 
p= − aq= − b
˙r ˘
f ( p, q ) I ( x + p, y + q ) ˇ max 
ˆ r 
(4.11)

Here, d = (1, 2, 3, …, rmax − 1) is the third dimension of matrix Nd (x, y, d), which rep-
resents the number of the produced feature maps obtained from applying the kernel
function f(p, q) along with different values of the scaling factor (r). Here, the values
of the scaling factor (r) are empirically chosen to be in the range between 2 and 9.
The kernel function f(p, q) operates by block processing on (7×7) neighboring pixels
of the face image, and calculating the fractal dimension value of each pixel from its
surrounding neighbors, as follows:

a b
pmax − pmin ˆ
f ( p, q) = floor ˛˙˝
p= − aq= − b
r ˘+1
ˇ
(4.12)

Here, a and b are non-negative integer variables, which are used to center the kernel
function on each pixel in the face image, and are defned as a and b = ceil ((n − 1)/ 2).
pmax and pmin are the highest- and lowest-intensity values of neighboring pixels in
the processing block. The size of the kernel function was determined empirically,
noting that increasing its size can affect the accuracy of the calculated fractal dimen-
sion, causing the obtained image to become less distinct, while decreasing its size
can result in an insuffcient number of surrounding pixels to calculate the fractal
dimension value accurately. Once the Nd (x, y, d) matrix is obtained, each element
A Multi-biometric Face Recognition System 103

from each array in Nd (x, y, d) will be stored in a new row vector (V). In other words,
the frst element in all arrays of Nd (x, y, d) will compose vector (V1), and all second
elements will compose (V2), etc., as follows:

° p111 p112 ° ° p11rmax ˙

° V1 ˙ ˝
˝ V ˇ ˝ p121 p122 ° ° p12 rmax ˇ
ˇ
˝ 2 ˇ =˝ ˜ ˜ ˛ ˜ ˇ (4.13)
˝ ˜ ˇ ˝
˝˛ V ˇ ˝ ˜ ˜ ˛ ˜ ˇ
M×N ˆ
ˇ
˝˛ p pMN 2 ° ° pMNrmax ˇˆ
MN1

Finally, the fractal dimension value at each pixel (x, y) is computed as the fractal
slope of the least square linear regression line as follows:

Sv ( x, y)
FD ( x, y) = (4.14)
Sr

Here, (Sv) and (Sr) are the sums of squares and can be computed as follows:

)−  
˙ j
ˇ˙ j
ˇ
j log(r)˘ ˆ v( x, y), r ˘
 log(r) (v(
ˆ
Sv ( x, y) = x , y ), r
r =1 r =1
(4.15)
r =1
j

2
˙
 ˇ
j
j
ˆ log(r)˘
Sr ( x, y) =  (log(r)) −
2 r =1
(4.16)
r =1
j

Here, j = rmax − 1 and a fractal transformed face image is illustrated in Figure 4.6.

FIGURE 4.6 (a) Original face images. (b) FDT fractal-transformed images.
104 Deep Learning in Computer Vision

4.4.1.2 Discriminative Restricted Boltzmann Machines

In the classifcation stage, a single RBM that has two sets of visible units is trained
as a nonlinear classifer to effciently model the joint distribution of inputs and target
classes. In addition to the visible units that represent a data vector, there is a Softmax
label unit that represents the predicted classes. After training, each possible label is
tried in turn and the one that gives lowest free energy is chosen as the most likely
class. As reported in the literature, it is intractable to compute the vihjmodel in Eq. 4.4.
Therefore, the contrastive divergence (CD) algorithm [59] has become the standard
learning method to update the RBM parameters by sampling k steps from the RBM
distribution to approximate the second term in Eq. 4.4. Effcient training requires
optimized training parameters (e.g., learning rate, the weight-cost, the momentum,
the initial values of the weights, the size of each mini-batch, the number of hidden
units, etc.). Other decisions need to be made, including the types of hidden units to
be used, how many times to update the binary states of the hidden units for each
training epoch, how to monitor the progress of learning, and when to terminate the
training process. One step of a CD algorithm for a single sample can be implemented
as follows:

1. Visible units (vi) are initialized using training data, and the probabilities of
hidden units are computed with Eq. 4.2. Then, a hidden activation vector
(hj) is sampled from this probability distribution.
2. Compute the outer product of (vi) and (hj), which refers to the positive phase.
3. Sample a reconstruction (vi˜) of the visible units from (hj) with Eq. 4.3, and
then from (vi˜) resample the hidden units’ activations (hi˜). (One Gibbs sam-
pling step.)
4. Compute the outer product of (vi˜) and (hi˜), which refers to the negative
phase.
5. Update weights matrix and biases with Eq. 4.4, Eq. 4.5, and Eq. 4.6.

The computation steps of the CD-1 algorithm are graphically shown in Figure 4.7. In
the CD learning algorithm, k is usually set to 1 for many applications. In this work,
the weights were initialized with small random values sampled from a zero-mean
normal distribution and standard deviation of 0.02. In the positive phase, the binary
states of the hidden units are determined by computing the probabilities of weights

FIGURE 4.7 A graphic illustration of a single step of the CD algorithm.

A Multi-biometric Face Recognition System 105

and visible units. Since the probability of training data is increased in this phase,
it is called positive phase. Meanwhile, the probability of samples generated by the
model is decreased during the negative phase. A complete positive-negative phase is
considered as one training epoch, and the error between produced samples from the
model and actual data vector is computed at the end of each epoch. Finally, all the
weights are updated by taking the derivative of the probability of visible units with
respect to weights, which is the expectation of the difference between positive phase
contribution and negative phase contribution. Algorithm 1 shows pseudo-code of the
procedure proposed to train the DRBM classifer using the CD algorithm.

Algorithm 1. Find the optimal DRBM classifer.

Input: Labeled dataset: Training Set (xi, yi) and Testing Set (xi, yi).
Output: Optimal_h = The best number of hidden units (h).
Initialization: Max_Accuracy= 0, LR = 10−2, WD = 0.0002, MM=0.9, and Epoch=300;
for h_now = 1000 : 100 : 5000
for Epoch = 1 : 300
• Train the DRBM classifer using the CD learning algorithm with one step of Gibbs
sampling.
• Apply the 10-fold cross-validation evaluation procedure on the Training Set.
• Calculate the AER over 10-fold using Eq. 4.17.
• Calculate the Validation_Accuracy = 1-AER;
• if Validation_Accuracy>Max_Accuracy
Max_Accuracy = Validation_Accuracy;
Optimal_h = h_now:
end
end
end
• Train the DRBM classifer on the whole Training Set using the Optimal_h.
• Evaluate the generalization ability of the fnal trained DRBM classifer using the Testing Set.

4.4.2 THE PROPOSED FACEDEEPNET SYSTEM

An effcient deep learning system is employed. This system, whose architecture is
based on a combination of a CNN and a DBN to extract discriminative facial fea-
tures from the input image of the same person, will be referred to as FaceDeepNet.
To the best of our knowledge, this is the frst work that investigates the potential use
the CNN and the DBN together in a unifed multi-biometric face recognition system
operating in both the identifcation and verifcation modes. Here, we argue that using
the local feature representations obtained from a powerful handcrafted-descriptor
instead of the pixel intensity representations as an input to the proposed deep learning
approaches can signifcantly improve the ability of the proposed FaceDeepNet system
to learn more discriminating facial features, with a shorter training time required to
obtain the fnal trained model. As depicted in Figure 4.4, the local facial features are
frstly extracted using the proposed FDT approach. Then, the extracted local features
are assigned to the feature extraction units of the proposed deep learning approaches
(e.g., CNN and DBN) to learn additional and complementary feature representations.
106 Deep Learning in Computer Vision

In this work, the architecture of the proposed CNN involves a combination of

convolutional layers and sub-sampling max-pooling. The top layers in the proposed
CNN are two fully connected layers for the multi-class classifcation purpose. Then,
the output of the last fully connected layer is fed into the Softmax classifer, which
produces a probability distribution over the N class labels. Finally, a cross-entropy
loss function, a suitable loss function for the classifcation task, is used to quantify
the agreement between the predicted class scores and the target labels, and to calcu-
late the cost value for different confgurations of CNN. On the other hand, the archi-
tecture of the proposed DBN stacks a number of RBMs (as hidden layers). The frst
N-1 layers of the network can be seen as feature detection layers. The last layer of
the network is trained as a discriminative model associated with Softmax classifer
for the classifcation task. Finally, the hidden layers of the proposed DBN are trained
using an unsupervised greedy layer-wise algorithm (e.g., the CD learning algorithm),
in which the stacked RBMs are trained one at a time in a bottom-to-top manner.
The proposed deep learning approaches (e.g., CNN and DBN) are trained using a
powerful training methodology equipped with a number of training strategies and
techniques (e.g., Back-propagation, AdaGrad algorithm, Dropout technique, Data
augmentation, etc.), in order to avoid overftting problems and increase the general-
ization ability of the proposed FaceDeepNet system. For more details on how both
the proposed CNN and DBN are trained, the reader is referred to [55, 7]. Finally,
as all the employed classifers produce the same type of output (similarity score),
and all of them have the same numeric range [0, 1], no normalization procedure is
needed before applying the score fusion method. In this work, the weighted sum rule
is used to fuse the matching scores produced from the proposed CNN and DBN.

4.5 EXPERIMENTAL RESULTS

In this section, the performances of the proposed unimodal approaches (e.g., using
only FDT-DRBM, DBN, and CNN) and multi-biometric approaches (e.g., using
FaceDeepNet) are tested on the most challenging face datasets currently available
in the public domain in order to demonstrate their effectiveness and compare their
performances with other existing face recognition approaches. A brief description of
the face datasets used in these experiments is given. Then, an elaborated evaluation
and comparison with the state-of-the-art face recognition approaches in both face
identifcation and verifcation tasks are presented. In addition, some insights and
fndings relevant to learning additional and complementary facial feature representa-
tions by training the proposed deep learning approaches on top of local feature rep-
resentations are presented. The system code was written to run in MATLAB R2018a
and later versions. The recognition time of the proposed approaches was measured
by implementing them on a personal computer with the Windows 10 operating sys-
tem, a 3.60 GHz Core i7-4510U CPU, and 16 GB of RAM. It should be noted that
the recognition times of the proposed approaches are proportional to the number of
registered subjects and their images in the database. However, the test time is less
than 4 seconds on average, which is fast enough to meet the demands of real-world
applications. Finally, the matching scores obtained from the multi-biometric face
recognition systems are combined using different fusion methods at the score level
A Multi-biometric Face Recognition System 107

FIGURE 4.8 Examples of face images in four face datasets: (a) SDUMLA-HMT, (b) FRGC
V 2.0, (c) UFI, and (d) LFW.

(e.g., using product, sum, weighted sum, max, and min rule) and rank level (e.g.,
using highest rank, Borda count, and logistic regression).

4.5.1 DESCRIPTION OF FACE DATASETS

All the experiments were conducted on four large-scale unconstrained face datasets:
SDUMLA-HMT [60], Face Recognition Grand Challenge Version 2.0 (FRGC V 2.0)
[61], Unconstrained Facial Images (UFI) [62], and Labelled Faces in the Wild (LFW)
[63]. Some examples of face images from each dataset are shown in Figure 4.8.

• SDUMLA-HMT Face dataset [60]: This includes 106 subjects; each

one has 84 face images taken from 7 viewing angles, and under differ-
ent experimental conditions including facial expressions, accessories,
108 Deep Learning in Computer Vision

poses, and illumination. The main purpose of this dataset is to simulate

real-world conditions during face image acquisition. The image size is
(640×480) pixels.
• FRGC V 2.0 dataset [61]: This is a large-scale challenging dataset collected
at the University of Notre Dame and is composed of two main partitions.
The frst part is the training partition; it was collected in the 2002–2003
academic year and is used for training still face recognition approaches.
This partition consists of 12,776 images from 222 subjects, with 6,388 con-
trolled still images and 6,388 uncontrolled still images. This large training
partition consists of 9 to 16 sessions per subject. The second part is the vali-
dation partition; it was collected during the 2003–2004 academic year and
contains images from 466 subjects collected in 4,007 subject sessions. This
partition consists of 1 to 22 subject sessions per subject. The image size is
either (1704×2272) pixels or (1200×1600) pixels, stored in JPEG format.
• Unconstrained Facial Images (UFI) dataset [62]: It contains real face
images obtained from a large set of images owned by the Czech News
Agency (CTK). The UFI dataset is used mainly for evaluating face identi-
fcation approaches; however, it is possible to be used in other related tasks
(e.g., face detection, verifcation, etc.). The UFI dataset is composed of two
partitions: cropped images (CI) and large images (LI). The CI partition
contains images of 605 subjects with an average of 7.1 images per subject
in the training set and one in the test set. The images are cropped to a size
of (128×128) pixels, and Figure 4.8 (c) shows some examples from the CI
partition. The LI partition contains a total of 530 subjects with an average
of 8.2 images per subject. The size of images in this partition is (384×384)
pixels, and some examples of this partition are shown in Figure 4.8 (c). The
main goal of the large images partition is to report the fnal performance
of the proposed face recognition system and compare it with other existing
systems.
• LFW dataset [63]: This contains a total of 13,233 images taken from 5,749
subjects, where 1,680 subjects appear in two or more images. In the LFW
dataset, all images were collected from Yahoo! News articles on the web,
with a high degree of intrapersonal variations in facial expression, illu-
mination conditions, occlusion from wearing hats and glasses, etc. It has
been used to address the problem of unconstrained face verifcation tasks in
recent years. The image size is (250×250) pixels, and some examples of the
matched pairs and mismatched pairs are shown in Figure 4.7 (d).

4.5.2 FACE IDENTIFICATION EXPERIMENTS

Initially, the robustness and effectiveness of the proposed face recognition approaches
were tested to address the unconstrained face identifcation problem using large-
scale unconstrained face datasets: SDUMLA-HMT, FRGC V 2.0, and UFI data-
set. The SDUMLA-HMT dataset was employed as the main dataset to fne-tune the
hyper-parameters of the proposed FDT-DRBM approach (e.g., number of hidden
units in the RBM) as well as the hyper-parameters of the proposed deep learning
A Multi-biometric Face Recognition System 109

approaches used in the FaceDeepNet system (e.g., number of layers, number of flters
per each layer, learning rate, etc.), as it has more images per person in its image gal-
lery than the other datasets. This allowed more fexibility in dividing the face images
into training, validation, and testing sets.
4.5.2.1 Parameter Settings of the FDT-DRBM Approach
The most important hyper-parameter in the proposed FDT-DRBM approach is the
number of the hidden units in the DRBM classifer. The other hyper-parameters (e.g.,
learning rate, the number of epochs, etc.) were determined based on domain knowl-
edge from the literature to be a learning rate of 10 −2, a weight decay of 0.0002,
momentum of 0.9, and 300 epochs, as in [7]. The weights were initialized with small
random values sampled from a zero-mean normal distribution and standard devia-
tion of 0.02. In this work, the number of the hidden units was determined empirically
by varying its values from 1,000 units to 5,000 units in steps of 100 units, using the
CD learning algorithm with one step of Gibbs sampling to train the DRBM as a
nonlinear classifer, as described in Section 4.1.2. Hence, 41 experiments were con-
ducted using 60% randomly selected samples per person for the training set, and the
remaining 40% for the testing set. In these experiments, the parameter optimization
process is performed on the training set using the 10-fold cross-validation procedure
that divides the training set into k subsets of equal size. Consecutively, one subset
is used to assess the performance of the DRBM classifer trained on the remaining
k-1 subsets. Then, the average error rate (AER) over 10 trials is computed as follows:
k

°Error
1
AER = i (4.17)
K i=1

Here, Errori refers to the error rate per trial. After fnding the best number of hidden
units with higher validation accuracy rate (VAR), the DRBM classifer was trained
using the whole training set, and its performance in predicting unseen data prop-
erly was then evaluated, using the testing set. Figure 4.9 shows the VAR produced
throughout 41 experiments. One can see that the higher VAR was obtained when the
3,000 units were employed in the hidden layer of the DRBM classifer. To verify the
proposed FDT-DRBM approach, its performance was compared with two current
state-of-the-art face recognition approaches, namely the Curvelet-Fractal approach
[7] and the Curvelet Transform-Fractional Brownian Motion (CT-FBM) approach
[64], using the same dataset. Figure 4.10 shows the cumulative match characteris-
tic (CMC) curve to visualize their performance. It can be seen in Figure 4.10 that
the proposed DFT-DRBM approach has the potential to outperform the CT-FBM
approach, as the Rank-1 identifcation rate has dramatically increased from 0.90 to
0.95 using the CT-FBM to more than 0.9513 to 1.0 using the DFT-DRBM approach.
On the other hand, despite the Curvelet-Fractal approach achieving slightly a higher
Rank-1 identifcation rate, it is observed that after Rank-10 identifcation rate the
accuracy of the Curvelet-Fractal approach drops compared to the FDT-DRBM
approach.
As mentioned previously, the main structure of the FaceDeepNet system is based
on building deep learning representations and fusing the matching scores obtained
110 Deep Learning in Computer Vision

FIGURE 4.9 The VAR produced throughout 41 experiments for fnding the best number of
hidden units.

FIGURE 4.10 Performance comparison between the FDT-DRBM, Curvelet-Fractal, and

CT-FBM approaches on the SDUMLA-HMT dataset.
A Multi-biometric Face Recognition System 111

from two discriminative deep learning approaches (e.g., CNN and DBN). It is well
known in the literature that the major challenge of using deep learning approaches is
the number of model architectures and hyper-parameters that need to be tested, such
as the number of layers, the number of units per layer, the number of epochs, etc. In
this section, a number of extensive experiments are performed to fnd the best model
architecture for both CNN and DBN, along with studying and analyzing the infu-
ence of the different training parameters on the performance of the proposed deep
learning approaches. All these experiments were conducted on the SDUMLA-HMT
face dataset using the same sizes of the training and testing sets described in Section
5.2.1, and the parameters with the best performance (e.g., highest VAR and best gen-
eralization ability) were kept to be used later in fnding the best network architecture.
For an initial CNN architecture, the architecture of our IrisConvNet system as
described in Al-Waisy et al. [55] was used to test its generalization ability and adap-
tation to another new application (e.g., face recognition). This initial CNN archi-
tecture will be denoted as FaceConvNet-A through the subsequent experiments.
The main structure of the IrisConvNet system is shown in Figure 4.3. Initially, the
FaceConvNet-A system was trained using the same training methodology along with
the same hyper-parameters described in [55]. However, it was observed that when
500 epochs were evaluated, the obtained model started overftting the training data,
and poor results were obtained on the validation set. Thus, a number of experiments
were conducted to fne-tune this parameter. Firstly, an initial number of epochs was
set to 100 epochs, and then larger numbers of epochs were also investigated, includ-
ing 200, 300, and 400. Figure 4.11 (a) shows the CMC curves used to visualize the
performance of the fnal obtained model on the validation set. It can be seen that as
long as the number of epochs is increased, the performance of the last model gets
better and the highest accuracy rate was obtained when 300 epochs were evaluated.
Furthermore, it was also observed that a better performance can be obtained by adding
a new convolution layer (C0= 10@64×64) before the frst one (C1= 6@64×64), using
the same flter size and increasing the number of flters per these two layers from 6 to

FIGURE 4.11 CMC curves for (a) epoch number parameter evaluation of CNN and (b)
performance comparison between FaceConvNet-A and FaceConvNet-B on the testing set of
the SDUMLA-HMT dataset.
112 Deep Learning in Computer Vision

10 learnable flters. This new CNN architecture was denoted as the FaceConvNet-B
system, and it was used in our assessment procedure for all remaining experiments
instead of the FaceConvNet-A system. The CMC curves of the FaceConvNet-A and
FaceConvNet-B systems are shown in Figure 4.11 (b) to visualize their performances
on the testing set of the SDUMLA-HMT dataset. Table 4.1 summarizes the details of
the FaceConvNet-B system and its hyper-parameters.
Following, the training methodology described in [7], the DBN approach was
greedily trained using input data acquired from the FDT-DRBM approach. Once the
training of a given RBM (hidden layer) is completed, its weights matrix is frozen,
and its activations are served as input to train the next RBM (next hidden layer) in
the stack. Unlike [7], in which the DBN models are composed of only 3 layers, here
larger and deeper DBM models have been investigated by stacking fve RBM mod-
els. As shown in Table 4.2, fve different 5-layer DBN models were greedily trained
in a bottom-up manner using different numbers of hidden units only in the last three
layers. For the frst four hidden layers, each one was trained separately as an RBM
model in an unsupervised way, using the CD learning algorithm with one step of
Gibbs sampling. Each individual RBM was trained for 100 epochs with the learn-
ing rate set to be 0.01, a momentum value of 0.95, a weight-decay value of 0.0005,
and a mini-batch size of 100. The weights of each model were initialized with small
random values sampled from a zero-mean normal distribution and standard devia-
tion of 0.02. The last RBM model was trained in a supervised way as a nonlinear
classifer associated with Softmax units for the multi-class classifcation purpose. In
this supervised phase, the last RBM model was trained using the same values of the
hyper-parameters used to train the frst four models. Finally, in the fne-tuning phase,
the whole DBN model was trained in a top-down manner using the Back-propagation
algorithm equipped with the Dropout technique, to fnd the optimized parameters
and to avoid overftting. The Dropout ratio was set to 0.5 and the number of epochs
through the training set was determined using an early stopping procedure, in which

TABLE 4.1
Details of Hyper-Parameters for the Proposed Deep Learning Approaches
(e.g., CNN and DBN)
FaceConvNet-B 5-Layer DBN Model
Hyper-parameter Value Hyper-parameter Value

No. of Conv. Layers 5 CD Learning Algorithm 1 step of Gibbs sampling

No. of Max-Pooling Layers 3 No. of Layers 5 RBMs
Optimization Method Adagrad No. of Epochs for each RBM 100
Momentum 0.9 Momentum 0.95
Weight-Decay 0.0005 Weight-Decay 0.0005
Dropout 0.5 Dropout 0.5
Batch Size 100 Batch Size 100
Learning Rate 0.01 Learning Rate 0.01
Total No. of Epochs 300 Total No. of Epochs (BW) 1000
A Multi-biometric Face Recognition System 113

TABLE 4.2
Rank-1 Identifcation Rates Obtained
for Different DBN Architectures
Using Validation Set
DBN Model Accuracy Rate

4096-4096-2048-1024-512 0.92
4096-4096-2048-1024-1000 0.94
4096-4096-2048-2000-1000 0.95
4096-4096-3000-2000-1000 0.94
4096-4096-3000-2000-2000 0.97

the training process is stopped as soon as the classifcation error on the validation
set starts to rise again. As can be seen in Table 4.2, the last DBN model provided
signifcantly better recognition rate compared to the other hidden layers investigated.
More information on the FaceConvNet-B and its hyper-parameters are presented in
Table 4.1.
Moreover, a few experiments were also carried out using the SDUMLA-HMT
dataset to fnd out how the performance of the proposed deep learning approaches
can be greatly improved by assigning the local feature representations to the visible
units of the DNNs instead of the raw data, as a way of guiding the neural network to
learn only the useful facial features. As shown in Figure 4.4, the local facial features
are frstly extracted using the proposed FDT-DRBM approach. Then, these extracted
local facial features are assigned to the feature extraction units of the CNN and DBN
to learn additional and complementary feature representations. As expected and as
shown in Table 4.3, our experiments demonstrate that training the proposed deep
learning approaches on the top of the FDT-DRBM approach’s output can signif-
cantly increase the ability of the DNNs to learn more discriminating facial features
with a shorter training time. Generally, the CNN produced the highest recognition
rate compared to the DBN, due to its ability to capture the local dependencies and
extract elementary features in the input image (e.g., edges, curves, etc.), which are
proven to be important for face recognition. Thus, in the fusion phase when the

TABLE 4.3
Rank-1 Identifcation Rates Obtained for the Proposed Deep Learning
Approaches (CNN and DBN) Using Different Types of Input Data
Raw Data Local Facial Features

Approach Accuracy Rate Training Time Accuracy Rate Training Time

CNN 0.89 12 h & 40 min 0.96 4 h & 10 min

DBN 0.85 15 h & 25 min 0.94 4 h & 40 min
114 Deep Learning in Computer Vision

weighted sum rule is applied, a higher weight was assigned to the proposed CNN
compared to the DBN approach, such that the maximum recognition rate is achieved.

4.5.2.2 Testing the Generalization Ability of the Proposed Approaches

To further examine the robustness of the proposed face recognition approaches, a
number of comprehensive experiments were conducted on the FRGC V 2.0, and UFI
datasets, and the results obtained were compared with the current state-of-the-art
approaches. For a fair comparison, the performance of the FDT-DRBM approach
was evaluated using the standard evaluation protocols of the FRGC V 2.0 and UFI
datasets, described in [61, 62], respectively. It is worth noting that DNNs need to
be trained on a massive amount of training data to achieve satisfactory prediction
and prevent overftting. Thus, to increase the generalization ability of the proposed
deep learning approaches (e.g., CNN and DBN), the data augmentation procedure
as described in [7] was applied to the gallery set of these two datasets. Then, their
performances during the learning process were observed on a separate validation
set taken from the fully augmented gallery set. According to the standard evaluation
protocol, the FRGC V 2.0 dataset consists of six experiments. The experiments used
to evaluate the performance of the 3D face recognition systems are not relevant to the
objectives of this work. Hence, the focus will be on experiments 1, 2, and 4. In exper-
iment 1, both the gallery and testing sets consist of a single controlled still image
of each person. Experiment 2 addresses the effect of using multiple still images per
person on performance. The gallery in this experiment consists of four images for
each person, taken in the same subject session. Likewise, a testing set is composed
of four images per person. Experiment 4 addresses the problem of uncontrolled face
recognition. In experiment 4, the gallery is composed of a single controlled still
image, and the testing set is composed of a single uncontrolled still image. In the
standard UFI evaluation protocol, both partitions (CI and LI) are divided into train-
ing and testing sets. For comparison purposes, the face images in the cropped images
were kept at the same size: (128×128) pixels.
Firstly, the matching scores are combined using different fusion methods at the
score level (e.g., using product, sum, weighted sum, max, and min rule) and rank
level (e.g., using highest rank, Borda count, and logistic regression). As shown in
Figure 4.4, three matching scores are produced when the face trait is submitted by
the user at identifcation point. One matching score is obtained from the FDT-DRBM
approach and two matching scores from the FaceDeepNet system. In the implemen-
tation of weighted fusion methods, a higher weight was assigned to the proposed
FaceDeepNet system compared to the FDT-DRBM approach, to ensure that the
maximum recognition rate is achieved. The Rank-1 identifcation rates, using differ-
ent fusion methods at the score-level and rank-level fusion, are listed in Tables 4.4
and 4.5, respectively. In general, the results obtained show a noticeable improvement
in the performance of the proposed multi-biometric face recognition system, com-
pared to applying only the FDT-DRBM approach or the FaceDeepNet system, by
achieving higher Rank-1 identifcation rates in most cases. From Tables 4.4 and 4.5,
one can see that the highest Rank-1 identifcation rates on all the testing sets of all
the face datasets were achieved by employing the weighted sum rule and the highest
rank method at the score-level and rank-level fusion, respectively. In this work, the
A Multi-biometric Face Recognition System 115

TABLE 4.4
Rank-1 Identifcation Rate of the Proposed Multi-Biometric Face Recognition
System on Three Different Face Datasets Using Score-Level Fusion SR and
PR are referred to the Sum Rule and Product Rule, respectively
Score Fusion Method

Dataset FDT-DRBM FaceDeepNet SR WSR PR Max Min

FRGC V 2.0 Exp. 1 94.11 97.62 95.12 97.34 94.92 95.34 95.21
Exp. 2 95.10 99.54 97.33 100 95.54 97.82 96.32
Exp. 4 96.21 98.33 97.13 99.12 97.13 97.12 97.12
UFI CI 92.23 93.60 94.32 96.66 92.12 92.31 93.21
LI 93.12 95.29 96.12 97.78 93.64 95.12 94.12
SDUMLA-HMT 95.13 95.13 98.89 100 97.53 96.23 96.21

TABLE 4.5
Rank-1 Identifcation Rate of the Proposed Multi-Biometric Face
Recognition System on Three Different Face Datasets Using
Rank-Level Fusion HR, BC, and LR are referred to Highest
Ranking, Borda Count, and Logistic regression, respectively
Rank Fusion Method

Dataset FDT-DRBM FaceDeepNet HR BC LR

FRGC V 2.0 Exp. 1 94.11 97.62 97.12 96.33 99.45

Exp. 2 95.10 99.54 98.31 97.1 100
Exp. 4 96.21 98.33 97.23 97.67 100
UFI CI 92.23 93.60 95.45 95.54 97.34
LI 93.12 95.29 96.23 96.11 98.88
SDUMLA-HMT 95.13 98.89 98.89 96.42 100

highest-ranking approach was adopted for comparing the performance of the pro-
posed multi-biometric face recognition system with other existing approaches, due
to its effciency compared to other fusion methods in exploiting the strength of each
classifer effectively and breaking the ties between the subjects in the fnal ranking
list. Table 4.6 compares the Rank-1 identifcation rates of the proposed approaches
and the current state-of-the-art face recognition approaches using the three experi-
ments with the FRGC V 2.0 dataset.
Although some approaches, such as the Partial Least Squares (PLS) approach
[65] achieved a slightly higher identifcation rate in experiments 1 and 2, they
obtained inferior results in the remaining experiments with the FRGC V 2.0 data-
set. In addition, the FaceDeepNet system and the fusion results obtained from the
proposed approaches achieved higher identifcation rates on all the experiments with
the FRGC V 2.0 dataset. Table 4.7 compares the Rank-1 identifcation rates of the
proposed approaches and the current state-of-the-art face recognition approaches
116 Deep Learning in Computer Vision

TABLE 4.6
Comparison of the Proposed
Multi-Biometric Face Recognition
System with the State-of-the-Art
Approaches on FRGC V 2.0 Dataset
FRGC V 2.0

Approach Exp.1 Exp.2 Exp.4

Deep C2D-CNN [67] – – 97.86

PLS [65] 97.5 99.4 78.2
LBP+HDM[68] – – 63.7
FDT-DRBM 94.11 95.10 96.21
FaceDeepNet 97.62 99.54 98.33
Fusion (LR) 99.45 100 100

TABLE 4.7
Comparison of the Proposed
Multi-Biometric Face
Recognition System with the
State-of-the-Art Approaches
on UFI Dataset
UFI

Approach CI LI

E-LBP [69] 65.28 –

MGM+LBP [70] 51.07 –
POEMmask-50 [71] 68.93 –
M- BNCC [72] 74.55 –
LBPHS [62] 55.04 31.89
LDPHS [62] 50.25 29.43
POEMHS [62] 67.11 33.96
FS-LBP [62] 63.31 43.21
FDT-DRBM 92.23 93.12
FaceDeepNet 93.60 95.29
Fusion (LR) 97.34 98.88

using these two partitions of the UFI dataset. It can be seen that we were able to
achieve a higher Rank-1 identifcation rate using only the FDT-DRBM approach
compared with other existing approaches, and better results were also achieved using
the FaceDeepNet system and the fusion framework on both partitions of the UFI
dataset. Finally, the authors in [66] have proposed a multimodal biometric system
A Multi-biometric Face Recognition System 117

TABLE 4.8
Comparison of the Proposed
Multi-Biometric Face
Recognition System with the
State-of-the-Art Approaches
on SDUMLA-HMT Dataset
Approach Accuracy Rate

MLP [66] 93.35

CLVQ [66] 96.54
CRBF[66] 92.25
FDT-DRBM 95.13
FaceDeepNet 98.89
Fusion (LR) 100

using only the face trait in the SDUMLA-HMT dataset. Therefore, for the purpose of
comparison, the best results obtained on the SDUMLA-HMT dataset using the high-
est-ranking approach were employed, and the results are listed in Table 4.8. It can
be seen that the accuracy rate of 100% achieved using the proposed system is higher
than the best results reported in [66], which is 96.54 % using the CLVQ approach.

4.5.3 FACE VERIFICATION EXPERIMENTS

In this section, LFW datasets were used to evaluate the robustness and effective-
ness of the proposed face recognition approaches to address the unconstrained face
verifcation problem. The face images in the LFW dataset were separated into two
distinct views. “View-1” is used for selecting and fne-tuning the parameters of the
face recognition model, while “View-2” is used to report the fnal performance of the
selected recognition model. In “View-2,” the face images are paired into 6,000 pairs,
with 3,000 pairs labeled as positive pairs, and the rest as negative pairs. As described
in [63], the fnal performance is reported by calculating the mean accuracy rate (m̂ m)
and the standard error of the mean accuracy (SE) over 10-fold cross-validation, with
300 positive and 300 negative image pairs per fold. For a fair comparison between
all face recognition approaches, six evaluation protocols were pre-defned by the
creators of the LFW dataset, as described in [73].
In these experiments, the “Image-Restricted, Label-Free Outside Data” pro-
tocol is followed, where only the outside data is used to train the proposed deep
learning approaches. Additionally, the aligned LFW-a* dataset is used, and the face
images were resized to (64×64) pixels after the face region has been detected using
pre-trained Viola-Jones† face detector. No data augmentation or outside data (e.g.,

* http://www.openu.ac.il/home/hassner/data/lfwa/
† Incorrect face detection results were assessed manually to ensure that all the subjects were included in

the subsequent evaluation of the proposed approaches.

118 Deep Learning in Computer Vision

creating additional positive/negative pairs from any other sources) was used in the
training phase of the proposed FDT-DRBM approach. The general fow of infor-
mation, when the proposed multi-biometric face recognition system operates in the
verifcation mode, is shown in Figure 4.12. The feature representations, fx and fy, of
a pair of two images, Ix and Iy, are obtained frstly by applying FDT approach, and
then a feature vector F for this pair is formed using element-wise multiplication (F =
fx ⊙ fx). Finally, this vector F (e.g., extracted from pairs of images) is used as input
data to the DBN and DRBM to learn additional feature representations and perform
face verifcation in the last layer, respectively. The fnal performance is reported in
Table 4.9 by calculating the mean accuracy rate (m̂ m ) and the standard error of the
mean accuracy (SE) over 10-fold cross-validation using different score-level fusion
methods, and the corresponding receiver operating characteristic (ROC) curves are
shown in Figure 4.13. From Table 4.9 and Figure 4.13, the highest accuracy rate was
obtained using the weighted sum rule as a fusion mothed in the score level, where a
higher weight was assigned to the proposed FaceDeepNet system compared to the
FDT-DRBM approach. The accuracy rate has been improved by 4.5% and 0.98%
compared to the FDT-DRBM approach and the FaceDeepNet system, respectively.
The proposed FDT-DRBM approach was compared to the current state-of-the-art
approaches on the LFW dataset, such as Curvelet-Fractal [7], DDML [34], LBP, Gabor
[74], and MSBSIF-SIEDA [75] using the same evaluation protocol (e.g., Restricted), as
shown in Table 4.10. It can be seen that the Curvelet-Fractal [7] has achieved slightly a
higher accuracy rate, but the proposed FDT-DRBM approach was able to achieve com-
petitive results with the other state-of-the-art face verifcation results reported on the
LFW dataset. The performance of the proposed face recognition approaches were also
compared with the current state-of-the-art deep learning approaches, such as DeepFace
[31], DeepID [76], ConvNet-RBM [77], ConvolutionalDBN [78], MDFR Framework
[7], and DDML [34]. The frst three approaches were mainly trained using a differ-
ent evaluation protocol, named the “Unrestricted, Labeled Outside Data” protocol,

FIGURE 4.12 Illustration of the proposed multi-biometric face recognition system operat-
ing in the verifcation mode.
A Multi-biometric Face Recognition System 119

TABLE 4.9
Performance Comparison between the Proposed Face Recognition
Approaches on LFW Dataset Using Different Score Fusion Methods
Acc. ( m
m̂ ± S E )
FDT-DRBM FaceDeepNet SR WSR PR Max Min

0.9542 ± 0.9899 ± 0.9918 ± 0.9992 ± 0.9837± 0.9758 ± 0.9880 ±

0.0192 0.0210 0.0110 0.0232 0.0164 0.0177 0.0211

FIGURE 4.13 ROC curves using different score-level fusion methods.

in which a private dataset consisting of a large number of training samples (>100K) is

used. As shown in Table 4.10, the accuracy rate has been further improved by 0.16%
and 1.09% compared to the next highest results reported by MDFR Framework [7]
using the FaceDeepNet system and the whole multi-biometric face recognition system
(e.g., using the weighted sum rule (WSR)). These promising results demonstrate the
good generalization ability of the proposed face recognition approaches, and their fea-
sibility for deployment in real-life applications.

4.6 CONCLUSIONS
In this chapter, a multi-biometric face recognition approach based on multimodal
deep learning termed FDT-DRBM is proposed. The approach is based on integrat-
ing the advantages of the fractal dimension and the DRBM classifer depending on
120 Deep Learning in Computer Vision

TABLE 4.10
Performance Comparison between the Proposed
Approaches and the State-of-the-Art Approaches
on LFW Dataset under Different Evaluation
Protocols

Approach Acc. ( m
m̂ ± S E ) Protocol

DeepFace [31] 0.9735 ±0.0025 Unrestricted

DeepID [76] 0.9745 ± 0.0026 Unrestricted
ConvNet-RBM[77] 0.9252 ± 0.0038 Unrestricted
VGG Net [79] 98.95 Unrestricted
ConvolutionalDBN [78] 0.8777 ± 0.0062 Restricted
DDML [34] 0.9068 ± 0.0141 Restricted
VMRS [80] 0.9110 ± 0.0059 Restricted
HPEN+HD-LBP+ DDML [74] 0.9257 ± 0.0036 Restricted
HPEN+HD-Gabor+ DDML [74] 0.9280 ± 0.0047 Restricted
Sub-SML+Hybrid+LFW3D [81] 0.9165 ± 0.0104 Restricted
MSBSIF-SIEDA [75] 0.9463 ± 0.0095 Restricted
3-layers DBN [7] 0.9353 ± 0.0165 Restricted
Curvelet-Fractal [7] 0.9622 ± 0.0272 Restricted
MDFR Framework [7] 0.9883 ± 0.0121 Restricted
FDT-DRBM 0.9542 ± 0.0192 Restricted
FaceDeepNet 0.9899 ± 0.0210 Restricted
Fusion (WSR) 0.9992 ± 0.0232 Restricted

extracting the face texture roughness and fuctuations in the surface under uncon-
strained environmental conditions and effciently modeling the joint distribution of
inputs and target classes, respectively. Furthermore, a novel FaceDeepNet system is
proposed to learn additional and complementary facial feature representations by
training two discriminative deep learning approaches (i.e., CNN and DBN) on top
of the local feature representations obtained from the FDT-DRBM approach. The
proposed approach was tested on four large-scale unconstrained face datasets (i.e.,
SDUMLA-HMT, FRGC V 2.0, UFI, and LFW) with high diversity in facial expres-
sions, illumination conditions, pose, noise, etc. A number of extensive experiments
were conducted, and a new state-of-the-art accuracy rate was achieved for both the
face identifcation and verifcation tasks by applying the proposed FaceDeepNet sys-
tem and the whole multi-biometric face recognition system (e.g., using the WSR)
on all the employed datasets. The obtained results demonstrate the reliability and
effciency of the FDT-DRBM approach by achieving competitive results with the
current state-of-the-art face recognition approaches (e.g., CLVQ, Curvelet-Fractal,
DeepID, etc.). For future work, it would be necessary to validate further the eff-
ciency and reliability of the proposed multi-biometric system using a larger multi-
modal dataset, containing more individuals with face images captured under more
challenging conditions.
A Multi-biometric Face Recognition System 121

BIBLIOGRAPHY
1. K. Nandakumar, A. Ross, and A. K. Jain, “Introduction to multibiometrics,” in
Appeared in Proc of the 15th European Signal Processing Conference (EUSIPCO)
(Poznan Poland), 2007, pp. 271–292.
2. A. S. Al-Waisy, “Ear identifcation system based on multi-model approach,”
International Journal of Electronics Communication and Computer Engineering, vol.
3, no. 5, pp. 2278–4209, 2012.
3. M. S. Al-ani and A. S. Al-Waisy, “Milti-view face datection based on kernel principal
component analysis and kernel,” International Journal on Soft Computing (IJSC), vol.
2, no. 2, pp. 1–13, 2011.
4. A. S. Al-Waisy, R. Qahwaji, S. Ipson, and S. Al-Fahdawi, “A fast and accurate Iris local-
ization technique for healthcare security system,” in IEEE International Conference on
Computer and Information Technology; Ubiquitous Computing and Communications;
Dependable, Autonomic and Secure Computing; Pervasive Intelligence and
Computing, 2015, pp. 1028–1034.
5. R. Jafri and H. R. Arabnia, “A survey of face recognition techniques,” Journal of
Information Processing Systems, vol. 5, no. 2, pp. 41–68, 2009.
6. C. Ding, J. Choi, D. Tao, and L. S. Davis, “Multi-directional multi-level dual-cross pat-
terns for robust face recognition,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 38, no. 3, pp. 518–531, 2016.
7. A. S. Al-Waisy, R. Qahwaji, S. Ipson, and S. Al-Fahdawi, “A multimodal deep learning
framework using local feature representations for face recognition,” Machine Vision
and Applications, vol. 29, no. 1, pp. 35–54, 2017.
8. A. S. Al-Waisy, R. Qahwaji, S. Ipson, and S. Al-fahdawi, “A multimodal bio-
metric system for personal identifcation based on deep learning approaches,” in
Seventh International Conference on Emerging Security Technologies (EST), 2017,
pp. 163–168.
9. A. Ross, K. Nandakumar, and J. K. Anil, “Handbook of multibiometrics,” Journal of
Chemical Information and Modeling, vol. 53, no. 9, pp. 1689–1699, 2006.
10. A. Ross and A. K. Jain, “Multimodal biometrics : An overview,” in 12th European
Signal Processing Conference IEEE, 2004, pp. 1221–1224.
11. A. Lumini and L. Nanni, “Overview of the combination of biometric matchers,”
Information Fusion, vol. 33, pp. 71–85, 2017.
12. L. Deng and D. Yu, “Deep learning methods and applications,” Signal Processing, vol.
28, no. 3, pp. 198–387, 2013.
13. J. Liu, C. Fang, and C. Wu, “A fusion face recognition approach based on 7-layer deep
learning neural network,” Journal of Electrical and Computer Engineering, vol. 2016,
pp. 1–7, 2016.
14. P. Fousek, S. Rennie, P. Dognin, and V. Goel, “Direct product based deep belief net-
works for automatic speech recognition,” in Proceedings of the IEEE International
Conference on Acoustics, Speech and Signal Processing, ICASSP, 2013, pp.
3148–3152.
15. H. Lee, P. Pham, Y. Largman, and A. Ng, “Unsupervised feature learning for audio
classifcation using convolutional deep belief networks,” in Advances in Neural
Information Processing Systems Conference, 2009, pp. 1096–1104.
16. R. Sarikaya, G. E. Hinton, and A. Deoras, “Application of deep belief networks for
natural language understanding,” IEEE Transactions on Audio, Speech and Language
Process., vol. 22, no. 4, pp. 778–784, 2014.
17. C. Ding and D. Tao, “A comprehensive survey on pose-invariant face recognition,”
ACM Transactions on Intelligent Systems and Technology (TIST), vol. 7, no. 3, pp.
1–40, 2016.
122 Deep Learning in Computer Vision

18. S. Biswas, G. Aggarwal, P. J. Flynn, and K. W. Bowyer, “Pose-robust recognition of

low-resolution face images,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 35, no. 12, pp. 3037–3049, 2013.
19. D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International
Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004.
20. L. Sirovich and M. Kirby, “Low-dimensional procedure for the characterization
of human faces.,” Journal of the Optical Society of America. A, Optics and Image
Science, vol. 4, no. 3. pp. 519–524, 1987.
21. K. Meethongjan and D. Mohamad, “A summary of literature review : Face recogni-
tion,” Review Literature and Arts of the Americas, vol. 2007, no. July, pp. 1–12, 2007.
22. M. Fischer, H. K. Ekenel, and R. Stiefelhagen, “Analysis of partial least squares for pose-
invariant face recognition,” in IEEE Fifth International Conference on Biometrics:
Theory, Applications and Systems (BTAS), 2012, pp. 331–338.
23. T. Berg and P. Belhumeur, “Tom-vs-Pete classifers and identity-preserving alignment
for face verifcation,” in Procedings of the British Machine Vision Conference, 2012,
pp. 129.1–129.11.
24. A. Li, S. Shan, and W. Gao, “Coupled bias – variance tradeoff for cross-pose face rec-
ognition,” IEEE Transactions on Image Processing, vol. 21, no. 1, pp. 305–315, 2012.
25. Z. Zhu, P. Luo, X. Wang, and X. Tang, “Multi-view perceptron : A deep model for
learning face identity and view representations,” Advances in Neural Information
Processing Systems, pp. 1–9, 2014.
26. A. S. Al-Waisy, “Detection and recognition of human faces based on hybrid techniques,”
International Journal of Applied Computing (IJAC), vol. 5, no. 2, pp. 115–126, 2012.
27. M. S. Al-Ani and A. S. Al-waisy, “Face recognition approach based on wavelet-curvelet
technique,” Signal & Image Processing: An International Journal (SIPIJ), vol. 3, no. 2,
pp. 21–31, 2012.
28. Y. Sun, X. Wang, and X. Tang, “Deep learning face representation by joint identifcation-
verifcation,” Advances in Neural Information Processing Systems, pp. 1988–1996, 2014.
29. L. Deng, “Three classes of deep learning architectures and their applications: A tutorial
survey,” APSIPA Transactions on Signal and Information Processing, 2013.
30. Y. Sun, X. Wang, and X. Tang, “Hybrid deep learning for face verifcation,” in
Proceedings of the IEEE International Conference on Computer Vision, 2013, pp.
1489–1496.
31. Y. Taigman, M. A. Ranzato, T. Aviv, and M. Park, “DeepFace: Closing the gap to
human-level performance in face verifcation,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, 2014, pp. 1701–1708.
32. C. Ding and D. Tao, “Robust face recognition via multimodal deep face representa-
tion,” IEEE Transactions on Multimedia, vol. 17, no. 11, pp. 2049–2058, 2015.
33. J. V. Davis, B. Kulis, P. Jain, S. Sra, and I. S. Dhillon, “Information-theoretic met-
ric learning,” in in Proceedings of the 24th International Conference on Machine
Learning, 2007, pp. 209–216.
34. J. Hu, J. Lu, and Y. P. Tan, “Discriminative deep metric learning for face verifcation
in the wild,” in Proceedings of the IEEE Computer Society Conference on Computer
Vision and Pattern Recognition, 2014, pp. 1875–1882.
35. Z. Cui, W. Li, D. Xu, S. Shan, and X. Chen, “Fusing robust face region descriptors via mul-
tiple metric learning for face recognition in the wild,” in Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern Recognition, 2013, pp. 3554–3561.
36. M. Guillaumin, J. Verbeek, C. Schmid, M. Guillaumin, J. Verbeek, C. Schmid, M.
Guillaumin, J. Verbeek, C. Schmid, and L. J. Kuntzmann, “Is that you ? Metric learn-
ing approaches for face identifcation To cite this version,” in IEEE 12th International
Conference on Computer Vision, 2009, pp. 498–505.
A Multi-biometric Face Recognition System 123

37. M. Kostinger, M. Hirzer, P. Wohlhart, P. M. Roth, and H. Bischof, “Large scale metric
learning from equivalence constraints,” in Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, 2012, pp. 2288–2295.
38. H. V Nguyen and L. Bai, “Cosine Similarity Metric Learning for Face Verifcation,” in
Asian Conference on Computer Vision. Springer Berlin Heidelberg, 2011, pp. 1–12.
39. C. Peng, X. Gao, N. Wang, and J. Li, “Graphical representation for heterogeneous
face recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 39, no. 2, pp. 301–312, 2017.
40. C. Peng, N. Wang, J. Li, and X. Gao, “DLFace: Deep local descriptor for cross-modal-
ity face recognition,” Pattern Recognition, vol. 90, pp. 161–171, 2019.
41. C. Peng, X. Gao, S. Member, N. Wang, and D. Tao, “Multiple representations-based
face sketch – photo synthesis,” IEEE Transactions on Neural Networks and Learning
Systems, vol. 27, no. 11, pp. 2201–2215, 2016.
42. Y. Gao, J. Ma, and A. L. Yuille, “Semi-supervised sparse representation based clas-
sifcation for face recognition with insuffcient labeled samples,” IEEE Transactions on
Image Processing, vol. 26, no. 5, pp. 2545–2560, 2017.
43. M. Hassaballah and S. Aly, “Face recognition: Challenges, achievements and future
directions,” IET Computer Vision, vol. 9, no. 4, pp. 614–626, 2015.
44. R. Lopes and N. Betrouni, “Fractal and multifractal analysis: A review, Medical Image
Analysis, vol. 13, no. 4, pp. 634–649, 2009.
45. B. Mandelbrot, The Fractal Geometry of Nature. Library of Congress Cataloging in
Publication Data, United States of America, 1983.
46. B. Mandelbrot, “Self-affnity and fractal dimension,” Physica Scripta, vol. 32, 1985,
pp. 257–260.
47. K. Lin, K. Lam, and W. Siu, “Locating the human eye using fractal dimensions,”
in Proceedings of International Conference on Image Processing, 2001, pp.
1079–1082.
48. M. H. Farhan, L. E. George, and A. T. Hussein, “Fingerprint identifcation using frac-
tal geometry,” International Journal of Advanced Research in Computer Science and
Software Engineering, vol. 4, no. 1, pp. 52–61, 2014.
49. G. Hinton, “A practical guide to training restricted boltzmann machines a practical
guide to training restricted boltzmann machines,” Neural Networks: Tricks of the
Trade. Springer, Berlin, Heidelberg, 2010, pp. 599–619.
50. H. Larochelle and Y. Bengio, “Classifcation using discriminative restricted Boltzmann
machines,” in Proceedings of the 25th International Conference on Machine Learning,
2008, pp. 536–543.
51. B. Abibullaev, J. An, S. H. Jin, S. H. Lee, and J. Il Moon, “Deep machine learning
a new frontier in artifcial intelligence research,” Medical Engineering and Physics,
vol. 35, no. 12, pp. 1811–1818, 2013.
52. G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for deep belief
nets,” Neural Computation, vol. 18, no. 7, pp. 1527–1554, 2006.
53. H. Khalajzadeh, M. Mansouri, and M. Teshnehlab, “Face recognition using convo-
lutional neural network and simple logistic classifer,” Soft Computing in Industrial
Applications. Springer International Publishing, pp. 197–207, 2014.
54. Y. Bengio, “Learning deep architectures for AI,” Foundations and Trends® in Machine
Learning, vol. 2, no. 1, pp. 1–127, 2009.
55. A. S. Al-Waisy, R. Qahwaji, S. Ipson, S. Al-Fahdawi, and T. A. M. Nagem, “A multi-
biometric iris recognition system based on a deep learning approach,” Pattern Analysis
and Applications, vol. 21, no. 3, pp. 783–802, 2018.
56. P. Viola, O. M. Way, and M. J. Jones, “Robust real-time face detection,” International
Journal of Computer Vision, vol. 57, no. 2, pp. 137–154, 2004.
124 Deep Learning in Computer Vision

57. A. R. Chowdhury, T.-Y. Lin, S. Maji, and E. Learned-Miller, “One-to-many face recog-
nition with bilinear CNNs,” in IEEE Winter Conference on Applications of Computer
Vision (WACV), 2016, pp. 1–9.
58. D. Wang, C. Otto, and A. K. Jain, “Face search at scale: 80 million gallery,” arXiv
Prepr. arXiv1507.07242, pp. 1–14, 2015.
59. G. E. Hinton, “Training products of experts by minimizing contrastive divergence,”
Neural Computation, vol. 14, no. 8, pp. 1771–1800, 2002.
60. Y. Yin, L. Liu, and X. Sun, “SDUMLA-HMT: A multimodal biometric database,”
in Chinese Conference Biometric Recognition, Springer-Verlag Berlin Heidelberg,
pp. 260–268, 2011.
61. P. J. Phillips, P. J. Flynn, T. Scruggs, K. W. Bowyer, J. Chang, K. Hoffman, J.
Marques, J. Min, and W. Worek, “Overview of the face recognition grand challenge,”
in Proceedings IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, CVPR, 2005, vol. I, pp. 947–954.
62. L. Lenc and P. Král, “Unconstrained facial images: Database for face recognition
under real-world conditions,” Lecture Notes in Computer Science (including subseries
Lecture Notes in Artifcial Intelligence and Lecture Notes in Bioinformatics), vol. 9414,
pp. 349–361, 2015.
63. G. B. Huang, M. Mattar, T. Berg, and E. Learned-miller, “Labeled faces in the wild:
A database for studying face recognition in unconstrained environments,” in Technical
Report 07-49, University of Massachusetts, Amherst, 2007, pp. 1–14.
64. A. S. Al-Waisy, R. Qahwaji, S. Ipson, and S. Al-Fahdawi, “A robust face recognition
system based on curvelet and fractal dimension transforms,” in IEEE International
Conference on Computer and Information Technology; Ubiquitous Computing
and Communications; Dependable, Autonomic and Secure Computing; Pervasive
Intelligence and Computing, 2015, pp. 548–555.
65. W. R. Schwartz, H. Guo, and L. S. Davis, “A robust and scalable approach to face iden-
tifcation,” in European Conference on Computer Vision, 2010, pp. 476–489.
66. M. Y. Shams, A. S. Tolba, and S. H. Sarhan, “A vision system for multi-view face rec-
ognition,” International Journal of Circuits, Systems and Signal Processing, vol. 10,
pp. 455–461, 2016.
67. J. Li, T. Qiu, C. Wen, K. Xie, and F.-Q. Wen, “Robust face recognition using the deep
C2D-CNN model based on decision-level fusion,” Sensors, vol. 18, no. 7, pp. 1–27,
2018.
68. J. Holappa, T. Ahonen, and M. Pietikäinen, “An optimized illumination normalization
method for face recognition,” in IEEE Second International Conference on Biometrics:
Theory, Applications and Systems, 2008, pp. 1–6.
69. P. Král, L. Lenc, and A. Vrba, “Enhanced local binary patterns for automatic face rec-
ognition,” arXiv Prepr. arXiv1702.03349, 2017.
70. J. Lin and C.-T. Chiu, “Lbp edge-mapped descriptor using mgm interest points for
face recognition,” in IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), 2017, pp. 1183–1187.
71. L. Lenc, “Genetic algorithm for weight optimization in descriptor based face recog-
nition methods,” in Proceedings of the 8th International Conference on Agents and
Artifcial Intelligence. SCITEPRESS-Science and Technology Publications, 2016, pp.
330–336.
72. J. Gaston, J. Ming, and D. Crookes, “Unconstrained face identifcation with multi-scale
block-based correlation,” in IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), 2017, pp. 1477–1481.
73. G. B. Huang and E. Learned-miller, “Labeled faces in the wild : Updates and new
reporting procedures,” Department of Computer Science, University of Massachusetts
Amherst, Amherst, MA, USA, Technical Report, pp. 1–14, 2014.
A Multi-biometric Face Recognition System 125

74. X. Zhu, Z. Lei, J. Yan, D. Yi, and S. Z. Li, “High-fdelity pose and expression normal-
ization for face recognition in the wild,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2015, pp. 787–796.
75. A. Ouamane, M. Bengherabi, A. Hadid, and M. Cheriet, “Side-information based
exponential discriminant analysis for face verifcation in the wild,” in 11th IEEE
International Conference and Workshops on Automatic Face and Gesture Recognition
(FG), 2015, pp. 1–6.
76. Y. Sun, X. Wang, and X. Tang, “Deep learning face representation from predict-
ing 10,000 classes,” in Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, 2014, pp. 1891–1898.
77. Y. Sun, X. Wang, and X. Tang, “Hybrid deep learning for face verifcation,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 10, pp. 1997–
2009, 2016.
78. G. B. Huang, H. Lee, and E. Learned-Miller, “Learning hierarchical representations
for face verifcation with convolutional deep belief networks,” in Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
2012, pp. 2518–2525.
79. O. M. Parkhi, A. Vedaldi, and A. Zisserman, “Deep face recognition,” in British
Machine Vision Conference, 2015, pp. 1–12.
80. O. Barkan, J. Weill, L. Wolf, and H. Aronowitz, “Fast high dimensional vector multi-
plication face recognition,” in Proceedings of the IEEE International Conference on
Computer Vision, 2013, pp. 1960–1967.
81. T. Hassner, S. Harel, E. Paz, and R. Enbar, “Effective face frontalization in uncon-
strained images,” in IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2015, pp. 4295–4304.
 5 Deep LSTM-Based
Sequence Learning
Approaches for Action
and Activity Recognition
Amin Ullah, Khan Muhammad, Tanveer Hussain,
Miyoung Lee, and Sung Wook Baik

CONTENTS
5.1 Introduction .................................................................................................. 127
5.1.1 Recurrent Neural Network (RNN) ................................................... 129
5.1.2 Long Short-Term Memory (LSTM) ................................................. 130
5.1.3 Multi-layer LSTM............................................................................. 131
5.1.4 Bidirectional LSTM.......................................................................... 131
5.1.5 Gated Recurrent Unit........................................................................ 132
5.2 Action and Activity Recognition Methods ................................................... 132
5.3 Results and Discussion ................................................................................. 139
5.4 Future Research Directions .......................................................................... 143
5.4.1 Benefting from Image Models......................................................... 144
5.4.2 Multiple Actions and Activity Recognition ...................................... 145
5.4.3 Utilization of Large-Scale Datasets.................................................. 145
5.5 Conclusion .................................................................................................... 145
Acknowledgment ................................................................................................... 145
Bibliography .......................................................................................................... 146

5.1 INTRODUCTION
Human action and activity recognition is one of the challenging areas in computer
vision due to the increasing number of cameras installed everywhere, which enables
its numerous applications in various domains. The enormous amount of data gener-
ated through these cameras requires human monitoring for identifcation of differ-
ent activities and events. Smarter surveillance, through which normal and abnormal
activities can be automatically identifed using artifcial intelligence and computer
vision technologies, is needed in this era. It has vast demand due to its importance
in real-life problems like detection of abnormal and suspicious activities, intelli-
gent video surveillance, elderly care, patient monitoring in healthcare centers, and

127
128 Deep Learning in Computer Vision

video retrieval based on different actions [1, 2]. In the literature, many researchers
have analyzed action and activity recognition through various sensors data including
RGB, grayscale, night vision, Kinect depth, and skeleton data [3].
Smartphone and smartwatch sensors are also utilized by some researchers for
activity analysis and ftness applications [4]. Categorizations of action and activity
recognition methods using available sensors are given in Figure 5.1. In this chapter,
we will study only vision RGB-based action and activity recognition methods in
which long short-term memory (LSTM) has been deployed for sequence learning.
Processing visual contents of video data for activity recognition is very challeng-
ing due to the inter-similarity of visual and motion patterns, changes in viewpoint,
camera motion in sports videos, scale, action poses, and different illumination con-
ditions [5].
Recently, convolutional neural network (CNN) outperformed state-of-the-art
techniques in various domains such as image classifcation, object detection, and
localization. However, an activity is recorded in sequences of images, so process-
ing a single image surely cannot give us good results to capture the whole idea of
the activity. Therefore, in videos, the whole activity can be recognized by analyzing
the motions and other visual changes of human body parts like hands and legs in a
sequence of frames. For instance, jumping for a headshot in football and skipping
rope have the same action pose in the initial frame. However, the difference between
these actions can be easily recognized in a sequence of frames. Another challenge for
analyzing long-term activities in video data is that each activity has parts that overlap
with other activities; for example, diving for swimming is one activity, but whenever
the swimmer dives into the pool he also jumps in the air, which is another activity.
So, when two activities like jumping and diving appear in the same sequence, it
can lead to false predictions because the sequence is interrupted. To address these
issues, many researchers [6, 7] have introduced different kinds of techniques to rep-
resent sequences and learn those sequences for activity recognition. For instance, the
current literature on human activity recognition highlights some popular methods

FIGURE 5.1 Categorization of action and activity recognition using available sensor data.
Deep LSTM-Based Sequence Learning Approaches 129

including volume motion templates [6], spatio-temporal features [7], motion history
volume [8], interest point descriptors [9], bag-of-words [10], optical fow trajectories
[11], and CNN-based learned features [12, 13] for sequence representation followed
by some machine learning algorithms such as support vector machine (SVM), deci-
sion tree, recurrent neural network (RNN), and LSTM. However, these methods
are not enough for sequence representation and learning. Recent studies prove that
the LSTM is one of the most effective sources of sequence learning. Therefore, in
this chapter, we will explore sequence learning concepts integrated with LSTM for
action and activity recognition.

5.1.1 RECURRENT NEURAL NETWORK (RNN)

RNNs are a type of artifcial neural network that has turned out to be more popular
in recent years because of its power to learn time series data. The RNN is a special
network that consists of recurrent connections, unlike traditional feedforward neural
networks. The major beneft is that with its network of recurrent connections, it is
able to process arbitrary time series sequences of input data. RNNs are presented
for investigating hidden sequential patterns in both temporal sequential and spatial
sequential data [14]. A standard RNN is shown in Figure 5.2 (a). It processes data by
taking input at different time steps over time. At each time step, it has two inputs: the
frst input comes from the hidden state of the previous RNN, and the second is from
the time series data.
From Figure 5.2 (a) it can be seen that RNN is taking input at each time step and
giving output as well. This process of taking input at each recurrent step disturbs
the patterns for sequences in time series data, because the neural network performs
multiplication of weights with data and addition with bias terms. Therefore, mul-
tiplication of data many times with several hidden states causes the effect of early
sequential data in the series to be forgotten. This issue, known as the vanishing
gradient problem, is shown in Figure 5.2 (b), where the effect of data is shown as
gray shading, which decreases when the sequence is going long. In practice, the
range of sequential information that standard RNNs access is limited to approxi-
mately ten time steps between the relevant input and target output. Video data is also
sequential where changes in visual contents are represented in many frames such

FIGURE 5.2 (a) A standard unrolled RNN network; (b) vanishing gradient problem
in RNN.
130 Deep Learning in Computer Vision

that a sequence of frames helps in understanding the context of an action. In long-

term sequences, RNNs can interpret such sequences but forget the earlier inputs of
the sequence.

5.1.2 LONG SHORT-TERM MEMORY (LSTM)

The vanishing gradient problem can be solved by utilizing a special type of RNN
known as LSTM [15]. It is capable of learning long-term sequential dependencies. Its
special structure, with input, output, and forget gates, controls the long-term sequen-
tial pattern identifcation in time series data. The gates are adjusted by a sigmoid
function connected to each unit, which learns during training those states in which
it needs to be open and those in which it should be closed. Eq. (1.1) to Eq. (1.7) [16]
explain the operations performed in the LSTM unit, where xt is the input at time t,
where input is fed in the form of chunks from the long sequence. ft is the forget gate
at time t, which clears information from the memory cell when needed and keeps
a record of the previous frame, whose information needs to be cleared from the
memory. The output gate ot keeps information about the upcoming step, where g is
the recurrent unit with activation function “tanh” and is computed from the input
of the current frame and the state of the previous frame st-1. The hidden state of an
RNN step is calculated through tanh activation and memory cell ct. The output of
each state can be used in other tasks; however, action recognition does not need
the intermediate output of the LSTM, so the fnal decision is made by applying the
Softmax classifer to the fnal state of the LSTM network, which gives probabilities
for the predicted activity.

(
it = s ( xt + st −1 ) W i + bi ) (5.1)

(
ft = s ( xt + st −1 ) W f + b f ) (5.2)

(
ot = s ( xt + st −1 ) W o + bo ) (5.3)

(
g = tanh ( xt + st −1 ) W g + bg ) (5.4)

ct = ct −1 ˛ ft + g ˛it (5.5)

st = tanh(ct )° ot (5.6)

Predictions = soft max(Vst ) (5.7)

Recent studies have proven [17–21] that huge amounts of data increase the accuracy
of a machine learning task. Therefore, learning complex sequential patterns in video
data of large-scale datasets is very challenging, and a single LSTM is not effective
enough to do it. Therefore, researchers have recently utilized multi-layer LSTM,
which is constructed by stacking multiple LSTM cells to learn long-term dependen-
cies in video data.
Deep LSTM-Based Sequence Learning Approaches 131

FIGURE 5.3 Two-layer LSTM network [22].

5.1.3 MULTI-LAYER LSTM

The performance of most deep neural networks can be boosted in terms of accuracy
by increasing the number of layers in the neural network models. The same strategy
is followed in LSTM by stacking multiple LSTM layers to the network. By add-
ing deep layers, it captures a higher level of sequence information [16]. In standard
RNN, data is fed to a single layer for activation and processing before output, but in
time sequence problems, we need to process data on several layers to capture hier-
archical patterns. By stacking LSTM layers, each layer in the network is a hierarchy
that receives the hidden state of the previous layer as input. Figure 5.3 shows a two-
layer LSTM. Layer 1 receives input from data xt, while the input of layer 2 is from
its previous time step st-1(2), the output of the current time step of layer 1, st(1). The
computation of the multi-layer LSTM cell is the same as simple LSTM cell, but only
each layer’s information is added to the superscript of each it, ft, ot, ct, and st. The
procedure for calculating the state of the layer is as follows:

stl = tanh (ctl ) ˝ otl (5.8)

5.1.4 BIDIRECTIONAL LSTM

In bidirectional LSTM, the output at time t is dependent not only on the previous
frames in the sequence, but also on the upcoming frames [23]. Bidirectional LSTMs
are quite simple, having two recurrent units, one stacked on top of the other. One
recurrent unit goes in the forward direction and the other goes in the backward
direction. The combined output is then computed based on the hidden state of both
recurrent units. The input data is fed to the bidirectional LSTM, and the hidden
states of forward pass and backward pass are combined in the output layer. The cost
132 Deep Learning in Computer Vision

function is adjusted after the output layer, weights, and biases are adjusted through
back-propagation.

5.1.5 GATED RECURRENT UNIT

Two dominant variants of RNNs have been introduced by researchers: LSTM [22]
and GRU [24]. The LSTM contains gated recurrent units, including the input, for-
get, output gates, and memory cell, while the GRU consists of the reset gate, update
gate, and activation unit. The structure of the LSTM is more complex and contains
gates and memory cell, resulting in more computation to process a sequence. On the
other hand, GRU involves only two gates, which makes it applicable for processing
real-time sequential data. However, many researchers have reported LSTM as more
effective than GRU [24].

5.2 ACTION AND ACTIVITY RECOGNITION METHODS

Over the last decade, researchers have presented many handcrafted and deep-nets-
based approaches for action recognition. The earlier work was based on handcrafted
features for non-realistic actions, where an actor used to perform some actions in a
scene with a simple background. Such systems extract low-level features from the
video data and then feed them to a classifer such as SVM, decision tree, and k-Nearest
Neighbors (KNN) for action recognition. However, current methods utilize realistic
video data for action and activity recognition tasks. For instance, an action recogni-
tion framework is proposed by Ullah et al. [22], where they used CNN features fol-
lowed by deep bidirectional LSTM (DB-LSTM) for sequence learning. They extracted
features from every six frames of the input video to reduce the time complexity and
redundancy. The sequential information extracted from these frames is inputted to the
DB-LSTM network. Multiple layers in the DB-LSTM are stacked together in both
forward and backward pass in order to increase the network depth and make it capable
of learning long-term sequences. The framework of DB-LSTM is shown in Figure 5.4.
Another recent method proposed by Ullah et al. [1] is for activity recognition in
industrial surveillance environments. Their framework is visualized in Figure 5.5.
First, they divided the continuous surveillance video stream into important shots
using the proposed CNN-based human saliency features. Next, the temporal features
of activity in the sequence of frames are extracted by utilizing the convolutional lay-
ers of a FlowNet2 CNN model. Finally, a multi-layer LSTM is presented for learning
long-term sequences in the temporal optical fow features for activity recognition.
The method proposed by Li et al. [25] selected discriminative parts in spatial
dimension, as visualized in Figure 5.6. They split the videos into multiple clips with
fxed-length frames in the temporal direction and used the sliding windows to gen-
erate local cubes in the spatial direction. The cubes are extracted using the C3D
network. Then, they designed the part selection method to select useful cubes and
utilized the max-pooling function to generate the representation of clips. Finally, the
representation of clips was fed into two layers of the LSTM network to train the mul-
tiple LSTM and the Softmax layer, which yielded the prediction by using the feature
vector generated from the LSTM network.
Deep LSTM-Based Sequence Learning Approaches 133

FIGURE 5.4 The framework for action recognition in videos using CNN and deep bidirec-
tional LSTM [22].

An adaptive recurrent-convolutional hybrid network for long-term action recogni-

tion is introduced by Xin et al. [7]. Xin et al. introduced a framework learning from
static, non-static, and sequential mixed features to aim at different problems such as
domain variation and temporal diversities. They represented human actions by frst
reducing data based on cognitive techniques and a hybrid “network-over-network”
architecture. Li et al. in [4] present end-to-end sequence learning for action rec-
ognition based on video LSTM. To exploit the spatial correlations, they hardwired
convolutions in soft-Attention LSTM architecture. They also introduced motion-
based attention, as motion informs about action contents. The authors performed
experiments on UCF101, HMDB51, and THUMOS13 datasets to present the better
accuracy of their system. All of the statistics of the exiting LSTM-based action and
activity recognition models are briefy given in Table 5.1. A group activity recogni-
tion framework using temporal dynamics is presented by Ibrahim et al. [26], where
they used an LSTM-based deep model to capture these dynamics. Their framework
is shown in Figure 5.7. As dynamics of group activity can be inferred on the basis
of individual persons’ activities, the authors presented a two-stage deep temporal
model to recognize the group activity occurring in video sequences. The LSTM
model in their method is designed especially to represent action dynamics of indi-
vidual people in a sequence and the second LSTM model take the average of person-
level information in order to understand the group activity.
Ma et al. [33] presented a framework to improve accuracy by advancing the train-
ing process of temporal deep learning models for activity detection and recogni-
tion tasks. Their proposed framework is shown in Figure 5.8. Conventional methods
based on RNNs and LSTMs consider only classifcation errors. But in this method,
they studied two problems: activity detection and then activity recognition. To
detect activity, the authors recognized the category of the activity and its initial and
fnal end points and then classifed the activity as visualized in Figure 5.8. A novel
 134

FIGURE 5.5 The framework of activity recognition in industrial systems. In the frst stage, the surveillance stream is preprocessed using CNN-based
salient feature extraction mechanism for shot segmentation. The second stage extracts CNN-based temporal optical fow features from the sequence of
frames. Finally, multi-layer LSTM is presented for sequence learning and the recognition of activities [1].
Deep Learning in Computer Vision
Deep LSTM-Based Sequence Learning Approaches 135

FIGURE 5.6 The framework of selected spatio-temporal features via bidirectional LSTM
for action recognition [25].

approach based on the convolutional features of ResNet-101-v2 model is proposed

by Chen et al. [29], where the input video is divided into segments of equal dura-
tion. Their proposed framework is shown in Figure 5.9. Given an input video, they
divided it into T segments (T = 3 in this case for succinct illustration) of equal dura-
tion and randomly sampled one short snippet from its corresponding segment. The
short snippet is either a single RGB image (spatial stream) or a sequence of 10 optical
flow fields (temporal stream). After feeding the selected short snippets through the
ResNet-101-v2 conv1 layer to the ResNet-101-v2 conv5 2 layer, they had two sepa-
rated branches.
The local branch (top dotted box) captures local appearance and short-term
motion features from snippets and generates a video-level feature representation
using temporal average pooling. The non-local branch (bottom dotted box) is for
relation reasoning using the relational LSTM module. Eventually, they obtained an
overall video-level feature representation by concatenating feature vectors gener-
ated by the two branches, added one fully connected layer, and optimized it using a
standard Softmax with cross entropy classifcation loss. Recently, encoder-decoder
frameworks based on attention mechanisms have been applied to action recognition
and have produced results comparable to those obtained with state-of-the-art meth-
ods, but they have always faced a bottleneck when it comes to distinguishing similar
actions. To tackle this challenge, Yang et al. [36] proposed a novel framework based
on recurrent attention CNN (RACNN). RACNN integrates CNNs and LSTMs along
with an attention mechanism.
Figure 5.10 contains four major components: (1) convolutional feature extraction;
(2) bidirectional LSTM sequence model and spatial attention model; (3) temporal
attention model; and (4) sequence recognition. CNNs encode the input video frames
into feature sequences. AttLSTM means hierarchical LSTM with spatial attention
model. The output of bidirectional AttLSTM is designed to be integrated and sent
to the temporal attention model. After refining the input, the architecture takes the
output of the last LSTM unit and calculates the resulting features extracted, which
are encoded and forward propagated to the second module to seek spatial attention,
followed by another module for temporal attention, which is advanced to decoder
 136

TABLE 5.1
Description of LSTM-Based Approaches with Details about the Nature of the Activity
Sequence representation Sequence learning Nature of activity

CNN/Deep Handcrafted Bidirectional Multi-layer Individual Group action/

Method features features RNN LSTM LSTM LSTM action/activity activity

Ibrahim et al. [26] ✓ – – – – ✓ – ✓

Ullah et al. [22] ✓ – – – ✓ – ✓ –
Ullah et al. [1] ✓ – – ✓ – – ✓ –
Xin et al. [7] ✓ – ✓ – – – ✓ –
Li et al. [27] ✓ – – – – – ✓ –
Ma et al. [28] ✓ – ✓ ✓ – – ✓ –
Chen et al. [29] ✓ – – – ✓ ✓ –
Du et al. [30] – ✓ – ✓ – – ✓ –
Donahue et al. [31] ✓ – – ✓ – – ✓ –
Huang et al. [32] ✓ – – ✓ – – ✓ –
Ma et al. [33] ✓ – – ✓ – – ✓ –
Sun et al. [34] ✓ – – – – ✓ ✓ –
Li et al. [25] ✓ – – – – ✓ ✓ –
Veeriah et al. [35] – ✓ ✓ – – – ✓ –
Yang et al. [36] ✓ – – – ✓ – ✓ –
Deep Learning in Computer Vision
Deep LSTM-Based Sequence Learning Approaches 137

FIGURE 5.7 Group activity recognition via a hierarchical model [26]. Each person in a
scene is modeled using a temporal model that captures his/her dynamics; these models are
integrated into a higher-level model that captures scene-level activity.

FIGURE 5.8 Framework of early activity detection [33]. At each video frame, their model
frst computes deep CNN features (named as fc7), and then the features are fed into the LSTM
to compute detection scores of activities.

and sequence recognition for fnal action prediction. The LSTM used by the authors
is bidirectional with hierarchical structure, and they employ spatial-temporal atten-
tion in their method. Finally, they fnd the integrated spatio-temporal features that
are extracted using three-dimensional CNNs (3DCNNs) to advance the relationship
mined in each frame.
Similarly, there are many other methods that utilize sequence learning based
on RNN or LSTM with different confgurations to recognize different actions and
activities from video data. Furthermore, researchers (see [25] and [29–36]) have
138 Deep Learning in Computer Vision

FIGURE 5.9 Network architecture for relational LSTM for video action recognition [29].

FIGURE 5.10 The architecture of the RACNN model for action recognition.

introduced different types of LSTM and RNN confgurations for precise action and
activity recognition. Mainstream LSTM-based action recognition literature follows
convolutional features for the representation of sequences and extracts deep features
as a prerequisite step for different versions of LSTM or RNN. The sequence learn-
ing step for action recognition based on LSTM uses different favors of it, such as
bidirectional, multi-layer, and simple single layer. The majority of the action recog-
nition literature uses a simple LSTM structure. The discussed LSTM-based meth-
ods are limited to individual action or activity recognition and localization in video
frames. They process all the frames instead of taking different clues from individual
in frame based on simple LSTM to predict the activity for a group.
Deep LSTM-Based Sequence Learning Approaches 139

5.3 RESULTS AND DISCUSSION

In this section, we compare the results achieved by different action and activity rec-
ognition methods on benchmark datasets. Detailed comparisons of state-of-the-art
techniques are given in Table 5.2. The listed LSTM-based approaches are given
in chronological order, starting with the most recent frst. As can be seen from
Table 5.2, the latest approach [36] used recurrent attention CNN, incorporating CNNs
and LSTM with an attention mechanism. Using this method, the authors obtained
94.8% and 71.9% accuracy scores on UCF-101 [37] and HMDB-51 [38], respectively.
The CNN-based temporal optical fow features followed by multi-layer LSTM [1]
achieved accuracy scores of 94.45% and 72.1% on UCF-101 and HMDB51 datasets,
respectively. This method improved the computational performance of the activity
recognition system by utilizing the lightweight FlowNet2 CNN model for feature
extraction. Another approach [29] uses relational LSTM, which captures long- and
short-range relations among the objects in the video data. This work obtained accu-
racy scores of 94.8% and 71.9% on UCF-101 and HMBD-51, respectively. A list of
popular action and activity datasets is given in Table 5.3, and sample categories are
given in Figure 5.11. These datasets contain human actions collected from YouTube
and different other resources, presenting real-life action categories.
Ullah et al. [1] used the AlexNet CNN model for frame-level feature extraction
followed by DB-LSTM and obtained 91.21%, 87.64%, and 92.84% accuracy for UCF-
101, HMDB51, and YouTube action datasets, respectively. This method achieved
higher accuracy for the HMDB51 dataset. The differential RNN (dRNN) [35] used
a differential gating scheme with the LSTM network in order to consider the impact
of spatio-temporal correspondences to salient motion patterns for action recognition.
This method was evaluated on a very old and simple dataset, KTH2 [39], obtaining
92.12% accuracy. An end-to-end long-term RNN [31] reported 65% accuracy on

TABLE 5.2
Comparison of Different LSTM-Based Action and Activity
Recognition Methods on Different Challenging Benchmark Datasets
Method UCF-101 YouTube HMDB-51

3DCNNs and bidirectional hierarchical LSTM [36] 94.8% – 71.9%

Temporal optical fow with multi-layer LSTM [1] 94.45 95.8% 72.21
TS-LSTM and temporal inception [28] 91.1% – 69.0%
Relational LSTM [29] 94.8% – 71.4%
Deep bidirectional LSTM [22] 91.21% 92.84% 87.64%
Video LSTM [27] 92.2% – 73.3%
RSTAN [30] 95.1% – 76.9%
ARMA [32] 85.2% –
Bidirectional LSTM [25] 94.2% – 70.4%
Lattice-LSTM [34] 93.6% – 66.2%
ARCH [7] 85.3% – 58.2%
Long-term RNN [31] 65.4% – –
140 Deep Learning in Computer Vision

TABLE 5.3
List of Popular Benchmark Action and Activity
Recognition Datasets
Number Number of action/
Dataset of videos activity classes

UCF101 [37] 13,320 101

HMDB51 [38] 7,000 51
Hollywood2 [40] 3,669 12
ActivityNet [41] 28,000 203
Kinetics [42] 500,000 600
AVA [43] 57,600 80
YouTube [44] 8,000,000 4,716
Sports-1M [45] 1,133,158 487
Moments in Time [46] 1,000,000 339

FIGURE 5.11 Sample activities from UCF-101, HMDB51, and YouTube action datasets [22].

UCF-101. After this technique, an improvement of 20% on UCF-101 was achieved

by Xin et al. [7], which solves the problem of the spatial and temporal domains
using sequential and dynamic features. The accuracy scores obtained were 85%,
95.2%, and 58.2% on UCF-101, KTH2, and HMDB51, respectively. Apart from all
these methods, other approaches presented in Table 5.2 proposed various procedures
using LSTMs and RNNs to precisely recognize the activities. The majority of these
methods challenged the performance over the UCF-101 and HMDB51 datasets and
improved the accuracy by 3% to 5%. The highest accuracy on the UCF-101 dataset
till now is 95.1%, achieved by RSTAN [30], and a high score of 87% was obtained
for the HMDB51 dataset by DB-LSTM [22].
The confusion matrix, also known as the error matrix, is a popular evaluation
matrix for recognition tasks. Usually, it is a table layout where the diagonal cells
represent the true position predictions. However, in recognition tasks for large-scale
categories, making a large table is not an attractive representation. Therefore, for
Deep LSTM-Based Sequence Learning Approaches 141

this purpose, a color-coded confusion matrix is used to show the error matrix of the
algorithm: values near 100% are brighter and lower values are darker. The confusion
matrices for the DB-LSTM approach on the HMDB51 and UCF-101 datasets are
given in Figure 5.12, where the intensity of true positives (diagonal) is high for each
category, proving the effciency of this method. A recent approach by Ullah et al. [1]
evaluated their method on fve benchmark datasets. Confusion matrices from their
method for UCF-101 and HMDB51 datasets are shown in Figure 5.13. They achieved
high accuracy for UCF-101 and UCF-50 datasets; therefore, from Figure 5.13 (a)
and Figure 5.13 (d), it can be seen that for each class the true positive is very high.
Similarly, for two challenging datasets, namely HMDB51 and Hollywood2, they
achieved average performance; therefore, in Figure 5.13 (b) and Figure 5.13 (c) the
diagonals are not as bright in the confusion matrix. Figure 5.13 (e) is the confusion
matrix for a YouTube action dataset that contains only 11 action categories; therefore
the results are good for each class and the true positives are very bright in the confu-
sion matrix.
Category-wise accuracy is another metric for large-scale dataset evolution. It
shows the overall performance of the algorithm for all categories in the dataset.
We have compared the class-wise accuracies obtained by DB-LSTM [22] and tem-
poral optical fow with multi-layer LSTM [1] methods for the UCF-101 dataset in
Figure 5.14 (a) and Figure 5.14 (b). The horizontal axis shows classes, and the verti-
cal axis represents percentage accuracies for the corresponding classes for test sets
of the UCF-101 dataset. It is evident from Figure 5.14 (a) and Figure 5.14 (b) that the
best accuracies are above 80%, approaching 100% for some categories.
However, in the results for DB-LSTM shown in Figure 5.14 (a), many categories
are less than 50%, which shows that the performance of the method is not effective
for this dataset. On the other hand, the temporal optical fow method with multi-layer
LSTM [1] given in Figure 5.14 (b) achieves higher than 70% accuracy for each class,
which shows the robustness of this approach for different kinds of activity recogni-
tion in real-world scenarios. The class-wise accuracies achieved by DB-LSTM [22]

FIGURE 5.12 Confusion matrices of HMDB51 and UCF-101 datasets obtained by deep
features of AlexNet model for frame-level representation followed by DB-LSTM for action
recognition [22].
 142

FIGURE 5.13 Confusion matrices for testing fve datasets used in the evaluation of CNN-based temporal optical fow features followed by multi-layer
LSTM. (a) UCF101 dataset, (b) HMDB51 dataset, (c) Hollywood2 dataset, (d) UCF50 dataset, and (e) YouTube dataset [1, 22].
Deep Learning in Computer Vision
Deep LSTM-Based Sequence Learning Approaches 143

FIGURE 5.14 Comparison of category-wise accuracy on the UCF101 dataset for DB-LSTM
and temporal optical fow with multi-layer LSTM approaches [1, 22].

and temporal optical fow with multi-layer LSTM [1] approach for HMDB51 dataset
are shown in Figure 5.15 (a) and Figure 5.15 (b). respectively.
HMDB51 is one of the most challenging datasets; for this dataset, the method of
[1] achieved 71% and the method of [22] obtained 87% overall accuracy. However,
the class-wise accuracy for the method shown in Figure 5.15 (a) the method of [22]
is better for maximum class, but a few of them are still less than 10%, while from
the method of [1], shown in Figure 5.15 (b), we can see that a few category scores
are above 90%, but most of them are within the range of 70% to 90%. The accuracy
for all classes is greater than 45% except kicking a football, picking, and swinging
a baseball bat. This is due to the overlapping of visual contents and similarity of
motion information with other classes. Therefore, predictions of these classes are
confused with others.

5.4 FUTURE RESEARCH DIRECTIONS

It has been observed that many researchers have addressed the problem of action
and activity recognition by applying different strategies in LSTM according to the
need in a particular domain. The current literature shows that many researchers have
144 Deep Learning in Computer Vision

FIGURE 5.15 Comparison of category-wise accuracy on the HMDB51 dataset for

DB-LSTM and temporal optical fow with multi-layer LSTM approaches [1, 22].

used different kinds of features for sequence representations followed by RNN or

LSTM. For instance, some researchers also used deep features for sequence repre-
sentation; however, the time complexity of those methods is very high. So, there is
a need for state-of-the-art techniques that can make tradeoffs between accuracy and
time complexity. Recommendations and future directions about actions and activity
recognition are as follows.

5.4.1 BENEFITING FROM IMAGE MODELS

Image classifcation has achieved state-of-the-art results using CNN and beaten
human error rates in recent deep models [47]. These existing image models can
be utilized for frame-level representation in sequential video data. There are many
lightweight CNN models, such as MobileNet [48], SqueezeNet [49], and ShuffeNet
[50], that can help in real-time processing for action and activity recognition. These
CNN models have already achieved state-of-the-art results in other tasks such as
object detection and localization [51, 52], image and video retrieval [53], video
Deep LSTM-Based Sequence Learning Approaches 145

summarization [54], etc. The dominance of these models is that they are trained
on large-scale image datasets, which helps to learn discriminative hidden local and
global representations from visual data.

5.4.2 MULTIPLE ACTIONS AND ACTIVITY RECOGNITION

Several action recognition methods exist, but we can see that these methods are
applicable for processing single-person action data effectively. The existing methods
are trained on UCF101, UCF sports, and HMDB51 datasets; in these datasets, only
one person is performing an activity in sample videos. From Table 5.1, we can see
that only one LSTM-based method addresses the problem of multiple activity recog-
nition in video data. In CCTV, multiple activities are performed in a single scene. So,
in the future, this task needs to be tackled by detecting and localizing the individual
in the video data and then further processing it for activity recognition. Localization
and tracking are very challenging tasks in group activity recognition, for which we
can use existing deep learning-based object detection and localization models such
as YOLO [55], faster RCNN [56], MobileNet SSD [48], etc.

5.4.3 UTILIZATION OF LARGE-SCALE DATASETS

The existing datasets utilized by researchers are very small and are not suffcient for
training deep learning models from scratch, although there exist several large-scale
video action and activity recognition datasets that can be utilized for effective rec-
ognition. Table 5.3 shows some of the popular datasets in the action and activity rec-
ognition literature. The frst three datasets are already used by many researchers and
have achieved good results, as shown in Table 5.2. However, by utilizing ActivityNet
[41], Kinetics [42], AVA [43], YouTube [44], Sports-1M [45], and Moments in Time
[46], we can recognize a larger number of activities and actions through a single
recognition model.

5.5 CONCLUSION
Human action and activity recognition is the problem of analyzing and predicting the
movements of human body parts in a sequence of video frames. In this chapter, we
have described the concept of sequence learning for action and activity recognition
using LSTM and its variants such as multi-layer or deep LSTM and bidirectional
LSTM networks. Moreover, recent LSTM-based methods for action and activity rec-
ognition are surveyed, and its achievements and drawbacks are discussed. We also
explained the working of traditional RNNs and their limitations, along with why
LSTM is better than RNN. Finally, the chapter concluded with recommendations for
future directions that may help overcome the present challenges of action and activ-
ity recognition using LSTM.

ACKNOWLEDGMENT
This work was supported by the National Research Foundation of Korea (NRF)
grant funded by the Korea Government (MSIP) (No. 2019R1A2B5B01070067).
146 Deep Learning in Computer Vision

BIBLIOGRAPHY
1. A. Ullah, K. Muhammad, J. Del Ser, S. W. Baik, and V. Albuquerque, “Activity recog-
nition using temporal optical fow convolutional features and multi-layer LSTM,” IEEE
Transactions on Industrial Electronics, vol. 66, no. 12, pp. 9692–9702, 2019.
2. I. U. Haq, K. Muhammad, A. Ullah, and S. W. Baik, “DeepStar: Detecting starring
characters in movies,” IEEE Access, vol. 7, pp. 9265–9272, 2019.
3. Y. Liu, L. Nie, L. Han, L. Zhang, and D. S. Rosenblum, “Action2Activity: Recognizing
complex activities from sensor data,” in IJCAI, 2015, pp. 1617–1623.
4. Y. Liu, L. Nie, L. Liu, and D. S. Rosenblum, “From action to activity: Sensor-based
activity recognition,” Neurocomputing, vol. 181, pp. 108–115, 2016.
5. A. Ullah, K. Muhammad, I. U. Haq, and S. W. Baik, “Action recognition using opti-
mized deep autoencoder and CNN for surveillance data streams of non-stationary envi-
ronments,” Future Generation Computer Systems, vol. 96, pp. 386–397, 2019.
6. M.-C. Roh, H.-K. Shin, and S.-W. Lee, “View-independent human action recognition
with volume motion template on single stereo camera,” Pattern Recognition Letters,
vol. 31, pp. 639–647, 2010.
7. M. Xin, H. Zhang, H. Wang, M. Sun, and D. Yuan, “Arch: Adaptive recurrent-convo-
lutional hybrid networks for long-term action recognition,” Neurocomputing, vol. 178,
pp. 87–102, 2016.
8. D. Weinland, R. Ronfard, and E. Boyer, “Free viewpoint action recognition using
motion history volumes,” Computer Vision and Image Understanding, vol. 104, pp.
249–257, 2006.
9. M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and A. Baskurt, “Action classifca-
tion in soccer videos with long short-term memory recurrent neural networks,” in
International Conference on Artifcial Neural Networks, 2010, pp. 154–159.
10. A. Kovashka, and K. Grauman, “Learning a hierarchy of discriminative space-time
neighborhood features for human action recognition,” in 2010 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2010, pp. 2046–2053.
11. M. Sekma, M. Mejdoub, and C. B. Amar, “Human action recognition based on multi-
layer fsher vector encoding method,” Pattern Recognition Letters, vol. 65, pp. 37–43,
2015.
12. J. Hou, X. Wu, Y. Sun, and Y. Jia, “Content-attention representation by factorized
action-scene network for action recognition,” IEEE Transactions on Multimedia, vol.
20, pp. 1537–1547, 2018.
13. F. U. M. Ullah, A. Ullah, K. Muhammad, I. U. Haq, and S. W. Baik, “Violence detec-
tion using spatiotemporal features with 3D convolutional neural network,” Sensors, vol.
19, p. 2472, 2019.
14. K.-i. Funahashi, and Y. Nakamura, “Approximation of dynamical systems by continu-
ous time recurrent neural networks,” Neural Networks, vol. 6, pp. 801–806, 1993.
15. S. Hochreiter, and J. Schmidhuber, “Long short-term memory,” Neural Computation,
vol. 9, pp. 1735–1780, 1997.
16. H. Sak, A. Senior, and F. Beaufays, “Long short-term memory recurrent neural network
architectures for large scale acoustic modeling,” in Fifteenth Annual Conference of the
International Speech Communication Association, 2014.
17. M. Sajjad, S. Khan, K. Muhammad, W. Wu, A. Ullah, and S. W. Baik, “Multi-grade
brain tumor classifcation using deep CNN with extensive data augmentation,” Journal
of Computational Science, vol. 30, pp. 174–182, 2019.
18. J. Ahmad, K. Muhammad, S. Bakshi, and S. W. Baik, “Object-oriented convolu-
tional features for fne-grained image retrieval in large surveillance datasets,” Future
Generation Computer Systems, vol. 81, pp. 314–330, 2018/04/01/ 2018.
Deep LSTM-Based Sequence Learning Approaches 147

19. K. Muhammad, J. Ahmad, Z. Lv, P. Bellavista, P. Yang, and S. W. Baik, “Effcient deep
CNN- based fre detection and localization in video surveillance applications,” IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 49, no. 7, pp. 1419–1434,
2019.
20. K. Muhammad, R. Hamza, J. Ahmad, J. Lloret, H. H. G. Wang, and S. W. Baik, “Secure
surveillance framework for IoT systems using probabilistic image encryption,” IEEE
Transactions on Industrial Informatics, vol. 14, no. 8, pp. 3679–3689, 2018.
21. M. Sajjad, A. Ullah, J. Ahmad, N. Abbas, S. Rho, and S. W. Baik, “Integrating salient
colors with rotational invariant texture features for image representation in retrieval
systems,” Multimedia Tools and Applications, vol. 77, pp. 4769–4789, 2018.
22. A. Ullah, J. Ahmad, K. Muhammad, M. Sajjad, and S. W. Baik, “Action recognition in
video sequences using deep Bi-directional LSTM with CNN features,” IEEE Access,
vol. 6, pp. 1155–1166, 2018.
23. A. Ogawa, and T. Hori, “Error detection and accuracy estimation in automatic
speech recognition using deep bidirectional recurrent neural networks,” Speech
Communication, vol. 89, pp. 70–83, 2017.
24. J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of gated recurrent
neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014.
25. W. Li, W. Nie, and Y. Su, “Human action recognition based on selected spatio-temporal
features via bidirectional LSTM,” IEEE Access, vol. 6, pp. 44211–44220, 2018.
26. M. S. Ibrahim, S. Muralidharan, Z. Deng, A. Vahdat, and G. Mori, “A hierarchical deep
temporal model for group activity recognition,” In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 1971–1980.
27. Z. Li, K. Gavrilyuk, E. Gavves, M. Jain, and C. G. Snoek, “VideoLSTM convolves,
attends and fows for action recognition,” Computer Vision and Image Understanding,
vol. 166, pp. 41–50, 2018.
28. C.-Y. Ma, M.-H. Chen, Z. Kira, and G. AlRegib, “TS-LSTM and temporal-inception:
Exploiting spatiotemporal dynamics for activity recognition,” Signal Processing:
Image Communication, vol. 71, pp. 76–87, 2019.
29. Z. Chen, B. Ramachandra, T. Wu, and R. R. Vatsavai, “Relational long short-term
memory for video action recognition,” arXiv preprint arXiv:1811.07059, 2018.
30. W. Du, Y. Wang, and Y. Qiao, “Recurrent spatial-temporal attention network for action
recognition in videos,” IEEE Transactions on Image Processing, vol. 27, pp. 1347–
1360, 2018.
31. J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, K.
Saenko, et al., “Long-term recurrent convolutional networks for visual recognition and
description,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 2625–2634.
32. Y. Huang, X. Cao, Q. Wang, B. Zhang, X. Zhen, and X. Li, “Long-short term fea-
tures for dynamic scene classifcation,” IEEE Transactions on Circuits and Systems for
Video Technology, vol. 29, no. 4, pp. 1038–1047, 2019.
33. S. Ma, L. Sigal, and S. Sclaroff, “Learning activity progression in lstms for activity
detection and early detection,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2016, pp. 1942–1950.
34. L. Sun, K. Jia, K. Chen, D.-Y. Yeung, B. E. Shi, and S. Savarese, “Lattice long short-
term memory for human action recognition,” in ICCV, 2017, pp. 2166–2175.
35. V. Veeriah, N. Zhuang, and G.-J. Qi, “Differential recurrent neural networks for action
recognition,” in Proceedings of the IEEE International Conference on Computer
Vision, 2015, pp. 4041–4049.
36. H. Yang, J. Zhang, S. Li, and T. Luo, “Bi-direction hierarchical LSTM with spatial-
temporal attention for action recognition,” Journal of Intelligent & Fuzzy Systems, vol.
36, no. 1, pp. 775–786, 2019.
148 Deep Learning in Computer Vision

37. K. Soomro, A. R. Zamir, and M. Shah, “UCF101: A dataset of 101 human actions
classes from videos in the wild,” arXiv preprint arXiv:1212.0402, 2012.
38. H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre, “HMDB: A large video
database for human motion recognition,” in 2011 IEEE International Conference on
Computer Vision (ICCV), 2011, pp. 2556–2563.
39. C. Schuldt, I. Laptev, and B. Caputo, “Recognizing human actions: A local SVM
approach,” in Proceedings of the 17th International Conference on Pattern Recognition,
ICPR 2004, 2004, pp. 32–36.
40. M. Marszalek, I. Laptev, and C. Schmid, “Actions in context,” in IEEE Conference on
Computer Vision and Pattern Recognition, CVPR 2009, 2009, pp. 2929–2936.
41. F. Caba Heilbron, V. Escorcia, B. Ghanem, and J. Carlos Niebles, “Activitynet: A large-
scale video benchmark for human activity understanding,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 2015, pp. 961–970.
42. W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, et al.,
“The kinetics human action video dataset,” arXiv preprint arXiv:1705.06950, 2017.
43. C. Gu, C. Sun, S. Vijayanarasimhan, C. Pantofaru, D. A. Ross, G. Toderici, et al., “AVA:
A video dataset of spatio-temporally localized atomic visual actions,” arXiv preprint
arXiv:1705.08421, vol. 3, p. 6, 2017.
44. S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici, B. Varadarajan, et
al., “Youtube-8m: A large-scale video classifcation benchmark,” arXiv preprint
arXiv:1609.08675, 2016.
45. A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei, “Large-
scale video classifcation with convolutional neural networks,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 1725–1732.
46. M. Monfort, B. Zhou, S. A. Bargal, A. Andonian, T. Yan, K. Ramakrishnan, et al.,
“Moments in time dataset: One million videos for event understanding,” arXiv preprint
arXiv:1801.03150, 2018.
47. S. Khan, K. Muhammad, S. Mumtaz, S. W. Baik, and V. H. C. de Albuquerque,
“Energy-effcient deep CNN for smoke detection in foggy IoT environment,” IEEE
Internet of Things Journal, vol. 6, no. 6, pp. 9237–9245, 2019.
48. A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, et al.,
“Mobilenets: Effcient convolutional neural networks for mobile vision applications,”
arXiv preprint arXiv:1704.04861, 2017.
49. F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer,
“Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb model
size,” arXiv preprint arXiv:1602.07360, 2016.
50. N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “Shuffenet v2: Practical guidelines for eff-
cient cnn architecture design,” arXiv preprint arXiv:1807.11164, vol. 5, 2018.
51. M. Sajjad, M. Nasir, F. U. M. Ullah, K. Muhammad, A. K. Sangaiah, and S. W. Baik,
“Raspberry Pi assisted facial expression recognition framework for smart security in
law-enforcement services,” Information Sciences, vol. 479, pp. 416–431, 2019.
52. M. Sajjad, S. Khan, T. Hussain, K. Muhammad, A. K. Sangaiah, A. Castiglione, et
al., “CNN-based anti-spoofng two-tier multi-factor authentication system,” Pattern
Recognition Letters, vol 126, pp. 123–131, 2019.
53. J. Ahmad, K. Muhammad, J. Lloret, and S. W. Baik, “Effcient conversion of deep fea-
tures to compact binary codes using Fourier decomposition for multimedia big data,”
IEEE Transactions on Industrial Informatics, vol. 14, pp. 3205–3215, July 2018.
54. K. Muhammad, T. Hussain, and S. W. Baik, “Effcient CNN based summarization
of surveillance videos for resource-constrained devices,” Pattern Recognition Letters,
2018.
Deep LSTM-Based Sequence Learning Approaches 149

55. J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unifed, real-
time object detection,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2016, pp. 779–788.
56. S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detec-
tion with region proposal networks,” in Advances in Neural Information Processing
Systems, 2015, pp. 91–99.
 6 Deep Semantic
Segmentation in
Autonomous Driving
Hazem Rashed, Senthil Yogamani,
Ahmad El-Sallab, Mahmoud Hassaballah, and
Mohamed ElHelw

CONTENTS
6.1 Introduction .................................................................................................. 152
6.2 Convolutional Neural Networks ................................................................... 154
6.2.1 Overview of Deep CNN ................................................................... 154
6.2.2 Types of CNN Layers ....................................................................... 155
6.3 Semantic Segmentation ................................................................................ 157
6.3.1 Automotive Scene Segmentation ...................................................... 159
6.3.2 Common Automotive Datasets......................................................... 160
6.3.2.1 Real Datasets ..................................................................... 160
6.3.2.2 Synthetic Datasets.............................................................. 161
6.4 Motion and Depth Cues................................................................................ 161
6.4.1 Depth Estimation .............................................................................. 162
6.4.2 Optical Flow Estimation................................................................... 163
6.5 Proposed Fusion Architectures..................................................................... 164
6.5.1 Single-Input Network........................................................................ 164
6.5.2 Early-Fusion Network....................................................................... 165
6.5.3 Mid-Fusion Network......................................................................... 165
6.6 Experiments.................................................................................................. 166
6.6.1 Experimental Settings ...................................................................... 166
6.6.2 Results Analysis and Discussion ...................................................... 167
6.6.2.1 Depth Experiments ............................................................ 167
6.6.2.2 Optical Flow Experiments ................................................. 167
6.7 Conclusion .................................................................................................... 174
Bibliography .......................................................................................................... 178

151
152 Deep Learning in Computer Vision

6.1 INTRODUCTION
The frst autonomous cars (ACs) developed by Carnegie Mellon University and
ALV appeared in 1984 [1]. Later, in 1987, Mercedes-Benz collaborated with
Bundeswehr University Munich in a project called the Eureka Prometheus Project
[2], which was the largest R&D project ever in autonomous driving (AD). Since
then, numerous car manufacturers have developed prototypes for their self-driv-
ing cars [3, 4]. The frst truly autonomous car, VaMP, developed in the 1990s
by Mercedes-Benz, was a 500 SEL Mercedes-Benz that had the capabilities to
control the steering wheel, throttle, and brakes by computers in real time. The
car relied only on four cameras for localization, as there was no GPS at that time.
The vehicle was able to drive more than 1,000 km in normal traffc at speeds up to
130, kmh moving from lane to another and passing other cars autonomously after
the approval of the safety driver [5]. Recently, AD has gained great attention with
signifcant progress in computer vision algorithms [6, 7], and it is considered one
of the highly trending technologies all over the globe. Within the next 5–10 years,
AD is expected to be deployed commercially. Currently, most automotive original
equipment manufacturers (OEMs) over the world, such as Volvo, Daimler, BMW,
Audi, Ford, Nissan, and Volkswagen, are working on development projects focus-
ing on the AD technology [8, 9].
In addition, every year large numbers of people die or are injured due to car acci-
dents [10]. Autonomous cars (ACs) have great potential to signifcantly reduce these
numbers. One of the potential benefts of ACs is the higher safety levels for both
drivers and surrounding pedestrians [11]. It has been reported by consulting frm
McKinsey & Company that deploying AD technology could eliminate 90% of acci-
dents in the USA and save up to US 190 billion due to damage prevention annually.
Besides, autonomous vehicles are predicted to have positive impact on traffc fow,
with higher speed limits permitted for increased roadway capacity [12]. AD can also
provide a higher level of satisfaction for customers, especially for the elderly and the
disabled, introduce new business models such as autonomous taxis as a service, and
reduce crime.
Additionally, when vehicles are self-aware, they will have the information needed
to avoid numerous accidents [13]. For example, vehicles may recognize when they
are being stolen or the driver is being threatened. Self-aware vehicles may recognize
children being kidnapped or weapons being transported, or they may spot drunk
drivers and consequently take proper action [14, 15]. Generally, there are fve levels
of autonomy [16]:

• Level 0 is “no automation”, where everything has to be controlled by the

driver.
• Level 1 is “driver assistance”, where the driver controls everything; how-
ever, separate systems provide assistance for one or more tasks related to
steering and acceleration/deceleration. This takes place on a condition that
the system controls either steering or acceleration/deceleration, but not both
together. A good example is the lane-keeping assist system, in which the
car monitors the lanes ahead and constantly applies torque to the steering
Deep Semantic Segmentation in Autonomous Driving 153

system to keep the car in the center of the lane. Another example is adaptive
cruise control (ACC), which controls the acceleration/deceleration accord-
ing to the distance between the ego-vehicle and the car ahead with a pre-
defned speed limit [17].
• Level 2 is “partial automation”, where the car can do maneuvers control-
ling both speed and steering together; however, the driver must keep his/
her hands on the steering wheel for any required intervention. Most of the
current commercial autonomous driving systems are of Level 2.
• Level 3 is “conditional automation”, where the car manages almost all of
the tasks but the system prompts the driver for intervention when there is a
scenario it cannot navigate through. Unlike Level 2, the driver can remove
his hands from the steering wheel; however, he/she should be available to
take over at any time.
• Level 4 is “high automation”, where the car can operate without the
driver’s supervision but under conditional circumstances for the envi-
ronment. Otherwise, the driver may take control. Google’s Firefy pro-
totype is an example for Level 4 autonomy; however, speed is limited
to 25 mph.
• Level 5 is “full automation”, where the car can operate in any condition and
the driver’s role is only providing the destination. This level of autonomy
has not been reached yet.

Nevertheless, like any new technology, AD faces various technical and non-techni-
cal challenges. The technical challenges include sensor accuracy compared to cost;
that is, the sensors have to be reasonable in terms of price [18]. Therefore, effcient
algorithms have to be implemented to deal with map generation errors. There are
localization errors that take place, as data from GPS and intertial measurement units
(IMUs) is not always stable due to connectivity issues and even odometry informa-
tion has errors, especially when the car travels long distances. Path planning and
navigation should not depend on imperfect inputs and should deal with various prac-
tical cases [19]. The complexity of a system must be acceptable for the purpose of
producing commercial cars, which adds another limitation in terms of the hardware
used for production. The non-technical challenges include the urgent need to take
tough decisions into implementation [20]. For instance, the AC has to be equipped
with a decision-making system for unavoidable accidents; it has to decide to choose
hitting a woman or a child if it has to, and which one to protect the most, the child
outside the car or the car occupant [21]. In this regard, a mobile robot must be able to
perform four key tasks to be able to navigate autonomously [22, 23].

• Environment perception and map creation using different types of sen-

sors. This task includes a lot of smaller subtasks, such as object recognition
and classifcation, object tracking, criticality defnition, etc. [24].
• Localization within the map where an accurate location for the robot must
be calculated using sensors such as GPS and further algorithms utilized to
improve accuracy and provide location even in areas where GPS signals are
not available including tunnels and garages.
154 Deep Learning in Computer Vision

• Path planning and decision making according to both map and position
information. In addition to path planning, there are some constraints and
regulations to be followed especially for a self-driving car, e.g., observ-
ing traffc lights and following rules in special paths such as roundabouts,
intersections, and U-turns.
• Vehicle control through sending signals to actuators over the vehicle’s net-
work for taking a desired action.

6.2 CONVOLUTIONAL NEURAL NETWORKS

Recently, convolutional neural networks (CNNs) and in particular deep learning have
played a great role in computer vision, and they are exploited in various applications
for AD. In this section, a brief introduction about deep CNN is presented, i.e., how
neural networks are used in machine learning, and the idea behind it. Afterwards,
semantic segmentation, which is one of the main applications of deep learning in
AD, is discussed.

6.2.1 OVERVIEW OF DEEP CNN

Deep learning is a specialized form of machine learning. Generally, a machine
learning algorithm starts after the phase of manual feature selection in a given input,
whether it is an image or any type of input. The features are then used to create a
model that performs the classifcation task according to the input data [25]. In the
deep learning workfow, relevant features are automatically extracted from images.
From this comes the power of deep learning, as it has the capability to extract hidden
relationships and features that cannot be easily perceived by humans. In addition,
deep learning performs end-to-end learning, where a network is given raw data and a
task to perform, such as classifcation, and it learns how to do this automatically [26].
There are several types of deep neural networks; one of these, built to deal with
input images, is called a CNN. In CNN algorithms, each level in the deep network
learns to transform its input data into a slightly more abstract and composite repre-
sentation [27]. For example, in an image recognition application, the raw input is a
matrix of pixels. The frst layer may abstract the pixels and encode edges; the second
layer may encode arrangements of edges. The third layer may encode specifc fea-
tures like a nose or eyes, and the fourth layer may recognize that the image contains a
face. The whole pipeline takes place without manual feature selection as opposed to
conventional algorithms, where specifc characteristics are being searched in order
to recognize which given images contain faces. In fact, deep learning has gained
great attention in the 2010s after the development of AlexNet [28] in the ImageNet
challenge, where an error of 16% was achieved for the image classifcation task in
the ImageNet challenge.
The CNNs are used for various tasks, like depth estimation, object detection,
localization, classifcation, semantic segmentation, and several others. The goal of
a CNN is to learn higher-order features in the data by using convolutions. In a con-
ventional neural network, an input 3-RGB channel image of 300 × 300 pixels would
create 270,000 connection weights per hidden neuron if it is fully connected, which
Deep Semantic Segmentation in Autonomous Driving 155

is not practical from an implementation perspective. Additionally, unlike any other

input vector, images are spatially correlated, where pixels belonging to the same
object typically will have close pixel values. This piece of information is very impor-
tant and can help the network to provide better outputs. Neural networks make use
of this piece of information through full connectivity. CNNs, on the other hand, are
able to utilize computation power better than fully connected neurons. With CNNs,
one can arrange the neurons in a three-dimensional structure. These attributes are
consistent with the image structure, as it has image width, image height, and 3 chan-
nels representing the RGB information. Neurons in a layer are connected to only a
small region of neurons in the layer before it, where each layer transforms the 3D
input volume from the previous layer into another 3D output volume of neuron acti-
vations with some differentiable function [29].

6.2.2 TYPES OF CNN LAYERS

The term “deep” in deep CNN refers to the number of hidden layers through which
the data representation is transformed [30]. The common types of layers used in a
typical deep CNN can be summarized as follows:
Input layers: This is where we load the raw input data of the image to be pro-
cessed by the network. The input is a 3D volume of image width, height, and depth.
Usually the depth represents the three RGB layers of an image.
Convolutional layers: These layers are the core building blocks of any CNN
architecture. Convolutional layers transform the input data by using a patch of
locally connected neurons from the previous layer. Generally, convolution is defned
as a mathematical operation describing how to merge two sets of information. It
computes a dot product between the region of neurons in the previous layer and the
weights to which they are locally connected in the output layer. Figure 6.1 shows a
visualization for neuron activation function, where each neuron holds a set of param-
eters that form a flter to be multiplied with the input volume as a dot product and

FIGURE 6.1 Visualization of the CNN activation function.

156 Deep Learning in Computer Vision

added to bias. The convolution input can be raw pixels or an output feature map from
the previous layer. The convolution is known in deep learning as a feature detector
for the CNN where the values in kernels defne an operation for feature detection.
A kernel with ones in the middle columns is used to detect vertical edges within
the input patch. The output of each convolution operation is called a feature map.
Stacking the output feature maps for different kernels generates a 3D output volume.
Convolution flters defne a small bounded region to output activation maps from
the input volume through connection with only a subset of the input volume. This
allows us to have high-quality feature extraction utilizing the spatial correlation
between pixels besides reduction the number of parameters per layer compared to
fully connected neurons [31]. Since the depth of the input volume is 3, the convolution
layer as well should have a depth of 3. Convolution kernels connect to the bounded
region on the image in terms of width and height; however, the depth always has to
be the same. CNNs use parameter sharing to control the total number of parameters,
where the same weights being learned using a certain flter are used while the flter
scans the whole image or feature map, instead of defning new parameters. There are
several parameters to control using a flter in CNNs such as receptive feld, output
depth, stride, and zero-padding.
Receptive feld: This term refers to the spatial flter size, in terms of width and
height. A flter with receptive feld of 5 × 5 × 3 means that it has a 5-pixel width, a
5-pixel height, and 3 channels of depth. Usually, flters of the frst convolutional layer
have depths of 3 to accept an RGB image input volume of 3 channels.
Output depth: The neuron count in the convolutional layer connected to the
same region of the input volume, meaning that it refers to the number of flters used
in the convolutional layer.
Stride: This refers to the step used by the flter to scan through the input volume.
A stride of 1 means that the flter moves horizontally and vertically with one-pixel
steps, resulting in a dense output volume of a large number of depth columns, as
there is a large overlap between receptive felds.
Zero-padding: This parameter is used to control the spatial size of the output
volume. This is very helpful if the output volume is required to be same as the input
volume. Additionally, if convolutions are done without zero-padding and the size is
reduced, the information at boundaries will be lost quickly.
Batch normalization layer: This layer is used in deep learning to speed up the
training by making normalization part of the network. For example, if input values
are in a range from 1 to 1,000 and from 1 to 10, the change from 1 to 2 in the sec-
ond input represents a 10% change in the input value, but actually represents only
a 0.1% change in the frst input. Normalizing the inputs will help the network learn
faster and more accurately without biasing certain dimensions over the others. Batch
normalization uses a very similar approach, however, for the hidden layers changing
the outputs all the time. Since the activations of a previous layer are the inputs of
the next layer, each layer in the neural network faces a problem due to input distribu-
tion change with each step. This problem is termed “covariate shift”. The basic idea
behind batch normalization is to limit covariate shift by normalizing the activations
of each layer through transforming the inputs to be mean 0 and unit variance. This
Deep Semantic Segmentation in Autonomous Driving 157

allows each layer to learn on a more stable distribution of inputs, and would acceler-
ate the training of the network. Limiting the covariate shift helps as well in avoid-
ing the vanishing gradient problem, where gradients become very small when the
input is large in a sigmoid activation function, for example, as the distribution keeps
changing during training and might be large enough to create the vanishing gradient
problem [32].
Pooling layers: These layers are periodically inserted after convolutional layer
in a typical ConvNet architecture. Their function is to progressively reduce the spa-
tial size of the input representation providing multiple advantages: (1) Obtaining an
abstract representation of the input data where only the most important features are
being considered; hence, this helps to avoid overftting the model to the input data.
(2) Reducing the computational cost as the number of learnable parameters becomes
less than those for corresponding networks not using pooling. (3) Providing basic
translation invariance to the input representation. The pooling layer operates inde-
pendently on each depth slice of the input representation. The most common form of
pooling is called max-pooling, where a 2 × 2 window scans the image with a stride
of 2, keeping the max value of that window only, resulting in keeping only 25% of
the input volume [33].
Fully connected layers: Neurons in a fully connected layer have connections
to all output activations from the previous layer, which is always good to be able to
model complex representations. The fully connected layer provides all the possible
combinations of the features in the previous layers, as the convolutional layer relies
only on the local spatial correlation between pixels. However, fully connected layers
contribute strongly to the network’s complexity [34].
Output layer: In a normal classifcation task where the objects belong to one
class only, the sigmoid layer is used to fnally generate the output. Let’s say we
are trying to classify an object among fve classes. The Softmax function pro-
vides an output probability for the likelihood that this object belongs to each class.
These output probabilities will sum up to 1, which is very helpful because the next
step would be classifcation of the object according to the maximum probability
generated.

6.3 SEMANTIC SEGMENTATION

Semantic image segmentation has witnessed tremendous progress recently, espe-
cially with the appearance of deep learning. It provides dense pixel-wise labeling
of the image, which leads to scene understanding. Semantic segmentation has been
used in many applications such as robotics [35–37], virtual reality [38], and more
importantly autonomous driving [39, 40], which is one of the main applications using
semantic segmentation. Figure 6.2 shows an example of the output of a deep convo-
lutional network for an automated driving scene.
Pixel-wise annotation of large datasets is very hard and labor intensive. Thus,
there is a lot of research with the aim of reducing the complexity of segmenta-
tion networks by using knowledge from other domains and incorporating other
cues wherever possible. To this end, most of the current semantic segmentation
158 Deep Learning in Computer Vision

FIGURE 6.2 Semantic segmentation for a typical automotive scene.

networks are implemented to use only appearance cues without considering sev-
eral other cues that might be helpful to a specifc application. In this context,
deep learning has played a great role in semantic segmentation tasks, ideally
because of the automatic feature selection advantage that deep learning has over
conventional methods. There are three main areas explored in deep learning for
semantic segmentation. The frst one uses patch-wise training, where, the input
image is fed into the Laplacian pyramid, each scale is forwarded through a three-
stage network for hierarchical feature extraction, and patch-wise classifcation is
used [41].
The second area is end-to-end training, in which the CNN network is trained to
take raw images as input and directly output dense pixel-wise classifcation instead
of patches classifcation. In [42], famous classifcation networks, namely AlexNet
[43], VGG-net [44], and GoogleLeNet [45], are adapted into fully convolutional net-
works, and the learned knowledge is transferred to the semantic segmentation task.
Skip connections between shallow and deep features are implemented to avoid los-
ing resolution. The features learned within the network are upsampled using decon-
volution to get dense predictions. In [46], deeper deconvolution is implemented, in
which stacked deconvolution and unpooling layers are utilized. In [47], an encoder-
decoder network is implemented where the decoder upsampling made use of the
pooling indices computed in the max-pooling step of the corresponding encoder for
nonlinear upsampling.
Recent works have focused on multiscale semantic segmentation [48]. In [41], the
scale issue has been dealt with using multiple rescaled versions of the image; how-
ever, after the emergence of end-to-end learning, the skip connections implemented
in [42] helped to merge feature maps from different resolutions, as loss of resolution
takes place due to image downsampling. Dilated convolution has been introduced
in [49], where the receptive feld has been expanded with zero values in between
Deep Semantic Segmentation in Autonomous Driving 159

depending on the dilation factor in order to avoid losing resolution. This helped for
the multiple scales issue. In the attention model [50, 51], the multiscale image seg-
mentation model is adapted so that it learns to softly weight the multiscale features
for objects with different scales.

6.3.1 AUTOMOTIVE SCENE SEGMENTATION

Since this work mainly focuses on autonomous driving applications, we provide a
summary of automotive environment conditions and information specifc to these
applications that can be utilized to enhance the performance of tasks on hand.
Modern cars are equipped with a multi-camera network, usually of 4 cameras
but up to 10 cameras for future generations. Figure 6.3 shows an example of 360°
surround-view system using four fsheye cameras installed for viewing all the sur-
roundings. Using such a system, there is a spatio-temporal relationship between the
outputs of the four cameras. For instance, a moving car will be shown in one camera
only, then overlap with another one, and so on. There is still no signifcant work
targeting automotive fsheye setup. However, in [52], a method is used to enhance
the way convolution captures information from neighbor pixels. Restricted deform-
able convolution is utilized as learnable dilated convolution. The idea is to initialize
a dilated convolution kernel, then learn offset parameters through training, which
will change the kernel’s center to enhance the learning of neighbor pixels. This tech-
nique is used to learn semantic segmentation of four fsheye cameras in automotive
surround-view systems.
There is no semantic segmentation annotation provided for fsheye cameras. To
solve this problem, a method for data augmentation named zoom augmentation is
implemented in [53], and focal length has been changed to generate fsheye images
to train CNNs for semantic segmentation. The problem of depth estimation for fsh-
eye cameras using sparse LiDAR data is investigated for automotive surround-view

FIGURE 6.3 Sample images from the surround-view camera network demonstrating near-
feld sensing and wide feld of view.
160 Deep Learning in Computer Vision

systems. The difference between viewpoints of both camera and LiDAR caused
occluded points to be projected incorrectly to the image plane. Then, an occlusion
correction method is presented to remove the incorrectly mapped points, and dense
depth map for the fsheye camera is generated using the LiDAR depth information
as supervised signals.
In a generic structure of automotive scenes, different types of prior informa-
tion that are common in any automotive road scene can be utilized, e.g., spatial
priors such as the road pixels usually taking almost all of the bottom half of the
image. The lanes are seen as thick white lines expanding until the road pixels
end. Other priors include the color priors for traffc lights, roads, and lanes.
Additionally, there is strong geometric structure in the scene as the road is fat,
and all the objects stand vertically above it. This is explicitly exploited in the
formulation of a commonly used objects-depth representation, namely Stixels
[54]. There are static and moving objects in addition to the vehicle ego-motion,
which may provide good temporal information. Although there is a lot in the
automotive scene structure that can be used as prior information, the system’s
complexity has to be taken into consideration in the automotive applications,
given the computational bounds in memory and processing power of the embed-
ded systems on board.

6.3.2 COMMON AUTOMOTIVE DATASETS

There are many datasets built specifcally for testing autonomous driving applica-
tions. They are divided into two main categories: real and synthetic. The most com-
monly used automotive datasets are discussed below.

6.3.2.1 Real Datasets

• CamVid [55] is one of the frst datasets released for automotive applica-
tions. It consists of 700 images from a 10-minute video sequence annotated
for semantic segmentation.
• KITTI [56] (Karlsruhe Institute of Technology and Toyota Technological
Institute) dataset was released in 2012 for several tasks like stereo, optical
fow, visual odometry, 3D object detection, and tracking. Different anno-
tations are provided such as GPS/IMU sensor readings for ego-motion
odometry, 3D objects, tracking information, and LiDAR data. The dataset
provides annotation for 200 images only for pixel-wise semantic segmenta-
tion training and another 200 images for testing.
• Cityscapes [57], introduced by Daimler, consists of 20,000 images of real-
istic urban road scenes with coarse annotations. Its images have high-res-
olution that reaches 2,048 × 1,024. A set of 5,000 images is provided with
fne annotations. The labels include 30 semantic classes, of which only 19
are used in evaluation as per the offcial evaluation script provided on the
website. Stereo images are provided with disparity map annotation using
the SGM algorithm [58] besides GPS coordinates and ego-motion data from
vehicle odometry. To obtain absolute disparity values, a method is provided
with the dataset for such calculations.
Deep Semantic Segmentation in Autonomous Driving 161

• Mapillary Vistas [59] is a large-scale dataset containing 25,000 high-res-

olution images annotated into 66 object categories in addition to instance
annotation. The dataset consists of images from all around the world for
different weather conditions and seasons. It is built through collaborative
work all over the globe and is the most diverse open dataset geographically.
• ApolloScape [60], provided by Baidu, is a large-scale dataset containing
high-resolution images of 3,384 × 2,710 pixels. The scene parsing dataset
includes 146,997 images with corresponding pixel-level poses, and depth
map annotations. However, depth maps are provided only for static objects
where the scene is scanned multiple times, the generated point clouds are
aligned, and then points based on temporal consistency are removed.

6.3.2.2 Synthetic Datasets

• Virtual KITTI [61] consists of 50 high-resolution videos consisting of
21,260 image frames of 1,242 × 375 resolution, generated using a unity
game engine in an urban environment. Different annotations are provided
with the dataset like semantic and instance segmentation, dense optical
fow (DOF), depth maps, and 2D and 3D object detection bounding boxes.
The depth map is a PNG16 image of the same size of the input image. The
value of each pixel represents the z coordinate of the point in camera coor-
dinate space. Each pixel takes values from 0 to 65,535, which corresponds
to 655.36 m from the camera image plane. Points at infnity, like sky, are
clipped to a depth of 655.3 m. The DOF image is a representation of dense
optical fow between two consecutive images. Semantic and instance infor-
mation is given per frame as a unique color for each class. Fourteen classes
are annotated: Building, Car, GuardRail, Misc, Pole, Road, Sky, Terrain,
TraffcLight, TraffcSign, Tree, Truck, Van, and Vegetation.
• SYNTHIA [62] consists of 200,000 images from video sequences with a
wide range of diversity regarding weather conditions and scene structure.
360° views are provided by simulating 8 RGB cameras. Depth maps are
provided as well.
• NUSCENES [63] is a large-scale dataset providing the most sensor mea-
surements. It contains 1,000 videos of 20 seconds each with several types
of annotation data such as LiDAR, radar, camera, IMU, and GPS data. It
also provides 3D bounding boxes over 25 classes of objects annotated at
2 Hz.
• SYNSCAPES [64] is very close to reality and provides both 1,440 × 720
and 2,048 × 1,024 resolution images, depth maps, semantic and instance
segmentation labels. 2D and 3D bounding boxes are also provided. Each
image of the 25,000 images has its own unique structure, which gives a
very wide range of variations and features combinations.

6.4 MOTION AND DEPTH CUES

This section gives brief information about motion and depth cues, which are utilized
to enhance semantic segmentation performance in our experiments.
162 Deep Learning in Computer Vision

6.4.1 DEPTH ESTIMATION

Depth is a very important information cue that implies the distance between sur-
rounding objects and the vehicle, and it is critical for autonomous driving applica-
tions. Knowing only semantics without depth is not enough for full environment
perception [65]. In addition to depth sensors that have their own pros and cons, depth
can be estimated using several approaches utilizing cameras only without further
more expensive sensors (e.g., LiDAR). The most common method for camera-based
depth estimation is based on stereo images using the SGM algorithm [58], where pix-
els from the left image are searched for in the right image across the epipolar lines.
Disparity is calculated according to a cost function that also guarantees the smooth-
ness of the output surface. Disparity is the difference between positions of a feature
that is viewed from two different cameras looking at the same scene with different
viewpoints. Depth has a nonlinear relationship with disparity and the absolute value
can be calculated if camera calibration information is available. However, disparity
maps using classical approaches are usually noisy; hence, depth estimation itself
is topic of research and deep learning has been used extensively through multiple
architectures for the purpose of providing better depth estimation.
Depth estimation using deep learning has been tackled through several approaches
[66, 67], which can be classifed as supervised, semi-supervised, and unsupervised
methods. Each approach has its pros and cons. For example, supervised approaches
generally provide better performance than unsupervised ones. Unsupervised
approaches have the fexibility to be trained without explicit depth annotations; how-
ever, they usually provide less accuracy than supervised ones. Generally, supervised
approaches require accurate depth annotation, which requires usage of synthetic
datasets. In [68], ego-motion information is used to exploit temporal information.
The depth network takes only the target view as input and a pixel-wise depth map
is generated. The pose network takes the target view and the source views as input
and generates relative camera poses. The outputs are used to inverse-warp the source
views to reconstruct the target view in an unsupervised manner. View synthesis is
used for that task, where a target image can be synthesized given a pixel-wise depth
map of that image, in addition to pose and visibility in a nearby view. Hence, unsu-
pervised training is performed. Godard et al. [69] formulated depth estimation as an
unsupervised learning problem where epipolar geometry constraints are exploited
to generate disparity images by training the network for image reconstruction loss.
This approach has the advantage of using only stereo information for training in
the unsupervised manner, while in cases of inference, only monocular images are
needed.
Apart from color, depth is another dimension for which the impact on semantic
segmentation has not yet been explored extensively. Most of the previous work using
depth for semantic segmentations made use of RGB-D cameras for indoor scenes.
Hazirbas et al. [70] investigated incorporating depth into semantic segmentation for
indoor use only using RGB-D cameras on the SUN RGB-D dataset. In [71], depth
and RGB were combined for multi-view semantic segmentation where depth was
leveraged to rewarp different views. Lin et al. [72] used a fully convolutional network
(FCN) with branch predictors to process layered discretized depth maps to match
Deep Semantic Segmentation in Autonomous Driving 163

them with the corresponding regions in the RGB image. The approach is used for
indoor scenes using the NYU dataset. Cao et al. [73] provided a way to learn depth
map and empirically proved enhanced performance for 2% on the VOC2012 dataset.

6.4.2 OPTICAL FLOW ESTIMATION

In addition to geometric information, there is a very important information prior that
can be used in autonomous driving applications (i.e., the motion information) [74]. In
a normal driving scenario, there are a lot of moving objects including cars, buses, and
pedestrians as well as the vehicle ego-motion. Optical fow provides a visualization
for the motion vectors of features in both x and y directions of the image. Moreover,
optical fow is already computed within the automotive vision platform and has been
a standard module for more than ten years. Due to hardware restrictions, sparse opti-
cal fow was used, but recently it has been replaced by dense optical fow.
In classical approaches, the most common way to generate sparse optical fow is
the Lucas-Kanade method [75], which assumes similar motion vectors on a small
window of 3 × 3 size as an example for every pixel at hand to be able to solve the
optical fow equation:

f x u + f y v + ft = 0 (6.1)

Where u and v denote the change in the pixels’ x and y positions with respect to
time, which implies the motion of a pixel in both directions; fx and f y denote image
derivatives along x and y directions; and ft is the image difference over time. The
Lucas-Kanade method is computed over previously selected features where the fnal
output will be sparse and hence provides computationally effcient solution. On the
contrary, Farneback’s method [76] computes optical fow, but for every pixel in the
image, and provides dense output where every pixel in the source frame is associ-
ated with a 2D vector showing values for pixel motion in x and y directions called u
and v. Both Lucas-Kanade and Farneback methods provide good performance only
when there is little motion between corresponding pixels in both images, and they
fail when the displacement is large.
On the other hand, deep learning has been used to formulate optical fow estima-
tion as a learning problem where temporally sequential images are used as input to
the network and the output is a dense optical fow. FlowNet [77] used a large-scale
synthetic dataset, namely “Flying Chairs”, as a ground truth for dense optical fow.
Two architectures were proposed. The frst one accepts an input of six channels for
two temporally sequential images stacked together in a single frame. The second is
a Mid-Fusion implementation in which each image is processed separately on one
stream and 2D correlation is done to output a DOF map. This work has been later
extended to FlowNet v2 [78] focusing on small displacement motion, where an addi-
tional specialized subnetwork is added to learn small displacements. Results are then
fused with the original architecture using another network to learn the best fusion. In
[79], the estimated fow is used to warp one of the two temporally sequential images
to the other, and the loss function is implemented to minimize the error between the
164 Deep Learning in Computer Vision

warped image and the original image. Junhwa and Stefan [80] implemented joint
estimation of optical fow and semantic estimation such that each task benefts from
the other.

6.5 PROPOSED FUSION ARCHITECTURES

The objective of this study is to experimentally investigate the impact of fusion of
both depth and optical fow with RGB images for enhanced semantic segmentation
of typical autonomous driving scenes. For this purpose, we implement three dif-
ferent segmentation network architectures: Single-Input network, Early-Fusion net-
work, and Mid-Fusion network.

6.5.1 SINGLE-INPUT NETWORK

An encoder-decoder architecture is employed to perform semantic segmentation.
For the encoder, the commonly used VGG-16 network [81] is used with the pre-
trained weights to initialize the encoder weights. A fully convolutional network
architecture, FCN8s [42], is adopted for semantic segmentation, and the four fully
connected layers of VGG16 are transformed to fully convolutional layers used to
learn the upsampling weights and restore the original image size. The fully convo-
lutional network is exploited for its fexibility with any input image size. In contrast
to fully connected networks, the fully convolutional ones preserve spatial relation-
ships between pixels at lower computational cost. Additionally, FCN8s skip connec-
tions combine deep semantic information with shallow ones to improve accuracy
as reported in [42].
In the proposed architecture, the convolutional layers are used for feature
extraction. Feature downsampling is performed using max-pooling layers, and
layer activation is done using rectifed linear unit (ReLU) activation function.
Afterwards, the segmentation decoder follows the FCN architecture and 1 × 1 con-
volutional layers are used to adapt depth channels, followed by three transposed
convolution layers to perform upsampling. In order to beneft from high-resolution
features, skip connections are utilized to add features from the shallow layers to
the partially upsampled deep ones to beneft from the high-resolution appearance.
The architecture of this network illustrated in Figure 6.4 is used to conduct experi-
ments for evaluation of a single information cue on semantic segmentation. The
same architecture is used for a single input of RGB-only, depth-only, and DOF-
only, which refers to dense optical fow.

FIGURE 6.4 Single-Input network utilized for semantic segmentation given only one piece
of information (e.g., RGB-only, depth-only, or DOF-only).
Deep Semantic Segmentation in Autonomous Driving 165

FIGURE 6.5 Early-Fusion Network: the pixel-level fusion is done on raw pixels before any
processing or feature extraction.

6.5.2 EARLY-FUSION NETWORK

Using this network, we investigate early fusion for multiple cues. The term “early
fusion” refers to pixel-level fusion, where the raw inputs are fused together before
extracting any features using the deep network. The input to the network in this
case is a 3D volume containing the original RGB image concatenated with the DOF
image, or the depth map. The RGB image consists of 3 layers, and the DOF map
consists of 1 layer. Concatenation is performed across the input volume’s depth axis,
so after concatenation we obtain an input 3D volume with the same image resolution,
and with a depth of 4 layers. In this network, the same architecture of the Single-
Input network is used and the VGG pre-trained weights are utilized to initialize the
network. However, the frst layer is adapted so that it accepts an input of four lay-
ers, and the corresponding weights are initialized randomly. Figure 6.5 shows the
concatenation of the 3-layer RGB with a 1-layer depth map for a 4-layer-total 3D
volume to be fed to the network. The same network is used to investigate the effect of
early fusion with DOF, where several optical fow representations have been studied
experimentally to fnd the best-performing input, namely, color wheel representa-
tion in 3 channels, magnitude and direction in 2 channels, and magnitude only in 1
channel.

6.5.3 MID-FUSION NETWORK

In this architecture, a two-stream network of two parallel VGG16 encoders for
extracting information from two different cues is implemented inspired by the works
[82–84]. The feature maps from both streams are fused together using a fusion
junction producing encoded features of the same dimensions as the output from
the Single-Input network encoder. The same decoder described in the Single-Input
network, which consists of three transposed convolution layers, is used to perform
upsampling. Figure 6.6 illustrates an example of a mid-level feature fusion approach
where the information from both streams is fused together using a summation junc-
tion after feature vector extraction. Following the same approach as the single-
stream network, skip connections are utilized to beneft from high-resolution feature
maps. This approach provides more fexibility for each stream to learn separately.
Accordingly, the outputs of both are then fused together, which has the advantage of
the ability to initialize both encoders with the VGG pre-trained weights. However,
the model is computationally more complex than the Early-Fusion single-stream
network.
166 Deep Learning in Computer Vision

FIGURE 6.6 Mid-Fusion Network: information of both streams is fused together.

6.6 EXPERIMENTS
6.6.1 EXPERIMENTAL SETTINGS
This section discusses experimental settings. For all experiments, the Adam opti-
mizer is used with a learning rate of value e−5. The L2 regularization is used in
the loss function, with a factor of value 5e −4 to overcome overftting the training
data. The VGG pre-trained weights are utilized for initialization. Dropout with
probability 0.5 is utilized for 1 × 1 convolution layers. Intersection over Union
(IoU) is used to evaluate class-wise accuracy, and the mean IoU is computed
over all the classes. In this work, the TensorFlow—an open-source library devel-
oped by Google—is utilized. TensorFlow can run on CPUs as well as GPUs, and
it is available for Windows, Linux, and macOS. In our experiments, we used
TensorFlow v1.0.0 on a Python 2.7 platform under a Linux Ubuntu v14.04 operat-
ing system, and the implemented networks are executed on TitanX GPU, Maxwell
architecture.
The datasets used in this work are chosen based on three factors: First, they
have to be automotive datasets containing urban road scenes, as our application
is mainly focused on autonomous driving. Second, the datasets should consist of
video sequences to allow us to compute DOF and use the motion information for
semantic segmentation. Third, there should be a way to calculate depth information
from the chosen dataset. Two well-known automotive datasets fulflling the require-
ments, namely, Virtual KITTI [61] and Cityscapes [57], are used in the experi-
ments. For the Virtual KITTI, we shuffed the dataset randomly and split the frames
to 80% training data, and 20% for test. In Cityscapes, we use the split provided
by the dataset, which is 3,474 images for training and 1,525 images for testing.
The optical fow annotation is not provided with the Cityscapes dataset; however,
there are video sequences containing temporal information about the scenes. Thus,
we exploit these sequences to generate different representations of dense optical
fow using two methods. We train the network to output all of the classes provided
in dataset; however, evaluation is done on 19 classes only, as mentioned by the
Cityscapes team where the offcial evaluation script published for the dataset is
exploited to report results.
Deep Semantic Segmentation in Autonomous Driving 167

6.6.2 RESULTS ANALYSIS AND DISCUSSION

6.6.2.1 Depth Experiments
Table 6.1 illustrates that depth augmentation consistently improves results. There is
an improvement of 5.7% in mean IoU. Although the overall improvement is incre-
mental, there is a large improvement for certain classes; for example, trucks, vans,
buildings, and traffc lights show improvements of 32%, 28%, 9%, and 8% respec-
tively. Table 6.2 shows quantitative evaluation for the Cityscapes dataset, and it had
a relatively moderate improvement of 1% in IoU. Results show that even noisy depth
maps with invalid values due to depth estimation errors can improve semantic seg-
mentation. Results on Virtual KITTI show that using concatenated feature maps
provides better results than using added feature maps. In Cityscapes, results for both
approaches are quite close, although the results from the monoDepth [69] estimator
using concatenated feature maps show slightly higher performance than those from
added feature maps, which leads to the same conclusion of the KITTI dataset.
Figure 6.11 illustrates qualitative results for the Cityscapes dataset. Figure 6.9
(g) shows better detection of the white van in the background, which has a uniform
texture; depth augmentation has helped to detect it better. Figure 6.11 (f) illustrates
better detection of the bus compared to RGB-only detection and better detection of
vegetation compared to depth-only, while showing degraded performance for the
poles in the background, due to the noise shown in the depth map; the input image
is shown in Figure 6.11 (b). Figure 6.11 (g) shows more immunity to noise regarding
the background poles compared to Figure 6.11 (f) due to greater fexibility obtained
from concatenated feature maps.
Depth-only network is reported to study the performance of depth cue alone
on semantic segmentation. Surprisingly, depth provides good results especially for
road, vegetation, vehicle, and pedestrians. This is also consistent with the results
obtained by [70], when depth-only is tested for indoor scenes. Although the depth-
only network shows good performance, the joint network shows only incremental
improvement. It is likely that using a simple network architecture does not effec-
tively combine depth and RGB, as they are completely different modalities with
different properties. Another possible reason is that the network trained on RGB has
the ability to learn geometric scene structure only implicitly.

6.6.2.2 Optical Flow Experiments

In this section, experiments were performed using different optical fow repre-
sentations to help us decide which one gives the best performance. We found that
results depend on the architecture used. In Cityscapes, we generated DOF using the
Farneback [76] algorithm, which provides noisy optical fow. The experiments were
conducted using FlowNet [78], the current state of the art for DOF computation,
which is based on a deep network trained on virtual datasets. Using the Single-Input
network, we attempted to investigate the relationship between motion cue alone and
semantic segmentation. Using optical fow only, the network learned to segment the
scene with good performance for moving objects. For static objects, the network
TABLE 6.1
168

Performance of Depth-Augmented Semantic Segmentation on the Virtual KITTI Dataset Using IoU Metric
Type Mean Truck Car Van Road Sky Vegetation Building Guardrail TraffcSign TraffcLight Pole

RGB 66.47 33.66 85.69 29.04 95.91 93.91 80.92 68.15 81.82 66.01 65.07 40.91
D (GT) 55 67.68 58.03 56.3 73.81 94.38 53.64 43.95 14.61 53.97 56.51 42.67
RGBD (GT) 66.76 65.34 91.74 56.93 95.46 94.41 79.17 54.91 73.42 60.21 46.09 30.46
RGB+D (GT-add) 68.6 43.38 91.59 29.19 69.01 94.32 85.17 77.6 80.13 69.54 72.73 32.09
RGB+D (GT-concat) 72.13 62.84 93.32 38.42 96.33 94.2 90.46 79.04 90.85 72.22 67.83 34.4
D (monoDepth) 46.1 36.05 75.46 33.2 77.3 87.3 39.3 32.3 6.8 42.14 45.9 15.9
RGB+D (monoDepth-add) 67.05 42.9 86.9 43.5 96.2 94.1 88.07 65.94 85.4 65.7 51.25 30.13
RGB+D (monoDepth-concat) 68.92 40.57 86.1 50.3 95.95 93.82 81.63 70.43 86.3 68.66 67.58 35.94

TABLE 6.2
Performance of Depth-Augmented Semantic Segmentation on the Cityscapes Dataset Using IoU Metric
Type Mean Bicycle Person Rider Motorcycle Bus Car Fence Building Road Sidewalk Sky TraffcSign

RGB 62.47 63.52 67.93 40.49 29.96 62.13 89.16 44.53 87.86 96.22 74.98 89.79 59.88
D (SGM) 47.8 39.84 54.99 29.04 11.29 48.1 82.36 34.32 78.42 95.15 67.78 81.18 27.96
RGBD (SGM) 55.5 56.68 60.27 34.64 21.18 58 86.94 36.47 84.7 94.84 70.39 84.64 45.48
RGB+D (SGM-add) 63.48 66.46 67.85 42.31 41.37 63.1 89.77 46.28 88.1 96.38 75.66 90.23 60.78
RGB+D (SGM-concat) 63.13 65.32 67.79 39.14 37.27 69.71 90.06 42.75 87.44 96.6 76.35 91.06 59.44
D (monoDepth) 40.89 36.63 44.6 18.5 7.3 37.5 77.78 16.16 77.01 92.83 54.87 89.33 24.67
RGB+D (monoDepth-add) 61.39 66.23 67.33 39.9 44.01 55.7 89.1 40.2 87.34 69.47 75.7 88.7 57.7
RGB+D (monoDepth-concat) 63.03 65.85 67.44 41.33 46.24 66.5 89.7 33.6 87.25 96.01 73.5 90.3 59.8
Deep Learning in Computer Vision
Deep Semantic Segmentation in Autonomous Driving 169

learned only the generate the scene structure without any further details. This result
motivated us toward the concept of fusion, which should help the network under-
stand the scene semantics from the RGB image and enhance performance using
motion information from optical fow. We used color wheel representation for this
task, encoding both fow magnitude and direction.
Table 6.3 shows an evaluation of four architectures on the Virtual KITTI data-
set. There is an improvement of 4% in mean IoU with the Mid-Fusion approach,
with larger improvements in moving classes like Truck, Car, and Van (38%, 6%,
and 27%, respectively), which demonstrates that the Mid-Fusion network provides
the best capability of capturing the motion cues and thus the best semantic seg-
mentation accuracy. The RGBF network provides very close performance to the
RGB-alone network, but with signifcant improvement in moving classes such as
Truck and Van. For all the results mentioned in Tables 6.3 and 6.4, we make use of
a color wheel representation format for fow. Table 6.4 illustrates the performance
on the Cityscapes dataset, indicating a marginal improvement of 0.1% in overall
IoU. There is a signifcant improvement in moving object classes like motorcycle
and train by 17% and 7%, even using noisy fow maps. The improvement obtained
from moving objects is signifcant compared to overall IoU, which is dominated by
sky and road pixels.
Quantitative comparisons between different DOF representations in the Early-
Fusion network RGBF and the Mid-Fusion RGB+F architectures for both Virual
KITTI and Cityscapes dataset are reported in Tables 6.5 and 6.6, respectively.
Results show that the color wheel Mid-Fusion network provides the best perfor-
mance, as color wheel representation encodes both magnitude and direction, while
the Mid-Fusion network provides the capacity for each stream to learn separately.
Hence, we obtained the best results using this approach. RGBF results are also
interesting: they show that the network is incapable of providing good output seg-
mentation when the input channels are increased to 6 layers including 3 for RGB
and 3 for color wheel DOF. However, augmenting only one channel for magnitude
DOF provides better results for moving classes such as Truck and Van in Virtual
KITTI. This might be benefcial for an embedded system with limited resources. In
the Cityscapes, the RGBF is incapable of providing better segmentation for several
reasons, one of which is the simple network architecture used and the large num-
ber of classes in the Cityscapes dataset trained for. Additionally, the Cityscapes has
only 3,475 images for training, while the Virtual KITTI has 17,008 frames. The
Virtual KITTI results show the need for more research in that direction, especially
for limited hardware resources. Figures 6.7 and 6.8 illustrate the qualitative results
of the four architectures on Virtual KITTI and Cityscapes, respectively. Figure 6.7
(f) shows better detection of the van, which has a uniform texture and which fow cue
has helped to detect more accurately. Figure 6.7 (h) shows that FlowNet-only results
provides good segmentation for the moving van with the ability to capture generic
scene structures like pavement, sky, and vegetation without color information.
However, fusion in Figure 6.7 (i) still needs to be improved. Figure 6.8 (f) and (h)
illustrate better detection of the bicycle and the cyclist after fusion with DOF. These
examples visually verify the accuracy improvement shown in Tables 6.3 and 6.4.
 170

TABLE 6.3
Performance of Flow-Augmented Semantic Segmentation Results on the Virtual KITTI Dataset Using IoU Metric
Type Mean Truck Car Van Road Sky Vegetation Building Guardrail TraffcSign TraffcLight Pole

RGB 66.47 33.66 85.69 29.04 95.91 93.91 80.92 68.15 81.82 66.01 65.07 40.91
F (GT) 42 36.2 55.2 20.7 62.6 93.9 19.54 34 15.23 51.5 33.2 29.3
RGBF (GT) 65.328 70.74 80.2 48.34 93.6 93.3 70.79 62.05 67.86 55.14 55.48 31.9
RGB+F (GT) 70.52 71.79 91.4 56.8 96.19 93.5 83.4 66.53 82.6 64.69 64.65 26.6
F (FlowNet) 28.6 24.6 47.8 14.3 57.9 68 13.4 4.9 0.8 31.8 18.5 6.6
RGB+F (FlowNet) 68.84 60.05 90.87 40.54 96.05 91.73 84.54 68.52 82.43 65.2 63.54 26.54

TABLE 6.4
Performance of Flow-Augmented Semantic Segmentation Results on the Cityscapes Dataset Using IoU Metric
Type Mean Bicycle Person Rider Motorcycle Bus Car Train Building Road Truck Sky TraffcSign

RGB 62.47 63.52 67.93 40.49 29.96 62.13 89.16 44.19 87.86 96.22 48.54 89.79 59.88
F (Farneback) 34.7 34.48 37.9 12.7 7.39 31.4 74.3 11.35 72.77 91.2 19.42 79.6 11.4
RGBF (Farneback) 47.8 52.6 55.8 31.1 22.4 39.34 82.75 22.8 80.43 92.24 20.7 81.87 44.08
RGB+F (Farneback) 62.56 63.65 66.3 39.65 47.22 66.24 89.63 51.02 87.13 96.4 36.1 90.64 60.68
F (FlowNet) 36.8 32.9 50.9 26.8 5.12 25.99 75.29 15.1 65.16 90.75 25.46 50.16 29.14
RGBF (FlowNet) 52.3 54.9 58.9 34.8 26.1 53.7 83.6 40.7 79.4 94 28.1 79.4 45.5
RGB+F (FlowNet) 62.43 64.2 66.32 40.9 40.76 66.05 90.03 41.3 87.3 95.8 54.7 91.07 58.21
Deep Learning in Computer Vision
TABLE 6.5
Semantic Segmentation Results on the KITTI for Different DOF (GT) Representations
Type Mean Truck Car Van Road Sky Vegetation Building Guardrail TraffcSign TraffcLight Pole

RGBF (GT-color wheel) 59.88 41.7 84.44 40.74 93.76 93.6 66.3 49.43 52.18 62.21 49.61 21.52
RGBF (GT-Mag & Dir) 58.85 45.12 82.3 30.04 90.25 94.1 60.97 56.48 51.48 58.74 49.7 26.01
RGBF (GT-Mag only) 65.32 70.73 80.16 48.33 93.59 93.3 70.79 62.04 67.86 55.13 55.48 31.92
RGB+F (GT-3 layers Mag only) 67.88 35.7 91.02 24.78 96.47 94.06 88.72 74.4 84.5 69.48 68.95 34.28
RGB+F (GT-color wheel) 70.52 71.79 91.4 56.8 96.19 93.5 83.4 66.53 82.6 64.69 64.65 26.6

TABLE 6.6
Semantic Segmentation Results on the Cityscapes for Different DOF (Farneback) Representations
Deep Semantic Segmentation in Autonomous Driving

Type Mean Bicycle Person Rider Motorcycle Bus Car Train Building Road Truck Sky TraffcSign

RGBF (Mag only) 47.8 52.63 55.82 31.08 22.38 39.34 82.75 22.8 80.43 92.24 20.7 81.87 44.08
RGBF (Mag & Dir) 54.6 57.28 58.63 33.56 18.49 56.44 84.6 41.15 84.41 95.5 31.8 87.86 44.26
RGBF (color wheel) 57.2 61.47 62.18 35.13 22.68 54.87 87.45 36.69 86.28 95.94 40.2 90.07 51.64
RGB+F (3 layers Mag only) 62.1 65.15 65.44 32.59 33.19 63.07 89.48 43.6 87.88 96.17 57.2 91.48 55.76
RGB+F (color wheel) 62.56 63.65 66.3 39.65 47.22 66.24 89.63 51.02 87.13 69.4 36.11 90.64 60.68
171
172 Deep Learning in Computer Vision

FIGURE 6.7 Flow augmentation: qualitative comparison of semantic segmentation outputs

from four architectures on the KITTI dataset.

Unexpectedly, Table 6.4 shows that Farneback and FlowNet v2 provide simi-
lar results; however, FlowNet shows better results for some classes like Truck. In
Figure 6.8, fusion with DOF in a two-stream approach enhances results of semantic
segmentation for both Farneback and FlowNet. Farneback DOF is noisy; however,
the result Figure 6.8 (f) shows that the network has good immunity towards noise.
Farneback has an advantage over FlowNet in fne details, where it is shown visually
that the rider can be separated from the bicycle; however, FlowNet doesn’t provide
this level of detail. As a result, fusion with Farneback provides slightly higher perfor-
mance than fusion with FlowNet. In both datasets, we observed some degradation in
accuracy of static objects after fusion despite the improvement obtained for moving
objects. This highlights the motivation towards multimodal fusion to combine RGB,
motion, and depth together, where depth signal is expected to improve the accuracy
of static objects. Moreover, there is a need for deeper investigation of fusion strategy
to maximize the beneft from each modality without loss in accuracy of other classes
(Tables 6.7 and 6.8).
Deep Semantic Segmentation in Autonomous Driving 173

FIGURE 6.8 Flow augmentation: qualitative comparison of semantic segmentation outputs

from four architectures on the Cityscapes dataset.
174 Deep Learning in Computer Vision

FIGURE 6.9 Depth augmentation: qualitative comparison of semantic segmentation out-

puts from four architectures on the KITTI dataset.

FIGURE 6.10 Depth augmentation: qualitative comparison of semantic segmentation out-

puts using the monoDepth estimator on the KITTI dataset.

6.7 CONCLUSION
This chapter focuses on computer vision tasks for autonomous driving applications.
It explores the autonomous driving feld and the role of recent computer vision
algorithms in self-driving car development, such as deep learning CNN, which
showed signifcant improvement in terms of accuracy for various tasks in different
Deep Semantic Segmentation in Autonomous Driving 175

FIGURE 6.11 Depth augmentation: qualitative comparison of semantic segmentation out-

puts from four architectures on the Cityscapes dataset.

applications. The main goal is to construct simple architectures to demonstrate the

beneft of fow and depth augmentation to standard CNN-based semantic segmen-
tation networks. The impact of both motion and depth information on semantic
segmentation was experimentally studied based on a simple network architecture.
Results of experiments on two public datasets demonstrate reasonable improvement
in overall accuracy. From the conducted experiments, it is shown that CNN can
generalize well even when using noisy maps such as SGM and Farneback; however,
the quality of the input map in terms of object structure largely affects the out-
put. Furthermore, the obtained experimental results show that there is still a lot of
176 Deep Learning in Computer Vision

FIGURE 6.12 Depth augmentation: qualitative comparison of semantic segmentation out-

puts using the monoDepth estimator on the Cityscapes dataset.

scope for improvement, as depth, fow, and color are different modalities, so more
research should be conducted to construct better fusion networks. Future research
aims to

• Conduct deeper investigation to implement more robust multimodal net-

works to combine valuable pieces of information.
• Understand the effect of different network architectures with varying com-
plexities in the multimodal fusion task.
• Incorporate sparsity invariant metrics to handle missing fow and depth
estimates.
• Migrate from pixel-level to object-level classifcation and make use of
depth per object, as it is more practical than pixel-wise depth, especially
in an AD application where the objective is to avoid collision with full
objects.
• Investigate multimodal fusion for different tasks such as moving and static
object detection, especially those employing structural constraints [85] and
compressive sensing techniques [86], which will be very helpful in the driv-
ing scenes.
• Train the network with fewer classes than the total number provided by the
dataset. In a normal driving scenario, information about all classes is not
critical for a vehicle to move. Dataset classes can be minimized to the ones
that are needed only by an AC.
• Train on multiple real and synthetic datasets with fne-tuning to solve the
problem of limited numbers of annotated images.
 TABLE 6.7
Comparison between Baseline, Depth-Augmented, Flow-Augmented, and Depth-Flow-Augmented Semantic
Segmentation Using Ground Truth
Type Mean Truck Car Van Road Sky Vegetation Building Guardrail TraffcSign TraffcLight Pole

RGB-only 66.47 33.66 85.69 29.04 95.91 93.91 80.92 68.15 81.82 66.01 65.07 40.91
RGB+D (GT) 72.13 62.84 93.32 38.42 96.33 94.2 90.46 79.04 90.85 72.22 67.83 34.4
RGB+F (GT) 70.52 71.79 91.4 56.8 96.19 93.5 83.4 66.53 82.6 64.69 64.65 26.6
RGB+D+F (GT) 71.88 71.688 92.08 61.44 95.85 94.83 84.32 71.86 83.42 64.69 60.67 31.08

TABLE 6.8
Comparison between Baseline, Depth-Augmented, Flow-Augmented, and Depth-Flow-Augmented Semantic Segmentation
Deep Semantic Segmentation in Autonomous Driving

Using Classical Estimators

Type Mean Bicycle Person Rider Motorcycle Bus Car Train Building Road Truck Sky TraffcSign

RGB-only 62.47 63.52 67.93 40.49 29.96 62.13 89.16 44.19 87.86 96.22 48.54 89.79 59.88
RGB+D (SGM) 63.48 66.46 67.85 42.31 41.37 63.1 89.77 46.28 88.1 96.38 75.66 90.23 60.78
RGB+F (Farneback) 62.56 63.65 66.3 39.65 47.22 66.24 89.63 51.02 87.13 96.4 36.1 90.64 60.68
RGB+D+F 62.58 65.46 68.18 43.09 37.5 64.4 88.34 57.13 86.86 96.13 41.45 87.96 59.23
177
178 Deep Learning in Computer Vision

BIBLIOGRAPHY
1. Richard S. Wallace, Anthony Stentz, Charles E. Thorpe, Hans P. Moravec, William
Whittaker, and Takeo Kanade. First results in robot road-following. In International
Joint Conferences on Artifcial Intelligence (IJCAI), pages 1089–1095. Citeseer, 1985.
2. Sebastian Thrun. Toward robotic cars. Communications of the ACM, 53(4): 99–106,
2010.
3. Matthew A. Turk, David G. Morgenthaler, Keith D. Gremban, and Martin Marra.
VITS-A vision system for autonomous land vehicle navigation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 10(3): 342–361, 1988.
4. Kichun Jo, Junsoo Kim, Dongchul Kim, Chulhoon Jang, and Myoungho Sunwoo.
Development of autonomous car-Part II: A case study on the implementation of an
autonomous driving system based on distributed architecture. IEEE Transactions on
Industrial Electronics, 62(8): 5119–5132, 2015.
5. Ernst Dieter Dickmanns. Dynamic Vision for Perception and Control of Motion.
Springer-Verlag, London, 2007.
6. Mahdi Rezaei and Reinhard Klette. Computer Vision for Driver Assistance. Springer,
Cham Switzerland, 2017.
7. Mahmoud Hassaballah and Khalid M. Hosny. Recent Advances in Computer Vision:
Theories and Applications, volume 804. Springer International Publishing, 2019.
8. David Michael Stavens. Learning to Drive: Perception for Autonomous Cars. Stanford
University, 2011.
9. Yuna Ro and Youngwook Ha. A factor analysis of consumer expectations for autono-
mous cars. Journal of Computer Information Systems, 59(1): 52–60, 2019.
10. Adil Hashim, Tanya Saini, Hemant Bhardwaj, Adityan Jothi, and Ammannagari Vinay
Kumar. Application of swarm intelligence in autonomous cars for obstacle avoidance.
In Integrated Intelligent Computing, Communication and Security, pages 393–404.
Springer, 2019.
11. Angelos Amanatiadis, Evangelos Karakasis, Loukas Bampis, Stylianos Ploumpis, and
Antonios Gasteratos. ViPED: On-road vehicle passenger detection for autonomous
vehicles. Robotics and Autonomous Systems, 112: 282–290, 2019.
12. Yusuf Artan, Orhan Bulan, Robert P. Loce, and Peter Paul. Passenger compartment vio-
lation detection in HOV/HOT lanes. IEEE Transactions on Intelligent Transportation
Systems, 17(2): 395–405, 2016.
13. Dorsa Sadigh, S Shankar Sastry, and Sanjit A. Seshia. Verifying robustness of human-
aware autonomous cars. IFAC-PapersOnLine, 51(34): 131–138, 2019.
14. Bhargava Reddy, Ye-Hoon Kim, Sojung Yun, Chanwon Seo, and Junik Jang. Real-time
driver drowsiness detection for embedded system using model compression of deep
neural networks. In IEEE Conference on Computer Vision and Pattern Recognition
Workshops, pages 121–128, 2017.
15. Theocharis Kyriacou, Guido Bugmann, and Stanislao Lauria. Vision-based urban navi-
gation procedures for verbally instructed robots. Robotics and Autonomous Systems,
51(1): 69–80, 2005.
16. SAE International Committee. Taxonomy and defnitions for terms related to on-
road motor vehicle automated driving systems. Technical Report J3016–201401, SAE
International, 2014. http://doi.org/ 10.4271/J3016_201806.
17. Jonathan Horgan, Ciaran Hughes, John McDonald, and Senthil Yogamani. Vision-
based driver assistance systems: Survey, taxonomy and advances. In International
Conference on Intelligent Transportation Systems (ITSC), pages 2032–2039.
IEEE, 2015.
18. Zhentao Hu, Tianxiang Chen, Quanbo Ge, and Hebin Wang. Observable degree analy-
sis for multi-sensor fusion system. Sensors, 18(12): 4197, 2018.
Deep Semantic Segmentation in Autonomous Driving 179

19. Shan Luo, Joao Bimbo, Ravinder Dahiya, and Hongbin Liu. Robotic tactile perception
of object properties: A review. Mechatronics, 48: 5467, 2017.
20. Guilherme N. DeSouza and Avinash C. Kak. Vision for mobile robot navigation: A survey.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(2): 237–267, 2002.
21. Oliver Pink, Jan Becker, and Soren Kammel. Automated driving on public roads:
Experiences in real traffc. IT-Information Technology, 57(4): 223–230, 2015.
22. Thomas Braunl. Embedded Robotics: Mobile Robot Design and Applications with
Embedded Systems. Springer Science & Business Media, 2008.
23. Gregory Dudek and Michael Jenkin. Computational Principles of Mobile Robotics.
Cambridge University Press, 2010.
24. Uwe Handmann, Thomas Kalinke, Christos Tzomakas, Martin Werner, and Werner von
Seelen. Computer vision for driver assistance systems. In Enhanced and Synthetic Vision
1998, volume 3364, pages 136–148. International Society for Optics and Photonics, 1998.
25. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):
436, 2015.
26. Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and Andrew
Y. Ng. Multimodal deep learning. In International Conference on Machine Learning
(ICML-11), pages 689–696, 2011.
27. Li Deng and Dong Yu. Deep learning: Methods and applications. Foundations and
Trends® in Signal Processing, 7(3–4): 197–387, 2014.
28. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classifcation with
deep convolutional neural networks. In Advances in Neural Information Processing
Systems, pages 1097–1105, 2012.
29. Josh Patterson and Adam Gibson. Deep Learning: A Practitioner’s Approach. O’Reilly
Media, Inc., 2017.
30. Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.
31. Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely
connected convolutional networks. In IEEE Conference on Computer Vision and
Pattern Recognition, pages 4700–4708, 2017.
32. Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network
training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
33. Risto Miikkulainen, Jason Liang, Elliot Meyerson, Aditya Rawal, Daniel Fink, Olivier
Francon, Bala Raju, Hormoz Shahrzad, Arshak Navruzyan, Nigel Duffy, et al. Evolving
deep neural networks. In Artifcial Intelligence in the Age of Neural Networks and
Brain Computing, pages 293–312. Elsevier, 2019.
34. Michael Hauser, Sean Gunn, Samer Saab Jr, and Asok Ray. State-space representations
of deep neural networks. Neural Computation, 31(3): 538–554, 2019.
35. Abhinav Valada, Gabriel L. Oliveira, Thomas Brox, and Wolfram Burgard. Deep
multispectral semantic scene understanding of forested environments using multi-
modal fusion. In International Symposium on Experimental Robotics, pages 465–477.
Springer, 2016.
36. Taigo M. Bonanni, Andrea Pennisi, D.D Bloisi, Luca Iocchi, and Daniele Nardi.
Human-robot collaboration for semantic labeling of the environment. In 3rd Workshop
on Semantic Perception, Mapping and Exploration, pp. 1–6, 2013.
37. Abhijit Kundu, Yin Li, Frank Dellaert, Fuxin Li, and James M. Rehg. Joint semantic
segmentation and 3D reconstruction from monocular video. In European Conference
on Computer Vision, pages 703–718. Springer, 2014.
38. Ondrej Miksik, Vibhav Vineet, Morten Lidegaard, Ram Prasaath, Matthias Nießner,
Stuart Golodetz, Stephen L. Hicks, Patrick Perez, Shahram Izadi, and Philip H.S. Torr.
The semantic paintbrush: Interactive 3d mapping and recognition in large outdoor
spaces. In 33rd Annual ACM Conference on Human Factors in Computing Systems,
pages 3317–3326. ACM, 2015.
180 Deep Learning in Computer Vision

39. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler,
Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes data-
set for semantic urban scene understanding. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), June 2016.
40. Gabriel J. Brostow, Julien Fauqueur, and Roberto Cipolla. Semantic object classes in
video: A high-defnition ground truth database. Pattern Recognition Letters, 30(2):
88–97, 2009.
41. Clement Farabet, Camille Couprie, Laurent Najman, and Yann LeCun. Learning
hierarchical features for scene labeling. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 35(8): 1915–1929, 2013.
42. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks
for semantic segmentation. In IEEE Conference on Computer Vision and Pattern
Recognition, pages 3431–3440, 2015.
43. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E . Imagenet classifcation with deep
convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q.
Weinberger, editors, Advances in Neural Information Processing Systems, 25, pages
1097–1105. Curran Associates, Inc., 2012.
44. Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-
scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
45. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper
with convolutions. In IEEE Conference on Computer Vision and Pattern Recognition,
pages 1–9, 2015.
46. Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network
for semantic segmentation. In IEEE International Conference on Computer Vision,
pages 1520–1528, 2015.
47. Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolu-
tional encoder-decoder architecture for image segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 39(12): 2481–2495, 2017.
48. Jun Mao, Xiaoping Hu, Xiaofeng He, Lilian Zhang, Liao Wu, and Michael J. Milford.
Learning to fuse multiscale features for visual place recognition. IEEE Access, 7:
5723–5735, 2019.
49. Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions.
arXiv preprint arXiv:1511.07122, 2015.
50. Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L. Yuille. Attention to
scale: Scale-aware semantic image segmentation. In EEE Conference on Computer
Vision and Pattern Recognition, pages 3640–3649, 2016.
51. Boyu Chen, Peixia Li, Chong Sun, Dong Wang, Gang Yang, and Huchuan Lu. Multi
attention module for visual tracking. Pattern Recognition, 87: 80–93, 2019.
52. Liuyuan Deng, Ming Yang, Hao Li, Tianyi Li, Bing Hu, and Chunxiang Wang.
Restricted deformable convolution based road scene semantic segmentation using sur-
round view cameras. arXiv preprint arXiv:1801.00708, 2018.
53. Varun Ravi Kumar, Stefan Milz, Christian Witt, Martin Simon, Karl Amende,
Johannes Petzold, Senthil Yogamani, and Timo Pech. Monocular fsheye camera depth
estimation using sparse lidar supervision. In International Conference on Intelligent
Transportation Systems (ITSC), pages 2853–2858. IEEE, 2018.
54. Marius Cordts, Timo Rehfeld, Lukas Schneider, David Pfeiffer, Markus Enzweiler,
Stefan Roth, Marc Pollefeys, and Uwe Franke. The stixel world: A medium-level repre-
sentation of traffc scenes. Image and Vision Computing, 68: 40–52, 2017.
55. Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation
and recognition using structure from motion point clouds. In European Conference on
Computer Vision, pages 44–57. Springer, 2008.
Deep Semantic Segmentation in Autonomous Driving 181

56. Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driv-
ing? The kitti vision benchmark suite. In Conference on Computer Vision and Pattern
Recognition, pp. 3354–3361, 2012.
57. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler,
Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes data-
set for semantic urban scene understanding. arXiv preprint arXiv:1604.01685, 2016.
58. Heiko Hirschmuller. Accurate and effcient stereo processing by semi-global match-
ing and mutual information. In IEEE Conference on Computer Vision and Pattern
Recognition, volume 2, pages 807–814. IEEE, 2005.
59. Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder.
The mapillary vistas dataset for semantic understanding of street scenes. In IEEE
International Conference on Computer Vision, pages 4990–4999, 2017.
60. Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang,
Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving.
In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages
954–960, 2018.
61. Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as
proxy for multi-object tracking analysis. In IEEE Conference on Computer Vision and
Pattern Recognition, pages 4340–4349, 2016.
62. German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M.
Lopez. The synthia dataset: A large collection of synthetic images for semantic seg-
mentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 3234–3243, 2016.
63. NuTonomy. Nuscenes dataset. 2012. https://www.nuscenes.org/.
64. Magnus Wrenninge and Jonas Unger. Synscapes: A photorealistic synthetic dataset for
street scene parsing. arXiv preprint arXiv:1810.08705, 2018.
65. Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan
L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous
convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 40(4): 834–848, 2018.
66. Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single
monocular images using deep convolutional neural felds. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 38(10): 2024–2039, 2016.
67. Yuanzhouhan Cao, Zifeng Wu, and Chunhua Shen. Estimating depth from mon-
ocular images as classification using deep fully convolutional residual networks.
IEEE Transactions on Circuits and Systems for Video Technology, 28(11): 3174–
3182, 2018.
68. Tinghui Zhou, Matthew Brown, Noah Snavely, and David G. Lowe. Unsupervised
learning of depth and ego-motion from video. In IEEE Conference on Computer Vision
and Pattern Recognition, pages 1851–1858, 2017.
69. Clement Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular
depth estimation with left-right consistency. In IEEE Computer Vision and Pattern
Recognition, volume 2, page 7, 2017.
70. Caner Hazirbas, Lingni Ma, Csaba Domokos, and Daniel Cremers. Fusenet:
Incorporating depth into semantic segmentation via fusion-based cnn architecture. In
Asian Conference on Computer Vision, pages 213–228. Springer, 2016.
71. Lingni Ma, Jorg Stückler, Christian Kerl, and Daniel Cremers. Multi-view deep
learning for consistent semantic mapping with RGB-d cameras. arXiv preprint
arXiv:1703.08866, 2017.
72. Di Lin, Guangyong Chen, Daniel Cohen-Or, Pheng-Ann Heng, and Hui Huang.
Cascaded feature network for semantic segmentation of rgb-d images. In IEEE
International Conference on Computer Vision (ICCV), pages 1320–1328. IEEE, 2017.
182 Deep Learning in Computer Vision

73. Yuanzhouhan Cao, Chunhua Shen, and Heng Tao Shen. Exploiting depth from single
monocular images for object detection and semantic segmentation. IEEE Transactions
on Image Processing, 26(2): 836–846, 2017.
74. Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic infor-
mation and deep matching for optical fow. In European Conference on Computer
Vision, pages 154–170. Springer, 2016.
75. Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with
an application to stereo vision. In Imaging Understanding Workshop, pages 121–130.
Vancouver, British Columbia, 1981.
76. Gunnar Farneback. Two-frame motion estimation based on polynomial expansion. In
Scandinavian Conference on Image Analysis, pages 363–370. Springer, 2003.
77. Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas,
Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet:
Learning optical fow with convolutional networks. In IEEE International Conference
on Computer Vision, pages 2758–2766, 2015.
78. Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and
Thomas Brox. Flownet 2.0: Evolution of optical fow estimation with deep networks.
In IEEE Conference on Computer Vision and Pattern Recognition, pages 2462–2470,
2017.
79. Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha.
Unsupervised deep learning for optical fow estimation. In Thirty-First AAAI
Conference on Artifcial Intelligence, pp. 1495–1501, 2017.
80. Junhwa Hur and Stefan Roth. Joint optical fow and temporally consistent semantic
segmentation. In European Conference on Computer Vision, pages 163–177. Springer,
2016.
81. Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-
scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
82. Mennatullah Siam, Heba Mahgoub, Mohamed Zahran, Senthil Yogamani, Martin
Jagersand, and Ahmad El-Sallab. Modnet: Moving object detection network with
motion and appearance for autonomous driving. arXiv preprint arXiv:1709.04821,
2017.
83. Suyog Dutt Jain, Bo Xiong, and Kristen Grauman. Fusionseg: Learning to combine
motion and appearance for fully automatic segmention of generic objects in videos.
arXiv preprint arXiv:1701.05384, 2017.
84. Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for
action recognition in videos. In Advances in Neural Information Processing Systems,
pages 568–576, 2014.
85. Mennatullah Siam and Mohammed Elhelw. Enhanced target tracking in uav imagery
with pn learning and structural constraints. In Proceedings of the IEEE International
Conference on Computer Vision Workshops, pages 586–593, 2013.
86. Ahmed Salaheldin, Sara Maher, and Mohamed Helw. Robust real-time tracking with
diverse ensembles and random projections. In Proceedings of the IEEE International
Conference on Computer Vision Workshops, pages 112–120, 2013.
 7 Aerial Imagery
Registration Using
Deep Learning for
UAV Geolocalization
Ahmed Nassar, and Mohamed ElHelw

CONTENTS
7.1 Introduction .................................................................................................. 183
7.2 An Integrated Framework for UAV Localization......................................... 185
7.2.1 Framework Architecture................................................................... 185
7.2.2 Calibration ........................................................................................ 187
7.2.3 Registration and Detection Using Semantic Segmentation .............. 188
7.2.4 Registration by Shape Matching....................................................... 190
7.2.5 Detection by Semantic Segmentation............................................... 193
7.2.6 Registration of Sequence Frames ..................................................... 193
7.3 Results and Analysis..................................................................................... 194
7.3.1 Platform ............................................................................................ 194
7.3.2 Datasets............................................................................................. 194
7.3.3 Earth Observation Image Registration ............................................. 198
7.3.4 Semantic Segmentation ....................................................................200
7.3.5 Semantic Segmentation Registration Using Shape
Matching Results ..............................................................................205
7.4 Conclusion ....................................................................................................206
Bibliography ..........................................................................................................207

7.1 INTRODUCTION
The proliferation of drones, also known as unmanned aerial vehicles (UAVs), has
shifted from being used for military operations to being available for the public mar-
ket. This conception came through the advancement and availability of mature hard-
ware components such as Bluetooth, Wi-Fi, radio communication, accelerometers,
gyroscopes, high-performance processors with low consumption, and effcient bat-
tery technologies. This rise can be attributed to the increasing smartphone demand,
which led to the race of manufacturing better hardware components, which brought
benefts to other sectors such as UAVs [1]. The smartphone market has led to the

183
184 Deep Learning in Computer Vision

affordability of these elements since it reduced prices and the size of the hardware.
Now UAVs can beneft from all the different types of components and modules in
many various applications, such as photography, mapping, agriculture, surveillance,
rescue operations, and the recently famous delivery systems. The UAV applications
mentioned previously and the ones available in the current market are currently
being controlled by radio communications through a controller, a software that plans
the route, or by using the on-board camera.
Conventionally, UAV navigation has relied on inertial sensors, on altitude meters,
and sometimes on a global positioning system within their solution to feed com-
mands to their motor controller. The latter modality uses the Global Navigation
Satellite System (GNSS) composed of different regional systems such as GPS (which
is owned by the United States), GLONASS, Galileo, Beidou, and others located at
the Medium Earth Orbit with the purpose of providing autonomous geo-spatial posi-
tioning [2]. This positioning is provided in latitude, longitude, and altitude data. Both
terms, “GPS” and “GNSS,” are commonly used interchangeably in UAV naviga-
tion literature. The signal transmitted from the GPS satellites could be affected by
the distance and the ambient/background noise emanating in numerous places with
higher vulnerability to interference and/or blockade in urban areas with buildings
causing loss of line of sight (LOS) to navigational satellites. However, some areas are
GPS denied, meaning that the GPS signal is not available or weak, and this neces-
sitates fnding alternative approaches to UAV guidance.
A plethora of work has been proposed for UAV localization by using inertial
measurement unit (IMU) readings and on-board camera to acquire features that are
stored in a feature database. Scaramuzza et al. [3], Chowdhary et al. [4], and Rady
et al. [5] offer examples of these methods. During UAV fight, information in the
pre-acquired feature database is used to approximate UAV location based on match-
ing currently extracted features with those in the database. Similar work [6] uses
online mapping services and feature extraction to be able to localize a UAV as well
as terrain classifcation. Mantelli et al. [7] used abBRIEF for global localization and
estimated the UAV pose in 4 degrees of freedom.
Currently, the abundance of satellite and aerial high-resolution imagery [8] cov-
ering most of the globe has facilitated the emergence of new applications such as
autonomous vision-only UAV navigation. In this case, UAV images are compared
with aerial/satellite imagery to infer the drone location as shown in Figure 7.1.
However, processing these images is computationally demanding and even more
involved if data labeling is needed to process and extract actionable information. The
ability to correlate two images of the same location under different viewing angles,
illumination conditions, occlusions, object arrangements, and orientations is crucial
to enable robust autonomous UAV navigation. In traditional computer vision, image
entropy and edges are used as features and compared for similarity as presented by
Fan et al. [9] and Sim et al. [10]. Other works use local feature detectors [11], such
as Scale-Invariant Feature Transform (SIFT) [12] or Speeded-Up Robust Features
(SURF) [13], to produce features that are then matched, providing affne transfor-
mations as investigated by Koch et al. [14] and Shukla [15]. Using Simultaneous
Localization And Mapping (SLAM), UAV navigation is achieved by acquiring and
correlating distinct features of indoor and/or outdoor regions in order to map the
Deep Learning for UAV Geolocalization 185

FIGURE 7.1 Registration of two images from different domains.

UAV to a specifc location. This method is more suitable to constrained environ-

ments and involves mainly micro UAVs and quadcopters. Conte et al. [16] and Yol
et al. [17] use template matching with cross-correlation where the aerial image is
used as a template to match another georeferenced image.
Recently, deep learning has been extensively used for aerial image analysis. To
this end, semantic segmentation based on fully convolutional networks is applied
on satellite imagery in order to classify each pixel within the image to a common
class such as surfaces (pavements, hard surfaces), buildings, vegetation, trees, cars,
clutter/background, and roads, among others [18, 19, 20, 21]. In previous work [22],
the comparison of multiple images is achieved by using a Siamese Network [23]. In
this case, a Siamese Network [24] was trained with two types of images: a satellite
image and a perspective-transformed panoramic image. Subsequently, features from
both images were extracted and the distance between the features of each image was
calculated to fnd out if they are similar.

7.2 AN INTEGRATED FRAMEWORK FOR UAV LOCALIZATION

7.2.1 FRAMEWORK ARCHITECTURE
A framework is proposed with the aim of geolocalizing a UAV while fnding out
its orientation and height using only its on-board camera [25]. While designing the
framework, the process is conceptualized based on how a person from a bird’s eye or
aerial view would be able to describe his/her location. Several components are used
to emulate this process using traditional computer vision and recent state-of-the-art
deep learning methods. To demonstrate the framework, Figure 7.3 is used to explain
every module and component, and they will be described in further detail in the
upcoming sections.
The primary goal is to localize the frames from the UAV with the earth obser-
vation images. The UAV sequence frames will be denoted as S = (S(1),S(2),…Sn) and
186 Deep Learning in Computer Vision

FIGURE 7.2 Left panel is a Sn, with its corresponding Sn ROI from Bing Maps in Famagusta,
Cyprus on the right panel.

FIGURE 7.3 The pipeline of the UAV localization framework.

their labels will be represented by L = (L(1), L(2),…Ln). In this literature, the Earth
Observation (EO) image will be referred to as the reference map, its notation will be
R and its label is M. The extents of r on the map and its pixel GPS coordinates are
known. Therefore, registering Sn onto r will provide back the GPS coordinates of the
Sn. Initially, Sn and its GPS location coordinates on r are assumed to be known, since
the person or system deploying the UAV knows its initial location.
Two main inputs are fed to the localization framework. The frst corresponds to
prerecorded top-down view footage S acquired from 1) a UAV fying over a specifc
region and 2) a simulated overfight generated by using Google Earth. The second,
is a high-resolution EO imagery map obtained from free public online services such
Deep Learning for UAV Geolocalization 187

as Google Maps, Bing Maps, or existing datasets such as ISPRS’s Potsdam, which is
used for r as demonstrated in Figure 7.2. Consequently, the different components of
the framework deal with those two inputs, and the techniques used will be evaluated
on how they deal with these inputs to localize the aerial vehicle.

7.2.2 CALIBRATION
Calibration is an integral part of the framework and is called upon at S(1) and every
fve iterations (S(n+5)) frames. It is responsible for fnding the initial location, the dif-
ference in scale, and orientation between the Sn and its corresponding feld of view
in r, and to autocorrect any drift that might occur. To accomplish this task, SIFT is
used to make sure that the localization estimation is checked and updated every fxed
period. That is, under the condition of being the frst frame, or the third sequence
frame, S and r are passed to the SIFT registration, which gives out a homography
that is applied to r and also updates the drone location as shown in Figure 7.4. SIFT
is robust; however, it is time consuming and thus it is not used frequently to reduce
computational overhead.
As stated before, knowing the initial GPS coordinates, a region of interest r
is cropped from r around the coordinate, given that r = R(i, ws+b) where i is the
coordinate at the center and the width is determined by ws+b, which is equal to the
width of S plus a margin b. This is done for the purpose of limiting the search for
the features extracted from Sn in r. As seen in Figure 7.5, r covers a wider region
than the one in S. Furthermore, to get the pixel location i from the coordinate, this
conversion is calculated by Equations 7.1 and 7.2, which maps the GPS bounds of R
to its pixel size, which returns for each pixel its actual GPS location. This process
results in reducing the search space and eliminates the possibility for the matching
to go astray.

(width max − width min )(lon − lon w )

pix x = (7.1)
(lon e − lon w )

(height max − height min )(lat − lat n )

pix y = (7.2)
(lat s − lat n )

lat n + (lat n − lat s )(pix h − height min )

lat = (7.3)
(height max − height min )

lon w + (lon e − lon w )(pix w − width min )

lon = (7.4)
(width max − width min )

Where pixx: Pixel’s location in x axis, pix y: Pixel’s location in y axis, heightmin: r
height minimum (0), heightmax: r height maximum, widthmin: r width minimum (0),
widthmax: r width maximum, lon: current longitude/coordinate, lat: current latitude
coordinate, latn: North latitude bound of r, lats: South latitude bound o r, lone: East
latitude bound of r, and lonw: West latitude bound of r.
188 Deep Learning in Computer Vision

FIGURE 7.4 Calibration using SIFT registration, where location estimation is checked and
updated every fxed period.

Using the search space created, features are extracted from both r and Sn.
Afterward, the features are matched and used to estimate a homography matrix that
in turn is used to extract transformation components. Forthwith, the rotation and
scale components are maintained and are incremented upon or decremented depend-
ing on the next frame or transformation. This helps calculate the orientation of the
UAV using the rotation and altitude by using scale in relation to r.

7.2.3 REGISTRATION AND DETECTION USING SEMANTIC SEGMENTATION

Using down-looking images, drone localization is improved by using available key
visual landmarks such as roads and buildings. Semantic segmentation is utilized to
help extract specifc shape types that are subsequently matched for enhanced results.
Deep Learning for UAV Geolocalization 189

FIGURE 7.5 Registration using SIFT for calibration. The circles represent feature key
points extracted, and the lines represent the matching of them. The polygon represents the
boundaries of the Sn (left) in r (right) after the transformations.

To be able to accomplish this as seen in Figure 7.6, Sn and r are fed into a U-Net
network and the segmented outputs are subsequently matched.
The Segmentation Network: The task to segment input images (r, S) provides
classifcation of input image pixels and is accomplished using U-Net network whose
architecture is yet another encoding and decoding model that follows the network archi-
tecture of Ronneberger et al. [26]. The segmentation network is able to classify image
regions into buildings, roads, and vegetation. Due to the size of images and their num-
ber, each class was trained separately, resulting in a model for each class in each dataset.
Data Pre-processing: The images that are used to train the network, presented
in Figure 7.3.2, each have a large size exceeding 2000 × 2000 pixels per image.
Therefore, the images are split into smaller patches of size 224 × 224 pixels, similar
to the sizes used by VGG [27] or ResNet [28]. Simple normalization is applied on
each patch by dividing the pixel value by its channel’s highest value. After that, the
patches are stitched to form the full image. Consequently, the same is applied to the
ground-truth images (M,L).

FIGURE 7.6 Two U-Net networks take in Sn and r and provide segmented shapes.
190 Deep Learning in Computer Vision

Moreover, data augmentation is applied to these patches before training. The

patches are rotated randomly, which helps the model to distinguish different objects
with different orientations. They are also fipped horizontally and vertically. On the
whole, these techniques create more data to be used in training, which increases the
model accuracy.
Optimization Algorithm: The optimization algorithm chosen was the Adaptive
Moment Estimator (ADAM) [29]. It’s known to be an update to RMSProp [30] and
is widely acknowledged to be superior to it. By keeping the exponentially decaying
average of all past gradients, ADAM is able to calculate an adaptive learning rate for
all network parameters. In early experiments, ADAM converged much faster than
stochastic gradient descent and NADAM [31].
Loss Function: The loss function used was Dice’s coeffcient, which is also
known as the Sørensen index [32]. Basically, Dice’s coeffcient fnds the similarity
of objects in image segmentation by fnding the overlap between a ground-truth and
the output provided by a method.

2TP
Dicescore = (7.5)
2TP + FP + FN
Where
TP: True Positive
FP: False Positive, and
FN: False Negative.

As shown in Equation 7.2.3, the Dice score is calculated by fnding the positives
and penalizing the false positives found as well as positives that the method could
not fnd. Dice is similar to the Jaccard index [33], which is another commonly used
loss function for image segmentation. However, Dice is currently being favored due
to being more intuitive than Jaccard, which only counts true positives once in both
the numerator and denominator.
Regularization and Normalization: To avoid overftting, regularization using
dropout layers was added after every convolutional layer with a value of 0.05. After
every dropout layer, a batch normalization layer was also added to normalize the
activation of each batch.
Network Initialization Using Transfer Learning: Training from scratch is
computationally expensive in terms of processing and time. Transfer learning is
especially relied upon to reduce the training time and in case of small datasets.
In fact, it is diffcult to have a dataset that will be capable of catering to truly deep
networks. Models trained on large datasets, such as ImageNet [27], contain generic
weights that can be used in different tasks. The pre-trained model is trained using a
random image generator that also creates the ground-truth for those images.

7.2.4 REGISTRATION BY SHAPE MATCHING

After processing Sn and r through the network, two images are generated contain-
ing only buildings and roads as seen in Figure 7.7. The two images are then each
split into two layers: buildings and roads. Using computer vision techniques such as
Deep Learning for UAV Geolocalization 191

FIGURE 7.7 Buildings segmented using U-Net (the left panel is an EO image, in the mid-
dle panel is the result of segmentation, and the right panel is an EO image overlaid with
segmentation).

contouring, the boundaries of different objects (buildings or roads) are determined.

Subsequently, region features such as area, shape, and location of each object are
computed. Consequently, the contour of every segment in each image is selected
and compared to objects in the other image for a match; this process is illustrated in
steps in Figure 7.8.
All matches are given a score based on contour area, spatial distance, radius,
Hausdorff distance [34], centroid, and orientation using Hu moment [35] as shown
in Figure 7.9. The Hausdorff distance measures the distance between two sets of
points created from shapes. The Hu moment is used to calculate the orientation and
centroid of the shape. The Hausdorff distance and Hu moments are used as features
in order to fnd the similarity between the shapes.

FIGURE 7.8 The segmented images are processed to fnd their properties such as area, dis-
tance, location, image moments, etc., which for each shape or object acts as a feature. These
features are matched to create a homography.
192 Deep Learning in Computer Vision

FIGURE 7.9 The left panel shows segmented Sn images, and the right panel shows seg-
mented r images. The contours are computed around a segmented image along with the area,
and radius is calculated.

Afterwards, the matches with the highest scores are chosen as presented in
Figure 7.10. However, matches are eliminated if the score is close and/or within a
short distance, since this might be an unreliable match between 2 objects that are
of the same size and within the same distance. For example, a grid of buildings that
look the same would confuse the matches. To conclude, as done with the calibra-
tion step, using a threshold the top matches that are found to be reliable are used to
calculate another homography. This homography, which is estimated from semantic
segmentation, is applied to calculate the current location of the UAV accurately.

FIGURE 7.10 Matches of segmented buildings.

Deep Learning for UAV Geolocalization 193

7.2.5 DETECTION BY SEMANTIC SEGMENTATION

An additional feature that semantic segmentation adds to this work is that it provides
contextual knowledge of the UAV surroundings. Knowing if the UAV is fying above
a building or road is very benefcial in surveillance, delivery, and search and rescue
missions. For instance, building and road detection is achieved by using the classi-
fed image with a box drawn around each contour; having defned the pixel values of
each class, each contour can be identifed.

7.2.6 REGISTRATION OF SEQUENCE FRAMES

Seeking a real-time motion estimation technique for S, the same method performed
in the calibration component (Section 7.2.2). Differently, the image registration pro-
cess is being done between Sn and Sn+f with f being the frame rate. The frame rate is
used to skip in-between frames since it will be exhausting to register all consecutive
frames with insignifcant changes. In fact, registering two consecutive images from a
sequence of the same domain can exploit the fact that given smooth platform motion,
the difference between the frames is insignifcant, as seen in Figure 7.12. Therefore,
ORB features were an appropriate choice for this task since they are effcient alter-
natives to SURF and SIFT. It is important to note that ORB cannot deal with scale
invariance effectively, and thus SIFT is used in the intermittent calibration step
(check Section 7.2.2) for updating UAV scale and height. As shown in Figure 7.11,
the UAV location is updated depending on the homography estimated from consecu-
tive frames.
Likewise, ORB produces a homography that is used to apply a translation that
is added to the current pixel location to calculate the fnal UAV coordinate using
Equations. 7.3 and 7.4. It is important to note that in case of insuffcient matches,
the calibration step, discussed in Section 7.2.2, is called upon for enhanced
reliability.

FIGURE 7.11 Two consecutive images from the UAV footage are taken and a homography
is estimated from them to update the current UAV location.
194 Deep Learning in Computer Vision

FIGURE 7.12 Registration of S using ORB.

7.3 RESULTS AND ANALYSIS

7.3.1 PLATFORM
To be able to carry out the experiments, many libraries and confgurations were
tried before settling on the current setup. Naturally, the operating system chosen was
Ubuntu*, a Debian-based Linux operating system that is available for free and is sup-
ported on many laptops, servers, and embedded development kits such as NVIDIA
Jetson Tx1/Tx2†. As deep learning is involved in this work, the GPU performance
was harnessed by using CUDA 8‡ and cuDNN v5§. The code was mostly written
in C++ to be able to cross-compile the code on multiple platforms and obtain fast
execution [36]. The Python programming language was also used to integrate with
deep learning libraries through Python wrappers. To train the network, IRISA-UBS
Lab Cluster ¶ was used. The lab has the following specifcations: Intel Xeon CPU
E5-2687W (3.10 GHz × 20 cores) and NVIDIA 900 GTX TitanX.

7.3.2 DATASETS
Due to lack of established datasets to benchmark UAV path navigation using cam-
eras, datasets had to be collected and ground-truth manually created to be able to
produce performance metrics. The datasets used in this project can be split into
two categories: aerial datasets and satellite imagery datasets. Multiple sources are
used to cater to different views (rooftops, roads, color). A dataset was also created

* https://www.ubuntu.com
† https://www.nvidia.com/en-us/autonomous-machines/embedded-systems-dev-kits-modules/
‡ https://www.geforce.com/hardware/technology/cuda
§ https://developer.nvidia.com/cudnn
¶ http://cluster-irisa.univ-ubs.fr/wiki/doku.php/accueil
Deep Learning for UAV Geolocalization 195

using images obtained from online mapping services for different locations and their
ground-truth created by hand using OpenStreetMap (OSM)*.
ISPRS WG II/4 Potsdam: This dataset [20] is provided by the ISPRS Working
Group II/4 for the sole purpose of detecting objects and aiding in 3D reconstruc-
tion in urban scenes. The dataset covers two cities: Vaihingen (Germany), a small
village with small buildings, and Potsdam (Germany), a dense city with big blocks
of large buildings and narrow streets. In this work, we focus on the Potsdam data-
set because it contains RGB information and the images refect a state-of-the-art
airborne image dataset in a very high resolution. Figure 7.13 shows samples of the
dataset.
The datasets come with ground-truth or label images to classify six classes:
Impervious Surfaces, Buildings, Low Vegetation, Trees, Cars, and Clutter/
Background. The Potsdam dataset is made of 38 patches of the same size trimmed
from a larger mosaic and a Digital Surface Model (DSM). The resolution of both
data types is 5 cm. The images come in different channel compositions: IRRG (3
channels IR-R-G), RGB (3 channels R-G-B), and RGBIR (4 channels R-G-B-IR).
This dataset has proven very benefcial throughout this work due to semantically
segmented labels and high resolution.
Potsdam Extended: In this work, Google Satellite Images (GSI) are used and
thus images from GSI with the same extents and zoom level for 12 of the original
ISPRS dataset locations were obtained. Subsequently, the ground-truth from ISPRS
is used and updated if there is a difference between recent GSI and ISPRS. Along

FIGURE 7.13 A sample of the tiles available from the ISPRS Potsdam 2D Semantic
Labeling Challenge.

* https://www.openstreetmap.org/
196 Deep Learning in Computer Vision

with the corresponding ground-truth from ISPRS, several tiles were generated from
areas outside the ISPRS dataset and ground-truth produced using OSM as shown in
Figure 7.13. The difference between both ISPRS datasets and Google Satellite Image
can be seen in Figures 7.14 and 7.15 (Table 7.1).
UAV Videos: One of the major diffculties that faced this work was the unavail-
ability of multiple data sources for one location, especially the top-down view video
from a UAV. To address this problem, previously acquired UAV footage from the
Internet was used.
The frst video used is shot in Famagusta (Cyprus), acquired with top-down view
and was available on YouTube by the name of “Above Varosha: one of the most
famous ghost cities” [37]. The UAV motion in the video was smooth as it was gliding
seamlessly with no vibrations or sudden changes in direction or camera orientation
(Figure 7.16). Figure 7.17 shows frame samples from the video. For every 32 frames,
ground-truth position was manually computed as GPS coordinates by comparison
with the other GSI map. The excerpt of the video that was good enough to be used
was 43 secs in length.
Having high-resolution aerial imagery from ISPRS of Potsdam inspired the idea
of simulating a UAV footage of this area. This was implemented by using Google
Earth 3D and navigating a camera with top-down view. The corresponding video
footage was captured while gliding over a region in Potsdam. The captured video
length is 2 minutes.

FIGURE 7.14 A sample of the Potsdam Extended dataset. On the top row, the left panel is
the corresponding Google Satellite Image copying ISPRS’s tile and using its ground-truth.
The middle and right panels are tiles produced from other areas in Potsdam. The bottom row
is OSM as the ground-truth.
Deep Learning for UAV Geolocalization 197

FIGURE 7.15 A comparison between and ISPRS Potsdam (left panel) and a Google Satellite
image (right panel).

TABLE 7.1
Comparison of the Different Datasets
Dataset name Resolution Area (Sq. km) Classes

Potsdam 5 2.16 6
Potsdam Extended 9.1 1.24 3
Famagusta 19 1.13 3

FIGURE 7.16 A sample of frames acquired from the Famagusta UAV video.

FIGURE 7.17 A sample of frames acquired from the Potsdam UAV video.
198 Deep Learning in Computer Vision

7.3.3 EARTH OBSERVATION IMAGE REGISTRATION

The calibration component is responsible for registering two images from two
sources together (S, r) to obtain the transformation describing UAV motion. The
registration component that is trying to register consecutive images S from the on-
board camera has a much easier job, since the images are from the same source with
minor differences. It is important to note that this experiment is carried out without
the corrective assistance of the semantic segmentation registration. Furthermore, the
experiment is evaluated on a ground-truth manually created.
The Calibration Component Experiment: The calibration step starts frst on
S(1) where the GPS location is known, being the source, and goes through the reg-
istration process with r cropped from r. After every fve frames processed by the
framework, this component is set to action. The Famagusta Sequence is 1374 frames
in total. After every 32 frames (video captured at 32 fps), a frame is processed,
since it is ineffcient and time consuming to process every frame. Subsequently, the
calibration step is called nine times. On the other hand, the Potsdam Sequence is
composed from 10,800 frames, and so the calibration component is called every 60
frames (video captured at 60 fps). As explained before, since ground-truth does not
exist, the registration had to be evaluated using the following algorithm:

1. Manually, another path is created from the center points of the UAV
sequence frames, which provides our ground-truth.
2. Features are extracted from the Sn and r.
3. If 70% of the matches are found have a distance of 0.2 or less, then a homog-
raphy is created.
4. Apply homography to the reference map ROI, which results in a warped
image.
5. A GPS coordinate is estimated for the center pixel of the warped frame
using the method explained in Section 7.2.2 and Equation 7.2.2.
6. The distance between the estimated GPS coordinate and the ground-truth is
calculated by using Haversine Distance Formula, as illustrated by Figure 7.18.

The UAV Sequence Experiment: As explained in Section 7.2.6, the registration of

the S is done using ORB features due to their effciency. Carrying over from the cali-
bration step, the image registration process done between Sn and Sn+f yields a transla-
tion that is added to the location estimated from the calibration step. Therefore, the
registration of S is called for the same number of frames. In fact, since the UAV is
not likely to travel that far in one second, Sn and Sn+f are likely to cover nearly identi-
cal scene areas.
In our experiment, these steps are taken to ensure that a registration is satisfactory
before taking it into account:

1. A path is created from the center points of the UAV sequence frames to set
up the ground-truth.
2. Features are then extracted from both consecutive frames passed from the
sequence.
Deep Learning for UAV Geolocalization 199

FIGURE 7.18 The left panel is Sn, and the right panel is the estimated r. The actual UAV
GPS coordinate is {52.404758, 13.053362}, while the estimated GPS coordinate is {52.404737,
13.053338}.

3. A higher tolerance is set here in comparison to Section 7.3.3 since it is an

undemanding process; therefore a homography is created if only 80% of the
matches found have a distance of 0.15 or less.
4. The translation component provided from the homography is added to the
previous GPS component, to estimate the current location.
5. The estimated locations create an estimated fight path of the UAV.

Local Feature Registration Results: The output of the pipeline excluding the seman-
tic segmentation pipeline performs relatively well overall. As shown in Table 7.2,
the estimated GPS coordinates are deviating 8.3 m from the actual fight path. As
expected, the SIFT calibration step helps the system to “catch up” when it is lagging
due to differences in scale (altitude) or photogrammetric differences. Without the
SIFT calibration step, longer videos or fights continue to drift until system failure or
until the system is no longer capable of comparing or estimating the location.
Qualitatively as seen in Figures 7.19 and 7.20 using this method alone, the devia-
tion to the eye does not seem signifcant at that height (300 m). Also, it is apparent in

TABLE 7.2
Deviation in Meters Comparing the Two Approaches
Dataset Without local features With local features

Famagusta 26 m 10.4 m
Potsdam 52.1 m† 8.3 m
† Only 60% of the Potsdam Sequence was used to produce this result
before localization drifted way and was not capable of fnding accurate
position.
200 Deep Learning in Computer Vision

FIGURE 7.19 The actual path of the UAV and the estimated one visualized on a map using
their GPS coordinates for the Famagusta Sequence.

FIGURE 7.20 The actual path of the UAV and the estimated one visualized on a map using
their GPS coordinates for the Potsdam Sequence. The intersecting rounded line represents
where the experiment ended due to accumulated drift.

several cases that the pipeline performs well if the buildings are exactly perpendicu-
lar from a top-down view and their facades are not visible.

7.3.4 SEMANTIC SEGMENTATION

As explained in Section 7.2.3, the semantic segmentation component was devised
for two reasons. The frst is to be able to replicate the methodology humans would
Deep Learning for UAV Geolocalization 201

use to navigate by comparing two aerial or satellite images together trying to fnd
matching structures to correct position. The second reason is to avail more infor-
mation to the UAV, hence making it aware of the types of objects it is fying over,
which could be benefcial in route planning or decision making. This component
builds upon the previous steps to increase the accuracy of localization and reduce
the deviation or drift of the UAV from its planned trajectory. In this section, the setup
of the experiment will be presented with all the different components and choices
made for the network including the training process. Afterward, the metrics chosen
for evaluation are presented along with a discussion of the results and the outcome
of the experiments.
The Semantic Segmentation Experiment: As a result of hardware constraints,
many models were pre-trained for this experiment. Each region (Famagusta,
Potsdam) had its separate model, and each class (buildings, roads, etc.) was trained
separately using U-Net (check Section 7.2.3). The pre-trained model was trained
with a random image generator [38] that was used to initially start training the dif-
ferent models. The data was split into 224 × 224 pixel patches, and then the patches
were augmented using vertical, horizontal, and random rotation and fed to the U-Net
network.
Training: Many confgurations and different runs have been performed for each
model with different hyper-parameters. The training is stopped instantly when the
average F1-score on the validation dataset stopped providing better scores. The aver-
age number of epochs trained for each model was 6 and each epoch took almost 3.5
hours to train.
Metrics: To evaluate the semantic segmentation method, the average F1-score is
used as an error metric on each trained class. Precision is the outcome of how many
positive predictions were predicted in fact positive. Recall is the measure of how
many positive predictions are predicted correctly. The F1-score is the average or the
harmonic mean of both precision and recall.

precision × recall
F1 = 2 (7.6)
precision + recall

TP
Precision = (7.7)
TP + FP

TP
Recall = (7.8)
TP + FN

Semantic Segmentation Results: The experiment was run over two regions,
Potsdam and Famagusta. The performance of this step was evaluated experimen-
tally using two ground-truth values since each city has its UAV image sequence
and its satellite or aerial reference map as presented in Section 7.3.2. The perfor-
mance of the semantic segmentation in our pipeline was evaluated in terms of
localization.
To evaluate the model fairly, the ISPRS Potsdam 2D Semantic Labeling Challenge
was used as a benchmark to compare how our model performs against other work.
202 Deep Learning in Computer Vision

However, this work was compared to work that solely uses RGB images in their
training and doesn’t utilize DSM (Digital Surface Models) (Table 7.3).
Experiment I was trained purely on ISPRS Potsdam dataset using 19 images
for training and 5 for validation. The purpose of this experiment was to fnd
out how well U-Net works in comparison to the other work based on bench-
mark scores. The experiment was run only for the building and roads classes,
which are the most important landmarks for this region since it is an urban area.
In general, the Experiment I model performed better across the 2 classes and
resulted in an average score of 84.7%. Unfortunately, the building class score
lagged behind, but this experiment was the basis on which the other experiments
were built (Table 7.4).
After testing the model generated from Experiment I on the UAV sequence with
source images from GSI, the results were not satisfactory. So Experiment II was
trained on the Potsdam Extended dataset (Section 1.3.2) using Experiment I model
for pre-training. The frst convolutional layer was frozen while training this experi-
ment since the frst layer contains the basic shapes and edges. The scores were satis-
factory when tested and the predictions provided sharp edges with hollow shapes as
shown in Figure 7.21.

TABLE 7.3
Comparison to Other Work Using ISPRS
Potsdam 2D Semantic Labeling
Challenge (RGB Images Only)*
Method Average Building Roads

CASIA2a [39] 95.5% 97% 93.3%

Experiment I 84.7% 85.1% 84.3%
Kaiser et al. [40] 83.85% 91.3% 76.4%

* The frst score is highlighted in bold.

a This entry is added for a full picture. It is the highest
score available on the ISPRS Potsdam 2D Challenge;
however, this method uses DSM data.

TABLE 7.4
Semantic Segmentation Results Using Our Models*
Method Average Building Roads Vegetation

Experiment II 84.43% 87.98% 85.2% 80.1%

Experiment III 85.7% 88% 83.4% –
Experiment IV 76% 76.3% 74.7% 77.4%

* The frst score is highlighted in bold, the second is underlined, and

the third is in italics.
Deep Learning for UAV Geolocalization 203

FIGURE 7.21 A sample of the predictions made using Experiment II. The left column
shows the target images, the middle column shows the predictions, and the right column is
the combination of the previous columns. These target images are Potsdam’s (frst 3 rows) and
Famagusta’s (last row) S. Buildings are labeled cyan, and roads are labeled red.

Experiment III was carried out to validate if freezing Experiment II’s frst con-
volutional layers was a good idea. Therefore, in this experiment the frst convolu-
tional layer was unfrozen. The Experiment III model provided a fuller shape but with
inaccurate edges in comparison to Experiment II as shown in Figure 7.22. However,
this model provided the highest score. To demonstrate the procedure of only training
from the online map services using OSM as a ground-truth, Experiment IV was
204 Deep Learning in Computer Vision

FIGURE 7.22 A sample of the predictions made using Experiment III. The left column
is the target image, the middle column is the prediction, and the right column is the pre-
diction overlaying the target image. These target images are Potsdam’s S. Buildings are
labeled in cyan.

arranged. As expected, the results were behind Experiments II and III by nearly
10%. This proves that high-resolution data with pixel-wise accurate ground-truths
defnitely had an effect (Table 7.5).
In general, our task was not to come up with the state-of-the-art segmentation
method, but to choose the best segmentation method for the proposed framework.
Although Experiment III provided the highest score, practically Experiment II was
the model used due to its sharper shapes, whereas hollow shapes were remedied
using morphological operations. Qualitatively, the buildings that are close to each
other are challenging to segment. There is also a discrepancy in segmenting build-
ings from sidewalks, which are sometimes considered part of the building, resulting
in added diffculty for the matching process during the registration step. Another

TABLE 7.5
The Specifcations That Trained Our Network, after Trying
Many Variations in Terms of Filter Size or Optimizers
Method Filter size Batch size LR Optimizer Epochs

Experiment I 3×3 224 0.001 ADAM 6

Experiment II 3×3 224 0.001 ADAM 6
Experiment III 3×3 224 0.001 ADAM 7
Experiment IV 3×3 224 0.001 ADAM 8
Deep Learning for UAV Geolocalization 205

point to consider when using OSM as a ground-truth is that OSM labels treat trees
and vegetation as one class. This inconsistency or misclassifcation between vegeta-
tion and tree pixels was accounted for by merging. However, for our application the
difference between trees and vegetation is not that vital. The result of the segmenta-
tion step usually contains inaccuracies that have to be remedied using morphological
techniques such as dilation and erosion, which are applied to fll empty hollows in
the shapes and remove small regions.

7.3.5 SEMANTIC SEGMENTATION REGISTRATION USING SHAPE MATCHING RESULTS

Similarly to the evaluation provided in Section 7.3.3(c) for the local feature-based
registration, the same is done for the registration using shape matching. This method
takes Sn and its corresponding S(r) and matches the equivalent shapes segmented to
register them. Once again, the path estimated by our framework is compared to the
ground-truth of GPS coordinates manually labeled. Practically, the shape-matching
step reduces drift errors and signifcantly improves localization.
As seen in Table 7.6 the deviation from the ground-truth is reduced to a 5.1-
meter error for Famagusta, and considerably for Potsdam to 3.6 meters as shown in
Figure 7.23 and illustrated in Figure 7.25. This result could be explained by the low
quality of Famagusta’s r. As for Potsdam, the quality of both r and S contributed to
better localization as shown in Figure 7.24 and illustrated in Figure 7.26. However,

TABLE 7.6
Deviation in Meters with the Addition of
Semantic Segmentation to the Pipeline
Without With local With semantic
Dataset local features features segmentation

Famagusta 26 m 10.4 m 5.1 m

Potsdam 52.1 m 6.3 m 3.6 m

FIGURE 7.23 This graph shows the deviation from the actual GPS coordinate in meters for
every time the calibration function is called for the Famagusta Sequence.
206 Deep Learning in Computer Vision

FIGURE 7.24 This graph shows the deviation from the actual GPS coordinate in meters for
every time the calibration function is called for the Potsdam Sequence.

FIGURE 7.25 The ground-truth path of the UAV and the estimated path visualized on a
map using their GPS coordinates using registration by shape matching in Famagusta.

it has been observed that in Potsdam, building blocks are tightly spaced, so when
segmented they create huge blobs that were very diffcult to match at a low altitude.
Better localization would be achieved in less dense areas or with fner segmentation.
In this section, the feasibility of our framework was investigated. The require-
ments of the system were frst laid out and the framework was introduced. The pro-
cess of acquiring datasets, their specifcations, and the creation of their ground-truths
were presented along with justifcation for picking specifc datasets. Afterward, the
main registration components were broken down and evaluated focusing on their
effectiveness within the framework for localizing the UAV separately.

7.4 CONCLUSION
The essence of this work is to demonstrate the capability of using computer vision to
geolocalize a UAV without using any inertial sensors or GPS but using only an on-board
camera. The proposed method is then demonstrated in its entirety and the process from
start to end is explained in an orderly fashion. Each component is broken down into
Deep Learning for UAV Geolocalization 207

FIGURE 7.26 The ground-truth path of the UAV and the estimated path visualized on a
map using their GPS coordinates using registration by shape matching in Potsdam.

its subcomponents and its purpose is clarifed and evaluated. The proposed method
is advantageous over other methods that use only local feature detectors to match the
UAV image to an offine reference map, even when searching within a window as the
experiments have shown. We believe that segmentation and matching shapes helps
avoid matching the wrong features in areas that look similar or have an urban pattern.
However, a major disadvantage of the method is areas with no recognizable landmarks
or features, but this is a common disadvantage of all methods that depend on vision
solely. Future work is to experiment with larger datasets that cover multiple cities with
different visual features, which would increase the robustness of the proposed method.
Another element is to use a trained neural network to compare the shapes of the seg-
mented objects and replace the demanding dictionary search and heuristics employed
by the current method.

BIBLIOGRAPHY
1. K. D. Atherton, “Senate hearing: Drones are “basically fying smart-phones”.” https://
www.popsci.com/technology/article/2013–03/ how-drone-smartphone, March 2013.
2. B. Hofmann-Wellenhof, H. Lichtenegger, and E. Wasle, GNSS-Global Navigation
Satellite Systems: GPS, GLONASS, Galileo, and More. Springer Science & Business
Media, 2007.
208 Deep Learning in Computer Vision

3. D. Scaramuzza, M. C. Achtelik, L. Doitsidis, F. Friedrich, E. Kosmatopoulos, A.

Martinelli, M. W. Achtelik, M. Chli, S. Chatzichristofs, L. Kneip, et al., “Vision-
controlled micro fying robots: From system design to autonomous navigation and map-
ping in gps-denied environments,” IEEE Robotics & Automation Magazine, vol. 21,
no. 3, pp. 26–40, 2014.
4. G. Chowdhary, E. N. Johnson, D. Magree, A. Wu, and A. Shein, “Gps-denied indoor
and outdoor monocular vision aided navigation and control of unmanned aircraft,”
Journal of Field Robotics, vol. 30, no. 3, pp. 415–438, 2013.
5. S. Rady, A. Kandil, and E. Badreddin, “A hybrid localization approach for uav in GPS
denied areas,” in IEEE/SICE International Symposium on System Integration, pp.
1269–1274, IEEE, 2011.
6. A. Masselli, R. Hanten, and A. Zell, “Localization of unmanned aerial vehicles using
terrain classifcation from aerial images,” in Intelligent Autonomous Systems, 13, pp.
831–842, Springer, 2016.
7. M. Mantelli, D. Pittol, R. Neuland, A. Ribacki, R. Maffei, V. Jorge, E. Prestes, and M.
Kolberg, “A novel measurement model based on abbrief for global localization of a UAV
over satellite images,” Robotics and Autonomous Systems, vol. 112, pp. 304–319, 2019.
8. V. Sowmya, K. Soman, and M. Hassaballah, “Hyperspectral image: Fundamentals and
advances,” in Recent Advances in Computer Vision, pp. 401–424, Springer, 2019.
9. B. Fan, Y. Du, L. Zhu, and Y. Tang, “The registration of UAV down-looking aerial
images to satellite images with image entropy and edges,” in Intelligent Robotics and
Applications, pp. 609–617, 2010.
10. D.-G. Sim, R.-H. Park, R.-C. Kim, S. U. Lee, and I.-C. Kim, “Integrated position esti-
mation using aerial image sequences,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 24, no. 1, pp. 1–18, 2002.
11. M. Hassaballah and A. I. Awad, “Detection and description of image features: An
introduction,” in Image Feature Detectors and Descriptors, pp. 1–8, Springer, 2016.
12. D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International
Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004.
13. H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool, “Speeded-up robust features (SURF),”
Computer Vision and Image Understanding, vol. 110, no. 3, pp. 346–359, 2008.
14. T. Koch, P. d’Angelo, F. Kurz, F. Fraundorfer, P. Reinartz, and M. Korner, “The tum-
dlr multimodal earth observation evaluation benchmark,” in IEEE Conference on
Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016.
15. P. Shukla, S. Goel, P. Singh, and B. Lohani, “Automatic geolocation of targets tracked by aer-
ial imaging platforms using satellite imagery,” International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences, vol. 40, no. 1, p. 381, 2014.
16. G. Conte and P. Doherty, “Vision-based unmanned aerial vehicle navigation using geo-
referenced information,” EURASIP Journal on Advances in Signal Processing, vol.
2009, p. 10, 2009.
17. A. Yol, B. Delabarre, A. Dame, J.-E. Dartois, and E. Marchand, “Vision-based abso-
lute localization for unmanned aerial vehicles,” in 2014 IEEE/RSJ International
Conference on Intelligent Robots and Systems, pp. 3429–3434, IEEE, 2014.
18. N. Audebert, B. Le Saux, and S. Lefvre, “Semantic segmentation of earth observa-
tion data using multimodal and multi-scale deep networks,” in Asian Computer Vision
Conference, pp. 180–196, Springer, Cham, November 2016.
19. D. Marmanis, J. D. Wegner, S. Galliani, K. Schindler, M. Datcu, and U. Stilla, “Semantic
segmentation of aerial images with an ensemble of CNNs,” ISPRS Annals of the
Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. 3, p. 473, 2016.
20. F. Rottensteiner, G. Sohn, J. Jung, M. Gerke, C. Baillard, S. Benitez, and U. Breitkopf,
“The ISPRS benchmark on urban object classifcation and 3d building reconstruc-
tion,” ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information
Sciences, vol. 1, no. 3, pp. 293–298, 2012.
Deep Learning for UAV Geolocalization 209

21. V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and

M. Riedmiller, “Playing atari with deep reinforcement learning,” arXiv preprint
arXiv:1312.5602, 2013.
22. S. Lefvre, D. Tuia, J. D. Wegner, T. Produit, and A. S. Nassar, “Toward seamless mul-
tiview scene analysis from satellite to street level,” Proceedings of the IEEE, vol. 105,
pp. 1884–1899, October 2017.
23. S. Chopra, R. Hadsell, and Y. LeCun, “Learning a similarity metric discriminatively,
with application to face verifcation,” in IEEE International Conference on Computer
Vision and Pattern Recognition, vol. 1, pp. 539–546, IEEE, 2005.
24. J. Bromley, I. Guyon, Y. LeCun, E. Sackinger, and R. Shah, “Signature verifcation
using a “siamese” time delay neural network,” in Advances in Neural Information
Processing Systems, pp. 737–744, 1994.
25. A. Nassar, K. Amer, R. ElHakim, and M. ElHelw, “A deep CNN-based framework
for enhanced aerial imagery registration with applications to uav geolocalization,”
in IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp.
1513–1523, 2018.
26. O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedi-
cal image segmentation,” arXiv preprint arXm:1505.0Ą597, 2015.
27. K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale
image recognition,” arXiv preprint arXm:1Ą09.1556, 2014.
28. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in
IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778, 2016.
29. D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint
arXiv:1412.6980, 2014.
30. T. Tieleman and G. Hinton, “Lecture 6.5-rmsprop: Divide the gradient by a run-
ning average of its recent magnitude,” COURSERA: Neural Networks for Machine
Learning, vol. 4, no. 2, pp. 26–31, 2012.
31. T. Dozat, “Incorporating nesterov momentum into adam,” in International Conference
on Learning Representations, 2016.
32. T. Sørensen, “A method of establishing groups of equal amplitude in plant sociology
based on similarity of species and its application to analyses of the vegetation on danish
commons,” Biol. Skr., vol. 5, pp. 1–34, 1948.
33. P. Jaccard, “Le coeffcient generique et le coeffcient de communaute dans la fore
marocaine.” Mémoires de la Société Vaudoise des Sciences Naturelles, 2: 385–403,
1926.
34. D. P. Huttenlocher, G. A. Klanderman, and W. J. Rucklidge, “Comparing images
using the hausdorff distance,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 15, no. 9, pp. 850–863, 1993.
35. M.-K. Hu, “Visual pattern recognition by moment invariants,” IRE Transactions on
Information Theory, vol. 8, no. 2, pp. 179–187, 1962.
36. R. Hundt, “Loop recognition in C++/Java/Go/Scala,” Proceedings of Scala Days, p. 38,
2011.
37. A. FPOHOB, “Above varosha: One of the most famous ghost cities.” https://youtu.be/
AYKC0dsrh4U, February 2016.
38. ZFTurbo, “ZF UNET 224 pretrained model.” https://github.com/ZFTurbo/ZF_UNET_
224_Pretrained_Model, 2017.
39. Y. Liu, B. Fan, and C. Pan, “Method description for potsdam: 2D labelling challenge.”
http://www2.isprs.org/commissions/comm3/wg4/ 2d-sem-label-potsdam.html.
40. P. Kaiser, J. D. Wegner, A. Lucchi, M. Jaggi, T. Hofmann, and K. Schindler, “Learning
aerial image segmentation from online maps,” IEEE Transactions on Geoscience and
Remote Sensing, vol. 55, no. 11, pp. 6054–6068, 2017.
 8 Applications of Deep
Learning in Robot Vision
Javier Ruiz-del-Solar and Patricio Loncomilla

CONTENTS
8.1 Introduction .................................................................................................. 211
8.2 Requirements in Robot Vision...................................................................... 212
8.3 Robot Vision Methods .................................................................................. 213
8.3.1 Object Detection and Categorization................................................ 213
8.3.2 Object Grasping and Manipulation .................................................. 216
8.3.3 Scene Representation and Classifcation .......................................... 218
8.3.4 Spatiotemporal Vision ...................................................................... 221
8.4 Future Challenges ......................................................................................... 222
8.5 Conclusions...................................................................................................224
Acknowledgments.................................................................................................. 225
Bibliography .......................................................................................................... 225

8.1 INTRODUCTION
Deep learning is a hot topic in the pattern recognition, machine learning, and com-
puter vision research communities. This fact can be seen clearly in the large number
of reviews, surveys, special issues, and workshops that are being organized, and
the special sessions in conferences that address this topic (e.g., [1–9]). Indeed, the
explosive growth of computational power and training datasets, as well as technical
improvements in the training of neural networks, has allowed a paradigm shift in
pattern recognition, from using hand-crafted features (e.g., Histograms of Oriented
Gradients, Scale Invariant Feature Transform, and Local Binary Patterns) [10]
together with statistical classifers to the use of data-driven representations, in which
features and classifers are learned together. Thus, the success of this new paradigm
is that “it requires very little engineering by hand” [4], because most of the parame-
ters are learned from the data, using general-purpose learning procedures. Moreover,
the existence of public repositories with source code and parameters of trained deep
neural networks (DNNs), as well as that of specifc deep learning frameworks/tools
such as Tensorfow [11], has promoted the use of deep learning methods.
The use of the deep learning paradigm has facilitated addressing several com-
puter vision problems in a more successful way than with traditional approaches.
In fact, in several computer vision benchmarks, such as the ones addressing image
classifcation, object detection and recognition, semantic segmentation, and action

211
212 Deep Learning in Computer Vision

recognition, just to name a few, most of the competitive methods are now based on
the use of deep neural networks. In addition, most recent presentations at the fagship
conferences in this area use methods that incorporate deep learning.
Deep learning has already attracted the attention of the robot vision community.
However, given that new methods and algorithms are usually developed within the
computer vision community and then transferred to the robot vision one, the ques-
tion is whether or not new deep learning solutions to computer vision and recog-
nition problems can be directly transferred to robotics applications. This transfer
is not straightforward considering the multiple requirements of current solutions
based on deep neural networks in terms of memory and computational resources,
which in many cases include the use of GPUs. Furthermore, following [12], it must
be considered that robot vision applications have different requirements from
standard computer vision applications, such as real-time operation with limited
on-board computational resources, and the constraining observational conditions
derived from the robot geometry, limited camera resolution, and sensor/object
relative pose.
Currently, there are several reviews and surveys related to deep learning [1–7].
However, they are not focused on robotic vision applications, and they do not con-
sider their specifc requirements. In this context, the main motivation of this chapter
is to address the use of this new family of algorithms in robotics applications. This
chapter is intended to be a guide for developers of robot vision systems. Therefore,
the focus is on the practical aspects of the use of deep neural networks rather than on
theoretical issues. It is also important to note that as convolutional neural networks
(CNNs) [7–13] are the most commonly used deep neural networks in vision applica-
tions, the analysis provided in this chapter will be focused on them.
This chapter is organized as follows: In Section 2, requirements for robot vision
applications are discussed. In Section 3, relevant papers related to robot vision are
described and analyzed. In Section 4, current and future challenges of deep learning
in robotic applications are discussed. Finally, in Section 5, conclusions related to
current robot vision systems and its usability are drawn.

8.2 REQUIREMENTS IN ROBOT VISION

The aim of robotics is to build machines that can behave, i.e., execute tasks, autono-
mously while interacting with their environment. For reaching autonomy, a robot
must be able both to perceive the environment using sensors, and to process the sen-
sorial data for obtaining information relevant to the task to be solved. Robot vision
encompasses algorithms that use data acquired from cameras for performing several
tasks, such as object detection and categorization, object grasping and manipulation,
representation and classifcation of scenes, and spatiotemporal vision, among others.
These capabilities are needed for performing complex tasks needed in applications
like pick-and-place robotics, 3D mapping, automated harvesting, service robotics,
and self-driving cars, just to name a few.
Robotics imposes signifcant restrictions in the algorithms that can be used: (i)
most autonomous robots use batteries as power sources, so hardware capabilities
must be adapted for increasing available operation time; (ii) robots must operate in
Applications of Deep Learning in Robot Vision 213

real time and be responsive; and (iii) several processes alongside robot vision must
be executed in parallel. These conditions are needed for a robot to behave and to
complete tasks autonomously in the real world, as these tasks commonly involve
time constraints for being solved successfully. Then, the computational workload
is a crucial component that must be considered when designing solutions for robot
vision tasks.
In the following sections, recent advances based on deep learning related to robot
vision and future challenges are described. As stated before, deep neural networks
require a considerable amount of computational power, which currently limits their
usability in robotics applications. However, the development of new processing plat-
forms for massive parallel computing with low power consumption will enable the
adoption of a wide variety of deep learning algorithms in robotic applications.

8.3 ROBOT VISION METHODS

In this section, recently published studies presented at fagship journals, robotics
conferences, and symposia are described and analyzed. The works have been cat-
egorized in four application areas: Object Detection and Categorization, Object
Grasping and Manipulation, Scene Representation and Classifcation, and
Spatiotemporal Vision. The most representative works and the observed tendencies
related to the use of deep learning for addressing robot vision problems in each area
are described in the next sections. It is important to note that many of the described
works have not been validated in real robots operating under real-world conditions,
but only in databases corresponding to robot applications. Therefore, in these cases
it is not clear whether the proposed methods can work in real time using a real robot
platform.

8.3.1 OBJECT DETECTION AND CATEGORIZATION

Object detection and categorization is a fundamental ability in robotics. It enables
a robot to execute tasks that require interaction with object instances in the real
world. Deep learning is already being used for general-purpose object detection and
categorization, as well as for pedestrian detection, and for detecting objects in robot-
ics soccer. State-of-the-art methods used for object detection and categorization are
based on generating object proposals, and then classifying them using a deep neural
network (DNN), enabling systems to detect thousands of different object categories.
As will be shown, one of the main challenges for the application of DNNs for object
detection and characterization in robotics is real-time operation. It must be stressed
that obtaining the required object proposals for feeding the DNNs is not real time
in the general case, and that, on the other hand, general-purpose object detection
and categorization methods based on DNNs are not able to run in real time in most
robotics platforms. These challenges are addressed by using task-dependent methods
for generating few, fast, and high-quality proposals for a limited number of pos-
sible object categories. These methods are based on using other information sources
for segmenting the objects (depth information, motion, color, etc.), and/or by using
object-specifc, non-general-purpose, weak detectors for generating the required
214 Deep Learning in Computer Vision

proposals. Also, smaller DNN architectures can be used when dealing with a limited
number of object categories.
It is worth mentioning that most of the reported studies do not indicate the frame
rate needed for full object detection/categorization, or they show frame rates that
are far from being real time. In generic object detection methods, computation of
proposals using methods like Selective Search or EdgeBoxes takes most of the time
[14]. Systems like Faster R-CNN [15] that compute proposals using CNNs are able
to obtain higher frame rates, but require the use of high-end GPUs to work (see
Figure  8.1). Also, methods derived from YOLO (You Only Look Once) [16] offer
a better runtime, but they both impose limits on the number of objects that can be
detected, and have a slightly lower performance. The use of task-specifc knowl-
edge-based detectors on depth [17–20], motion [21], color segmentation [22], or weak
object detectors [23] can be useful for generating fewer, faster proposals, which is the
key for achieving high frame rates on CPU-based systems. Methods based on fully
convolutional networks (FCNs) cannot achieve high frame rates on robotic platforms
with low processing capabilities (no GPU available) because they process images
with larger resolutions than normal CNNs. Thus, FCNs cannot be used trivially for
real-time robotics on these kinds of platforms.
First, the use of complementary information sources for segmenting the objects
will be analyzed. Robotic platforms usually have range sensors, such as Kinects/
Asus or LIDAR (Light Detection and Ranging) sensors, that are able to extract depth
information that can be used for boosting the object segmentation. Methods that use
RGBD data for detecting objects include those presented in [17–19, 24, 25], while
methods that use LIDAR data include those of [20, 26]. These methods are able
to generate proposals by using tabletop segmentation or edge/gradient information.

FIGURE 8.1 Faster-RCNN is an object detector that uses a region proposal network and a
classifcation network that share the frst convolutional layers [15].
Applications of Deep Learning in Robot Vision 215

Also, they generate colorized images from depth for use in a standard CNN pipeline
for object recognition. These studies use CNNs and custom depth-based proposals,
and then their speed is limited by the CNN model. For instance, in [27], a system is
described that runs at 405 fps on a Titan X GPU, 67 fps on a Tegra X1 GPU, and 62
fps on a Core i7 6700 K CPU [27]. It must be noted that while Titan X GPU and Core
i7 CPU processors are designed for desktop computers, Tegra X1 is a mobile GPU
processor for embedded systems aimed at low power consumption, and so it can be
used in robotic platforms.
Second, two applications that use specifc detectors for generating the object pro-
posals will be presented: pedestrian detection, and detection of objects in robotics
soccer. Detecting pedestrians in real time with high reliability is an important ability
in many robotics applications, especially for autonomous driving. Large differences
in illumination and variable, cluttered backgrounds are hard issues to be addressed.
Methods that use CNN-based methods for pedestrian detection include [23, 28–34].
Person detectors such as LDCF [23] or ACF [35] are used for generating pedestrian
proposals in some methods. As indicated in [23], the system based on AlexNet [36]
requires only 3 msec for processing a region, and proposal computation runs at 2
fps when using LDCF, and at 21.7 fps when using ACF in an NVIDIA GTX980
GPU, with images of size 640 × 480. The lightweight pedestrian detector is able
to run at 2.4 fps in an NVIDIA Jetson TK1, and beats classical methods by a large
margin [23]. A second approach is to use an FCN for detecting parts of persons [37].
This method is tested on an NVIDIA Titan X GPU, and it is able to run at 4 fps when
processing images of size 300 × 300. KITTY pedestrians is a benchmark used by
state-of-the-art methods. Current leading methods in the KITTY benchmark whose
algorithms are described (many of the best-performing methods do not describe their
algorithms), are shown in Table 8.1. Note that, as the objective of the benchmark is to
evaluate accuracy, most of the methods cannot run in real time (~30 fps). This is an
aspect that needs further research.

TABLE 8.1
Selected Methods for Pedestrian Detection from KITTY Benchmark*
Method Moderate Easy Hard Runtime Computing environment

RRC [30] (code 75.33% 84.14% 70.39% 3.6 s GPU @ 2.5 Ghz (Python +
available) C/C++)
MS-CNN [31] 73.62% 83.70% 68.28% 0.4 s GPU @ 2.5 Ghz (C/C++)
(code available)
SubCNN [32] 71.34% 83.17% 66.36% 2s GPU @ 3.5 Ghz (Python +
C/C++)
IVA [33] (code 70.63% 83.03% 64.68% 0.4 s GPU @ 2.5 Ghz (C/C++)
available)
SDP+RPN [34] 70.20% 79.98% 64.84% 0.4 s GPU @ 2.5 Ghz (Python +
C/C++)

* Note that methods not describing their algorithms are not included.
216 Deep Learning in Computer Vision

Robotics soccer is an interesting area because it requires real-time vision algo-

rithms that are able to work with very limited computational resources. Deep learning
techniques developed in robot soccer can therefore be transferred to other resource-
constrained platforms, such as smartphones and Unmanned Aerial Vehicles (UAVs).
Current applications for robotics soccer include soccer player detection [22, 38], ball
detection [39], and visual navigation [40]. State-of-the-art soccer player detection [38]
uses a model based on SqueezeNet using only two convolutional layers, an extended
fre layer and a fully connected layer, while processing regions of size 28 × 28. This
method requires only 1 msec on a NAO robot for processing a region, using only an
Intel Atom CPU. Another application is ball detection, which is addressed in [39].
This method uses an architecture based on three convolutional layers and two fully
connected layers. However, its runtime inside a Nao robot is not reported. Finally,
in [40] a system based on image processing and DNNs that is able to detect players,
balls, and the orientation of the robot inside the playing feld is proposed. The system
is able to work in real time inside Nao robots, without using any color information.
Currently, CNNs have been used for solving very specifc tasks in robotics soccer
successfully, and it is expected that new algorithms (and hardware) will enable the
use of deep learning in more global, high-level tasks.

8.3.2 OBJECT GRASPING AND MANIPULATION

The ability to grasp objects is of paramount importance for autonomous robots.
Classical algorithms for grasping and manipulating objects require detecting both
the gripper and the object in the image, and using graph-based or Jacobian-based
algorithms for computing the movements for the gripper. Deep learning has gener-
ated a paradigm change since tasks like grasping, manipulation, and visual servoing
can be solved directly, without the need of hand-crafting intermediate processes like
detecting features, detecting objects, detecting the gripper, or performing graph-
based optimizations. As problems are solved as a whole from raw data, the perfor-
mance limit imposed by inaccuracies of intermediate processes can be surpassed,
and, therefore, novel approaches for solving tasks related to grasping, manipulation,
and visual servoing are being developed. Also, the new algorithms are able to be
generalized to novel objects and situations. The performance of this new generation
of methods is currently limited by the size of the training datasets and the com-
puting capabilities, which becomes relevant when using hardware-limited robotic
platforms.
Grasping objects directly, without the need of detecting specifc objects, is a
diffcult problem because it must generalize over different kinds of objects, includ-
ing those not available when training the system. This problem has recently been
addressed successfully by using deep learning. Studies that are based on selecting
grasp locations include [41–43] and [44], which used self-supervision during 700
hours of trial and error in a Baxter robot. That work has been extended for learn-
ing eye-hand coordination in [45], which is the state of the art. The latter system
is trained without supervision by using over 800,000 grasp attempts collected over
the course of two months, using between 6 and 14 manipulators at any given time.
The proposed network processes two inputs: an image (size 512 × 512), and a motor
Applications of Deep Learning in Robot Vision 217

command. The system computes the probability that a given motor command will
produce a successful grasp, conditioned on the image, which enables the system to
control the robot. The network is composed by seven convolutional layers, whose
output is concatenated with the motor command. The concatenation layer is fur-
ther processed by an additional eight convolutional layers and two fully connected
layers. The system achieves effective real-time control; the robot can grasp novel
objects successfully and correct mistakes by continuous servoing. It also lowers
the failure rate from 35% (hand designed) to 20%. However, learned eye-hand
coordination depends on the particular robot used for training the system. The
computing hardware used in this work is not fully described. It must be noted that
these kinds of solutions are able to run successfully in real time using GPUs.
Systems able to infer manipulation trajectories for object-task pairs using DNNs
have been developed recently. This task is diffcult because generalization over dif-
ferent objects and manipulators is needed. Proposed solutions include [46, 47]. These
systems are able to learn from manipulation demonstrations and generalize over new
objects. In Robobarista [46], the input of the system is composed of a point cloud,
a trajectory, and a natural language space, which is mapped to a joint space. Once
trained, the trajectory generating the best similarity to a new point cloud and natural
language instruction is selected. Computational hardware (GPU) and runtime are not
reported in this work. In [47], the proposed system is able to learn new tasks by using
image-based reinforcement learning. That system is composed by an encoder and
a decoder. The encoder is formed by three convolutional layers and two fully con-
nected layers. Computational hardware and runtime are not reported in this work.
Visual servoing systems have also benefted from the use of deep learning tech-
niques. In Saxena et al. [48], a system that is able to perform image-based visual
servoing by feeding raw data into convolutional networks is presented. The system
does not require the 3D geometry of the scene or the intrinsic parameters from the
camera. The CNN takes a pair of monocular images as input, representing the cur-
rent and desired pose, and processes them by using four convolutional layers and a
fully connected layer. The system is trained for computing the transformation in 3D
space (position and orientation) needed for moving the camera from its current pose
to the desired one. The estimated relative motion is used for generating linear and
angular speed commands for moving the camera to the desired pose. The system is
trained on synthetic data and tested both in simulated environments and by using a
real quadrotor. The system takes 20 msec for processing each frame when using an
NVIDIA Titan X GPU, and 65 msec when using a Quadro M2000 GPU. A wif con-
nection was used for transferring data between the drone and a host PC.
Multimodal data delivers valuable information for object grasping. In [49], a hap-
tic object classifer that is fed by visual and haptic data is proposed. The outputs of a
visual CNN and a haptic CNN are fused by a fusion layer, followed by a loss function
for haptic adjectives. The visual CNN is based on the (5a) layer of GoogleNet, while
the haptic CNN is based on three one-dimensional convolutional layers followed by
a fully connected layer. Examples of haptic adjectives are “absorbent,” “bumpy,” and
“compressible.” The trained system is able to estimate haptic adjectives of objects
by using only visual data and it is shown to improve over systems based on classical
features. Computing hardware and runtime of the method are not reported.
218 Deep Learning in Computer Vision

8.3.3 SCENE REPRESENTATION AND CLASSIFICATION

Scene representation is an important topic since it is required by mobile robots for
performing a variety of tasks that depend on semantic and/or geometric information
from the current scene. In particular, visual navigation is an important ability that is
needed by autonomous robots for performing more general tasks that involve moving
autonomously. Classical approaches for representing scenes from monocular images
are based on extraction of hand-crafted features, while visual navigation requires the
robot to have a metric or topological map of its environment, as well as an estimation
of the depth of the obstacles/objects in the current frame. Deep learning has enabled
solving a variety of scene-related tasks directly, ranging from encoding scene infor-
mation to performing navigation from raw monocular data, without the need of
hand-crafting sub-processes like feature extraction, map estimation, or localization.
These novel task-oriented algorithms are able to solve diffcult problems success-
fully that were not directly affordable before. In the following paragraphs the use
of deep learning in applications such as place recognition, semantic segmentation,
visual navigation, and 3D geometry estimation is analyzed.
Place recognition is a task that has been solved successfully by using deep learn-
ing-based approaches. Studies related to this problem are [50, 55]. These methods are
able to recognize places under severe appearance changes caused by weather, illumi-
nation, and changes in viewpoint. Also, in [52], data from LIDAR is used alongside
a CNN for building a tridimensional semantic map. The CNN used is based on
AlexNet and requires 30 msec to run on an NVIDIA Quadro K4000 GPU. There are
several datasets for testing algorithms, such as Sun [56] and Places [57, 58], and also
several variants derived from them. As there is not a dominant benchmark for place
recognition, different studies use different datasets, sometimes customized, and thus
most methodologies cannot be compared directly. Also, the best-performing meth-
ods in ILSVRC, scene classifcation challenge [59] and Places challenges [60], do not
describe their algorithms in detail, but achieve an impressive 0.0901 top-5 classifca-
tion error. Computing hardware and runtime related to those challenges [59, 60] is
not reported. Nevertheless, the related datasets used in scene classifcation, and the
best-performing methods in each case, are shown in Table 8.2.
Semantic segmentation consists of classifying each pixel in the image, enabling
to solve tasks that require a pixel-wise level of precision. Studies related to semantic

TABLE 8.2
Datasets Used for Scene Classifcation
Dataset Results

Places [58] 50.0% [58] (CaffeNet), 14.56 msec @ Pascal Titan X

85.16% [54] (AlexNet for scene classifcation, 8 conv for semantic segmentation)
Sun [56, 61] 38% [56] (non deep)
66.2% [58] (AlexNet), 14.56 msec @ Pascal Titan X
Sun-RGBD [54] 41.3% [54] (input 81 × 81, 5 conv + 3 fc classifcation, 8 conv semantic
segmentation)
Applications of Deep Learning in Robot Vision 219

segmentation of scenes are [26, 54, 62–70]. Also, multimodal information is used in
[52, 71, 72]. These studies are useful for tasks that require pixel-wise precision like
road segmentation and other tasks in autonomous driving. However, runtimes of
methods based on semantic segmentation are usually not reported. PASCAL VOC
and Cityscapes are two benchmarks used for semantic segmentation. While images
from PASCAL VOC are general-purpose, Cityscapes is aimed at autonomous driv-
ing. Best-performing methods in both databases are reported in Tables 8.3 and 8.4.
It can be noted that DeepLabv3 [64] and PSPNet [65] (see Figure 8.2) achieve good
average precision at both benchmarks, and also code from [65] is available. Also,
the increasing performance in average precision has a deep impact on autonomous
driving, as this application requires confdent segmentation for working because of
the risks involved on it.
Visual navigation using deep learning techniques is an active research topic since
new methodologies are able to solve navigation directly, without the need for detect-
ing objects or roads. Examples of studies related to visual navigation are [74–78].
These methods are able to perform navigation directly over raw images captured by
RGB cameras. For instance, the forest trail follower in [74] (see Figure 8.3) has an
architecture composed of seven convolutional layers and one fully connected layer. It
is able to run at 15 fps on an Odroid-U3 platform, enabling it to be used in a real UAV.

TABLE 8.3
PASCAL VOC Segmentation Benchmark*
Name AP Runtime (msec/frame)

DeepLabv3-JFT [64] 86.9 n/a

DIS [69] 86.8 140 msec (GPU model not reported)
CASIA_IVA_SDN [70] 86.6 n/a
IDW-CNN [73] 86.3 n/a
PSPNet [65] (code available) 85.3 n/a

* Only entries with reported methods are considered. Repeated entries are not
considered.

TABLE 8.4
Cityscapes Benchmark*
Name IoU class Runtime (ms/frame)
Deeplabv3 [64] 81.3 n/a
PSPNet [65] 81.2 n/a
ResNet-38 [66] 80.6 6 msec (minibatch size 10) @ GTX 980 GPU
TuSimple_Coarse [67] (code available) 80.1 n/a
SAC-multiple [68] 78.2 n/a

* Only entries with reported methods are considered. Repeated entries are not considered.
220 Deep Learning in Computer Vision

FIGURE 8.2 Pyramid Scene Parsing Network (PSPNet) is a scene-parsing algorithm based
on the use of convolutional layers, followed by upsampling and concatenation layers [65].

FIGURE 8.3 A deep neural network is used to decide one of three possible actions (turn
left, go straight, turn right) from a single image. This system is used as guide in an autono-
mous quadrotor [74].

In [79], a visual navigation system for Nao robots based on deep reinforcement learn-
ing is proposed. The system consists of two neural networks—an actor and a critic—that
are trained by using Deep Deterministic Policy Gradients. Both networks are conformed
by convolutional, fully connected, and Long Short-Term Memory (LSTM) layers. Once
trained, the system is able to run in real time in the Nao internal CPU.
The use of CNN methodologies for estimating 3D geometry is another active
research topic in robotics. Studies related to 3D geometry estimation from RGB
images include [80–82]. Also, depth information is added in [71, 72]. These methods
provide a new approach to dealing with 3D geometry, and they are shown to over-
come the classical structure-from-motion approaches, as they are able to infer depth
even from only a single image.
A functional area is a zone in the real world that can be manipulated by a
robot. Functional areas are classifed by the kind of manipulation action the robot
can realize on them. In [83], a system for localizing and recognizing functional
areas by using deep learning is presented. The system enables an autonomous
robot to have a functional understanding of an arbitrary indoor scene. An ontol-
ogy considering 11 possible end categories is set. Functional areas are detected
by using selective-search region proposals, and then by applying a VGG-based
Applications of Deep Learning in Robot Vision 221

CNN trained on the functionality categories. The system is able to generalize

onto novel scenes. Neither hardware nor runtime are reported. However, its archi-
tecture is similar to RCNN [14], which requires 18 seconds for processing a frame
on a K20x GPU.

8.3.4 SPATIOTEMPORAL VISION

The processing of video and other spatiotemporal sources of information is of para-
mount importance in robotics. Classical methods related to spatiotemporal vision
require engineering subtasks such as motion detection, optical fow, local feature
extraction, or image segmentation. Deep learning has the potential to change the
current paradigm of spatiotemporal vision from hand-engineering complex subtasks
to solving tasks as a whole. The increment in both video datasets and computational
power will enable the analysis of spatiotemporal data at near real time, but only in
the future. Then, spatiotemporal vision will be a goal of future robot vision as well
as computer vision research.
The use of short video sequences for object recognition is an alternative for
improving current image-based techniques. In [84], a convolutional LSTM is pro-
posed for processing short video sequences containing about four frames, for robotic
perception. The system is shown to improve on the baseline (smoothing CNN results
from individual image frames), and it is able to run at 0.87 fps when using a GPU.
Human action recognition is an area that can beneft greatly from using video
sequences instead of isolated images. Reports of research dealing with recognition
of human actions include [85–93], in which driver activity anticipation is performed.
In various studies, information from multiple frames is integrated by using several
CNNs for each frame, by using LSTM or Dynamic Time Warping. The C3D method
[90] is able to run at 313 fps on a K40 GPU, processing videos with a resolution
of 320 × 240 pixels. Thus, those kinds of methods have the potential for running
in real time. However, current state-of-the-art methods are not real time, or their
runtimes are not described. The best-performing methods on the UCF-101 dataset
are reported in Table 8.5. Note that both the accuracy and runtime of C3D [90] are
lower than those reported by Wu et al. [92], and so a tradeoff between accuracy

TABLE 8.5
Current Results on the UCF-101 Video Dataset for Action
Recognition [91]
Name Accuracy Runtime (msec/frame)

C3D (1 net) + linear SVM [90] 82.3% 3.2 msec/frame @ Tesla K40 GPU
VGG-3D + C3D [90] 86.7% n/a
Ng et al. [92] 88.6% n/a
Wu et al. [93] 92.6% 1,730 msec/frame @ Tesla K80 GPU
Guo et al. [91] 93.3% n/a
Lev et al. [86] 94.0% n/a
222 Deep Learning in Computer Vision

FIGURE 8.4 TSC-DL is a system that is able to segment a suturing video sequence into
different stages (position needle, push needle, pull needle, and hand-off [94]).

and runtime is present. For selecting a method, it must be considered if real time is
needed in the specifc task to be solved.
Transfer of video-based computer vision techniques to medical robotics is an
active area of research, especially in surgery-related applications. The use of deep
learning in this application area has already begun. For instance, in [94], a system
that is able to segment video and kinematic data for performing unsupervised tra-
jectory segmentation of multimodal surgical demonstrations is implemented (see
Figure 8.4). The system is able to segment video-kinematic descriptions of surgi-
cal demonstrations successfully, corresponding to stages such as “position,” “push
needle,” “pull needle,” and “hand-off.” The system is based on a switching linear
dynamic system, which considers both continuous and discrete states composed of
kinematic and visual features. Visual features are extracted from the ffth convo-
lutional layer of a VGG CNN and then reduced to 100 dimensions by using PCA.
Clustering on the state space is applied by the system and the system is able to learn
transitions between clusters, enabling trajectory segmentation. The system uses a
video source with 10 frames per second; however, the frame rate of the full method
is not reported.

8.4 FUTURE CHALLENGES

As already mentioned, the increase in the use of deep learning in robotics will
depend on the possibility that DNN methods will be able to adapt to the require-
ments of robotics applications, such as real-time operation with limited on-board
computational resources, and the ability to deal with constrained observational
conditions derived from the robot geometry, limited camera resolution, and sensor/
object relative pose.
The current requirements of large computational power and memory needed for
most CNN models used in vision applications are relevant barriers to their adop-
tion in resource-constrained robotic applications such as UAVs, humanoid biped
robots, or autonomous vehicles. Efforts are already being made to develop CNN
architectures that can be compressed and quantized for use in resource-constrained
Applications of Deep Learning in Robot Vision 223

platforms [95, 96]. For instance, in [38] a CNN-based robot detector is presented that
is able to run in real time, in Nao robots while playing soccer. In addition, compa-
nies such as Intel, NVIDIA, and Samsung, just to name a few, are developing CNN
chips that will enable real-time vision applications [4]. For instance, mobile GPU
processors like NVIDIA Tegra K1 enable effcient implementation of deep learning
algorithms with low power consumption, which is relevant for mobile robotics [27].
It is expected that these methodologies will consolidate in the next few years and
will then be available to the developers of robot vision applications.
The ability of deep learning models to achieve spatial invariance respect to the
input data in a computationally effcient manner is another important issue. In order
to address this, the spatial transformer module (STM) was recently introduced [97].
It is a self-contained module that can be incorporated into DNNs, and that is able to
perform explicit spatial transformation of the features. The most important charac-
teristic of STMs is that its parameters (i.e., the transformations of the spatial trans-
formation) can be learned together with the other parameters of the network using
the backpropagation of the loss. The further development of the STM concept will
allow addressing the required invariance to the observational conditions derived
from the robot geometry, limited camera resolution, and sensor/object relative pose.
STMs are able to deal with rotations, translations, and scaling [97], and STMs are
already being extended to deal with 3D transformations [98]. Further development
of STM-inspired techniques is expected in the next few years.
Unsupervised learning is another relevant area of future development for deep
learning-based vision systems. Biological learning is largely unsupervised; ani-
mals and humans discover the structure of the world by exploring and observing it.
Therefore, one would expect that similar mechanisms could be used in robotics and
other computer-based systems. Until now, the success of purely supervised learning,
which is based largely on the availability of massive labeled data, has overshad-
owed the application of unsupervised learning in the development of vision sys-
tems. However, when considering the increasing availability of non-labeled digital
data, and the increasing number of vision applications that need to be addressed,
one would expect further development of unsupervised learning strategies for DNN
models [4, 7]. In the case of robotics applications, a natural way of addressing this
issue is by using deep learning and reinforcement learning together. This strategy
has already been applied for the learning of games (Atari, Go, etc.) [99, 100], for
solving simple tasks like pushing a box [47], and for solving more complex tasks
such as navigating in simulated environments [77, 78], but it has not been used for
robots learning in the wild. The visual system of most animals is active, i.e., animals
decide where to look, based on a combination of internal and external stimuli. In a
robotics system having a similar active vision approach, reinforcement learning can
be used for deciding where to look according to the results of the interaction of the
robot with the environment.
A related relevant issue is that of open-world learning [101], i.e., to learn to detect
new classes incrementally, or to learn to distinguish among subclasses incrementally
after the main one has been learned. If this can be done without supervision, new
classifers can be built based on those that already exist, greatly reducing the effort
required to learn new object classes. Note that humans are continuously inventing
224 Deep Learning in Computer Vision

new objects, fashion changes, etc., and therefore robot vision systems will need to be
updated continuously, adding new classes and/or updating existing ones [101]. Some
recent work has addressed these issues, based mainly on the joint use of deep learn-
ing and transfer learning methods [102, 103].
Classifcation using a very large number of classes is another important challenge
to address. AlexNet, like most CNN networks, was developed for solving problems
that contain ~1,000 classes (e.g., ILSVRC 2012 challenge). It uses fully connected
units and a fnal softmax unit for classifcation. However, when working on problems
with a very large number of classes (10,000+) in which dense data sampling is avail-
able, nearest neighbors could work better than other classifers [104]. Also, the use
of a hierarchy-aware cost function (based on WordNet) could enable providing more
detailed information when the classes have redundancy [104]. However, in current
object recognition benchmarks (like ILSVRC 2016), the number of classes has not
increased, and no hierarchical information has been used. The ability of learning
systems to deal with a high number of classes is an important challenge that needs to
be addressed for performing open-world learning.
Another important area of research is the combination of new methods based
on deep learning with classical vision methods based on geometry, which has been
very successful in robotics (e.g., visual SLAM). This topic has been addressed at
recent workshops [105] and conferences [106]. There are several ways in which these
important and complementary paradigms can be combined. On the one hand, geom-
etry-based methods such as Structure from Motion and Visual SLAM can be used
for the training of deep learning-based vision systems; geometry-based methods can
extract and model the structure of the environment, and this structure can be used
for assisting the learning process of a DNN. On the other hand, DNNs can also be
used for learning to compute the visual odometry automatically, and eventually to
learn the parameters of a visual SLAM system. In a very recent study, STMs were
extended to 3D spatial transformations [98], allowing end-to-end learning of DNNs
for computing visual odometry. It is expected that STM-based techniques will allow
end-to-end training for optical fow, depth estimation, place recognition with geo-
metric invariance, small-scale bundle adjustment, etc. [98].
A fnal relevant research topic is the direct analysis of video sequences in robot-
ics applications. The analysis of video sequences by using current techniques based
on deep learning is very demanding of computational resources. It is expected that
the increase of computing capabilities will enable the analysis of video sequences in
real time in the near future. Also, the availability of labeled video datasets for learn-
ing tasks such as action recognition will enable the improvement of robot percep-
tion tasks, which are currently addressed by using independent CNNs for individual
video frames.

8.5 CONCLUSIONS
In this chapter, several approaches based on deep learning that are related to robot
vision applications are described and analyzed. They are classifed into four cat-
egories of methods: (i) object detection and categorization, (ii) object grasping and
manipulation, (iii) scene representation and classifcation, and (iv) spatiotemporal
Applications of Deep Learning in Robot Vision 225

vision. (i) Object detection and categorization has been transferred successfully
into robot vision. While the best-performing object detectors are mainly variants
of Faster-RCNN, methods derived from YOLO offer a better runtime. (ii) Object
grasping, manipulation, and visual servoing are currently being improved by the
adoption of deep learning. While real-time processing is needed for hand-eye coor-
dination and visual servoing to work, it is not needed for open-loop object grasping
detection. Also, these methods are highly dependent on the robot used for collecting
data. (iii) Methods for representing, recognizing, or classifying scenes as a whole
(not pixel-wise) feed images directly into a single CNN, and then near-real-time
processing can be achieved. However, pixel-wise methods based on FCNs require a
GPU for running in near real time. (iv) Spatiotemporal vision applications need real-
time processing capability to be useful in robotics. However, current research using
deep learning is experimental, and the frame rates for most of the methods are not
reported. Platforms with low computing capability (only CPU) are not able to run
most of the methods at a useful frame rate. The adoption of general deep learning
methods in robot vision applications requires addressing several problems: (i) Deep
neural networks require large amounts of data for being trained. Obtaining large
amounts of data for specifc robotic applications is a challenge to address each time
a specifc task needs to be solved. (ii) Deep learning is able to deal with diffcult
problems. However, it cannot generalize to conditions that are not represented by
the available data. (iii) The need of low power consumption is a critical condition
to be considered. However, the availability of specialized hardware for computing
algorithms based on CNNs with low power consumption would make it possible to
broaden the use of deep neural networks in robotics.

ACKNOWLEDGMENTS
This work was partially funded by FONDECYT grant 1161500 and CONICYT PIA
grant AFB18004.

BIBLIOGRAPHY
1. Yanming Guo, Yu Liu, Ard Oerlemans, Songyang Lao, Song Wu, and Michael S. Lew.
Deep learning for visual understanding: A review. Neurocomputing, 187, 27–48, 2016.
2. Soren Goyal and PaulBenjamin. Object recognition using deep neural networks: A sur-
vey. http://arxiv.org/abs/1412.3684, 2014.
3. Jiuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai,
Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, and Tsuhan Chen. Recent advances
in convolutional neural networks. Pattern Recognition, 77, 354–377, 2018.
4. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521, 436–
444, 28 May 2015, doi:10.1038/nature14539,
5. Li Deng. A tutorial survey of architectures, algorithms, and applications for deep learn-
ing. APSIPA Transactions on Signal and Information Processing, 3, 1–29, 2014.
6. Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural Networks,
61, 85–117, 2015.
7. Suraj Srinivas, Ravi Kiran Sarvadevabhatla, Konda Reddy Mopuri, Nikita Prabhu,
Srinivas S. S. Kruthiventi, and R. Venkatesh Babu. A taxonomy of deep convolutional
neural nets for computer vision. Frontiers in Robotics and AI, 2, 2016.
226 Deep Learning in Computer Vision

8. Xiaowei Zhou, Emanuele Rodola, Jonathan Masci, Pierre Vandergheynst, Sanja Fidler,
and Kostas Daniilidis. Workshop geometry meets deep learning, ECCV 2016, 2016.
9. Yi Li, Yezhou Yang, Michael James, Danil Prokhorov. Deep Learning for Autonomous
Robots, Workshop at RSS 2016, 2016.
10. Awad, Ali Ismail, and Mahmoud Hassaballah. Image feature detectors and descriptors.
Studies in Computational Intelligence. Springer International Publishing, Cham, 2016.
11. Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey
Dean, et al. Tensorfow: A system for large-scale machine learning. 12th Symposium
on Operating Systems Design and Implementation (OSDI 16), pp. 265–283, 2016.
12. Patricio Loncomilla, Javier Ruiz-del-Solar, and Luz Martínez, Object recognition
using local invariant features for robotic applications: A survey, Pattern Recognition,
60, 499–514, 2016.
13. Y. LeCun, B. Boser, J. Denker, D. Henderson, R. Howard, W. Hubbard, and L. Jackel.
Backpropagation applied to handwritten zip code recognition. Neural Computation,
1(4), 541–551, 1989.
14. R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for
accurate object detection and semantic segmentation. IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 580–587, 2014.
15. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards
real-time object detection with region proposal networks. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 39, 6, 1137–1149, 2017.
16. Joseph Redmon and Ali Farhadi. YOLO9000: Better, faster, stronger. IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 6517–6525, 2017.
17. Max Schwarz, Hannes Schulz, and Sven Behnke. RGB-D object recognition and
pose estimation based on Pre-trained convolutional neural network features. IEEE
International Conference on Robotics and Automation (ICRA), 1329–1335, 2005.
18. Saurabh Gupta, Ross Girshick, Pablo Arbeláez, and Jitendra Malik. Learning
rich features from RGB-D images for object detection and segmentation. European
Conference on Computer Vision (ECCV), 345–360, 2014.
19. Andreas Eitel, Jost Tobias Springenberg, Luciano Spinello, Martin Riedmiller,
and Wolfram Burgard. Multimodal deep learning for robust RGB-D object recogni-
tion. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS),
681–687, 2015.
20. Joel Schlosser, Christopher K. Chow, and Zsolt Kira. Fusing LIDAR and images
for pedestrian detection using convolutional neural networks. IEEE International
Conference on Robotics and Automation (ICRA), 2198–2205, 2016.
21. G. Pasquale, C. Ciliberto, F. Odone, L. Rosasco, and L. Natale. Teaching iCub
to recognize objects using deep convolutional neural networks. 4th International
Conference on Machine Learning for Interactive Systems (MLIS’15), 43, 21–25, 2015.
22. Dario Albani, Ali Youssef, Vincenzo Suriani, Daniele Nardi, and Domenico
Daniele Bloisi. A deep learning approach for object recognition with NAO soccer
robots. RoboCup International Symposium, 392–403, 2016.
23. Denis Tomè, Federico Monti, Luca Baroffo, Luca Bondi, Marco Tagliasacchi, and
Stefano Tubaro. Deep convolutional neural networks for pedestrian detection. Signal
Processing: Image Communication, 47, C, 482–489, 2016.
24. Hasan F. M. Zaki, Faisal Shafait, and Ajmal Mian. Convolutional hypercube pyra-
mid for accurate RGB-D object category and instance recognition. IEEE International
Conference on Robotics and Automation (ICRA), 2016, 1685–1692, 2016.
25. Judy Hoffman, Saurabh Gupta, Jian Leong, Sergio Guadarrama, and Trevor
Darrell. Cross-modal adaptation for RGB-D detection. IEEE International Conference
on Robotics and Automation (ICRA), 5032–5039, 2016.
Applications of Deep Learning in Robot Vision 227

26. Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3D lidar using fully
convolutional network. Proceedings of Robotics: Science and System Proceedings,
2016.
27. Nvidia: GPU-Based deep learning inference: A performance and power analysis.
Whitepaper, 2015.
28. Jan Hosang, Mohamed Omran, Rodrigo Benenson, and Bernt Schiele. Taking
a deeper look at pedestrians. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 4073–4082, 2015.
29. Rudy Bunel, Franck Davoine, and Philippe Xu. Detection of pedestrians at far
distance. IEEE International Conference on Robotics and Automation (ICRA), 2326–
2331, 2016.
30. Jimmy Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Jiahao Pang, Qiong Yan,
Yu-Wing Tai, and Li Xu. Accurate single stage detector using recurrent rolling con-
volution. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
752–760, 2017.
31. Zhaowei Cai, Quanfu Fan, Rogerio S. Feris, and Nuno Vasconcelos. A unifed
multi-scale deep convolutional neural network for fast object detection. European
Conference on Computer Vision (ECCV), 354–370, 2016.
32. Yu Xiang, Wongun Choi, Yuanqing Lin, and Silvio Savarese. Subcategory-aware
convolutional neural networks for object proposals and detection. IEEE Winter
Conference on Applications of Computer Vision (WACV), 924–933, 2017.
33. Yousong Zhu, Jinqiao Wang, Chaoyang Zhao, Haiyun Guo, and Hanqing Lu.
Scale-adaptive deconvolutional regression network for pedestrian detection. Asian
Conference on Computer Vision, 416–430, 2016.
34. Fan Yang, Wongun Choi, and Yuanqing Lin. Exploit all the layers: Fast and accu-
rate CNN object detector with scale dependent pooling and cascaded rejection classi-
fers. IEEE International Conference on Computer Vision and Pattern Recognition
(CVPR), 2129–2137, 2016.
35. Piotr Dollar, Ron Appel, Serge Belongie, and Pietro Perona. Fast feature pyramids
for object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence,
36, 8, 1532–1545, 2014.
36. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classifcation
with deep convolutional neural networks. Advances in Neural Information Processing
Systems, 25, 1097–1105, 2012.
37. Gabriel L. Oliveira, Abhinav Valada, Claas Bollen, Wolfram Burgard, and
Thomas Brox. Deep learning for human part discovery in images. IEEE International
Conference on Robotics and Automation (ICRA), 1634–1641, 2016.
38. Nicolás Cruz, Kenzo Lobos-Tsunekawa, and Javier Ruiz-del-Solar. Using convolu-
tional neural networks in robots with limited computational resources: Detecting NAO
robots while playing soccer. RoboCup 2017: Robot World Cup XXI, 19–30, 2017.
39. Daniel Speck, Pablo Barros, Cornelius Weber, and Stefan Wermter. Ball localiza-
tion for robocup soccer using convolutional neural networks. RoboCup International
Symposium, 19–30, 2016.
40. Francisco Leiva, Nicolas Cruz, Ignacio Bugueño, and Javier Ruiz-del-Solar.
Playing soccer without colors in the SPL: A convolutional neural network approach.
RoboCup Symposium, 2018 (in press).
41. Ian Lenz, Honglak Lee, and Ashutosh Saxena. Deep learning for detecting robotic
grasps. The International Journal of Robotics Research, 34, 4–5, 705–724, 2015.
42. Joseph Redmon and Anelia Angelova. Real-time grasp detection using convolu-
tional neural networks. IEEE International Conference on Robotics and Automation
(ICRA), 1316–1322, 2015.
228 Deep Learning in Computer Vision

43. Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp
from 50K tries and 700 robot hours. IEEE International Conference on Robotics and
Automation (ICRA), 3406–3413, 2016.
44. Di Guo, Tao Kong, Fuchun Sun, and Huaping Liu. Object discovery and grasp
detection with a shared convolutional neural network. IEEE International Conference
on Robotics and Automation (ICRA), 2038–2043, 2016.
45. Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen. Learning hand-
eye coordination for robotic grasping with deep learning and large-scale data collection.
The International Journal of Robotics Research, 37, 4–5, 421–436, 2018.
46. Jaeyong Sung, Seok Hyun Jin, and Ashutosh Saxena. Robobarista: Object part
based transfer of manipulation trajectories from crowd-sourcing in 3D pointclouds.
Robotics Research, 3, 701–720, 2018.
47. Chelsea Finn, Xin Yu Tan, Yan Duan, Trevor Darrell, Sergey Levine, and Pieter
Abbeel. Deep spatial autoencoders for visuomotor learning. IEEE International
Conference on Robotics and Automation (ICRA), 512–519, 2016.
48. Aseem Saxena, Harit Pandya, Gourav Kumar, Ayush Gaud, and K. Madhava
Krishna. Exploring convolutional networks for end-to-end visual servoing. IEEE
International Conference on Robotics and Automation (ICRA), 3817–3823, 2017.
49. Yang Gao, Lisa Anne Hendricks, Katherine J. Kuchenbecker, and Trevor Darrell.
Deep learning for tactile understanding from visual and haptic data. IEEE International
Conference on Robotics and Automation (ICRA), 536–543, 2016.
50. Manuel Lopez-Antequera, Ruben Gomez-Ojeda, Nicolai Petkov, and Javier
Gonzalez-Jimenez. Appearance-invariant place recognition by discriminatively
training a convolutional neural network. Pattern Recognition Letters, 92, 89–95,
2017.
51. Yi Hou, Hong Zhang, and Shilin Zhou. Convolutional neural network-based image
representation for visual loop closure detection. IEEE International Conference on
Information and Automation (ICRA), 2238–2245, 2015.
52. Niko Sunderhauf, Feras Dayoub, Sean McMahon, Ben Talbot, Ruth Schulz, Peter
Corke, Gordon Wyeth, Ben Upcroft, and Michael Milford. Place categorization and
semantic mapping on a mobile robot. IEEE International Conference on Robotics and
Automation (ICRA), 5729–5736, 2006.
53. Niko Sunderhauf, Sareh Shirazi, Adam Jacobson, Feras Dayoub, Edward Pepperell,
Ben Upcroft, and Michael Milford. Place recognition with convnet landmarks:
Viewpoint-robust, condition-robust, training-free. Robotics: Science and System
Proceedings, 2015.
54. Yiyi Liao, Sarath Kodagoda, Yue Wang, Lei Shi, and Yong Liu. Understand Scene
Categories by objects: A semantic regularized scene classifer using convolutional neu-
ral networks. IEEE International Conference on Robotics and Automation (ICRA),
2318–2325, 2016.
55. Peter Uršic, Rok Mandeljc, Aleš Leonardis, and Matej Kristan. Part-based room
categorization for household service robots. IEEE International Conference on
Robotics and Automation (ICRA), 2287–2294, 2016.
56. Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba.
Sun database: Large-scale scene recognition from abbey to zoo. IEEE Conference on
Computer Vision and Pattern recognition (CVPR), 3485–3492, 2010.
57. Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba.
Places: A 10 million image database for scene recognition. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 40, 6, 1452–1464, 2018.
58. Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva.
Learning deep features for scene recognition using places database. Advances in Neural
Information Processing Systems 27 (NIPS), 487–495, 2014.
Applications of Deep Learning in Robot Vision 229

59. ILSVRC 2016 scene classifcation, 2016. http://image-net.org/challenges/LSVRC/2016/

results.
60. Places Challenge 2016. http://places2.csail.mit.edu/results2016.html.
61. Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba.
SUN database: Large-scale scene recognition from abbey to zoo. IEEE Conference on
Computer Vision and Pattern recognition (CVPR), 3485–3492, 2010.
62. Caio Cesar Teodoro Mendes, Vincent Fremont, and Denis Fernando Wolf.
Exploiting fully convolutional neural networks for fast road detection. IEEE
International Conference on Robotics and Automation (ICRA), 3174–3179, 2016.
63. Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene
understanding benchmark suite. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 567–576, 2015.
64. Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam.
Rethinking atrous convolution for semantic image segmentation. Technical Report.
arXiv preprint arXiv:1706.05587, 2017.
65. Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia.
Pyramid scene parsing network. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 6230–6239, 2017.
66. Zifeng Wu, Chunhua Shen, and Anton van den Hengel. Wider or deeper: Revisiting
the resnet model for visual recognition. Pattern Recognition, 90, 119–133, 2019.
67. Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and
Garrison Cottrell. Understanding convolution for semantic segmentation. IEEE Winter
Conference on Applications of Computer Vision (WACV), 1451–1460, 2018.
68. Rui Zhang, Sheng Tang, Yongdong Zhang, Jintao Li, and Shuicheng Yan. Scale-
adaptive convolutions for scene parsing. IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2031–2039, 2017.
69. Ping Luo, Guangrun Wang, Liang Lin, and Xiaogang Wang. Deep dual learning
for semantic image segmentation. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2718–2726, 2017.
70. Jun Fu, Jing Liu, Yuhang Wang, and Hanqing Lu. Stacked deconvolutional network
for semantic segmentation. IEEE Transactions on Image Processing, 99, 2017.
71. Cesar Cadena, Anthony Dick, and Ian D. Reid. Multi-modal auto-encoders as joint
estimators for robotics scene understanding. Proceedings of Robotics: Science and
System Proceedings, 2016.
72. Farzad Husain, Hannes Schulz, Babette Dellen, Carme Torras, and Sven Behnke.
Combining semantic and geometric features for object class segmentation of indoor
scenes. IEEE Robotics and Automation Letters, 2 (1), 49–55, 2016.
73. Guangrun Wang, Ping Luo, Liang Lin, and Xiaogang Wang. Learning object
interactions and descriptions for semantic image segmentation. IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 5859–5867, 2017.
74. Alessandro Giusti, Jérôme Guzzi, Dan C. Ciresan, Fang-Lin He, Juan P. Rodríguez,
Flavio Fontana, Matthias Faessler, Christian Forster, Jürgen Schmidhuber, Gianni Di
Caro, Davide Scaramuzza, and Luca M. Gambardella. A machine learning approach
to visual perception of forest trails for mobile robots. IEEE Robotics and Automation
Letters, 1, 2, 661–667, 2016.
75. Dan C. Ciresan, Ueli Meier, and Jürgen Schmidhuber. Multi-column deep neural
networks for image classifcation. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 3642–3649, 2012.
76. Lei Tai and Ming Liu. A robot exploration strategy based on q-learning network. IEEE
International Conference on Real-time Computing and Robotics (RCAR), 57–62, 2016.
77. Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li
Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep
230 Deep Learning in Computer Vision

reinforcement learning. IEEE International Conference on Robotics and Automation

(ICRA), 3357–3364, 2017.
78. Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J. Ballard,
Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu,
Dharshan Kumaran, and Raia Hadsell. Learning to navigate in complex environments.
International Conference on Learning Representations (ICLR), 2017.
79. Kenzo Lobos-Tsunekawa, Francisco Leiva, and Javier Ruiz-del-Solar. Visual navi-
gation for biped humanoid robots using deep reinforcement learning. IEEE Robotics
and Automation Letters, 3, 4, 3247–3254, 2018.
80. Gabriele Costante, Michele Mancini, Paolo Valigi, and Thomas A. Ciarfuglia.
Exploring representation learning with CNNs for frame to frame ego-motion estima-
tion. IEEE Robotics and Automation Letters, 1, 1, 18–25, 2016.
81. Shichao Yang, Daniel Maturana, and Sebastian Scherer. Real-time 3D scene lay-
out from a single image using convolutional neural networks. IEEE International
Conference on Robotics and Automation (ICRA), 2183–2189, 2016.
82. Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for
camera relocalization. IEEE International Conference on Robotics and Automation
(ICRA), 4762–4769, 2016.
83. Chengxi Ye, Yezhou Yang, Cornelia Fermuller, and Yiannis Aloimonos. What
can i do around here? Deep functional scene understanding for cognitive robots. IEEE
International Conference on Robotics and Automation (ICRA), 4604–4611, 2017.
84. Ivan Bogun, Anelia Angelova, and Navdeep Jaitly. Object recognition from short
videos for robotic perception. arXiv:1509.01602v1 [cs.CV], 4 September 2015.
85. Jimmy Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Jiahao Pang, Qiong Yan,
Yu-Wing Tai, and Li Xu. Accurate single stage detector using recurrent rolling con-
volution. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
752–760, 2017.
86. Guy Lev, Gil Sadeh, Benjamin Klein, and Lior Wolf. RNN fsher vectors for action
recognition and image annotation. European Conference on Computer Vision (ECCV),
833–850, 2016.
87. Farzad Husain, Babette Dellen, and Carme Torra. Action recognition based on
effcient deep feature learning in the spatio-temporal domain. IEEE Robotics and
Automation Letters, 1, 2, 984–991, 2016.
88. Xiaochuan Yin and Qijun Chen. Deep metric learning autoencoder for nonlinear
temporal alignment of human motion. IEEE International Conference on Robotics and
Automation (ICRA), 2160–2166, 2016.
89. Ashesh Jain, Avi Singh, Hema S Koppula, Shane Soh, and Ashutosh Saxena.
Recurrent neural networks for driver activity anticipation via sensory-fusion architec-
ture. IEEE International Conference on Robotics and Automation (ICRA), 3118–3125,
2016.
90. Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri.
Learning spatiotemporal features with 3D convolutional networks. IEEE International
Conference on Computer Vision (ICCV), 4489–4497, 2015.
91. Huiwen Guo, Xinyu Wu, and Wei Fengab. Multi-stream deep networks for human
action classifcation with sequential tensor decomposition. Signal Processing, 140,
198–206, 2017.
92. Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol
Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for
video classifcation. IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 4694–4702, 2015.
Applications of Deep Learning in Robot Vision 231

93. Zuxuan Wu, Yu-Gang Jiang, Xi Wang, Hao Ye, and Xiangyang Xue. Multi-stream
multi-class fusion of deep networks for video classifcation. ACM Multimedia, 791–
800, 2016.
94. Adithyavairavan Murali, Animesh Garg, Sanjay Krishnan, Florian T. Pokorny,
Pieter Abbeel, Trevor Darrell, and Ken Goldberg. TSC-DL: Unsupervised trajec-
tory segmentation of multi-modal surgical demonstrations with deep learning. IEEE
International Conference on Robotics and Automation (ICRA), 4150–4157, 2016.
95. Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep
neural networks with pruning, trained quantization and huffman coding. International
Conference on Learning Representations (ICLR’16), 2016.
96. Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A. Horowitz,
and William J. Dally. EIE: Effcient inference engine on compressed deep neural net-
work. International Conference on Computer Architecture (ISCA), 243–254, 2016.
97. Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu.
Spatial transformer networks. Advances in Neural Information Processing Systems, 28,
2017–2025, 2015.
98. Ankur Handa, Michael Bloesch, Viorica Patraucean, Simon Stent, John McCormac,
and Andrew Davison. gvnn: Neural network library for geometric computer vision.
Computer Vision – ECCV 2016 Workshops, 67–82, 2016.
99. Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness,
Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg
Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen
King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-
level control through deep reinforcement learning. Nature, 518, 529–533, 2015.
100. David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George
van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam,
Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya
Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel,
and Demis Hassabis. Mastering the game of go with deep neural networks and tree
search. Nature, 529, 7587, 484–489, 2016.
101. Rodrigo Verschae and Javier Ruiz-del-Solar. Object detection: Current and future
directions. Frontiers in Robotics and AI, 29, 2, 2015.
102. Yoshua Bengio. Deep learning of representations for unsupervised and transfer
learning. International Conference on Unsupervised and Transfer Learning Workshop,
27, 17–37, 2012.
103. Marc-Andr Carbonneau, Veronika Cheplygina, Eric Granger, and Ghyslain
Gagnon. Multiple instance learning. Pattern Recognition, 77, 329–353, 2018.
104. Jia Deng, Alexander C. Berg, Kai Li, and Li Fei-Fei. What does classifying more
than 10,000 image categories tell us? European Conference on Computer Vision: Part
V (ECCV’10), 71–84, 2010.
105. Stefan Leutenegger, Thomas Whelan, Richard A. Newcombe, and Andrew J.
Davison. Workshop the future of real-time SLAM: Sensors, processors, representa-
tions, and algorithms, ICCV, 2015.
106. Ken Goldberg. Deep grasping: Can large datasets and reinforcement learning
bridge the dexterity gap? Keynote Talk at ICRA, 2016.
 9 Deep Convolutional
Neural Networks
Foundations and
Applications in Medical
Imaging
Mahmoud Khaled Abd-Ellah, Ali Ismail Awad,
Ashraf A. M. Khalaf, and Hesham F. A. Hamed

CONTENTS
9.1 Introduction .................................................................................................. 234
9.2 Convolutional Neural Networks (CNNs) ...................................................... 236
9.2.1 CNN Layers ...................................................................................... 236
9.2.1.1 Convolutional Layers ......................................................... 237
9.2.1.2 Max-Pooling Layer ............................................................ 238
9.2.1.3 Nonlinearity and ReLU ..................................................... 238
9.2.1.4 Fully Connected Layers (FCLs)......................................... 239
9.3 Optimization Approaches............................................................................. 239
9.3.1 Gradient-Based Optimization...........................................................240
9.3.2 Dropout Property..............................................................................240
9.3.3 Batch Normalization.........................................................................240
9.3.4 Network Depth.................................................................................. 241
9.4 Operational Properties.................................................................................. 241
9.4.1 Using Pretrained CNNs .................................................................... 241
9.4.1.1 Fine-Tuning ........................................................................ 241
9.4.1.2 CNN Activations as Features............................................. 242
9.4.2 Deep Learning Diffculties............................................................... 242
9.4.2.1 Overftting Reduction......................................................... 242
9.5 Commonly Used Deep Neural Networks ..................................................... 242
9.5.1 AlexNet............................................................................................. 243
9.5.2 GoogleNet......................................................................................... 243
9.5.3 VGG-19............................................................................................. 243
9.5.4 Deep Residual Networks (ResNet) ................................................... 243
9.5.5 Tools and Software ...........................................................................244

233
234 Deep Learning in Computer Vision

9.6 Convolutional Neural Network Types ..........................................................244

9.6.1 Region-Based Convolutional Neural Networks (R-CNN) ...............244
9.6.2 Fully Convolutional Networks (FCNs) ............................................. 245
9.6.3 Hybrid Learning Networks............................................................... 245
9.7 Deep Learning Applications in Medical Imaging........................................ 245
9.7.1 Deep Learning for Feature Extraction ............................................. 247
9.7.2 Deep Learning for Tumor Detection ................................................ 247
9.7.3 Deep Learning for Tumor Segmentation..........................................248
9.8 Brain Tumor Detection Using Deep Convolutional Neural Network...........248
9.8.1 DCNN Design................................................................................... 249
9.8.2 Simulation Experiments and Evaluation Criteria ............................. 250
9.8.2.1 Environment Setup ............................................................ 250
9.8.2.2 Evaluation Criteria ............................................................. 250
9.8.2.3 Experimental Results ......................................................... 252
9.9 Conclusion .................................................................................................... 253
Bibliography .......................................................................................................... 253

9.1 INTRODUCTION
Machine learning has become a primary focus and one of the most mainstream
subjects among research groups. The main uses of machine learning include feature
extraction, feature reduction, feature classifcation, forecasting, clustering, regres-
sion, and ensembles with different algorithms. Classic algorithms and techniques are
optimized to provide effective self-learning [1]. Since machine learning is applied
in a wide range of studies, many techniques have been provided, such as Bayesian
network, clustering, decision tree learning, and deep learning as shown in Figure 9.1.
Meanwhile, a few analysts in the machine learning feld have attempted to learn
models that combine the learning of features. These models normally consist of
different layers of nonlinearity. This led to the generation of the frst deep learning
models. Early models such as deep belief networks [2], stacked autoencoders (SAEs)
[3], and restricted Boltzmann machines [4] indicated guarantees on small datasets.
This was known as the “unsupervised pretraining” phase. It was believed that these
“pretrained” models would be a good initialization for supervised tasks, such as
classifcation.
Deep learning frst appeared in 2006 as a new subfeld of machine learn-
ing research. In Mosavi and Varkonyi-Koczy [5], it was named the hierarchical
learning algorithm, and it provided research felds related to pattern recogni-
tion. Deep learning of key factors mainly depends on supervised or unsuper-
vised learning and nonlinear processing in multiple layers [6]. With supervised
learning, the class target label is available; however, the absence of the class
target label means an unsupervised system. During the ImageNet large-scale
visual recognition challenge (ILSVRC) competition in 2012 [7], different scaled
algorithms were applied to a large dataset, which required, in addition to other
tasks, the grouping of an image into one of a thousand classes. Surprisingly,
a convolutional neural network (CNN) signifcantly reduced the error rate on
Deep Convolutional Neural Networks 235

FIGURE 9.1 Machine learning is the overall category that includes supervised learning,
unsupervised learning, reinforcement learning, neural networks, deep learning, and other
approaches.

that task, outperforming the traditional methodologies. Moreover, this could be

accomplished by applying end-to-end supervised training without unsupervised
pretraining. Throughout the following few years, “ImageNet uses deep neural
networks in classifcation” [8] has been the most powerful and poignant paper in
the computer vision feld.
Convolutional neural networks, a specifc type of deep learning model, have since
been broadly embraced by the computer vision community. Specifcally, the model
prepared by Alex Krizhevsky, famously called “AlexNet”, has been utilized and
adapted for different vision tasks. The output of a layer is provided as input to the
next layer in the algorithm and is called nonlinear processing in multiple layers. A
hierarchy is used between layers to consider and organize the importance of data. As
an example, applying a deep learning algorithm can perform grayscale image color-
ing automatically. Recurrent neural networks (RNNs) can add sound to a silent video
[1]. Consequently, this chapter principally discusses CNNs because these models are
more applicable to the computer vision feld. This chapter can also serve as a guide
for specialists starting in deep learning/computer vision.
Computer vision problems such as image characterization and object recogni-
tion have customarily employed handcrafted features [ 9] such as the scale-invariant
feature transform (SIFT) [10] and the histogram of oriented gradients (HOG) [11].
Specifcally, representations dependent on the bag-of-visual-words descriptor [12]
have achieved success in image classifcation. These works were normally followed
by learning algorithms, such as support vector machines (SVMs). Accordingly, the
implementation of these algorithms essentially depends on the features utilized,
implying that advancements in computer vision depend on building better arrange-
ments of features [13].
236 Deep Learning in Computer Vision

This chapter provides a comprehensive foundation for deep convolutional neu-

ral networks. The basic components of CNNs are summarized. Then, we present
some later developments in various aspects of CNNs, including the convolutional
layer, pooling layer, activation function, fully connected layers (FCLs), regulariza-
tion, optimization, normalization, and network depth, and we present the advances
in each phase. Next, some network operations are discussed, specifcally fne-tuning,
reducing overftting, and using CNN activations as features. Then, some commonly
used networks, such as AlexNet, GoogleNet, VGG-19, and ResNet, are presented.
The theoretical knowledge of CNN types along with the three most common archi-
tectures and learning algorithms are provided. Moreover, these networks are called
region-based CNNs (R-CNNs), fully convolutional networks (FCNs), and hybrid
learning networks.
CNNs are widely being applied in different domains because of their outstanding
performance, such as in image classifcation, object detection, speech recognition,
face detection, vehicle recognition, facial expression recognition, segmentation of
MR brain images, and many more. This chapter also introduces various processing
applications of deep convolutional neural networks in medical imaging, such as fea-
ture extraction, tumor detection, and tumor segmentation. This study will serve as a
reference and resource for researchers who are interested in this feld.
This chapter is arranged as follows. It frst presents the general foundations of
CNNs (Section 9.2). Section 9.3 is devoted to reporting effcient learning approaches.
The network operations and a brief introduction to the current programming tools
available for executing these algorithms are described in Section 9.4. An overview
of the commonly used CNN is documented in Section 9.5. Then, a discussion of
different CNN types for specifc issues is presented in Section 9.6. Section 9.7 is
dedicated to discussing deep learning applications in medical imaging. Section 9.8
describes the network design of the proposed system and provides simulation results
and evaluation criteria. Finally, conclusions are outlined in Section 9.9.

9.2 CONVOLUTIONAL NEURAL NETWORKS (CNNS)

The possibility of a CNN is not new. This model appeared to function admirably for
handwritten digit recognition [14] in early 1998. However, because of the inability
of these models to scale to substantially larger images, they gradually lost support.
This loss of support was to a great extent because of memory and equipment limita-
tions and the inaccessibility of large amounts of training data. With the expansion
in computational power due to the wide accessibility of GPUs and the emergence of
extensive datasets such as ImageNet [7], it became possible to train larger and more
complex models. This ability frst appeared with the mainstream AlexNet model,
which was discussed previously and was the start of using the deep neural network
in computer vision.

9.2.1 CNN LAYERS

In this section, we will discuss the essential building layers of CNNs. This discus-
sion assumes that the reader knows about traditional neural networks (TNNs), which
Deep Convolutional Neural Networks 237

FIGURE 9.2 Representation of CNN network weights. A ReLU nonlinearity is added after
every convolutional layer.

we will call “fully connected layers (FCLs)”. Figure 9.2 presents a portrayal of the
weights in the AlexNet network. Although convolutional layers are used in the frst
fve layers, the last three layers are fully connected.

9.2.1.1 Convolutional Layers

Utilizing TNNs for image classifcation is unfeasible for the following reason: the
number of input nodes will be very large. Consider a 200 × 200 image, for which we
would have 40,000 nodes. If the hidden layer has 20,000 nodes, the size of the input
weights matrix would be 40,000 × 20,000 = 800 million. This number is only for one
layer, and as we increase the number of layers, this number signifcantly increases
more quickly. Moreover, image vectorization ignores the image spatial structure.
How can we develop a network that overcomes these disadvantages?
Two-dimensional convolutions are used instead of matrix multiplications.
Additionally, it normally considers the 2D structure of the image. Meanwhile, con-
volutions can also be thought of as a TNN with two requirements: local connec-
tivity and weight sharing [15]. There is additional proof from visual neuroscience
that comparable calculations occur inside the human brain. Hubel and Wiesel [16]
discovered two types of cells in the essential visual cortex: the basic cells and the
complex cells. The basic cells reacted primarily to arranged edges and gratings,
which are reminiscent of Gabor features, an exceptional class of convolutional chan-
nels. The complex cells were also sensitive to these edges and gratings. CNNs are
profound learning networks that have been enhanced in recent decades. CNNs can
automatically learn features from inputs, particularly images, similar to the case in
this exploration [17]. The convolutional layers are utilized to convolve the image
with weights to acquire a feature map. The weights of the portions associate the
feature map to the previous layer. In the example shown in Figure 9.2, the numbers
11 × 11 × 3, along with 96 underneath C1, show that the frst layers have 96 flters,
every one a spatial measurement of 11 × 11, with one for each of the 3 RGB channels.
238 Deep Learning in Computer Vision

Therefore, the subsequent component initiation after convolution has 96 channels,

and we are now able to apply another arrangement of multichannel features for these
features as well. The 2D convolution can be defned as in Equation 9.1 [17].

y(i, j ) = x(i, j )* w(i, j ) = ˜˜x(m, n)w(i − m, j − n).

m n
(9.1)

where x is the input and w is the weight that is provided from the previous layer. The
convolution matrix size is m × n, and the convolutional layer output is y. The input
of the CNN and the convolutional kernel dimensions should be the same. A multidi-
mensional kernel should be used if the CNN is a multidimensional array [17].

9.2.1.2 Max-Pooling Layer

The Neocognitron display inspired the consideration of basic cells as convolutions.
Proceeding in a similar way, the complex cells can be considered a max-pooling
task. This task can be thought of as a maximum flter, where every n × n area is sup-
planted with its maximum value [18]. The max-pooling task requires the following:

1. Selection of the most astounding activation in a nearby region, in this way

providing a low level of spatial invariance. This is comparable to the task of
complex cells.
2. Size reduction of the following layer by a factor of n2. Fewer parameters
should be learned in the following layers when the activation size is reduced.

Different types of pooling include stochastic pooling [19], average pooling, and win-
ner-takes-all pooling [20]. However, these types of pooling are not as commonly
utilized as max-pooling. Other pooling types occasionally cannot extract good fea-
tures; for example, average pooling provides an average value by taking all features
into account, which is a very generalized computation. Therefore, poor accuracy
will be obtained with average pooling when the systems do not require all features
from the convolutional layer. Another example, winner-takes-all pooling, is similar
to max-pooling; however, it keeps the feature dimension fxed, and no downsampling
occurs. However, some classifcation tasks have used different pooling or mixed
pooling and outperform max-pooling. Thus, the selection of pooling type depends
on the database type.

9.2.1.3 Nonlinearity and ReLU

Nonlinearities are necessary between layers to ensure that the network is more
expressive than a linear one. This makes the network more powerful and includes
the capability to learn complicated and complex features from data. The nonlinearity
enables the network to generate nonlinear mappings from inputs to output; further-
more, it can cut-off or adjust the developed output. A nonlinear layer is employed to
reduce the generated output or saturate the output [21]. Thus, the convolutional layers
are followed by a nonlinear layer. Customary feedforward neural networks utilize
the sigmoid or tanh nonlinearities, as shown in Equations 9.2 and 9.3, respectively.
Therefore, the new convolutional systems utilize a nonlinear activation function. The
Deep Convolutional Neural Networks 239

most commonly applied activation function for deep learning applications in the lit-
erature is rectifed linear unit (ReLU) because of the following reasons:

1. It exhibits faster learning, which is characterized as in Equation 9.4 [17].

2. ReLU provides better performance and generalization compared to other
activation functions used in deep learning.
3. Sigmoid and tanh functions cause a problem called the vanishing gradient
in the backpropagation, which is solved with ReLU [21].

CNNs with this nonlinearity have been found to train quicker than others.

1
s( x ) = (9.2)
1+ e − x

e x − e− x
tanh( x ) = (9.3)
e x + e− x

Re LU( x ) = max(0, x ) (9.4)

Recently, Maas et al. [22] presented another type of nonlinearity, called the leaky-
ReLU. This nonlinearity is characterized as shown in Equation 9.5 [18].

leaky − ReLU( x ) = max(0, x) + a min(0, x) (9.5)

where α is a predecided parameter. He et al. [23] enhanced this by recommending

that the α parameter be additionally learned, prompting a considerably richer model.

9.2.1.4 Fully Connected Layers (FCLs)

The TNN was created by matrix multiplications rotated with nonlinearities such
as sigmoid. These layers are currently called fully connected layers, where every
unit in the previous layer is linked to each unit in the following layer. This is in
contrast to convolutional layers, where there is only a local connection. In cur-
rent networks, utilizing numerous FCLs is mainly avoided because it utilizes a
great number of parameters [24]. One or more FCLs can be used, as shown in
Figure 9.2.

9.3 OPTIMIZATION APPROACHES

A highly expressive network is of little use without a good learning algorithm to
effectively learn the network’s parameters. The generic CNN era endeavored to pro-
duce such effective algorithms. The tasks in computer vision show that a powerful
model can be learned by a very simple supervised training procedure. Minimizing
the loss function is the general approach for learning. The sigmoid cross-entropy or
softmax loss functions are used in classifcation tasks, while those including regres-
sion utilize the Euclidean error function. In Figrue 9.2, the FC8 layer output classifes
one of the thousand classes of the dataset.
240 Deep Learning in Computer Vision

9.3.1 GRADIENT-BASED OPTIMIZATION

The backpropagation algorithm is generally used to train neural networks [25],
which speeds up the gradient descent (GD) computation by the chain rule. However,
for large datasets with (at least thousands) images, utilizing GD is unfeasible. In
such cases, the stochastic gradient descent (SGD) is regularly utilized, where one
calculates gradients for individual images instead of the whole dataset. It has been
discovered that training with SGD is superior to training with GD. The main disad-
vantage of SGD is that it converges slowly [26]. To overcome this disadvantage, SGD
is commonly utilized with a minibatch, where the minibatch generally contains a
few data points (≈ 100) instead of all data points.
Momentum [27] has a place with a group of techniques that aim to accelerate
the SGD convergence. This technique is mainly used in deep network training
and is regularly considered a basic component. Augmentations such as Adam [28],
Nesterov’s accelerated GD [29], Adagrad [30], and Adadelta [31] are known to work
similarly well. For a point-by-point discussion on how these techniques function,
more details can be found in [32].

9.3.2 DROPOUT PROPERTY

This occurs when the model network has more terms than necessary or uses more
complicated techniques with numerous training parameters, especially with a small
amount of training data or limited data, to combat this effective regularization tech-
nique. It has been observed that traditional regularizers such as ˜1 or ˜2 regulariza-
tion on the weights of the neural network are inadequate in this view.
Another nice property of CNNs is named “dropout”, which is a simple regular-
ization technique to solve the overftting problem. Dropout is a strong regularization
method [33], and it has been shown to enhance generalization for extensive neu-
ral networks. In dropout, we arbitrarily drop neurons in training with a probabil-
ity p. Therefore, only an irregular subset of neurons in SGD iterations are trained.
Applying dropout properties is valuable only in the training stage. At test time, we
utilize all neurons, and every neuron activation is multiplied by p to represent the
scaling. Hinton et al. [33] demonstrated that this strategy was equivalent to training a
large range of neural networks with shared parameters and then utilizing their geom-
etry, intending to obtain a single expectation. Numerous augmentations to dropout,
such as DropConnect [34] and Fast Dropout [35], have appeared to work better in
specifc cases. Maxout [36] is a nonlinearity that enhances the execution of a net-
work that utilizes dropout.

9.3.3 BATCH NORMALIZATION

Batch normalization (BN) [37] is another helpful regularizer that enhances the gen-
eralization and that also defnitely accelerates convergence. This system handles the
issue of inward covariate shift, where the circulation of each layer’s data sources
changes persistently amid the training phase. This is the result of variable weights
for each layer, which results in a changing output activation distribution. This ability
Deep Convolutional Neural Networks 241

requires careful preparation and generally slows training. To overcome this issue, the
layer output activations are normalized by BN to guarantee that its value is inside a
small interval. In particular, minibatch normalization is performed with BN, which
applies the average of the mean-variance statistics. Extra move and scale parameters
are additionally learned to frustrate the normalization effect if necessary. Recently,
BN has been observed to be a fundamental part of extra deep network training.

9.3.4 NETWORK DEPTH

The universal approximation hypothesis [38] states that one hidden layer is adequate
to display any continuous function. However, it has been reported [6] that such sys-
tems require an exponentially extensive number of neurons compared to using many
hidden layers. Although the inspiration for making deep neural networks was clear,
for quite a while researchers did not have a method that could profciently train deep
neural networks (greater than 3 layers). Training much deeper networks became pos-
sible for researchers after the introduction of greedy layerwise pretraining [2], which
made deep learning networks the mainstream in machine learning.
Recently, many deep neural networks have appeared; for example, AlexNet [8]
has 7 layers. Recent models such as VGGnet [39] and GoogleNet [40] have 19 and
22 layers, respectively, and appeared to perform much better than AlexNet. The
main disadvantage of the much deeper networks is that such networks are very hard
to train. Srivastava et al. [41] used deep neural networks with 100 layers to present
highway systems. He et al. [24] utilized an adaptation of this method with 152 layers
to obtain cutting-edge results on the ILSVRC 2015 database.

9.4 OPERATIONAL PROPERTIES

9.4.1 USING PRETRAINED CNNS
The main advantage of the AlexNet network is that it can be used for different tasks
for which it was not originally intended. It is very easy to download a pretrained
model and then adjust the model to suit the current application. We explain two such
approaches to utilize networks in this way [18].

9.4.1.1 Fine-Tuning
This process refers to modifying a pretrained network that is used for image clas-
sifcation for an alternate task that is performed by utilizing the trained weights
as an initial condition and starting SGD again for the other task. The learning
rate will be much lower than that for the original network. When the new task is
the same as the original task, we can make the earlier layers fxed, and only the
later semantic layers should be relearned. However, if the new task is completely
different, we should either retrain all layers or train the network from scratch. The
number of layers to retrain also depends on the quantity of information available
for learning the new task. The larger the dataset is, the higher is the number of
layers that can be retrained. The reader is referred to Yosinski et al. [42] for more
information.
242 Deep Learning in Computer Vision

9.4.1.2 CNN Activations as Features

As mentioned previously, the later layers in AlexNet appear to learn visually seman-
tic characteristics. These representations are essential in making 1,000-class clas-
sifcations. Because these demonstrate a large number of classes, the FC7 layer can
be used as a global feature identifer. The provided features are better than the tradi-
tional features, such as SIFT or HOG, for different computer vision tasks. Donahue
et al. [43] frst presented the use of CNNs as features and performed tests to decide
their proportionality for different errands. Fully connected layer activations are addi-
tionally utilized for image recovery [44]. Indeed, these are used as generic features
and can be utilized in many tasks [45].

9.4.2 DEEP LEARNING DIFFICULTIES

9.4.2.1 Overftting Reduction
Training a deep learning model with a limited number of training images is a chal-
lenging task due to the number of learnable parameters. Various strategies have been
developed by researchers, including the following:

1. Applying a nonlinear activation function such as ReLU [46].

2. Using a good initialization and increasing the momentum slowly [47].
3. Applying dropout to randomly deactivate connections on each iteration
[48].
4. Performing normalization by using batch normalization layers [37].
5. Using denoising autoencoder layers to reconstruct the corrupted inputs [49].

9.5 COMMONLY USED DEEP NEURAL NETWORKS

Deep learning uses considerable amounts of computational energy and acquires
advanced results in multiple diffcult tasks. Among the various deep learning strate-
gies, methods based on convolutional neural networks (CNNs) are more appropriate
for tasks related to images. The main goal of this section is to provide an overview
of the commonly used CNN methods, such as AlexNet, VGG-19, deep residual net-
works, and GoogleNet, as reported in Table 9.1.

TABLE 9.1
Comparison among Well-Known Convolutional Neural
Networks Using ImageNet Dataset
Year CNN No. of layers No. of parameters Error rate

2012 AlexNet [8] 7 60 million 15.3%

2014 GoogleNet [40] 22 4 million 6.67%
2014 VGGNet [39] 19 138 million 7.3%
2015 ResNet [24] 152 50 million 3.6%
Deep Convolutional Neural Networks 243

9.5.1 ALEXNET
AlexNet has been discussed previously. Figure 9.2 shows max-pooling grouped with
layers 1, 2, and 5, while the last two fully connected layers are combined with drop-
out because they contain the most parameters. Local response normalization has
been provided to layers 1 and 2, which does not affect the network performance as
in [50]. The ILSVRC 2012 database, which has 1.2 million images for 1,000 classes,
is used to train the network. Two GPUs are used to train the network for one month.
With more powerful GPUs, the network can be trained in a few days [50]. Learning
algorithms for hyperparameters such as dropout, weight decay, momentum, and
learning rate were manually tuned. Blob-like features and Gabor-like oriented
edges are learned in the prior layers. The highest-order features, such as shapes, are
learned by the next layers. Semantic attributes such as wheels or eyes are learned by
the last layers.

9.5.2 GOOGLENET
For GoogleNet [40], the best method at ILSVRC-2014; the classifcation perfor-
mance is signifcantly improved with very deep networks. The number of parameters
has been increased due to the large number of layers, and a number of design tricks
are used, such as using a 1 × 1 convolutional layer after the traditional convolutional
layer, which reduces the number of parameters and increases the expressive power of
the CNN. It is believed that having at least one 1 × 1 convolutional layer is much the
same as having a multilayer perceptron network preparing the convolutional layer
output that goes before it. Another trick used by the authors is to involve the internal
layers of the network in the calculation of the target function rather than the fnal
softmax layer (as in AlexNet).

9.5.3 VGG-19
The VGG-19 [39] network design is one of the highest-performing deep CNN
(DCCN) networks. The VGG design provides a fascinating feature in that it divides
the large convolutional flters into small ones. The small flters are carefully selected
with a number of parameters approximately the same as the parameters in the larger
convolutional flters, which they should replace. The net impact of this design choice
is productivity and regularization-like impact on parameters because of the smaller
size of the included flters.

9.5.4 DEEP RESIDUAL NETWORKS (RESNET)

Deep residual networks (ResNet) [24] increase the number of trained layers to the
highest depth with 152–200 layers. The network used residual connections between
layers that were used to train network layers to adapt the residual map instead of
adapting to the underlying map. In other words, the layers did not learn a mapping of
h(x) directly, but it will learn f(x) = h(x)-x and then add x to be f(x)+x. This straightfor-
ward method was observed to be helpful for learning networks with several layers,
providing sophisticated accuracy in many tasks.
244 Deep Learning in Computer Vision

TABLE 9.2
Software List for Convolutional Neural Networks
Package Interfaces AutoDiff

Theano Python Yes

MatConvNet MATLAB No
Chainer Python No
Caffe Python, MATLAB, C++, command line No
Torch Torch, C Yes
TensorFlow C++, Python Yes
CNTK C++, command line No

9.5.5 TOOLS AND SOFTWARE

In this section, we will discuss different software programs that are available for
deep neural network learning. Table 9.2 includes the commonly used packages.
Caffe is the most commonly utilized software among computer vision applica-
tions. It is composed with C++ and is frequently utilized through the Python and
command line interfaces. The structure of this system makes it simple to utilize
pretrained neural systems for different tasks. In general, Theano, Torch, and
TensorFlow appear to be more prominent. These structures are intended to make
it simple to alter low-level subtle elements of neural system models. Theano and
TensorFlow are utilized in Python, while Torch is utilized in Lua. Likewise, a
couple of wrapper libraries exist for Theano, for example, Lasagne [51] and Keras
[52], which provide normally utilized functionalities and some help for pretrained
models.
In this specifc situation, automatic differentiation (AutoDiff) [53] is an essential
feature to have in such a structure. When determining new layers and regularizers
for neural systems, one also needs to indicate what the gradient through that layer
would resemble. AutoDiff calculates the derivative of complex functions automati-
cally in a closed shape using derivative defnitions of primitives.

9.6 CONVOLUTIONAL NEURAL NETWORK TYPES

9.6.1 REGION-BASED CONVOLUTIONAL NEURAL NETWORKS (R-CNN)
Most CNN networks are trained to classify images using images that include one
object. During testing, most of the images include multiple objects, but the CNN
network can recognize a single object. The main problem with CNN network design
is the need to cover more tasks than image classifcation, such as localization and
detection. The major aim of object detection is to detect different objects in each
image of various sizes. Meanwhile, a group of object-like regions is detected in a
specifc image, named region proposals, which have gained much attention [54].
Region proposal methods work at a low level and produce hundreds of objects such
as image patches at multiple levels. To appoint a network classifcation to cover the
Deep Convolutional Neural Networks 245

task of localizing objects, the network must search for image regions with different
scales simultaneously.
Girshick et al. [55] used region proposals to solve the problem of object local-
ization. The method obtained the best results on the PASCAL VOC 2012 detec-
tion dataset. It is named region-based CNN (R-CNN) because it uses regions of the
images followed by a CNN. Different works have used the R-CNN method to extract
features in small regions to solve many target applications in computer vision.

9.6.2 FULLY CONVOLUTIONAL NETWORKS (FCNS)

The achievements of convolutional neural networks in object detection [55] and clas-
sifcation tasks [8, 40] have motivated specialists to utilize deep neural networks for
more diffcult tasks, such as scene parsing and semantic segmentation. Scene pars-
ing and semantic segmentation assign a label for each pixel in the image to which it
belongs (e.g., table, sofa, road, and so forth), which is called per-pixel classifcation.
The region features are extracted to a support vector machine (SVM) to classify
them into classes.
Long et al. [56] provided a fully convolutional network (FCN) structure to learn
per-pixel tasks, such as semantic segmentation, in a comprehensive manner. FCN
layers perform a location-invariant activity. The FCN network does not have any
inner product or densely connected layers, as in CNN networks that are used for clas-
sifcation. The classifcation CNN networks have a restriction on input size because
the inner product layers have a limitation on the input size. However, FCN does not
include any of the inner product layers; it can basically work with any input size.
Long et al. [56] introduced a deconvolution layer that obtains the spatial resolution
back to the upsampling operation from the output of the sub-sampled image. They
performed the upsampling operation at intermediate layers, which are concatenated
to acquire features at the pixel level.

9.6.3 HYBRID LEARNING NETWORKS

Multitask learning is basically a machine learning model, wherein the goal is to
prepare the learning network to perform well on numerous tasks. Multitask learn-
ing systems will generally explore shared portrayals that exist among the functions
to obtain a superior generalization execution over their counterparts created for one
task. In CNNs, multitask learning is acknowledged utilizing distinctive methodolo-
gies. One class of methodologies uses a multitask loss function with hyperparameters
normally controlling the losses. For instance, Girshik et al. [57] utilize a multitask
loss to train the network for bounding-box and classifcation regression functions
along these lines, enhancing execution for object recognition.

9.7 DEEP LEARNING APPLICATIONS IN MEDICAL IMAGING

Over the past decades, we have seen the signifcance of medical imaging, e.g., mag-
netic resonance (MR), mammography, computed tomography (CT), X-ray, ultra-
sound, positron emission tomography (PET), and so forth, for the early treatment,
246 Deep Learning in Computer Vision

diagnosis, and detection of diseases [58]. In medical centers, the interpretation of

medical images has generally been performed by human specialists, for example,
radiologists and doctors. Because of vast varieties in pathology and the potential
exhaustion of human specialists, the use of computer-aided diagnosis systems has
increased. As medical imaging technologies improve, medical image analysis will
improve with the support of machine learning methods. Feature representation
or meaningful feature extraction is important for machine learning prosperity to
achieve the target tasks. Specialists, including radiologists and physicians, use their
knowledge about the object domains in designing task-related or meaningful fea-
tures, making it diffcult for non-specialists to exploit automated learning systems
for their own examinations [59].
However, deep learning has mitigated these obstacles by retaining the feature
extraction step in a learning step [60]. Deep learning requires only the input dataset.
Thus, feature extraction is performed internally in deep learning networks, which
enables non-specialists in machine learning to adequately utilize deep learning for
their own applications. Deep learning strategies are very successful due to the fol-
lowing reasons:

1. Developments in learning algorithms.

2. Advancements in graphics processing units (GPUs) and central processing
units (CPUs).
3. Dataset availability.

Generally, deep learning strategies are very successful when the available number of
image samples in the training stage is very large. For instance, 1 million images were
given in the ILSVRC competition. However, the medical feld has a limited number
of samples in the dataset, less than 1,000 samples. This increases the challenge of
developing a deep learning model with a low number of training samples without
experiencing overftting. Various strategies have been developed by researchers [59],
such as the following:

1. Reducing the dimension of the input by using 2D or 3D patches rather than

the full images.
2. Using data augmentation to train the network from scratch.
3. Using a pretrained model as a feature extractor and retraining it with the
target task.
4. Using pretrained model parameters and then fne-tuning them with the
related medical images.
5. Transforming the fully connected layer weights of a pretrained model into
convolutional kernels.

There are several medical imaging applications that depend on the type of medi-
cal images. Medical imaging applications can be categorized based on the human
parts (such as brain tumor, heart, organ/body part, cell, pulmonary nodules, lymph
nodes, interstitial lung disease, cerebral microbleeds, and noticeably sclerosis
lesion), medical image type (MRI, CT, X-ray, ultrasound, and PET), and application
Deep Convolutional Neural Networks 247

type (detection, segmentation, and classifcation). In this section, we present the

deep learning methods used for feature representation, structure detection, and
segmentation.

9.7.1 DEEP LEARNING FOR FEATURE EXTRACTION

The methods for medical image processing mainly characterize the local anatomical
characteristics based on representations of the morphological features designed by
experts. Moreover, the represented features of an image for a specifc problem are hardly
ever reusable for other types. For instance, the methods of brain MR image segmenta-
tion designed for 1.5-Tesla T1-weighted images are not viable for 7.0-Tesla T1-weighted
or T2-weighted images, or for different imaging modalities or distinct organs [61, 62].
Supervised learning methods require large amounts of training data that are manually
labeled by experts to fnd the most essential and relevant features for specifc tasks,
while the extracted features may deviate due to the complexity of the anatomical struc-
tures. Therefore, a change in image features makes the training process start again.
Wu et al. [61, 62] presented a general framework for feature extraction that detects
the essential characteristics for the accurate segmentation and detection of brain
regions and anatomical structures. Furthermore, this framework can be used for var-
ious types of medical images. A stacked autoencoder (SAE) is used with a sparsity
constraint to learn extracted features hierarchically from layer to layer, which con-
tains hierarchical encoding and decoding systems. A nonlinear deterministic map-
ping is used to map the input to an activation vector y. This procedure is repeated by
using the vector y as input for the next layer until the high-level feature presentations
are obtained. The decoding system is used to minimize the reconstruction errors
between the reconstructed patch from the decoder and the input patch.
From the application mentioned above, we found that

1. Feature extraction by deep learning represents the local characteristics of

the image very well.
2. Deep learning can learn the intrinsic features of new medical imaging
modalities for new image analysis methods.
3. It is fully adaptive to understand the data of the image and apply to different
medical imaging applications, such as prostate localization [63, 64], hippo-
campus segmentation [65], and brain tumor detection [17].

9.7.2 DEEP LEARNING FOR TUMOR DETECTION

Automatic detection aims to classify medical images as abnormal and normal due to
the presence or absence of abnormal or suspicious regions (tumors). Computer-aided
detection (CAD) is very important for medical image diagnosis. The main goal is to
maximize the detection rate. Typically, the traditional pipeline of CAD is as follows:

1. Image processing techniques are applied to the input images.

2. A set of features presents the preprocessed images, such as statistical or
morphological information.
248 Deep Learning in Computer Vision

3. A classifer such as support vector machine (SVM) is used to make a deci-

sion. Different detection applications are applied to different medical image
modalities to detect or segment suspicious or abnormal regions, such as
organ or body part detection (liver, neck, pelvis, lungs, and legs), cell detec-
tion, and brain tumor detection [66, 67].

Recently, deep learning models have improved detection performance in different

medical imaging applications, such as lymph node detection [68], pulmonary nodule
detection [69], interstitial lung disease classifcation [70], multiple sclerosis lesion
detections [71], cerebral microbleed detection [72], brain tumor detection [17, 73],
and cell detection [74].

9.7.3 DEEP LEARNING FOR TUMOR SEGMENTATION

Automatic segmentation is used for quantitative assessment of the brain at all
ages. Removing non-brain regions such as the skull is the main preprocessing step.
Kleesiek et al. [75] used 3D convolutional neural networks for skull extraction in
T1-weighted MRI images. Moeskops et al. [76] presented an accurate tissue seg-
mentation method. They used a multiscale CNN with multiple convolution sizes
and multiple patch sizes, which provided very good segmentation results with an
average Dice ratio from 0.82 to 0.91. The brain MRI segmentation of infants is
more diffcult than that of adults due to severe partial volume effects, reduced tis-
sue contrast, ongoing white matter (WM) myelination, and increased noise in the
infant’s brain [77].
Zhang et al. demonstrated multimodality MR segmentation using four CNN
architectures for infant brain tissues. Three convolutional layers were applied, fol-
lowed by a fully connected layer and a softmax layer for tissue classifcation [78].
More recently, in [79], the isointense-phase brain image was segmented by multiple
fully convolutional networks (MFCNs); the average Dice ratios were 0.873 for gray
matter (GM), 0.852 for cerebrospinal fuid (CSF), and 0.887 for WM. Moreover,
several methods for brain tumor segmentation can be found in the literature [73, 80,
81, 82, 83, 84, 85].

9.8 BRAIN TUMOR DETECTION USING DEEP

CONVOLUTIONAL NEURAL NETWORK
The combination of machine learning approaches and brain magnetic resonance
imaging (MRI) is crucial for the brain tumor diagnosis process by clinicians and
radiologists. Although most tumor diagnosis studies have focused on segmenta-
tion and localization operations, few studies have focused on tumor detection as
a time- and effort-saving operation in which normal images can be discarded
early. This study introduces a new network structure for accurate brain tumor
detection using a deep convolutional neural network (DCNN). The proposed
structure is designed to identify the presence or absence of a brain tumor in MRI
images.
Deep Convolutional Neural Networks 249

9.8.1 DCNN DESIGN

A CNN architecture is proposed for brain tumor detection using MRI images. A
block diagram of the proposed architecture is shown in Figure 9.3 The sequence of
processes occurs through the proposed structure as follows: brain MRI images are
fed to the network’s input, and then preprocessing is performed to reduce the calcu-
lation complexity and hence reduce the processing time. Because the database used
contains 2D slices, the images are resized to 96 × 96 pixels because of the varying
MRI image sizes.
Later, the network, which includes convolutional layers and fully connected
layers, is used to classify the input image as a normal or abnormal (contains
tumor) image. The convolutional layers are used to extract features in the feature
extraction path. Moreover, it contains 7 convolutional layers with a ReLU acti-
vation function and 7 max-pooling layers. The convolutional layers use a small
flter size of 5 × 5 pixels to provide features in the images. The max-pooling layer
is used after each convolutional layer to downsample the convolutional layer
output.
Furthermore, the output path contains a BN layer connected to a ReLU layer and
a fully connected layer that is followed by a dropout layer. Dropout with different
values was tested, and we found that a ratio of 0.8 performed with slightly better
outcomes. At the fnal stage, another fully connected layer with the softmax function
is used for the image classifcation process.

FIGURE 9.3 A block diagram of the proposed DCNN architecture for brain tumor detection.
250 Deep Learning in Computer Vision

9.8.2 SIMULATION EXPERIMENTS AND EVALUATION CRITERIA

9.8.2.1 Environment Setup
We extracted the dataset from the Reference Image Database to Evaluate Response
(RIDER) [86]. The dataset consists of 109 normal images and 240 abnormal images.
We show a sample image of the dataset used in the experimental work in Figure 9.4.
Table 9.3 provides the specifcations of the training and testing subsets. We devel-
oped the proposed DCNN code in a Jupyter notebook using Keras and TensorFlow
toolkits. The machine is equipped with an Intel Core i7 processor that works at 3.2
GHz with 24 GB of RAM and the Ubuntu desktop 64-bit experimental environment.

9.8.2.2 Evaluation Criteria

Several metrics can test classifcation algorithms. These metrics can be categorized
as scalar metrics, such as accuracy (ACC), geometric mean (GM), sensitivity (or
recall), specifcity (or inverse recall), and positive prediction value (PPV) (precision),
and graphical metrics, such as detection error trade-off (DET), precision-recall, and
receiver operating characteristics (ROC) curves. We present most of these metrics in
this section with details.

FIGURE 9.4 Sample images used for the experimental work. The image dataset is extracted
from the RIDER database. The top row shows normal brain MRI images, while abnormal
brain images are shown in the bottom row.

TABLE 9.3
Image Distribution in the Performance Evaluation Dataset
Image type No. of training sets No. of testing sets

RIDER Normal 45 64
Abnormal 77 163
Deep Convolutional Neural Networks 251

The accuracy (ACC) is a widely used metric for the classifcation performance,
which is the ratio between the correctly classifed samples to the total number of
samples, as shown in Equation 9.6.

TP + TN
ACC = . (9.6)
TP + TN + FP + FN

where TP, TN, FP, and FN indicate the true positive, true negative, false positive, and
false negative, respectively [66]
Sensitivity (SV), recall, hit rate, and true positive rate (TPR) represent the true
positive samples to the total number of positive samples, as shown in Equation 9.7.
Specifcity (SP), inverse recall, and true negative rate (TNR) express the correctly
classifed negative samples to the total number of negative samples, as shown in
Equation 9.8 [67].

TP
SV = . (9.7)
TP + FN

TN
SP = . (9.8)
TN + FP

The precision or positive prediction value (PPV) refects the ratio of correctly clas-
sifed positive samples to all classifed positive samples, as shown in Equation 9.9.
Negative predictive value (NPV) or inverse precision represents the ratio of cor-
rectly classifed negative samples to the entire classifed negative samples, as
shown in Equation 9.10. Equation 9.11 refers to the balanced classifcation rate
(BCR) or balanced accuracy (BA), which combines the sensitivity and specifcity
metrics [17].

TP
Precision = PPV = . (9.9)
TP + FP

TN
NPV = . (9.10)
TN + FN

SV + SP 1 ° TP TN ˙
BCR = BA = = ˝ + ˇ. (9.11)
2 2 TP + FN TN + FP ˆ
˛

The false positive rate (FPR) represents the ratio between the false positive classi-
fed samples to all negative samples, as shown in Equation 9.12. However, the false
negative rate (FNR) expresses the false negative classifed samples to all positive
samples, as shown in Equation 9.13.

FP
FPR = . (9.12)
TN + FP
252 Deep Learning in Computer Vision

FN
FNR = . (9.13)
TP + FN

In addition to the above metrics, there are many different metrics that can be con-
sidered for evaluating various applications. The new metrics include Youden’s index
(YI) or Bookmaker informedness (BM), Matthews correlation coeffcient (MCC),
discriminant power (DP), F-measure, markedness (MK), geometric mean (GM),
Jaccard, and optimization precision (OP), as shown in Equations 9.14, 9.15, 9.16,
9.17, 9.18, 9.19, 9.20, and 9.21, respectively [87].

YI = SV + SP −1. (9.14)

TP × TN − FP × FN
MCC = . (9.15)
(TP + FP)(TP + FN)(TN + FP)(TN + FN)

3° ° SV ˙ ° SP ˙ ˙
DP = log ˝ + log ˝ . (9.16)
p ˝˛ ˛ 1 − SP ˇˆ ˛ 1 − SV ˇˆ ˇˆ

2PPV × SV
F-measure = . (9.17)
PPV + SV

MK = PPV + NPV −1. (9.18)

GM = SV + SP. (9.19)

TP
Jaccard = . (9.20)
TP + FP + FN

SV − SP
OP = ACC − . (9.21)
SV + SP

9.8.2.3 Experimental Results

To test the proposed system, standard evaluation metrics such as accuracy, sen-
sitivity, and specifcity have been considered [17, 66]. Multiple tests were run to
confrm the performance of the proposed network in fulflling the tumor detec-
tion task. A comparison with other methods found in the literature is presented
in Table 9.4. The empirical works show that the proposed DCNN structure can
achieve better results in terms of accuracy, sensitivity, and specifcity compared
to other approaches in the literature. Although the network training time was not
considered, the average testing time per MRI image was measured to be 0.24
seconds.
Deep Convolutional Neural Networks 253

TABLE 9.4
Comparison of the Proposed Brain Tumor Detection
Model against Some Approaches in the Literature
Performance evaluation criteria (%)

Methods Sensitivity Specifcity Accuracy

Abd-Ellah et al. [66] 83.43 25.00 66.96

Mohsen et al. [88] 97 97 96.97
Proposed DCNN 98.43 97.54 97.79

9.9 CONCLUSION
The major advantage of deep learning over the traditional machine learning
techniques is that it can freely reveal signifcant features in a high-dimensional
database. Recently, CNNs have led to great developments in processing text,
speech, images, videos, and other applications. In this chapter, the basic under-
standing of CNNs and the recent enhancements of CNNs were explained and
presented. CNN improvements have been discussed from various prospects,
specifcally the general model of CNN, layer design, loss function, activation
function, optimization, regularization, normalization, fast computation, and net-
work depth, reviewing the benefts of each phase of CNN. Additionally, the use
of pretrained CNNs and deep learning diffculties have been discussed. This
chapter also introduced the commonly used deep convolutional neural networks,
namely, AlexNet, GoogleNet, VGG-19, and ResNet. It discussed the commonly
used programs for deep CNN learning. The chapter focused on different types of
CNNs, which are region-based CNNs, fully convolutional networks, and hybrid
learning networks.
This chapter outlined the main deep learning applications in medical imaging
processing in terms of feature extraction, tumor detection, and tumor segmentation.
A deep convolutional neural network (DCNN) structure was proposed for brain
tumor detection from MRI images. The proposed DCNN was evaluated using the
RIDER database using 122 and 227 images for training and testing, respectively. It
achieved accurate detection within a time of 0.24 seconds per image in the testing
phase. This chapter promises to provide comprehensive knowledge to researchers,
learners, and those who are interested in this feld.

BIBLIOGRAPHY
1. Marra, F., Poggi, G., Sansone, C., Verdoliva, L.: A deep learning approach for iris
sensor model identifcation. Pattern Recognition Letters 113, 4653 (2018), integrating
Biometrics and Forensics.
2. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief Nets.
Neural Computation 18(7), 1527–1554 (2006), https://doi.org/10.1162/neco.2006.18.7.1
527, pMID: 16764513.
254 Deep Learning in Computer Vision

3. Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P.A.: Stacked denoising
autoencoders: Learning useful representations in a deep network with a local denoising
criterion. Journal of Machine Learning Research 11, 3371–3408 (Dec 2010).
4. Hinton, G.E.: Training products of experts by minimizing contrastive divergence.
Neural Computation 14(8), 1771–1800 (2002), https://doi.org/10.1162/08997660276
0128018.
5. Mosavi, A., Varkonyi-Koczy, A.R.: Integration of machine learning and optimization
for robot learning. In: Jabloński, R., Szewczyk, R. (eds.) Recent Global Research and
Education: Technological Challenges, pp. 349–355. Springer International Publishing,
Cham (2017).
6. Bengio, Y.: Learning deep architectures for AI. Foundations and Trends in Machine
Learning 2(1), 1–127 (2009), http://dx.doi.org/10.1561/2200000006.
7. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy,
A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual rec-
ognition challenge. International Journal of Computer Vision (IJCV) 115(3), 211–252
(2015), https://doi.org/10.1007/s11263–015–0816-y.
8. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classifcation with deep convo-
lutional neural networks. In: Pereira, F., Burges, C.J.C., Bottou, L., Weinberger, K.Q.
(eds.) Advances in Neural Information Processing Systems 25, (NIPS 2012), pp. 1097–
1105. Curran Associates, Inc. (2012).
9. Hassaballah, M., Awad, A.I.: Detection and description of image features: An introduc-
tion. In: Awad, A.I., Hassaballah, M. (eds.) Image Feature Detectors and Descriptors:
Foundations and Applications, Studies in Computational Intelligence, Vol. 630, pp.
1–8. Springer International Publishing, Cham (2016).
10. Lowe, D.G.: Distinctive image features from scale-invariant key-points. International
Journal of Computer Vision 60(2), 91–110 (2004), https://doi.Org/10.1023/B:VISI.0000
029664.99615.94.
11. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: 2005
IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR’05). Vol. 1, pp. 886–893. IEEE (2005), https://doi.org/10.1109/CVPR.2005.177.
12. Yang, J., Jiang, Y.G., Hauptmann, A.G., Ngo, C.W.: Evaluating bag- of-visual-words
representations in scene classifcation. In: Proceedings of the International Workshop
on Workshop on Multimedia Information Retrieval. pp. 197–206. ACM, New York,
NY, USA (2007), https://doi.org/10.1145/1290082.1290111.
13. Awad, A.I., Hassaballah, M.: Image Feature Detectors and Descriptors: Foundations
and Applications, Studies in Computational Intelligence, Vol. 630. Springer
International Publishing, Cham, 1st edn. (2016).
14. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to
document recognition. Proceedings of the IEEE 86(11), 2278–2324 (1998), https://doi.
org/10.1109/5–726791.
15. Bishop, C.: Pattern Recognition and Machine Learning. Springer-Verlag New York, 1
edn. (2006).
16. Hubel, D.H., Wiesel, T.N.: Receptive felds, binocular interaction and functional archi-
tecture in the cat's visual cortex. The Journal of physiology 160(1), 106–154 (1962).
17. Abd-Ellah, M.K., Awad, A.I., Khalaf, A.A.M., Hamed, H.F.A.: Two-phase multi-model
automatic brain tumour diagnosis system from magnetic resonance images using con-
volutional neural networks. EURASIP Journal on Image and Video Processing 97(1),
1–10 (2018).
18. Srinivas, S., Sarvadevabhatla, R.K., Mopuri, K.R., Prabhu, N., Kruthiventi, S.S., Babu,
R.V.: Chapter 2—An introduction to deep convolutional neural nets for computer
vision. In: Zhou, S.K., Greenspan, H., Shen, D. (eds.) Deep Learning for Medical Image
Analysis, pp. 25–52. Academic Press (2017).
Deep Convolutional Neural Networks 255

19. Zeiler, M., Fergus, R.: Stochastic pooling for regularization of deep convolutional
neural networks. In: Proceedings of the International Conference on Learning
Representation (ICLR). pp. 1–9 (2013).
20. Srivastava, R.K., Masci, J., Kazerounian, S., Gomez, F., Schmidhuber, J.: Compete to
compute. In: Burges, C.J.C., Bottou, L., Welling, M., Ghahramani, Z., Weinberger,
K.Q. (eds.) Advances in Neural Information Processing Systems 26, (NIPS 2013), pp.
2310–2318. Curran Associates, Inc. (2013).
21. Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural
network. In: 2017 International Conference on Engineering and Technology (ICET).
pp. 1–6 (Aug 2017), https://doi.org/10.1109/ICEngTechnol.2017.8308186.
22. Maas, A.L.: Rectifer nonlinearities improve neural network acoustic models. In:
Dasgupta, S., McAllester, D. (eds.) Proceedings of the 30th International Conference
on Machine Learning. Proceedings of Machine Learning Research, Vol. 28. PMLR,
Atlanta, Georgia, USA (17–19 June 2013).
23. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifers: Surpassing human-level
performance on ImageNet classifcation. In: 2015 IEEE International Conference on
Computer Vision (ICCV). pp. 1026–1034. IEEE, Santiago, Chile (2015), https://doi.
org/10.1109/ICCV.2015.123.
24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In:
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–
778. IEEE, Las Vegas, NV, USA (June 2016), https://doi.org/10.1109/CVPR.2016.90.
25. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by back-prop-
agating errors. Nature 323, 333–336 (1986), https://doi.org/10.1038/323533a0.
26. Bottou, L.: Large-scale machine learning with stochastic gradient descent. In: Lechevallier,
Y., Saporta, G. (eds.) Proceedings of COMP-STAT’2010. pp. 177–186. Physica-Verlag
HD, Heidelberg, Paris, France (2010), https://doi.org/10.1007/978–3-7908–2604–3_16.
27. B.T.Polyak: Some methods of speeding up the convergence of iteration methods. USSR
Computational Mathematics and Mathematical Physics 4(5), 1–17 (1964), https://do
i.org/10.1016/0041–5553(64)90137–5.
28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980
(2014), http://arxiv.org/abs/1412.6980.
29. Nesterov, Y.: A method of solving a convex programming problem with convergence
rate O(1/sqr(k)). Soviet Mathematics Doklady 27(2), 372–376 (1983).
30. Duchi, J., Hazan, E., Singer, Y.: Adaptive subgradient methods for online learning and
stochastic optimization. Journal of Machine Learning Research 12, 2121–2159 (2011).
31. Zeiler, M.D.: ADADELTA: An adaptive learning rate method. CoRR abs/1212.5701
(2012), http://arxiv.org/abs/1212.5701.
32. Sutskever, I., Martens, J., Dahl, G., Hinton, G.: On the importance of initialization and
momentum in deep learning. In: Dasgupta, S., McAllester, D. (eds.) Proceedings of the
30th International Conference on Machine Learning. Proceedings of Machine Learning
Research, Vol. 28, pp. 1139–1147. PMLR, Atlanta, Georgia, USA (17–19 June 2013).
33. Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.:
Improving neural networks by preventing co-adaptation of feature detectors. CoRR
abs/1207.0580 (2012), http://arxiv.org/abs/1207.0580.
34. Wan, L., Zeiler, M., Zhang, S., Cun, Y.L., Fergus, R.: Regularization of neural networks
using DropConnect. In: Dasgupta, S., McAllester, D. (eds.) Proceedings of the 30th
International Conference on Machine Learning. Proceedings of Machine Learning
Research, Vol. 28, pp. 1058–1066. PMLR, Atlanta, Georgia, USA (17–19 June 2013).
35. Wang, S., Manning, C.: Fast dropout training. In: Dasgupta, S., McAllester, D. (eds.)
Proceedings of the 30th International Conference on Machine Learning. Proceedings
of Machine Learning Research, Vol. 28, pp. 118–126. PMLR, Atlanta, Georgia, USA
(17–19 June 2013).
256 Deep Learning in Computer Vision

36. Goodfellow, I., Warde-Farley, D., Mirza, M., Courville, A., Bengio, Y.: Maxout net-
works. In: Dasgupta, S., McAllester, D. (eds.) Proceedings of the 30th International
Conference on Machine Learning. Proceedings of Machine Learning Research, Vol.
28, pp. 1319–1327. PMLR, Atlanta, Georgia, USA (17–19 June 2013).
37. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by
reducing internal covariate shift. In: Proceedings of the 32nd International Conference
on International Conference on Machine Learning, Vol. 37. pp. 448–456. ICML’15,
JMLR.org (2015).
38. Hornik, K.: Approximation capabilities of multilayer feedforward networks. Neural
Networks 4(2), 251–257 (1991), https://doi.org/10.1016/0893–6080(91)90009-T.
39. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image
recognition. CoRR abs/1409.1556 (2014), http://arxiv.org/abs/1409.1556.
40. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke,
V., Rabinovich, A.: Going deeper with convolutions. In: 2015 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR). pp. 1–9. Boston, MA, USA (2015),
https://doi.org/10.1109/CVPR.2015.7298594.
41. Srivastava, R.K., Greff, K., Schmidhuber, J.: Training very deep networks. In:
Proceedings of the 28th International Conference on Neural Information Processing
Systems, Vol. 2. pp. 2377–2385. NIPS’15, MIT Press, Montreal, Canada (2015).
42. Yosinski, J., Clune, J., Bengio, Y., Lipson, H.: How transferable are features in deep
neural networks? In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D.,
Weinberger, K.Q. (eds.) Advances in Neural Information Processing Systems 27, pp.
3320–3328. Curran Associates, Inc. (2014).
43. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.: DeCAF:
A deep convolutional activation feature for generic visual recognition. In: Xing,
E.P., Jebara, T. (eds.) Proceedings of the 31st International Conference on Machine
Learning. Proceedings of Machine Learning Research, Vol. 32, pp. 647–655. PMLR,
Bejing, China (22–24 June 2014).
44. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image
retrieval. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision—
ECCV 2014. pp. 584–599. Springer International Publishing, Cham (2014), https://do
i.org/10.1007/978–3-319–10590–1_38.
45. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: CNN features off-the-shelf: An
astounding baseline for recognition. In: 2014 IEEE Conference on Computer Vision
and Pattern Recognition Workshops. pp. 512–519 (June 2014), https://doi.org/10.1109/
CVPRW.2014.131.
46. Glorot, X., Bordes, A., Bengio, Y.: Deep sparse rectifer neural networks. In: Gordon, G.,
Dunson, D., Dudík, M. (eds.) Proceedings of the Fourteenth International Conference
on Artifcial Intelligence and Statistics. Proceedings of Machine Learning Research,
Vol. 15, pp. 315–323. PMLR, Fort Lauderdale, FL, USA (11–13 Apr 2011).
47. Glorot, X., Bengio, Y.: Understanding the diffculty of training deep feedforward
neural networks. In: Teh, Y.W., Titterington, M. (eds.) Proceedings of the Thirteenth
International Conference on Artifcial Intelligence and Statistics. Proceedings of
Machine Learning Research, Vol. 9, pp. 249–256. PMLR, Chia Laguna Resort,
Sardinia, Italy (13–15 May 2010).
48. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: A
simple way to prevent neural networks from overftting. Journal of Machine Learning
Research 15, 1929–1958 (2014).
49. Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P.A.: Stacked denoising
autoencoders: Learning useful representations in a deep network with a local denoising
criterion. Journal of Machine Learning Research 11, 3371–3408 (2010).
Deep Convolutional Neural Networks 257

50. Chatfeld, K., Simonyan, K., Vedaldi, A., Zisserman, A.: Return of the devil in the
details: Delving deep into convolutional Nets. In: Valstar, M., French, A., Pridmore,
T. (eds.) Proceedings of the British Machine Vision Conference. BMVA Press (2014),
http://dx.doi.org/10.5244/C.28.6.
51. Lasagne: https://lasagne.readthedocs.io/en/latest/, Accessed: September 01, 2019.
52. Keras: https://keras.io/, Accessed: September 01, 2019.
53. Neidinger, R.: Introduction to automatic differentiation and MAT-LAB object-ori-
ented programming. SIAM Review 52(3), 545–563 (2010), https://doi.org/10.1137/
080743627.
54. Hosang, J., Benenson, R., Dollar, P., Schiele, B.: What makes for effective detection
proposals? IEEE Transactions on Pattern Analysis and Machine Intelligence 38(4),
814–830 (April 2016), https://doi.org/10.1109/TPAMI.2015.2465908.
55. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate
object detection and semantic segmentation. In: 2014 IEEE Conference on Computer
Vision and Pattern Recognition. pp. 580–587. IEEE, Columbus, OH, USA (June 2014),
https://doi.org/10.1109/CVPR.2014.81.
56. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic seg-
mentation. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR). pp. 3431–3440. IEEE, Boston, MA, USA (June 2015), https://doi.org/10.1109/
CVPR.2015.7298965.
57. Girshick, R.: Fast R-CNN. In: 2015 IEEE International Conference on Computer Vision
(ICCV). pp. 1440–1448. IEEE, Santiago, Chile (Dec 2015), https://doi.org/10.1109/
ICCV.2015.169.
58. Brody, H.: Medical imaging. Nature 502, S81 (10/30 2013).
59. Shen, D., Wu, G., Suk, H.I.: Deep learning in medical image analysis. Annual review of
biomedical engineering 19, 221–248 (June 21 2017).
60. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural Networks 61,
85–117 (2015).
61. Wu, G., Kim, M., Wang, Q., Munsell, B.C., Shen, D.: Scalable high-performance image
registration framework by unsupervised deep feature representations learning. IEEE
Transactions on Biomedical Engineering 63(7), 1505–1516 (July 2016), https://doi.org
/10.1109/TBME.2015.2496253.
62. Wu, G., Kim, M., Wang, Q., Gao, Y., Liao, S., Shen, D.: Unsupervised deep feature
learning for deformable registration of MR brain images. In: Mori, K., Sakuma, I.,
Sato, Y., Barillot, C., Navab, N. (eds.) Medical Image Computing and Computer-
Assisted Intervention—MICCAI 2013. Lecture Notes in Computer Science, Vol. 8150,
pp. 649–656. Springer Berlin Heidelberg, Berlin, Heidelberg (2013).
63. Liao, S., Gao, Y., Oto, A., Shen, D.: Representation learning: A unifed deep learning
framework for automatic prostate MR segmentation. In: Mori, K., Sakuma, I., Sato,
Y., Barillot, C., Navab, N. (eds.) Medical Image Computing and Computer-Assisted
Intervention—MICCAI 2013. Lecture Notes in Computer Science, Vol. 8150, pp. 254–
261. Springer Berlin Heidelberg, Berlin, Heidelberg (2013).
64. Guo, Y., Gao, Y., Shen, D.: Deformable MR prostate segmentation via deep feature
learning and sparse patch matching. IEEE Transactions on Medical Imaging 35(4),
1077–1089 (April 2016), https://doi.org/10.1109/TMI.2015.2508280.
65. Kim, M., Wu, G., Shen, D.: Unsupervised deep learning for hippocampus segmentation
in 7.0 Tesla MR images. In: Wu, G., Zhang, D., Shen, D., Yan, P., Suzuki, K., Wang,
F. (eds.) Machine Learning in Medical Imaging. Lecture Notes in Computer Science,
Vol. 8184, pp. 1–8. Springer International Publishing, Cham (2013), https://doi.org/10.1
007/978–3-319–02267–3_1.
258 Deep Learning in Computer Vision

66. Abd-Ellah, M.K., Awad, A.I., Khalaf, A.A.M., Hamed, H.F.A.: Classifcation of brain
tumor MRIs using a kernel support vector machine. In: Li, H., Nykanen, P., Suomi,
R., Wickramasinghe, N., Widen, G., Zhan, M. (eds.) Building Sustainable Health
Ecosystems, WIS 2016. Communications in Computer and Information Science, Vol.
636. Springer, Cham, Tampere, Finland (2016).
67. Abd-Ellah, M.K., Awad, A.I., Khalaf, A.A.M., Hamed, H.F.A.: Design and implemen-
tation of a computer-aided diagnosis system for brain tumor classifcation. In: 2016
28th International Conference on Microelectronics (ICM). pp. 73–76. IEEE, Giza,
Egypt (2016).
68. Roth, H.R., Lu, L., Liu, J., Yao, J., Seff, A., Cherry, K., Kim, L., Summers, R.M.:
Improving computer-aided detection using convolutional neural networks and random
view aggregation. IEEE Transactions on Medical Imaging 35(5), 1170–1181 (May
2016), https://doi.org/10.1109/TMI.2015.2482920.
69. Ciompi, F., de Hoop, B., van Riel, S.J., Chung, K., Scholten, E.T., Oudkerk, M., de Jong,
P.A., Prokop, M., van Ginneken, B.: Automatic classifcation of pulmonary peri-fssural
nodules in computed tomography using an ensemble of 2D views and a convolutional
neural network out-of-the-box. Medical Image Analysis 26(1), 195–202 (September
2015), https://doi.org/10.1016/j.media.2015.08.001.
70. Gao, M., Bagci, U., Lu, L., Wu, A., Buty, M., Shin, H.C., Roth, H., Papadakis, G.Z.,
Depeursinge, A., Summers, R.M., Xu, Z., Mollura, D.J.: Holistic classifcation of CT
attenuation patterns for interstitial lung diseases via deep convolutional neural net-
works. Computer methods in biomechanics and biomedical engineering: Imaging &
Visualization 6(1), 1–6 (2016), https://doi.org/10.1080/21681163.2015.1124249.
71. Brosch, T., Tang, L.Y.W., Yoo, Y., Li, D.K.B., Traboulsee, A., Tam, R.: Deep 3D convo-
lutional encoder networks with shortcuts for multi-scale feature integration applied to
multiple sclerosis lesion segmentation. IEEE Transactions on Medical Imaging 35(5),
1229–1239 (May 2016), https://doi.org/10.1109/TMI.2016.2528821.
72. Dou, Q., Chen, H., Yu, L., Zhao, L., Qin, J., Wang, D., Mok, V.C., Shi, L., Heng, P.:
Automatic detection of cerebral microbleeds from MR images via 3D convolutional
neural networks. IEEE Transactions on Medical Imaging 35(5), 1182–1195 (May 2016),
10.1109/TMI.2016.2528129.
73. Abd-Ellah, M.K., Awad, A.I., Khalaf, A.A.M., Hamed, H.F.A.: A review on brain
tumor diagnosis from MRI images: Practical implications, key achievements, and les-
sons learned. Magnetic Resonance Imaging 61, 300–318 (2019), https://doi.org/10.1
016/j.mri.2019.05.028.
74. Ciresan, D.C., Giusti, A., Gambardella, L.M., Schmidhuber, J.: Mitosis detection in
breast cancer histology images with deep neural networks. In: Mori, K., Sakuma,
I., Sato, Y., Barillot, C., Navab, N. (eds.) Medical Image Computing and Computer-
Assisted Intervention—MICCAI 2013. pp. 411–418. Springer International Publishing,
Berlin, Heidelberg (2013), https://doi.org/10.1007/978–3-642–40763–5_51.
75. Kleesiek, J., Urban, G., Hubert, A., Schwarz, D., Maier-Hein, K., Bendszus, M., Biller,
A.: Deep MRI brain extraction: A 3D convolutional neural network for skull stripping.
NeuroImage 129, 460–469 (2016).
76. Moeskops, P., Viergever, M.A., Mendrik, A.M., de Vries, L.S., Benders, M.J.N.L.,
Isgum, I.: Automatic segmentation of MR brain images with a convolutional neural
network. IEEE Transactions on Medical Imaging 35(5), 1252–1261 May 2016).
77. Weisenfeld, N.I., Warfeld, S.K.: Automatic segmentation of newborn brain MRI.
NeuroImage 47(2), 564–572 (2009).
78. Zhang, W., Li, R., Deng, H., Wang, L., Lin, W., Ji, S., Shen, D.: Deep convolutional
neural networks for multi-modality isointense infant brain image segmentation.
NeuroImage 108, 214–224 (2015).
Deep Convolutional Neural Networks 259

79. Nie, D., Wang, L., Gao, Y., Shen, D.: Fully convolutional networks for multi-modal-
ity isointense infant brain image segmentation. In: 2016 IEEE 13th International
Symposium on Biomedical Imaging (ISBI). pp. 1342–1345 (April 2016).
80. Zhao, X., Wu, Y., Song, G., Li, Z., Fan, Y., Zhang, Y.: Brain tumor segmentation using
a fully convolutional neural network with conditional random felds. In: Crimi, A.,
Menze, B., Maier, O., Reyes, M., Winzeck, S., Handels, H. (eds.) Brainlesion: Glioma,
Multiple Sclerosis, Stroke and Traumatic Brain Injuries. Lecture Notes in Computer
Science, Vol. 10154, pp. 75–87. Springer International Publishing, Cham (2016).
81. Casamitjana, A., Puch, S., Aduriz, A., Vilaplana, V.: 3D convolutional neural networks
for brain tumor segmentation: A comparison of multi-resolution architectures. In: Crimi,
A., Menze, B., Maier, O., Reyes, M., Winzeck, S., Handels, H. (eds.) Brainlesion: Glioma,
Multiple Sclerosis, Stroke and Traumatic Brain Injuries. Lecture Notes in Computer
Science, Vol. 10154, pp. 150–161. Springer International Publishing, Cham (2016).
82. Pereira, S., Oliveira, A., Alves, V., Silva, C.A.: On hierarchical brain tumor segmenta-
tion in MRI using fully convolutional neural networks: A preliminary study. In: 2017
IEEE 5th Portuguese Meeting on Bioengineering (ENBENG). pp. 1–4 (Feb 2017).
83. Havaei, M., Davy, A., Warde-Farley, D., Biard, A., Courville, A., Bengio, Y., Pal, C.,
Jodoin, P.M., Larochelle, H.: Brain tumor segmentation with deep neural networks.
Medical Image Analysis 35, 18–31 (2017), https://doi.Org/10.1016/j.media.2016.05.004.
84. Wang, G., Li, W., Ourselin, S., Vercauteren, T.: Automatic brain tumor segmenta-
tion using convolutional neural networks with test-time augmentation. In: Crimi,
A., Bakas, S., Kuijf, H., Keyvan, F., Reyes, M., van Walsum, T. (eds.) Brainlesion:
Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. pp. 61–72. Springer
International Publishing, Cham (2019), https://doi.org/10.1007/978–3-030–11726–9_6.
85. Abd-Ellah, M.K., Khalaf, A.A.M., Awad, A.I., Hamed, H.F.A.: TPUAR-Net: Two
parallel U-Net with asymmetric residual-based deep convolutional neural network for
brain tumor segmentation. In: Karray, F., Campilho, A., Yu, A. (eds.) Image Analysis
and Recognition. ICIAR 2019. Lecture Notes in Computer Science, Vol. 11663, pp.
106–116. Springer International Publishing, Cham (2019).
86. The Cancer Imaging Archive: RIDER NEURO MRI database. https://wiki.canceri
magingarchive.net/display/Public/RIDER+NEURO+MRI (2016), Accessed: February
5th, 2019.
87. Tharwat, A.: Classifcation assessment methods. Applied Computing and Informatics
(2018), https://doi.org/10.1016/j.aci.2018.08.003.
88. Mohsen, H., El-Dahshan, E.S.A., El-Horbaty, E.S.M., Salem, A.B.M.: Classifcation
using deep learning neural networks for brain tumors. Future Computing and
Informatics Journal 3, 68–71 (2018).
 10 Lossless Full-Resolution
Deep Learning
Convolutional Networks
for Skin Lesion Boundary
Segmentation
Mohammed A. Al-masni, Mugahed A. Al-antari,
and Tae-Seong Kim

CONTENTS
10.1 Introduction .................................................................................................. 262
10.2 Related Works............................................................................................... 263
10.2.1 Conventional Segmentation Methods ............................................... 263
10.2.2 Initial Deep Learning Segmentation Methods .................................264
10.2.3 Recent Deep Learning Segmentation Methods................................ 265
10.2.4 Previous Deep Learning Works on Skin Lesion Segmentation .......266
10.3 Materials and Methods ................................................................................. 267
10.3.1 Dataset .............................................................................................. 267
10.3.2 Data Preprocessing and Augmentation ............................................ 267
10.3.3 Full-Resolution Convolutional Networks ......................................... 269
10.3.4 Training and Testing......................................................................... 272
10.3.5 Evaluation Metrics............................................................................ 274
10.3.6 Key Component Optimization.......................................................... 275
10.3.6.1 The Effect of Training Data............................................... 275
10.3.6.2 The Effect of Network Optimizers .................................... 276
10.4 Results of Skin Lesion Segmentation ........................................................... 277
10.4.1 Segmentation Performance on the ISIC 2017 Test Dataset .............. 277
10.4.2 Network Computation....................................................................... 277
10.4.3 Comparisons of Segmentators ..........................................................280
10.5 Discussion..................................................................................................... 282
10.6 Conclusions................................................................................................... 283
Acknowledgments.................................................................................................. 283
Bibliography .......................................................................................................... 283

261
262 Deep Learning in Computer Vision

10.1 INTRODUCTION
Skin cancer is one of the most widespread diagnosed cancers in the world. Indeed,
melanoma (i.e., malignant skin tumor) usually starts when melanocyte cells begin to
grow out of control [1]. Among different types of skin cancers, melanoma is the most
aggressive type due to its greater capability of spreading into other organs and its
higher death rate [2]. According to the annual report of the American Cancer Society
(ACS), in the United States in 2018, it is estimated that about 99,550 cases were
diagnosed as new cases of skin cancer (excluding basal and squamous cells skin
cancers), and the estimated deaths from this disease reached up to 13,460 cases [3].
Melanoma makes up the majority among various skin cancer types, with estimated
new cases and expected death cases of 91.7% and 69.2%, respectively. Moreover, it
is reported that melanoma is the most fatal skin cancer, with deaths from melanoma
making up 1.53% of total cancer deaths, and it represents 5.3% of all new cancer
cases [3, 4]. Regarding the survival statistics from the Surveillance, Epidemiology,
and End Results (SEER) Cancer Statistics Review (CSR) report [5], the survival rate
of patients diagnosed early with melanoma over fve years is about 91.8%. However,
this rate decreases to 63% when the disease spreads into lymph nodes. Therefore, a
skin lesion detected and correctly diagnosed in its earliest stage is highly curable,
and such detection reduces the mortality rate. This highlights the signifcance of
timely diagnosis and appropriate treatment of melanoma for patients’ survival.
Visual inspection via the naked eye during the medical examination of skin can-
cers is hindered by the similarity among normal tissues and skin lesions, which
may produce an incorrect diagnosis [6, 7]. In order to allow better visualization of
skin lesions and improve the diagnosis of melanoma, different non-invasive imag-
ing modalities have been developed. Dermoscopy (also known as dermatoscopy or
epiluminescence microscopy) has become a gold standard imaging technology that
assists dermatologists to improve the screening of skin lesions through visualizing
prominent features present under the skin surface. The key idea of dermoscopy is
that it acquires a magnifed high-resolution image while reducing or fltering out
skin surface refections, utilizing polarized light. In clinical practice, skin lesion
diagnosis is based on visual assessment of dermoscopy images. Although dermos-
copy images improve diagnostic precision, examination of dermoscopy images by
dermatologists via visual inspection is still tedious, time consuming, complex, sub-
jective, and fault prone [6, 7]. Hence, automated computerized diagnostic systems
for skin lesions are highly demanded to support and assist dermatologists in clinical
decision-making.
Indeed, automatic segmentation of skin lesion boundaries from surrounding
tissue is a key prerequisite procedure in any computerized diagnostic system of
skin cancer. Accurate segmentation of the lesion boundaries plays a critical role in
obtaining more prominent, specific, and representative features, which are utilized
to distinguish between different skin lesion diseases [8–11]. However, automatic seg-
mentation of skin lesions is still a challenging task due to their large variations in
size, shape, texture, color, and location in the dermoscopy images. Some challeng-
ing examples of skin lesion dermoscopy images are shown in Figure 10.1. These
examples show the presence of artifacts for screening: artifcial segmentation such as
Lossless Deep Learning FrCN 263

FIGURE 10.1 Examples of some challenging cases of skin lesions such as (a) low contrast,
(b) irregular fuzzy boundaries, (c) color illumination, (d) blood vessels, (e) ruler mark artifact,
(f) bubbles, (g) frame artifact, and (h) hair artifact. White line contours indicate the segmen-
tation of the lesions by expert dermatologists.

ruler marks, ebony frames, air bubbles, and color illumination, and natural artifacts
such as hair and blood vessels.
This chapter presents a novel deep learning method for skin lesion boundary seg-
mentation called full-resolution convolutional networks (FrCN) [12]; this method
produces very accurate segmentation. The FrCN method is a resolution-preserving
model that leads to learning high-level features and improved segmentation perfor-
mance. It directly learns the full-resolution attributes of each individual pixel of the
input image. This is achieved by removing all the subsampling layers from the net-
work. The evaluation of the effciency and effectiveness of the FrCN segmentation
method is presented using the well-established public International Skin Imaging
Collaboration (ISIC) 2017 challenge dataset. In addition, the performance of the
FrCN method is compared against well-known deep learning techniques such as
U-Net, SegNet, and fully convolutional network (FCN) methods under the same
experimental conditions.

10.2 RELATED WORKS

10.2.1 CONVENTIONAL SEGMENTATION METHODS
In general, conventional segmentation methods for the skin lesion boundaries can be
categorized into the following fve groups:

• Thresholding methods assign one or more appropriate threshold values

to distinguish the skin lesions from the surrounding normal tissues. Pixel
values greater than a particular threshold are placed as the foreground
class, while the rest are placed as the background class. Improved adap-
tive threshold techniques such as histogram-based and Otsu threshold have
been introduced to select optimal threshold value [ 7, 13–16].
• Unsupervised clustering techniques use the color space features from
RGB dermoscopy images to acquire homogenous regions of the skin
264 Deep Learning in Computer Vision

lesions [17–20]. Cluster analysis methods include a connectivity model

such as distance clustering or a centroid model such as the k-means
clustering.
• Region-based methods utilize the region splitting or merging algorithms
to generate groups of homogenous pixels. However, edge-based methods
use the edge operator to provide edges between different regions [7, 21–23].
• Active contour methods (e.g., snakes active contours) exploit the evolution
curve technique to segment skin lesions [7, 24, 25].
• Supervised segmentation approaches segment the skin lesion boundaries
by learning multiple features and representations. Support vector machines
(SVMs), decision trees (DTs), and artifcial neural networks (ANN) are
examples of commonly utilized classifers [20, 26].

For further details of these traditional techniques, the latest comprehensive reviews
provide more details of the skin lesion boundary segmentation algorithms [8, 27–30].
All of these methods utilize low-level features that rely only on pixel-level attributes.
Hence, these classical segmentation techniques still do not deliver satisfactory per-
formance and cannot get over the challenges of hair artifacts and low contrast.

10.2.2 INITIAL DEEP LEARNING SEGMENTATION METHODS

The great success of deep learning approaches in supervised convolutional neu-
ral networks (CNNs) for object detection and image classifcation have motivated
researchers to develop models for the pixel-level segmentation problem. Therefore,
segmentation methods employing deep learning techniques have gotten major atten-
tion in the felds of object localization, semantic segmentation, and image classifca-
tion [31–37]. Deep leaning based on CNN is a powerful approach since it has the
ability to extract more prominent features or representations from a whole image
rather than hand-crafted features. Hence, it improves the performance of challeng-
ing segmentation tasks [38–40]. The initial CNN method for segmentation problems
was introduced by Ciresan et al. in 2012 [41]. The aim was to provide full automa-
tion segmentation of the biological neuron membranes while eliminating the user
intervention in the stacks of electron microscopy images. This method performed
pixel-wise segmentation, which determined if each predicted pixel does or does not
belong to the foreground object (i.e., binary classifcation). The process works frst
by dividing the input image into smaller regions of interest (ROIs) or patches and
then passing them into a CNN classifer. The prediction of each sliding window
represented the segmentation of the center pixel in the corresponding patch. At the
end of this process, each pixel of the input image was segmented as membrane or
non-membrane pixel. Similar work was proposed by Cernazanu-Glavan and Holban
in order to segment bone pixels from chest X-ray images [42]. Due to the large size
of the images, the segmentation process was performed only for some rib regions.
Melinščak et al. presented a comparable CNN technique to segment retinal vessels
as a binary classifcation task [43]. The prediction of each ROI determined the seg-
mentation of its central pixel. This technique still needs further enhancement since
the deep features extracted from each patch do not exactly refect the attributes of the
Lossless Deep Learning FrCN 265

center pixel. In addition, it requires complex computation and a lot of execution time,
since the network should be processed separately for each patch.

10.2.3 RECENT DEEP LEARNING SEGMENTATION METHODS

Lately, novel approaches utilizing deep CNN methods to improve the performance
of challenging segmentation tasks have been under development. The frst end-to-
end pixel-wise semantic segmentation approach was developed by Long et al. [44,
45]. Their segmentation method was inspired from the conventional CNN recog-
nizer, in which the authors adapted well-known contemporary classifcation models
such as GoogLeNet, AlexNet, and Visual Geometry Group (VGG)Net into their
CNN by replacing the fully connected neural network (FC-NN) layers with con-
volutional layers to produce spatial maps instead of prediction scores. Their model
was named fully convolutional networks (FCN). FCN is robust and has the capabil-
ity of yielding hierarchies of attributes. However, the subsampling operations after
each convolutional layer cause the loss of the spatial resolution of the input pixels.
Due to this, upsampling using deconvolution and bilinear interpolation process was
applied in the last convolutional layer in order to produce a map that contains dense
prediction with the same size as the original input image. FCN produces a signif-
cant enhancement of the pattern analysis, statistical modeling, and computational
learning in the visual object classes from the PASCAL VOC dataset as compared to
traditional techniques [46].
Another architecture, called U-Net, was developed to convert the CNN classif-
cation models to a suitable segmentation network [47]. U-Net was an extension and
amendment of the FCN architecture [45], in which it performed well with fewer
training images and produced more accurate segmentations on the stacks of elec-
tron microscopy images. In fact, U-Net consisted of two consecutive and symmet-
ric paths generating a U-shaped model. The contracting path, which is typical in
the FCN architecture, is responsible for extracting the relevant representations and
patterns using the convolution and subsampling operations. In the expansive path,
the upsampling process followed by a convolution operator was applied opposite
to the contracting path to produce some large number of feature channels. In each
convolutional layer, the feature map in the contracting path was concatenated with
the corresponding features in the expansive path in order to provide high-resolution
features. Similar to the FCN architecture, U-Net did not utilize fully connected lay-
ers; instead, it used the dense maps from the convolutional networks to segment the
pixels of the images. Apart from the FCN technique, a deep convolutional encoder-
decoder architecture for image segmentation was proposed by Badrinarayanan et
al. in 2015 and was named SegNet [48, 49]. The SegNet architecture consisted of
an encoder network that produces low-resolution feature maps, a corresponding
decoder network that transforms the features into a full resolution of the input using
a combination of convolution and upsampling layers, and a softmax pixel-wise clas-
sifcation layer. The encoder network used the VGG16 classifcation model, which
involves 13 convolutional layers. However, the same layers were mirrored in the
decoder network. The novelty of this segmentation method lies in the way that the
decoder upsamples the lower-resolution feature maps. Each upsampling layer in the
266 Deep Learning in Computer Vision

decoder network matched to the corresponding subsampling layer in the encoder

network, preserving the locations of the feature maps in an input image during the
upsampling process. The softmax classifer was utilized to generate predictions for
the input pixels, which represented the segmentation map.
Noh et al. evolved a deep learning semantic segmentation technique called
Deconvolution Network (DeconvNet) as an extension of FCN [50]. Analogous to
the SegNet architecture, DeconvNet contained convolution networks that extracted
representations from the input image and deconvolution networks that retrieved the
feature maps into higher resolution as the dimension of the input image. DeconvNet
was also a shape generator that produced a probability map of all input pixels and
had a higher spatial resolution compared to that generated by FCN. Thus, the fnal
output of the DeconvNet is the segmentation masks of the objects in the input
images. Furthermore, an image segmentation technique via deep convolutional nets
called DeepLab was introduced by Chen et al. [51, 52]. This method was proposed
to provide a solution to some of the challenges faced in the semantic segmentation
models. First, the authors replaced the last few subsampling layers in the network
by Atrous convolution, which is a dilated convolution that extends the flter size and
flls the hole locations with zeros to maintain the feature resolution. Second, they
used the Atrous spatial pyramid pooling for multi-scale image representations. Then,
they appended the DeepLab network with the fully connected conditional random
feld (CRF) to recover more precise boundaries of the segmented objects. Arguably,
all of the aforementioned segmentation methods take the well-known classifcation
networks and eliminate their FC-NN layers. Along with the encoder network, which
extracts low-resolution feature maps, these methods build a corresponding decoder
network to compensate for the loss of features and generate dense pixel-wise predic-
tions. More details of these segmentation architectures are explained in a review
paper [46].

10.2.4 PREVIOUS DEEP LEARNING WORKS ON SKIN LESION SEGMENTATION

Recently, these deep learning approaches such as FCN, U-Net, and SegNet have
been stratifed in the skin lesion segmentation feld. In 2018, Li and Shen proposed
a framework named the lesion indexing network (LIN) consisting of multi-scale
fully convolutional residual networks (FCRN) and the lesion index calculation unit
(LICU) [53]. Authors utilized the cropped center area of the skin lesion images from
the ISIC 2017 challenge dataset. The performance of the LIN method was evaluated
on the ISIC 2017 validation set: it achieved overall accuracy and Jaccard indices
(JACs) of 95.0% and 75.3%, respectively. Bi et al. presented a cascaded multi-phase
FCN followed by the parallel integration (PI) technique to segment lesion boundar-
ies from dermoscopy images [54]. The PI technique was used for further refne-
ment of the boundaries of the segmented lesions. They evaluated their segmentation
approach utilizing two public datasets: the PH2 and ISIC 2016 skin lesion chal-
lenge datasets. Their method achieved overall segmentation accuracies of 94.24%
and 95.51% and dice coeffcient indices of 90.66% and 91.18% from the two datas-
ets. Yu et al. built a two-stage approach for segmentation and classifcation of skin
lesions using deep residual networks [55]. Their FCRN was ranked second in the
Lossless Deep Learning FrCN 267

segmentation challenge with an overall accuracy of 94.9% when using the ISIC 2016
challenge dataset.
Yuan et al. developed a skin lesion segmentation technique via deep FCN [56].
In order to handle the imbalance among lesion-tissue pixels, the authors extended
the well-known FCN approach by employing the Jaccard distance as a loss func-
tion. Their technique achieved a dice index of 93.8% and an overall segmentation
accuracy of 95.5% with the PH2 and ISIC 2016 datasets, respectively. Goyal and Yap
presented a multi-class semantic segmentation via FCN, which was capable of seg-
menting three classes of skin lesions (i.e., melanoma, seborrheic keratoses [SK], and
benign nevi) from the ISIC 2017 challenge dataset [57]. Their method achieved dice
metrics of 65.3%, 55.7%, and 78.5% for melanoma, seborrheic keratosis, and benign
lesions, respectively. Lin et al. proposed a comparison between two skin lesion seg-
mentation methods, C-means clustering and U-Net-based histogram equalization
[58]. They evaluated their methodologies utilizing the ISIC 2017 challenge dataset.
The U-Net method obtained a dice coeffcient measure of 77.0%, which signifcantly
outperformed the clustering method’s result of only 61.0%. In 2017, Yading Yuan
developed a skin lesion segmentation utilizing deep convolutional-deconvolutional
neural networks (CDNN) [59]. The CDNN model was trained with various color
spaces of dermoscopy images using the ISIC 2017 dataset. CDNN approach was
ranked frst in the ISIC 2017 challenge, in which it obtained a Jaccard index (JAC)
of 76.5%.

10.3 MATERIALS AND METHODS

10.3.1 DATASET
In this study, a well-known and public dermoscopy dataset, called the IEEE
International Symposium on Biomedical Imaging (ISBI) 2017 segmentation chal-
lenge dataset, was used with the FrCN segmentation technique. This dataset is pro-
vided by the ISIC archive [60] and is available at [61]. The ISIC 2017 challenge
dataset includes 8-bit RGB dermoscopy images with various image sizes ranging
from 540 × 722 to 4,499 × 6,748 pixels. This dataset consists of 2,000 training images
along with validation and testing datasets of 150 and 600 images, respectively. All
of these dermoscopy images include skin lesions, which are labeled by expert der-
matologist as benign nevi, seborrheic keratosis (SK), or melanoma. Table 10.1 sum-
marizes the distribution of the dataset. Moreover, this dataset provided the original
dermoscopy images paired with their ground-truth segmented lesion boundaries in
the form of binary masks, which are manually annotated by expert dermatologists.
Figure 10.2 illustrates an example of these dermoscopy images from the ISIC 2017
challenge dataset with their ground-truth segmentation masks.

10.3.2 DATA PREPROCESSING AND AUGMENTATION

A normalization process is applied to each channel of the RGB images to facilitate
and increase the convergence speed of the segmentation method. Therefore, each
channel is normalized in a range from 0 to 1. Due to the differences in the image
268 Deep Learning in Computer Vision

TABLE 10.1
Distribution of the ISIC 2017 Challenge Dataset
Data Type Benign SK Melanoma Total

Training Data 1,372 254 374 2,000

Validation Data 78 42 30 150
Test Data 393 90 117 600
Total 1,843 386 521 2,750

FIGURE 10.2 (a) and (b) are an exemplary pair of the original dermoscopy image and its
segmentation mask, which was annotated by expert dermatologists. (c) illustrates the seg-
mented lesion boundaries.

sizes of the ISIC 2017 dataset (i.e., the majority ratio of height to width of 3:4), we
resized all of the dermoscopy images into 192 × 256 pixels utilizing bilinear interpo-
lation according to the report in [56], where the optimal segmentation achievement
was made with the image size of 192 × 256 among other different image sizes. In
fact, deep learning methods require a large amount of data for appropriate train-
ing. However, a limit on the sizes of medical image datasets, especially a limit on
reliable annotated ground-truths, is one of the challenges to adopting deep learning
approaches. Due to this, data augmentation was applied to enlarge the training data-
set. Augmentation is a process that generates augmented images utilizing the given
original images using various transformation techniques such as rotation, scaling,
and translation [62–64].
By taking the merit of the dermoscopy images (i.e., RGB color), the information
from three channels—hue-saturation-value (HSV)—could be derived in addition to
the original RGB images, producing more color space features. The augmented data-
set includes the RGB and HSV dermoscopy images and their rotated images with
angles of 0˚, 90˚, 180˚, and 270˚. Additionally, horizontal and vertical fipping was
applied. Thus, a total of 32,000 [i.e., (2,000 RGB + 2,000 HSV) × 4 rotations × 2 fip-
ping] images with their corresponding ground-truth binary masks were utilized to
train the FrCN segmentation technique. Figure 10.3 shows a sample of the original
RGB images in addition to the 15 augmented images. In fact, the data augmentation
process reduces the overftting problem of deep learning and improves the robust-
ness of the deep network.
Lossless Deep Learning FrCN 269

FIGURE 10.3 The augmentation processes for the training images. Top two rows represent
RGB images, while the bottom two rows are for HSV images. From each original RGB image
(top left corner), we generate a total of 16 different images including four rotations and four
horizontal and vertical fippings on the original RGB and HSV images.

10.3.3 FULL-RESOLUTION CONVOLUTIONAL NETWORKS

In general, CNN contains two essential parts to obtain pixel-wise classifcation for
segmentation: the frst includes convolutional and subsampling layers and the sec-
ond includes upsampling layers. Convolutional and subsampling layers represent the
frst part of these networks. The convolutional layers are in charge of extracting
deep feature representations from the input image utilizing different flter types of
various sizes. The subsampling layers are utilized to reduce the dimensionality of
the extracted feature maps, but they also reduce the resolution of the feature maps
[65, 66]. In image-wise classifcation tasks (i.e., image classifcation via CNNs), the
270 Deep Learning in Computer Vision

pooling or subsampling layer does promote the robustness of the classifer, since it
reduces overftting, eliminates the redundancy of features, and minimizes computa-
tion time [45, 54]. This is because all the reduced attributes represent a single class
label of the input image. However, in the pixel-wise segmentation tasks, the subsam-
pling layers cause loss of spatial resolution in the features of input images.
The second part of the segmentation networks is the upsampling layers followed
by the softmax classifer, which exploits the extracted features of each individual
pixel to distinguish it as a tissue or a lesion pixel. Recent deep learning segmentation
approaches have used complicated procedures to compensate for the missing fea-
tures of particular pixels, due to the reduced image size, via upsampling, deconvolu-
tion, bilinear interpolation, decoding, or Atrous convolution, increasing the number
of hyper-parameters [45, 47, 48, 50, 51]. Furthermore, these mechanisms require
further PI or CRF operations to refne boundaries of the segmented objects [51, 54].
These deep learning segmentation networks suffer from reduced spatial resolution
and loss of details. Because pixel-wise dependency is not adequately addressed
throughout the previous deep learning segmentation methods such as FCN, U-Net,
and SegNet, important information about some pixels is left out and is not easy to
retrieve.
The full-resolution convolutional networks (FrCN) [12] is an end-to-end super-
vised deep network that is trained via mapping the entire input image to its corre-
sponding ground-truth masks with no resolution loss, leading to better segmentation
performance of skin lesion boundaries. In fact, the FrCN method is a resolution-
preserving model that leads to high-level feature learning and improves the segmen-
tation performance. The full-resolution feature maps of the input pixels are reserved
by removing all the subsampling layers in the architecture. Figure 10.4 illustrates the
architecture of the FrCN segmentation method. The FrCN allows each pixel in the
input dermoscopy image to extract its own features utilizing the convolutional layers.

FIGURE 10.4 An illustration of the FrCN segmentation method. There are no subsampling
or fully connected layers.
Lossless Deep Learning FrCN 271

The FrCN architecture contains 16 convolutional layers, which are inspired by the
VGG-16 network [67].
It has multiple stacks of convolutional layers with different sizes of feature maps,
where each stack learns distinctive representations and causes an increase over the
network complexity. The last three FC-NN layers in Block-6 were substituted with
the convolutional layers to allow the network to process input images with arbitrary
sizes as well as to generate pixel-wise maps of dense prediction with the same sizes
as the input images. Details of the number of maps and the convolutional flter sizes
in each layer are given in Table 10.2. The early layers of FrCN learn localization
information, while the late layers learn subtle and fne feature representations of skin
lesion boundaries.
A convolutional network is originally translation invariant, in which its main
role is to extract different features from the entire dermoscopy images and then
produce feature maps using convolution operations [68]. In the deep learning net-
work, the flter kernels of the convolutional layers are represented by weights,
which are automatically updated during the training process. The basic compo-
nents of this convolutional network are the flter kernels W, convolution operator *,
and activation function ϕ(·). Therefore, the kth feature map of layer L is computed
as follows:

( )
FLk = f WLk * FLk−1 + bLk , (10.1)

where bLk is the bias applied to each feature map for each layer.
A nonlinear activation process gets used via rectifed linear units (ReLUs)
instantly after each convolutional layer to provide nonlinearity to the network. A
ReLU activation function is widely used rather than the traditional sigmoid and tanh
functions due to its capability to alleviate the vanishing gradient problem and to

TABLE 10.2
Details of Architecture Layers of the FrCN Method
Filter Size,
Block Layer Filter Size, Maps Block Layer Maps

Input data 192 × 256 × 3 – Block-5 Conv. 11 3 × 3, 512

Block-1 Conv. 1 3 × 3, 64 Conv. 12 3 × 3, 512
Conv. 2 3 × 3, 64 Conv. 13 3 × 3, 512
Block-2 Conv. 3 3 × 3, 128 Block-6 Conv. 14 7 × 7, 4096
Conv. 4 3 × 3, 128 Dropout 1 p = 0.5
Block-3 Conv. 5 3 × 3, 256 Conv. 15 1 × 1, 4096
Conv. 6 3 × 3, 256 Dropout 2 p = 0.5
Conv. 7 3 × 3, 256 Conv. 16 1 × 1, 2
Block-4 Conv. 8 3 × 3, 512 Softmax –
Conv. 9 3 × 3, 512 Output map 192 × 256 × 3 –
Conv. 10 3 × 3, 512
272 Deep Learning in Computer Vision

train the network with higher computational effciency [69, 70]. The ReLU activation
function is defned as follows:

˙ x, if x ˛ 0
˜( x ) = max(0, x ) = ˆ (10.2)
ˇ0, if x < 0.

In addition, a dropout is used with p = 0.5 after the convolutional layers Conv14 and
Conv15, as shown in the FrCN architecture in Figure 10.4 and Table 10.2. Dropout
is a process used in deep learning to tackle the overftting problem of deep layers
[71]. During training, it randomly eliminates some neural units with their connec-
tions to prevent overftting to the training data. Hence, every unit is retrained with a
particular probability, leading the outgoing weights of that unit to be multiplied by
p at the test time.
Finally, a softmax classifer known as multinomial logistic regression is fed in the
last layer of the FrCN architecture to classify each pixel in the dermoscopy image
into two binary classes (i.e., lesion and non-lesion). This logistic regression function
generates a map of predictions for each pixel, which involves the segmented map. As
a result, the resolutions of the spatial output maps are same as that of the input data.
Cross-entropy is used as a loss function, in which the overall loss H of each pixel is
minimized throughout the training process as defned by

H = −(y ˛ log(ŷ) + (1 − y)log(1 − ŷ)), (10.3)

where y and ŷ are the ground-truth delineation and predicted segmented map,
respectively. The cross-entropy loss function is utilized when a deep convolutional
network is applied for a pixel-wise recognition [45, 56].

10.3.4 TRAINING AND TESTING

Biomedical deep learning applications still suffer from the lack of medical data [62].
One reason is the cost and complexity of the labeling procedure during data acqui-
sition [72]. To overcome the challenge of insuffcient training data in the feld of
skin lesion dermoscopy images, we exploited two mechanisms, data augmentation
as described above and transfer learning [54, 57, 62, 72–75]. For transfer learning,
we utilized the pre-training network weights from the trained VGG-16 [67] using
the public ImageNet database [76]. Subsequently, the deep FrCN segmentation tech-
nique was fne-tuned or retrained using the augmented training data of the dermos-
copy images. Regarding the optimization of the FrCN, a separate validation dataset
consisting of 150 dermoscopy images was utilized. Therefore, the evaluations of
the optimization network and the fnal network performance were performed inde-
pendently utilizing the validation and test datasets, respectively. This procedure is
called an unbiased double cross-validation scheme, and it is more commonly used
and reliable than the trial-and-error process in the most recent deep learning meth-
ods [77, 78].
To ensure generalization of deep learning architectures and avoid overftting to
training data, deep learning networks must be trained, validated, and tested using
Lossless Deep Learning FrCN 273

separate datasets as shown in Figure 10.5. A training dataset (i.e., 72.7% of the whole
dataset) is used to train the deep learning network with hyper-parameters, which are
evaluated utilizing the validation dataset (5.5% of the whole dataset). The optimal
deep learning model is selected according to its high effciency on the validation set.
Then, further evaluation is performed on the test dataset (21.8% of the whole dataset)
to obtain the overall performance of the network [65, 79].
The objective of training deep learning approaches is to optimize the weight
parameters in each layer. The optimization procedure of a single cycle performs
as follows [38, 80]. First, the forward propagation pass sequentially computes the
output in each layer utilizing the training data. In the last output layer, the error
between the ground-truth and predicted labels gets computed using the loss function.
To reduce the training error, back-propagation is proceeded through the network lay-
ers. Consequently, the training weights of FrCN get updated using the training data.
Furthermore, the performance of FrCN is compared against the performance of the
recent deep learning approaches such as FCN, U-Net, and SegNet utilizing the test
subset of 600 dermoscopy images from the ISIC 2017 dataset. All the segmentation
models were trained using the superior Adadelta optimization method (details in
Section 10.3.6.2) with a batch size of 20. The learning rate was initially set as 0.2 and
then reduced with automated updating throughout the training process. The training
network tends to have convergence at about the 200th epoch.
The system implementations of all experiments were performed on a personal
computer (PC) with the following specifcations: a CPU of Intel® Core(TM) i7-6850
K @ 3.360 GHz with 16 GB RAM and a GPU of NVIDIA GeForce GTX 1080. This
work was conducted with Python 2.7.14 on Ubuntu 16.04 OS using the Keras and
Theano DL libraries [81, 82].

FIGURE 10.5 Functionality scheme of training, validation, and testing.

274 Deep Learning in Computer Vision

10.3.5 EVALUATION METRICS

To quantitatively evaluate the effciency and capability of the deep FrCN method
on skin lesion boundary segmentation, the following statistical evaluation mea-
sures were utilized. Sensitivity (a.k.a., recall or true positive rate) measures the
ratio of the skin lesion pixels that are correctly segmented, while the specifcity
(a.k.a., the true negative rate) indicates the proportion of the skin non-lesion pix-
els that are not correctly segmented [57, 83, 84]. The dice coeffcient, also known
as the F1-score, was also utilized to quantify the effciency of the FrCN segmen-
tation method. This measures the similarity between the segmented lesions and
the delineated ground-truths [85, 86]. The Jaccard index (JAC) is an intersection
over union of the ground-truth masks with the segmented lesions [87]. Accuracy
is also provided to show the overall pixel-wise segmentation performance. The
Matthew correlation coeffcient (MCC) measures the correlation between the seg-
mented and annotated skin lesion pixels and produces values in a range from −1
to +1 [87]. A higher MCC value implies a better segmentation performance. All
of these metrics were computed in relation to the elements of the confusion matrix
as follows.

TP
Sensitivity (SEN) = , (10.4)
TP + FN

TN
Specificity (SPE) = , (10.5)
TN + FP

2 ° TP
Dice Index (DIC) = , (10.6)
(2 ° TP ) + FP + FN
TP
Jaccard Index (JAC) = , (10.7)
TP + FN + FP

TP + TN
Accuracy (ACC) = , (10.8)
TP + FN + TN + FP

TP ° TN − FP ° FN
MCC = , (10.9)
( TP + FP )( TP + FN )( TN + FP )( TN + FN )
where TP and FP denote the true and false positives, while TN and FN indicate the
true and false negatives, respectively. If the lesion pixels were segmented correctly,
they were considered as TPs; otherwise, they were FNs. On the contrary, the non-
lesion pixels were considered as TNs if their segmentation was classifed correctly
as non-lesion; otherwise, they were FPs. An illustration of the sensitivity and speci-
fcity concept with a dermoscopy image for classifcation is shown in Figure 10.6.
Moreover, the curve of the receiver operator characteristic (ROC) with its area under
the curve (AUC) was used for further segmentation evaluation.
Lossless Deep Learning FrCN 275

FIGURE 10.6 Concepts of sensitivity and specifcity in terms of TP, FN, FP, and TN pixels.
The white contour indicates the ground-truth of the lesion boundary annotation by a derma-
tologist, while the red one is an exemplary segmented boundary.

10.3.6 KEY COMPONENT OPTIMIZATION

This section presents the effect of some key components of the performance of the
FrCN method over the skin lesion segmentation to develop a robust deep model
capable of overcoming the overftting and speeding up its convergence during the
training process. These components involve data augmentation and network optimi-
zation. For each component, some experiments have been performed with different
alternative settings, while the other components are fxed. The modifed models of
all these experiments are trained using the training data of the ISIC 2017 dataset,
while the models are evaluated using only the validation data (i.e., 150 dermoscopy
images) of the same dataset.

10.3.6.1 The Effect of Training Data

This section presents the effect of training the deep FrCN with different training
dataset sizes (i.e., original versus augmented training datasets). The original training
dataset consists of 2,000 dermoscopy images, while the augmented dataset contains
32,000 images. In general, the performance of deep learning approaches improves
as the size of the training dataset increases. The effect of training data size on the
FrCN segmentation method is shown in Table 10.3. The FrCN was frst trained with
the original training images (i.e., without augmentation), which provided a dice
coeffcient index of 79.59% on the validation dataset. Subsequently, the network
was trained with the augmented training dataset, which yielded an improvement
of 8.15% in terms of dice coeffcient index. The results show an improvement in
terms of Jaccard index in the case of the augmented dataset against the original,
with 73.58% and 63.36%, respectively. Significantly, the augmented data also affects
276 Deep Learning in Computer Vision

TABLE 10.3
Segmentation Performance with Different Optimization Components
Performance Evaluation (%)

Component Strategy SEN SPE DIC JAC MCC ACC

Data Augmentation Without Aug. 73.40 92.07 79.59 63.36 72.88 88.81
With Aug. 81.20 94.15 87.74 73.58 82.89 91.89
Optimizer SGD 74.05 96.44 77.57 63.36 73.22 92.53
Adam 78.60 97.35 82.23 69.83 78.81 94.08
Adadelta 81.20 94.15 87.74 73.58 82.89 91.89

the sensitivity, which refects the ability of the network to segment the skin lesion
correctly, with an improvement of 7.80% as reported in Table 10.3. Due to this, the
overall segmentation accuracy is increased from 88.81% to 91.89%. These results
ensure that training deep learning models with a larger training data achieves better
performance and provides more feasible and reliable models. The overftting prob-
lem is clearly shown in Figure 10.7 (a), which presents how the loss function diverges
over epochs on the validation dataset when the network is trained with the original
training dataset without data augmentation. In contrast, the loss function converged
rapidly in training the network with the larger augmented dataset.

10.3.6.2 The Effect of Network Optimizers

In this section, the Adadelta [88] optimization method was tested against Adam [89]
and Stochastic Gradient Descent (SGD) [90] optimizers. The learning rate of the
Adam and SGD optimizers was set as 0.003. Table 10.3 shows the statistical evalua-
tion of these three optimizers with various measures. Adadelta showed the superior
performance over the skin lesion segmentation against Adam and SGD methods,

FIGURE 10.7 Optimization performance in terms of loss function. Comparison between

the training and validation loss functions (a) with and without data augmentation, and (b)
using Adadelta, Adam, and SGD optimization methods.
Lossless Deep Learning FrCN 277

with overall Jaccard indices of 73.58%, 69.83%, and 63.36% on the validation
dataset, respectively. Furthermore, Figure 10.7 (b) illustrates how the loss function
declined during the training process on the training datasets. This fgure provides
an indication of the convergence speed of the network under training over epochs.
Clearly, the training and validation curves of Adadelta are signifcantly better than
those of Adam and SGD. In addition, these curves clarify the network performance
on the validation data, in which a small variation between them could give the FrCN
the capability to perform well with unseen test data. Although the loss function of
the training data using the Adam optimizer decreases rapidly over epochs, the loss
for the validation data diverges, leading reduced performance for skin lesion bound-
ary segmentation compared to Adadelta. However, the SGD optimizer provides the
worst performance with this dataset, as illustrated in Figure 10.7 (b).

10.4 RESULTS OF SKIN LESION SEGMENTATION

10.4.1 SEGMENTATION PERFORMANCE ON THE ISIC 2017 TEST DATASET
This section presents the skin lesion segmentation performance of the FrCN against
the latest deep learning methods, utilizing 600 dermoscopy images from the ISIC
2017 test dataset. Quantitatively, Tables 10.4 and 10.5 summarize the segmentation
performance results of the FrCN compared to FCN, U-Net, and SegNet. In this
work, the segmentation of FCN was performed utilizing a stride of 32. Table 10.4
explores the performances in terms of specifcity, sensitivity, and overall pixel-wise
segmentation accuracy, while the measurements in Table 10.5 are reported in terms
of the dice, Jaccard, and Matthew correlation indices for benign, SK, melanoma, and
overall cases. The FrCN method outperformed the FCN, U-Net, and SegNet, with
overall dice, Jaccard, and accuracy indices of 87.08%, 77.11%, and 94.03%, respec-
tively. The highest specifcity for overall skin lesion segmentation was achieved by
U-Net, with 97.24%.
An analysis of the segmentation performance on the diagnostic categories of the
skin lesions is also provided. Measurements of each diagnostic class including benign,
seborrheic keratosis (SK), and melanoma cases are shown in Tables 10.4 and 10.5.
The segmentation accuracies of the benign, SK, and melanoma cases in the ISIC 2017
test dataset for the FrCN were 95.62%, 91.29%, and 90.78%, respectively. Finally, the
ROC curves with their AUCs of the FrCN versus FCN, U-Net, and SegNet are illus-
trated in Figure 10.8 separately for each diagnostic class. The overall AUCs of FCN,
U-Net, SegNet, and FrCN were 88.32%, 82.19%, 87.71%, and 91.04%, respectively. For
qualitative evaluation, Figure 10.9 (a), (b), and (c) present some exemplar segmentation
results of the FrCN compared to the ground-truth contours for the benign, SK, and
melanoma cases from the ISIC 2017 test dataset. The relevant bars in Figure 10.9 indi-
cate the sensitivity, Jaccard index, and accuracy for each image. Obviously, the FrCN
segmentation method provides promising results as compared against other methods.

10.4.2 NETWORK COMPUTATION

The computation time for the training process was about 17.5 hours over 200
epochs. However, the decoding (i.e., inference time) for a single dermoscopy
 278

TABLE 10.4
Segmentation Performance (%) of the FrCN Compared to FCN, U-Net, and SegNet in Terms
of Sensitivity, Specifcity, and Accuracy for Benign, SK, Melanoma, and Overall Cases
Benign Cases SK Cases Melanoma Cases Overall

Method SEN SPE ACC SEN SPE ACC SEN SPE ACC SEN SPE ACC

FCN 85.25 96.95 94.45 75.10 95.91 90.96 70.67 96.18 88.25 79.98 96.66 92.72
U-Net 76.76 97.26 92.89 43.81 97.64 84.83 58.71 96.81 84.98 67.15 97.24 90.14
SegNet 85.19 96.30 93.93 70.58 92.50 87.29 73.78 94.26 87.90 80.05 95.37 91.76
FrCN 88.95 97.44 95.62 82.37 94.08 91.29 78.91 96.04 90.78 85.40 96.69 94.03

TABLE 10.5
Segmentation Performance (%) of the FrCN Compared to FCN, U-Net, and SegNet in Terms
of Dice, Jaccard, and MCC Indices for Benign, SK, Melanoma, and Overall Cases
Benign Cases SK Cases Melanoma Cases Overall

Method DIC JAC MCC DIC JAC MCC DIC JAC MCC DIC JAC MCC
FCN 86.77 76.63 83.28 79.81 66.40 74.26 78.89 65.14 71.84 83.83 72.17 79.30
U-Net 82.16 69.72 78.05 57.88 40.73 53.89 70.82 54.83 63.71 76.27 61.64 71.23
SegNet 85.69 74.97 81.84 72.54 56.91 64.32 79.11 65.45 71.03 82.09 69.63 76.79
FrCN 89.68 81.28 86.90 81.83 69.25 76.11 84.02 72.44 77.90 87.08 77.11 83.22
Deep Learning in Computer Vision
Lossless Deep Learning FrCN 279

FIGURE 10.8 ROC curves of different deep learning methods for skin lesion boundary seg-
mentation on the ISIC 2017 test dataset for (a) benign, (b) SK, (c) melanoma, and (d) overall
clinical cases.

image took about 9.7 seconds. This should make the FrCN segmentation approach
applicable for clinical practice. In addition, Table 10.6 shows the trainable param-
eters along with the training computation time per epoch and the test time per
single dermoscopy image for each of the segmentation techniques. Clearly, the
FrCN approach seems feasible for medical practices, since less than 10 seconds of
inference time is required to segment the suspicious lesion from the dermoscopy
image. The computational speed during the training process of the FrCN approach
was faster than other segmentation techniques. Table 10.6 shows the speed of the
FrCN approach in the training process time compared to the FCN model (i.e.,
from 651 to 315 seconds per epoch). Indeed, this is because FrCN was executed
without the need for the downsampling processes, but keeping the same convolu-
tion operations. In summary, the deep FrCN segmentation method outperforms
the most popular deep learning methods of FCN, U-Net, and SegNet on the ISIC
2017 datasets.
280 Deep Learning in Computer Vision

FIGURE 10.9 The lesion contours segmentation with statistical evaluation per each image
of the FrCN against FCN, U-Net, and SegNet for (a) benign, (b) seborrheic keratosis, and (c)
melanoma cases from the ISIC 2017 dataset. Segmentation results are depicted as follows:
ground-truth, white; FrCN, blue; FCN, green; U-Net, red; and SegNet, black.

TABLE 10.6
Measurements in Seconds of the Training Time
per Epoch and Test Time per Dermoscopy Image
Trainable Training Time/ Test Time/
Method Parameters Epoch Single Image

FCN 134.3 M 651 10.6

U-Net 12.3 M 577 10.9
SegNet 11.5 M 543 11.13
FrCN 16.3 M 315 9.7

10.4.3 COMPARISONS OF SEGMENTATORS

To show the robustness and effectiveness of the FrCN segmentation method, com-
parisons with the latest research works in the literature are shown in Table 10.7. All of
the segmentation models in this table were evaluated utilizing the overall diagnostic
cases in the ISIC 2017 challenge dataset. In fact, 21 teams participated in the ISIC
2017 challenge. Table 10.7 provides a comparison with the top fve models [59, 91–93]
in the ISIC 2017 challenge. These models were ranked according to the segmentation
performance of the Jaccard index. Clearly, the FrCN approach outperformed the top
CDNN model [59] with a marginal increment of 0.61% in terms of the Jaccard index.
In contrast, the deep ResNet method [92] showed a higher specifcity of 98.50% for
Lossless Deep Learning FrCN 281

TABLE 10.7
Performance (%) of the FrCN Method Compared to the Latest Studies in the
Literature on Skin Lesion Segmentation
Ranked in the
References ISIC Challenge SEN SPE DIC JAC ACC

Yuan et al. (CDNN) [59] Top 1 82.50 97.50 84.90 76.50 93.40
Berseth [91] Top 2 82.00 97.80 84.70 76.20 93.20
Bi et al. (ResNet) [92] Top 3 80.20 98.50 84.40 76.00 93.40
Bi et al. (Multi-scale ResNet) [92] Top 4 80.10 98.40 84.20 75.8 93.40
Menegola et al. [93] Top 5 81.70 97.00 83.90 75.40 93.10
Lin et al. (U-Net) [58] – – – 77.00 62.00 –
FrCN – 85.40 96.69 87.08 77.11 94.03

the skin lesion boundary segmentation. However, the FrCN method provides prom-
ising results with fewer layers (i.e., only 16 layers) compared to the ResNet model,
which employed a deeper network of 50 layers. Although the skin lesion U-Net-
based histogram equalization technique proposed in [58] outperformed the clustering
method, its segmentation performance was lower compared to the recent deep learn-
ing approach, with Jaccard and dice indices of 62.00% and 77.00%, respectively. The
distribution performance of the segmented skin lesion boundaries of the ISIC 2017
test dataset (i.e., 600 images) in terms of Jaccard index via different segmentation
networks is shown in Figure 10.10 and Table 10.7. This fgure illustrates the boxplot
performance of different models and also shows the median, lower, and upper Jaccard
measures on the test dataset. Obviously, the FrCN method achieves a slightly higher
median Jaccard value than the top method, with fewer counts below a Jaccard index
of 0.4. With regard to the overall pixel-wise segmentation results of the skin lesion
task, the FrCN method proves its feasibility and effectiveness compared to others

FIGURE 10.10 Boxplots of the Jaccard index for different segmentation networks on the
ISIC 2017 challenge test dataset.
282 Deep Learning in Computer Vision

with a segmentation accuracy of 94.03%. The higher segmentation achievement of

the FrCN model against others is due to FrCN directly learning the full-resolution
features of each pixel through the convolutional layers of the network. It should be
noted that although we utilized the fxed image size, which was a requirement for the
conventional segmentation techniques, one can use any whole image size in FrCN
without resizing. We conclude that the convolutional process in the segmentation task
could extract and learn better prominent and specifc features utilizing the full spatial
resolution without the need for the downsampling process.

10.5 DISCUSSION
Automatic delineation of skin lesion boundaries is highly demanded as a prerequisite
step for melanoma recognition. Accurate segmentation of skin lesions enables the clas-
sifcation CNNs to extract more specifc and representative features from only seg-
mented skin lesion areas instead of entire dermoscopy images. Generally, the overall
segmentation performance of deep learning approaches improves as the size of the
training dataset increases. The impact of various training dataset sizes is presented in
[31, 62, 84], in which the results showed the signifcance of using larger training sets. In
this work, the training datasets are augmented by utilizing two procedures. First, HSV
images are generated from the original RGB images. Second, the training dermoscopy
images are augmented by rotating the RGB and HSV images with four rotation angles.
Furthermore, the images are fipped by applying horizontal and vertical fipping. The
augmented training data should support the network to learn better features of the
skin lesion characteristics. The results present the capability of the FrCN technique to
segment skin lesions with accuracies higher than those obtained with the conventional
methods. This is because the FrCN method is trained with the full spatial resolution
of the input images. In our experiments, the segmentation performance of the FrCN
technique was compared against three well-known segmentation approaches, namely
FCN, U-Net, and SegNet, under the same test conditions. Quantitative evaluations with
different indices of the recent deep learning methods are reported in Tables 10.4 and
10.5 for the ISIC 2017 test dataset. Signifcantly, the FrCN method outperformed other
skin lesion segmentation approaches. The MCC index showed how the segmented
skin lesions were correlated with the annotated ground-truths. The FrCN architecture
obtained overall increments of 3.92%, 11.99%, and 6.43% in MCC as compared to
FCN, U-Net, and SegNet, respectively. These promising segmentation results indicate
the consistency and capability of the FrCN method.
In this work, further analysis was done to show the segmentation performance
for clinical diagnosis cases. In the ISIC 2017 test dataset, the benign cases achieved
a higher segmentation performance than the melanoma and SK cases, with Jaccard
indices of 81.28%, 72.44%, and 69.25%, respectively, as shown in Table 10.4. This
signifcant improvement for the benign cases is due to the larger existing percentage
of the cases in the training dataset (i.e., 1,372/2,000 cases) and in the test dataset (i.e.,
393/600), as numbered in the database distribution in Table 10.1. Note that due to the
inequality of the numbers of each diagnostic type, the segmentation results of the
overall indices shown in Tables 10.4 and 10.5 were not computed as direct averages
of all cases, but were computed as percentages of their presence.
Lossless Deep Learning FrCN 283

Figure 10.9 illustrates some examples of the segmentation results of the FrCN
versus FCN, U-Net, and SegNet compared to the ground-truth contours. This fgure
shows the segmentation results of each clinical diagnostic case in the ISIC 2017
dataset. All approaches showed reasonable segmentation performance on the skin
lesion cases such as the presented in Figure 10.9 (a). In contrast, FrCN achieved bet-
ter segmentation when the dermoscopy image contained hair obstacles, as shown in
Figure 10.9 (c). Also, Figure 10.9 (b) illustrates how the FrCN method segments one
of the most challenging skin lesion cases with very low contrast.
In fact, the FrCN method is able to take an input image of any arbitrary size since
there are no downsampling layers. However, there is a demand on the computational
resources (i.e., enough memory), since the original dermoscopy images come with
very large sizes (i.e., from 540 × 722 to 4,499 × 6,748). In this work, all images are
resized in a preprocessing step for two reasons. One is to compare the conventional
segmentation methods that require the fxed image size. The other is to take our
computation resources into account.

10.6 CONCLUSIONS
In this chapter, we presented a full-resolution convolutional networks (FrCN)
method for skin lesion segmentation. Unlike the previous well-known segmentation
approaches, the FrCN method is able to utilize full spatial resolution features for each
pixel of dermoscopy images, improving in the performance of pixel-wise segmenta-
tion. We have evaluated the FrCN method utilizing the ISIC 2017 challenge dataset.
The results show that the FrCN method outperformed the recent FCN, U-Net, and
SegNet techniques. In future work, a larger number of training dermoscopy images
should be used to improve the segmentation performance of each class. Developing
a computer-aided diagnostic system using segmented lesions is also necessary to
distinguish between benign, melanoma, and normal skin lesions.

ACKNOWLEDGMENTS
This work was supported by the International Collaborative Research and
Development Programme (funded by the Ministry of Trade, Industry and Energy
[MOTIE, Korea]) (N0002252). This work was also supported by the National
Research Foundation of Korea (NRF) grant funded by the Korean government
(MEST) (NRF-2019R1A2C1003713).

BIBLIOGRAPHY
1. American Joint Committee on Cancer, “Melanoma of the skin,” Cancer Staging
Manual, pp. 209–220, Springer, New York, NY, 2002.
2. M. E. Celebi, H. A. Kingravi, B. Uddin, H. Lyatornid, Y. A. Aslandogan, W. V.
Stoecker, and R. H. Moss, “A methodological approach to the classifcation of der-
moscopy images,” Computerized Medical Imaging and Graphics, vol. 31, no. 6, pp.
362–373, September, 2007.
3. R. L. Siegel, K. D. Miller, and A. Jemal, “Cancer statistics,” CA Cancer Journal for
Clinicians, vol. 68, no. 1, pp. 7–30, 2018.
284 Deep Learning in Computer Vision

4. American Cancer Society, “Cancer facts & fgures 2018, American Cancer Society,
Atlanta, 2018,” Accessed [September 3, 2018]; https://www.cancer.org/cancer/melan
oma-skin-cancer.html.
5. A. M. Noone, N. Howlader, M. Krapcho, D. Miller, A. Brest, M. Yu, J. Ruhl, Z.
Tatalovich, A. Mariotto, D. R. Lewis, H. S. Chen, E. J. Feuer, and K. A. Cronin, “SEER
cancer statistics review, 1975–2015,” National Cancer Institute, 2018.
6. M. E. Vestergaard, P. Macaskill, P. E. Holt, and S. W. Menzies, “Dermoscopy com-
pared with naked eye examination for the diagnosis of primary melanoma: A meta-
analysis of studies performed in a clinical setting,” British Journal of Dermatology,
vol. 159, no. 3, pp. 669–676, 2008.
7. M. Silveira, J. C. Nascimento, J. S. Marques, A. R. Marçal, T. Mendonça, S. Yamauchi,
J. Maeda, and J. Rozeira, “Comparison of segmentation methods for melanoma diagno-
sis in dermoscopy images,” IEEE Journal of Selected Topics in Signal Processing, vol.
3, no. 1, pp. 35–45, 2009.
8. M. E. Celebi, H. Iyatomi, G. Schaefer, and W. V. Stoecker, “Lesion border detection in
dermoscopy images,” Computerized Medical Imaging and Graphics, vol. 33, no. 2, pp.
148–53, March, 2009.
9. H. Ganster, A. Pinz, R. Rohrer, E. Wildling, M. Binder, and H. Kittler, “Automated
melanoma recognition,” IEEE Transactions on Medical Imaging, vol. 20, no. 3, pp.
233–239, March, 2001.
10. E. Meskini, M. S. Helfroush, K. Kazemi, and M. Sepaskhah, “A new algorithm for skin
lesion border detection in dermoscopy images,” Journal of Biomedical Physics and
Engineering, vol. 8, no. 1, pp. 117–126, 2018.
11. G. Schaefer, B. Krawczyk, M. E. Celebi, and H. Iyatomi, “An ensemble classifcation
approach for melanoma diagnosis,” Memetic Computing, vol. 6, no. 4, pp. 233–240,
December, 2014.
12. M. A. Al-Masni, M. A. Al-antari, M. T. Choi, S. M. Han, and T. S. Kim, “Skin lesion
segmentation in dermoscopy images via deep full resolution convolutional networks,”
Computer Methods and Programs in Biomedicine, vol. 162, pp. 221–231, August, 2018.
13. M. E. Yuksel, and M. Borlu, “Accurate segmentation of dermoscopic images by image
thresholding based on type-2 fuzzy logic,” IEEE Transactions on Fuzzy Systems, vol.
17, no. 4, pp. 976–982, August, 2009.
14. K. Mollersen, H. M. Kirchesch, T. G. Schopf, and F. Godtliebsen, “Unsupervised seg-
mentation for digital dermoscopic images,” Skin Research and Technology, vol. 16, no.
4, pp. 401–407, November, 2010.
15. M. E. Celebi, Q. Wen, S. Hwang, H. Iyatomi, and G. Schaefer, “Lesion border detection
in dermoscopy images using ensembles of thresholding methods,” Skin Research and
Technology, vol. 19, no. 1, pp. E252–E258, 2013.
16. F. Peruch, F. Bogo, M. Bonazza, V. M. Cappelleri, and E. Peserico, “Simpler, faster,
more accurate melanocytic lesion segmentation through MEDS,” IEEE Transactions
on Biomedical Engineering, vol. 61, no. 2, pp. 557–565, February, 2014.
17. H. Y. Zhou, G. Schaefer, A. H. Sadka, and M. E. Celebi, “Anisotropic mean shift based
fuzzy c-means segmentation of dermoscopy images,” IEEE Journal of Selected Topics
in Signal Processing, vol. 3, no. 1, pp. 26–34, February, 2009.
18. S. Kockara, M. Mete, V. Yip, B. Lee, and K. Aydin, “A soft kinetic data structure for
lesion border detection,” Bioinformatics, vol. 26, no. 12, pp. i21–i28, June 15, 2010.
19. S. Suer, S. Kockara, and M. Mete, “An improved border detection in dermoscopy
images for density based clustering,” BMC Bioinformatics, vol. 12, no. 10, p. S12.
20. F. Y. Xie, and A. C. Bovik, “Automatic segmentation of dermoscopy images using self-
generating neural networks seeded by genetic algorithm,” Pattern Recognition, vol. 46,
no. 3, pp. 1012–1019, March, 2013.
Lossless Deep Learning FrCN 285

21. M. E. Celebi, H. A. Kingravi, H. Iyatomi, Y. A. Aslandogan, W. V. Stoecker, R. H.

Moss, J. M. Malters, J. M. Grichnik, A. A. Marghoob, H. S. Rabinovitz, and S. W.
Menzies, “Border detection in dermoscopy images using statistical region merging,”
Skin Research and Technology, vol. 14, no. 3, pp. 347–353, August, 2008.
22. Q. Abbas, M. E. Celebi, and I. F. Garcia, “Skin tumor area extraction using an improved
dynamic programming approach,” Skin Research and Technology, vol. 18, no. 2, pp.
133–142, May, 2012.
23. Q. Abbas, M. E. Celebi, I. F. Garcia, and M. Rashid, “Lesion border detection in der-
moscopy images using dynamic programming,” Skin Research and Technology, vol.
17, no. 1, pp. 91–100, February, 2011.
24. B. Erkol, R. H. Moss, R. J. Stanley, W. V. Stoecker, and E. Hvatum, “Automatic lesion
boundary detection in dermoscopy images using gradient vector fow snakes,” Skin
Research and Technology, vol. 11, no. 1, pp. 17–26, February, 2005.
25. M. Mete, and N. M. Sirakov, “Lesion detection in demoscopy images with novel den-
sity-based and active contour approaches,” BMC Bioinformatics, vol. 11, no. 6, p. S23,
2010.
26. A. R. Sadri, M. Zekri, S. Sadri, N. Gheissari, M. Mokhtari, and F. Kolahdouzan,
“Segmentation of dermoscopy images using wavelet networks,” IEEE Transactions on
Biomedical Engineering, vol. 60, no. 4, pp. 1134–1141, 2013.
27. K. Korotkov, and R. Garcia, “Computerized analysis of pigmented skin lesions: A
review,” Artifcial Intelligence in Medicine, vol. 56, no. 2, pp. 69–90, October, 2012.
28. M. E. Celebi, Q. Wen, H. Iyatomi, K. Shimizu, H. Zhou, and G. Schaefer, “A state-of-
the-art survey on lesion border detection in dermoscopy images,” Dermoscopy Image
Analysis, pp. 97–129, CRC Press, 2015.
29. R. B. Oliveira, J. P. Papa, A. S. Pereira, and J. M. R. S. Tavares, “Computational meth-
ods for pigmented skin lesion classifcation in images: Review and future trends,”
Neural Computing & Applications, vol. 29, no. 3, pp. 613–636, February, 2018.
30. S. Pathan, K. G. Prabhu, and P. C. Siddalingaswamy, “Techniques and algorithms for
computer aided diagnosis of pigmented skin lesions-A review,” Biomedical Signal
Processing and Control, vol. 39, pp. 237–262, 2018.
31. M. A. Al-Masni, M. A. Al-Antari, J. M. Park, G. Gi, T. Y. Kim, P. Rivera, E. Valarezo,
M. T. Choi, S. M. Han, and T. S. Kim, “Simultaneous detection and classifcation of
breast masses in digital mammograms via a deep learning YOLO-based CAD system,”
Computer Methods and Programs in Biomedicine, vol. 157, pp. 85–94, 2018.
32. G. Carneiro, J. Nascimento, and A. P. Bradley, “Automated analysis of unregis-
tered multi-view mammograms with deep learning,” IEEE Transactions on Medical
Imaging, vol. 36, no. 11, pp. 2355–2365, 2017.
33. N. Dhungel, G. Carneiro, and A. P. Bradley, “A deep learning approach for the analysis
of masses in mammograms with minimal user intervention,” Medical Image Analysis,
vol. 37, pp. 114–128, April, 2017.
34. G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. van
der Laak, B. van Ginneken, and C. I. Sanchez, “A survey on deep learning in medical
image analysis,” Medical Image Analysis, vol. 42, pp. 60–88, December, 2017.
35. X. Zhao, Y. Wu, G. Song, Z. Li, Y. Zhang, and Y. Fan, “A deep learning model integrat-
ing FCNNs and CRFs for brain tumor segmentation,” Medical Image Analysis, vol. 43,
pp. 98–111, January, 2018.
36. E. Gibson, W. Li, C. Sudre, L. Fidon, D. I. Shakir, G. Wang, Z. Eaton-Rosen, R.
Gray, T. Doel, Y. Hu, T. Whyntie, P. Nachev, M. Modat, D. C. Barratt, S. Ourselin,
M. J. Cardoso, and T. Vercauteren, “NiftyNet: A deep-learning platform for medical
imaging,” Computer Methods and Programs in Biomedicine, vol. 158, pp. 113–122,
May, 2018.
286 Deep Learning in Computer Vision

37. M. A. Al-Antari, M. A. Al-Masni, M. T. Choi, S. M. Han, and T. S. Kim, “A fully

integrated computer-aided diagnosis system for digital X-ray mammograms via deep
learning detection, segmentation, and classifcation,” International Journal of Medical
Informatics, vol. 117, pp. 44–54, September, 2018.
38. Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp.
436–44, May 28, 2015.
39. M. A. Al-Masni, M. A. Al-Antari, J. M. Park, G. Gi, T. Y. Kim, P. Rivera, E. Valarezo,
S. M. Han, and T. S. Kim, “Detection and classifcation of the breast abnormalities in
digital mammograms via regional Convolutional Neural Network,” in Proceedings of
the 39th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), Jeju Island, Republic of Korea, 2017, pp. 1230–1233.
40. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classifcation with deep con-
volutional neural networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–90,
June, 2017.
41. D. Ciresan, A. Giusti, L. M. Gambardella, and J. Schmidhuber, “Deep neural networks
segment neuronal membranes in electron microscopy images,” in Advances in Neural
Information Processing Systems, 2012, pp. 2843–2851.
42. C. Cernazanu-Glavan, and S. Holban, “Segmentation of bone structure in X-ray
images using convolutional neural network,” Advances in Electrical and Computer
Engineering, vol. 13, no. 1, pp. 87–94, 2013.
43. M. Melinščak, P. Prentašić, and S. Lončarić, “Retinal vessel segmentation using deep
neural networks,” in 10th International Conference on Computer Vision Theory and
Applications (VISAPP 2015), pp. 577–582, 2015.
44. J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic seg-
mentation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
2015, pp. 3431–3440.
45. E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks for semantic seg-
mentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39,
no. 4, pp. 640–651, April, 2017.
46. A. Garcia-Garcia, S. Orts-Escolano, S. Oprea, V. Villena-Martinez, and J. Garcia-
Rodriguez, “A review on deep learning techniques applied to semantic segmentation,”
arXiv preprint arXiv:1704.06857, 2017.
47. O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedi-
cal image segmentation,” in International Conference on Medical Image Computing
and Computer-Assisted Intervention, 2015, pp. 234–241.
48. V. Badrinarayanan, A. Kendall, and R. Cipolla, “SegNet: A deep convolutional
encoder-decoder architecture for image segmentation,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 39, no. 12, pp. 2481–2495, December, 2017.
49. V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep convolutional encoder-
decoder architecture for image segmentation,” arXiv preprint arXiv:1511.00561, 2015.
50. H. Noh, S. Hong, and B. Han, “Learning deconvolution network for semantic segmen-
tation,” in Proceedings of the IEEE International Conference on Computer Vision,
2015, pp. 1520–1528.
51. L. C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “DeepLab:
Semantic image segmentation with deep convolutional nets, atrous convolution,
and fully connected CRFs,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 40, no. 4, pp. 834–848, April, 2018.
52. L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Deeplab: Semantic
image segmentation with deep convolutional nets, atrous convolution, and fully con-
nected CRFs,” CoRR abs/1606.00915, 2016.
53. Y. X. Li, and L. L. Shen, “Skin lesion analysis towards melanoma detection using deep
learning network,” Sensors, vol. 18, no. 2, February, 2018.
Lossless Deep Learning FrCN 287

54. L. Bi, J. Kim, E. Ahn, A. Kumar, M. Fulham, and D. Feng, “Dermoscopic image
segmentation via multistage fully convolutional networks,” IEEE Transactions on
Biomedical Engineering, vol. 64, no. 9, pp. 2065–2074, 2017.
55. L. Q. Yu, H. Chen, Q. Dou, J. Qin, and P. A. Heng, “Automated melanoma recogni-
tion in dermoscopy images via very deep residual networks,” IEEE Transactions on
Medical Imaging, vol. 36, no. 4, pp. 994–1004, 2017.
56. Y. D. Yuan, M. Chao, and Y. C. Lo, “Automatic skin lesion segmentation using deep
fully convolutional networks with jaccard distance,” IEEE Transactions on Medical
Imaging, vol. 36, no. 9, pp. 1876–1886, 2017.
57. M. Goyal, and M. H. Yap, “Multi-class semantic segmentation of skin lesions via fully
convolutional networks,” arXiv preprint arXiv:1711.10449, 2017.
58. B. S. Lin, K. Michael, S. Kalra, and H. R. Tizhoosh, “Skin lesion segmentation: U-Nets
versus clustering,” in 2017 IEEE Symposium Series on Computational Intelligence
(SSCI), 2017, pp. 1–7.
59. Y. Yuan, “Automatic skin lesion segmentation with fully convolutional-deconvolutional
networks,” arXiv preprint arXiv:1703.05165, 2017.
60. N. C. F. Codella, D. Gutman, M. E. Celebi, B. Helba, M. A. Marchetti, S. W. Dusza,
A. Kalloo, K. Liopyris, N. Mishra, H. Kittler, and A. Halpern, “Skin lesion analy-
sis toward melanoma detection: A challenge at the 2017 International Symposium on
Biomedical Imaging (ISBI), hosted by the International Skin Imaging Collaboration
(ISIC),” in 15th IEEE International Symposium on Biomedical Imaging (ISBI 2018),
2018, pp. 68–172.
61. International Skin Imaging Collaboration, “ISIC 2017: Skin lesion analysis towards mela-
noma detection,” Accessed [October 19, 2018]; https://challenge.kitware.com/#challenges.
62. Z. Jiao, X. Gao, Y. Wang, and J. Li, “A deep feature based framework for breast masses
classifcation,” Neurocomputing vol. 197, pp. 221–231, 2016.
63. T. Kooi, G. Litjens, B. van Ginneken, A. Gubern-Merida, C. I. Sancheza, R. Mann, A. den
Heeten, and N. Karssemeijer, “Large scale deep learning for computer aided detection of
mammographic lesions,” Medical Image Analysis, vol. 35, pp. 303–312, January, 2017.
64. H. R. Roth, L. Lu, J. M. Liu, J. H. Yao, A. Seff, K. Cherry, L. Kim, and R. M. Summers,
“Improving computer-aided detection using convolutional neural networks and ran-
dom view aggregation,” IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp.
1170–1181, May, 2016.
65. M. Lin, Q. Chen, and S. Yan, “Network in network,” in rXiv preprint arXiv:1312.4400,
pp. 1–10, 2013.
66. D. Scherer, A. Müller, and S. Behnke, “Evaluation of pooling operations in convo-
lutional architectures for object recognition,” in Artifcial Neural Networks–ICANN
2010, 2010, pp. 92–101.
67. K. Simonyan, and A. Zisserman, “Very deep convolutional networks for large-scale
image recognition,” arXiv preprint arXiv:1409.1556, 2014.
68. Q. Guo, F. L. Wang, J. Lei, D. Tu, and G. H. Li, “Convolutional feature learning and
Hybrid CNN-HMM for scene number recognition,” Neurocomputing, vol. 184, pp.
78–90, April 5, 2016.
69. A. L. Maas, A. Y. Hannun, and A. Y. Ng., “Rectifer nonlinearities improve neural
network acoustic models,” in Proceeding of 30th International Conference on Machine
Learning (ICML), 2013, p. 3.
70. X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifer neural networks,” in
Proceedings of the 14th International Conference on Artifcial Intelligence and
Statistics, 2011, pp. 315–323.
71. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout:
A simple way to prevent neural networks from overftting,” Journal of Machine
Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014.
288 Deep Learning in Computer Vision

72. S. Hoo-Chang, H. R. Roth, M. Gao, L. Lu, Z. Xu, I. Nogues, J. Yao, D. Mollura, and
R. M. Summers, “Deep convolutional neural networks for computer-aided detection:
CNN architectures, dataset characteristics and transfer learning,” IEEE Transactions
on Medical Imaging, vol. 35, no. 5, pp. 1285–1298, 2016.
73. Y. Bar, I. Diamant, L. Wolf, S. Lieberman, E. Konen, and H. Greenspan, “Chest pathol-
ogy identifcation using deep feature selection with non-medical training,” Computer
Methods in Biomechanics and Biomedical Engineering-Imaging and Visualization,
vol. 6, no. 3, pp. 259–263, 2018.
74. R. K. Samala, H. P. Chan, L. Hadjiiski, M. A. Helvie, J. Wei, and K. Cha, “Mass detec-
tion in digital breast tomosynthesis: Deep convolutional neural network with trans-
fer learning from mammography,” Medical Physics, vol. 43, no. 12, pp. 6654–6666,
December, 2016.
75. J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are features in deep
neural networks?” in Advances in Neural Information Processing Systems, 2014, pp.
3320–3328.
76. O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. H. Huang, A.
Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei, “ImageNet large scale
visual recognition challenge,” International Journal of Computer Vision, vol. 115, no.
3, pp. 211–252, December, 2015.
77. E. Szymanska, E. Saccenti, A. K. Smilde, and J. A. Westerhuis, “Double-check:
validation of diagnostic statistics for PLS-DA models in metabolomics studies,”
Metabolomics, vol. 8, no. 1, pp. S3–S16, June, 2012.
78. S. Smit, M. J. van Breemen, H. C. J. Hoefsloot, A. K. Smilde, J. M. F. G. Aerts, and
C. G. de Koster, “Assessing the statistical validity of proteomics based biomarkers,”
Analytica Chimica Acta, vol. 592, no. 2, pp. 210–217, June 5, 2007.
79. T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning; Data
Mining, Inference and Prediction, Second ed., Springer, New York, 2008.
80. S. Min, B. Lee, and S. Yoon, “Deep learning in bioinformatics,” Briefngs in
Bioinformatics, vol. 18, no. 5, pp. 851–869, September, 2017.
81. Image Segmentation Keras, “Implementation of Segnet, FCN, UNet and other models
in Keras,” Accessed [October 23, 2018]; https://github.com/divamgupta/image-segm
entation-keras.
82. Keras, “Keras: The python deep learning library,” Accessed [January 8, 2019]; https://
keras.io/.
83. M. Dong, X. Lu, Y. Ma, Y. Guo, Y. Ma, and K. Wang, “An effcient approach for
automated mass segmentation and classifcation in mammograms,” Journal of Digital
Imaging, vol. 28, no. 5, pp. 613–25, 2015.
84. M. A. Al-antari, M. A. Al-masni, S. U. Park, J. Park, M. K. Metwally, Y. M. Kadah,
S. M. Han, and T. S. Kim, “An automatic computer-aided diagnosis system for breast
cancer in digital mammograms via deep belief network,” Journal of Medical and
Biological Engineering, vol. 38, no. 3, pp. 443–456, June, 2018.
85. D. Zikic, Y. Ioannou, M. Brown, and A. Criminisi, “Segmentation of brain tumor tis-
sues with convolutional neural networks,” in MICCAI-BRATS, 2014, pp. 36–39.
86. S. Pereira, A. Pinto, V. Alves, and C. A. Silva, “Brain tumor segmentation using convo-
lutional neural networks in MRI images,” IEEE Transactions on Medical Imaging, vol.
35, no. 5, pp. 1240–1251, May, 2016.
87. D. M. Powers, “Evaluation: From precision, recall and F-measure to ROC, informed-
ness, markedness & correlation,” Journal of Machine Learning Technologies, vol. 2,
no. 1, pp. 37–63, 2011.
88. M. D. Zeiler, “ADADELTA: An adaptive learning rate method,” arXiv preprint
arXiv:1212.5701, 2012.
Lossless Deep Learning FrCN 289

89. D. P. Kingma, and J. L. Ba, “Adam: A method for stochastic optimization,” in

International Conference on Learning Representations (ICLR), arXiv preprint
arXiv:1412.6980, 2015.
90. T. Tieleman, and G. Hinton, “Lecture 6.5-rmsprop: Divide the gradient by a running
average of its recent magnitude,” 4, http://www.cs.toronto.edu/~tijmen/csc321/slides/le
cture_slides_lec6.pdf, 2012.
91. M. Berseth, “ISIC 2017-skin lesion analysis towards melanoma detection,” arXiv pre-
print arXiv:1703.00523v1, 2017.
92. L. Bi, J. Kim, E. Ahn, and D. Feng, “Automatic skin lesion analysis using large-scale
dermoscopy images and deep residual networks,” arXiv preprint arXiv:1703.04197, 2017.
93. A. Menegola, J. Tavares, M. Fornaciali, L. T. Li, S. Avila, and E. Valle, “RECOD titans
at ISIC challenge 2017,” arXiv preprint arXiv:1703.04819, 2017.
 11 Skin Melanoma
Classifcation Using
Deep Convolutional
Neural Networks
Khalid M. Hosny, Mohamed A. Kassem,
and Mohamed M. Foaud

CONTENTS
11.1 Introduction .................................................................................................. 291
11.2 Literature Review ......................................................................................... 293
11.3 Background................................................................................................... 294
11.3.1 Preprocessing.................................................................................... 294
11.3.2 Convolutional Neural Networks ....................................................... 296
11.3.2.1 Alex-Net CNN Model ........................................................ 296
11.3.2.2 Resnet50............................................................................. 297
11.3.3 Transfer Knowledge for Deep Neural Network................................ 298
11.4 Methodology................................................................................................. 299
11.4.1 The Augmentation Process............................................................... 299
11.4.2 Transfer Learning ............................................................................. 301
11.4.3 Dropout Layer...................................................................................302
11.5 Experimental Results....................................................................................304
11.5.1 Performance of the Alex-net.............................................................304
11.5.2 Performance of the Resnet50............................................................307
11.6 Discussion.....................................................................................................307
11.7 Conclusion .................................................................................................... 311
Bibliography .......................................................................................................... 312

11.1 INTRODUCTION
Computer-supported skin lesion inspection and dermal investigation have been
extended over the past decade to classify and model dermatological diseases and
human skin [1]. Automated systems for analyzing and synthesizing human skin face
different diffculties and complexities such as humidity and seasonal variation in
temperature, geographically based differences in diseases, environmental factors,
and hair presences [2]. Computerized assessment of skin lesions is an active research

291
292 Deep Learning in Computer Vision

area. The main target of this feld is to create systems able to analyze images of pig-
mented skin lesions for automated diagnosis of cancerous lesions or to assist derma-
tologists in this task [3]. The most important challenge in this context is classifying
the color images of the skin lesions with highly accurate classifers [4]. In terms of
dermoscopic images, skin lesions fall into different categories based on the disease
type, severity, and stage [5]. There are many algorithms used to extract skin lesion
features [6]. The “ABDC” rule is the most known algorithm, where A, B, D, and C
refer to lesion Asymmetry, irregularity of the Border, Color variation, and lesion
Diameter, respectively. COLOR is the most prevalent content that can be easily visu-
ally observed when retrieving images. This is due to its value in isolating diseased
skin from wholesome [7]. However, color-based features are highly infuenced by
the condition of the patient and his/her external environment such as variations in
natural skin complexion, lighting, and camera specifcations [8].
During the last decade, there have been many studies on classifying non-mela-
nocytic and melanocytic skin lesions. These studies vary in using statistical meth-
ods, traditional feature extraction, and classifcation methods [11–15]. Recently, deep
learning has been raised as a promising tool for feature extraction of images through
different models. In these models, many layers of neurons have been used to fulfll
different layers to extract the fne details of the image, where the new layer combines
the characteristics of the previous one. This comprises the newest generic of such
models, which are called convolutional neural networks (CNNs) [16]. Applying deep
learning techniques in recent healthcare research area has had positive results, as
suggested by Dubal et al. and Kumar et al. [17, 18]. There are many studies using
different learning techniques, as can be found in Nasr-Esfahan et al., Premaladha
and Ravichandran, and Kostopoulos et al. [19–22]. Various kinds of skin tumor have
been found, like squamous cell carcinoma, basal cell carcinoma, and melanoma;
the last of these is the most unpredictable. Jain et al. [23] suggested that detecting
melanoma at an early stage is helpful for curing it. Precisely analyzing each patient
sample is a tedious task for dermatologists; therefore, an effcient automated system
with a high classifcation rate is necessary to assess the dangers associated with the
given samples [24].
Martin et al. [25] showed that melanoma is a highly invasive and malignant
tumor that can reach the bloodstream easily and cause metastasis in affected
patients. Azulay [26] reported that one of the hardest challenges in dermatology
is to distinguish malignant melanotic lesions (melanoma) from non-malignant
melanotic lesions (nevus), since both diseases share morphological character-
istics. Melanomas may also take on features that may not be common to their
morphology, which can cause it to be confused with a completely different mela-
noma disease. Goodfellow et al. [16] concluded that CNNs are the best tool for
recognizing patterns in digital images. In this chapter, we assessed the existing
methods for skin lesion classifcation, especially deep learning-based methods.
It was observed that existing deep learning-based methods encounter two major
problems that negatively affect their performance and signifcantly reduce their
classifcation rates. First, color images of skin lesions are associated with compli-
cated backgrounds. Second, the available skin lesion datasets contain a limited
Skin Melanoma Classifcation Using Deep CNNs 293

number of images, while deep learning classifcation methods need a large num-
ber of images for good training.
These challenges motivated us to present a new method that signifcantly improves
the skin cancer classifcation rate and outperforms all previous methods. High skin
lesion detection and classifcation rates enable dermatologists and physicians to make
the right decisions at an earlier stage, which generally improves patients’ health and
saves their lives. The proposed method is a pretrained deep convolutional neural
network-based classifcation method. The well-known datasets of color images of
skin, MED-NODE, Dermatology Information System (DermIS), & DermQuest,
are used to prove the validity of the proposed method and evaluate its performance
against the other existing methods. The obtained results clearly show that the pro-
posed method achieves the best classifcation rate.

11.2 LITERATURE REVIEW

This section summarizes some of the state of the art. By the end of this section,
a table is presented to compare all the listed methods. Bunte et al. [7] proposed a
classifying melanoma system. In addition to frst, second, and third moment orders,
Bunte and his coauthors used correlograms, coherence vectors, and histograms for
the representation of color image features. A binary mask has been used by Chang
et al. [9] to perform principal component analysis (PCA) for the classifcation of
conventional photographs. They achieved a classifcation rate of 82.30%. In order to
improve the performance of their system, they extracted gray-level co-occurrence-
based features and used these features in classifying different benign skin lesions.
Their system achieves an accuracy of 90.86% [9]. Kundu et al. [10] used the local
binary pattern (LBP) for feature extraction and developed another classifcation sys-
tem that utilized multilayer perceptron, naïve Bayes, and a support vector machine
(SVM) for skin disease classifcation. Amelard et al. [11] proposed extracting a set
of morphological high-level intuitive features (HLIFs) to describe the amount of
lesion border irregularity. They incorporated the HLIFs into a set of low-level fea-
tures to get the more semantic meaning of the feature set and to allow the system to
provide an intuitive rationale for the classifcation decision. Their system achieved
the classifcation rate of 87.38%. An automated model for skin cancer classifcation
has been developed by Karabulut and Ibrikci [12]. They used the convolutional neu-
ral networks (CNNs) and the SVM. Texture analysis of block difference of inverse
probabilities and LBP have been used as a preprocessing step. They achieved a clas-
sifcation rate of 71.4%.
A computer-aided diagnosis (CAD) system has been proposed by Almaraz-
Damian et al. [13], where the ABCD is rule used to extract the features and SVM
is used in classifcation. Their system achieved a classifcation rate of 75.1%.
Giotis et al. [14] proposed a system that extracts the regions of the lesion by using
the descriptors of the texture and the color. Jafari et al. [15] describes a guided
fltering method in which the noise of the input images is reduced. Then, the fea-
tures of skin lesions are extracted using the ABCD rule. This method achieved a
classifcation rate of 79%. Implementation of deep learning with clinical images
294 Deep Learning in Computer Vision

has been proposed by Nasr-Esfahan et al. [19]. The accuracy of the system is
increased by preprocessing to correlate the illumination and the segmentation of
the region of interest (ROI). The proposed system sends the image to the CNN
after segmentation for feature extraction. The system achieves a classifcation
rate of 81%. Premaladha and Ravichandran [20] proposed a CAD system that
combines supervised and deep learning algorithms in skin image classifcation.
The contrast-limited adaptive histogram equalization technique (CLAHE) is used
to enhance the input image; then, the median flter with the normalized Otsu’s
segmentation (NOS) method is used to separate the affected lesion from normal
skin. This system achieved a classifcation rate of 90.12% with artifcial neu-
ral networks (ANN) and a rate of 92.89% with deep learning neural networks
(DLNN). Kostopoulos et al. [21] proposed a computer-based analysis of plain
photography using a probabilistic neural network (PNN) to extract the features
and decide if the lesion is melanocytic or melanoma. The system achieved a clas-
sifcation rate of 76.2%. Esteva et al. [22] classifed the skin lesion using a pre-
trained CNN called Inception v3. They achieved accuracy of about 71.2% when
they augmented the images in the dataset using rotations with random angles
from 0° and 359° and vertical fipping. Table 11.1 summarizes and compares these
state-of-the-art methods.
In Hosny et al. [42], transfer learning has been applied to a CNN model. To extract
ROI from images, images have been segmented to discard the background. To over-
come the limitation of low numbers of images, different methods of augmentation
have been carried on the segmented images. They achieved 97.70% accuracy. In
[43], the last layers are replaced in a pretrained CNN model, with different ways of
augmentation, and fne-tuned; an accuracy of 98.61% is achieved.

11.3 BACKGROUND
During the last few years, deep neural networks have gained researchers’ attention
around the world as an alternative approach to extracting features. The core idea of
deep neural networks is that these networks automatically select the most important
features without performing feature extraction [31]. This mechanism is a powerful
step in machine intelligence [33]. Through the following subsections, a brief descrip-
tion of the convolutional neural networks is given.

11.3.1 PREPROCESSING
In machine vision and image-processing applications, segmenting the ROI is a very
crucial task. Image segmentation usually improves the accuracy of classifcation,
since this allows the background of the original image to be ignored. We utilized the
segmentation method of Basavaprasad and Hegad [39] to segment the ROIs from the
color images of skin. In this method, a number of bins are used to coarsely represent
the input color image. The spatial information is used by coarse representation from
a histogram-based windowing process, and then the coarse data of the image is clus-
tered by hierarchical clustering.
TABLE 11.1
Comparison of Literature Review
Classifcation Accuracy
Method Dataset Enhancement Segmentation Image type method (%)

Bunte et al. [7] Real-world dataset University of Groningen Yes No RGB Image retrieval 84
Chang et al. [9] Collected by authors No No Gray level SVM 82.30
Kundu et al. [10] Collected by authors Gray level MLP 94.28
Amelard et al. [11] DermIS & DermQuest Yes Yes RGB SVM 87.38
Almaraz et al. [13] DermIS & DermQuest Yes Yes RGB SVM 75.1
Karabulut et al. [12] DermIS & DermQuest Yes Yes Black & white DCNN 71.4
Karabulut et al. [12] DermIS & DermQuest Yes Yes RGB SVM 71.4
Premaladha et al. [20] MED-NODE Yes Yes Black & white DLNN 92.89
Premaladha et al. [20] MED-NODE Yes Yes Black & white ANN 90.12
Skin Melanoma Classifcation Using Deep CNNs

Nasr-Esfahan et al. [19] MED-NODE Yes Yes RGB CNN 81

Giotis et al. [14] MED-NODE Yes Yes RGB Majority vote 81
M.H. Jafari et al. [15] MED-NODE Yes Yes RGB SVM 79
Kostopoulos et al. [21] MED-NODE Yes Yes RGB PNN 76.2
Esteva et al. [22] ISIC 2017 No Yes RGB DCNN 71.2
Hosny et al. [42] ISIC 2017 No Yes RGB DCNN 95.91
Hosny et al. [43] PH2 No Yes RGB DCNN 98.61
295
296 Deep Learning in Computer Vision

11.3.2 CONVOLUTIONAL NEURAL NETWORKS

A convolutional neural network (CNN) consists of stacked layers in which a given
layer has learned from the previous layer in a linear or nonlinear way [27]. Any CNN
model contains building blocks consisting of the following:

1. Convolutional layer: Applies a convolution operation to the input image by

passing the result to the next layer.
2. Pooling layer (also called subsampling or down-sampling): Reduces the
dimensionality of each feature map but retains the most important informa-
tion. Spatial pooling can be of different types: max, average, sum, etc.
3. Nonlinear rectifed linear unit (RELU) layer: Element-wise operation
(applied per pixel) that replaces all negative pixel values in the feature
map with zero and that is connected to a regular multilayer neural network
called a fully connected layer.
4. Connected layer: A traditional multilayer perceptron that uses a SoftMax
activation function in the output layer and in which every neuron in the
previous layer is connected to every neuron on the next layer.
5. Loss layer: Specifes how training penalizes the deviation between the pre-
dicted and true labels, being at the backend.

CNNs have been used successfully with signifcant improved performance in dif-
ferent applications such as natural language processing [28] and visual tasks [29].
Many deep convolution neural networks (DCNNs) are available, including LeNet,
Alex-net, ZFNet, GoogLeNet, VGGNet, and Resnet [30, 31].

11.3.2.1 Alex-Net CNN Model

Alex-net was developed by Krizhevsky et al. [29] as a DCNN model. It is used for the
ImageNet large-scale visual recognition challenge (ILSVRC) [30]. The input image
of Alex-net must be resized to the specifc size 227 × 227 × 3, where the width (W) is
227, the height (H) is 227, and the depth (D) is 3. The depth number refers to the red,
green, and blue channels. The input image is fltered by using the frst convolutional
layer with a number of kernels (K) equal to 96, with a flter (F) size of 11 × 11, and
a stride (S) equal to 4 pixels. The stride is defned as the distance in the kernel map
between responsive feld centers of neighboring neurons. The size of the convolu-
tion layer output volume is computed using the mathematical form (W−F+2P)/S+1,
where P refers to the number of padded pixels, here equal to zero. For input size of
227 × 227 × 3, the width W = 227, F = 11, S = 4, K = 96, and the size of the convolu-
tion layer output volume is (227–11+0)/4 + 1 = 55. Then, the size of the convolution
layer output volume is 55 × 55 × 96.
The computational load for the second convolutional layer is distributed to two
GPUs; therefore, the load for each GPU is 55 × 55 × 48. The output of the frst con-
volutional layer is the input to the second convolutional layer, which consists of max
pooling followed by convolution. The max pooling is 55/2 × 55/2 × 96 ≈ 27 × 27 × 96.
The input in this layer is fltered using 256 flters. The size of each flter is 5 × 5 × 48
Skin Melanoma Classifcation Using Deep CNNs 297

FIGURE 11.1 Architecture of Alex-net. C, P, N, FC, FZ, and DS refer to convolutional

layer, max pooling layer, local response normalization, fully connected layer, flter size, and
depth size, respectively.

with a stride of 2 pixels. The output of this layer is 27 × 27 × 256, which splits to two
GPUs, where each GPU works with 27 × 27 × 128.
The third, fourth, and ffth convolutional layers are created without any normal-
ization and pooling layers. The third layer of the convolutional network is connected
to the (pooled, normalized) outputs of the second convolutional layer with 384 ker-
nels, each of size 3 × 3 × 192. The fourth and ffth convolutional layers have 384 and
256 kernels of sizes 3 × 3 × 192 and 3 × 3 × 128, respectively. There are two fully con-
nected layers containing 4096 neurons for each one. The output features of the ffth
layer are the input of these two fully connected layers. Krizhevsky et al. [29] selected
1.2 million images from ImageNet [30] to construct 1000 classes to train the Alex-
net. An illustration of the architecture of the Alex-net is displayed in Figure 11.1.
Table 11.2 shows different layers with their associated parameters.

11.3.2.2 Resnet50
Resnet50 [44] is one of the deep residual networks (DRN). It is a very deep feed-
forward neural network (FFNN) with extra connections. A model of deep learning
based on the residual learning passes the input from one layer to a far layer by escap-
ing 2 to 5 layers; this method is called “skip connection”. Some models like Alex-net
try to fnd a solution by mapping some of the input to meet the same characteristics
in the output, but there is an enforcement in the residual network to learn how to map
some of the input to some of the output and input.
Resnet50 is a pretrained model consisting of 177 layers and containing a number
of connections equaling 192 × 2 to connect layers with each other. The input layer
restricts the size of the input images. The input must be resized to 224 × 224 × 3,
where the width is 224, the height is 224, and the depth is 3. The depth number
refers to the color space red, green, and blue channels. Resnet50 layers consist of
298 Deep Learning in Computer Vision

TABLE 11.2
Different Layers and Their Associated Parameters
Input Output Filter Depth Stride Padding
Layers size size size size size size

Conv#1+ RELU 227 × 227 × 3 55 × 55 × 96 11 × 11 96 4

Max pooling 55 × 55 × 96 27 × 27 × 96 3×3 2
Norm 27 × 27 × 96
Conv#2 + RELU 27 × 27 × 96 27 × 27 × 256 5×5 256 1 2
Max pooling 27 × 27 × 256 13 × 13 × 256 3×3 2
Norm 13 × 13 × 256
Conv#3 + RELU 13 × 13 × 256 13 × 13 × 384 3×3 384 1 1
Conv#4 + RELU 13 × 13 × 384 13 × 13 × 384 3×3 384 1 1
Conv#5 + RELU 13 × 13 × 384 13 × 13 × 256 3×3 256 1 1
Max pooling 13 × 13 × 256 6 × 6 × 256 3×3 2

convolutional, RELU, pooling, batch normalization, a number of additional layers

to maintain the skipped connection fully connected, and a classifcation layer. The
building block shown in Figure 11.2 is based on the following equation:

( )
y = F x, {Wi } + x (11.1)

where x and y are the input and the output vectors of the layers considered, Ƒ(x,{Wi})
and represents the residual mapping to be learned. The dimensions of x and Ƒ must
be equal.
Because Resnet50 contains 177 layers and 192 × 2 connections, a detailed part
of Resnet50 has been shown in Figure 11.3. Furthermore, Table 11.3 shows different
layers with their associated output size and the building block parameters.

11.3.3 TRANSFER KNOWLEDGE FOR DEEP NEURAL NETWORK

In traditional neural networks, utilization of these networks is restricted by the type
of the trained data. In other words, the neural networks are not able to classify data
if the tested and trained data are different. To overcome this problem, the transfer
knowledge, which also called transfer learning methodology, is used with the DCNN.
Transfer learning was developed as an alternative learning system to solve this prob-
lem [34]. In transfer learning, we must inquire about the knowledge that could be
transferred. For example, Alex-net [29] and Resnet50 [44] were trained using the
ImageNet [30] to classify images into 1000 classes. Transfer learning means the
ability to use Alex-net and Resnet50 to classify different kinds of images with dif-
ferent training classes. To achieve this process, we need to defne which knowledge
or information could be transferred and which learning algorithms should be used
in the knowledge transfer [35]. The relationship types of transfer learning and tradi-
tional machine learning are summarized in Table 11.4.
Skin Melanoma Classifcation Using Deep CNNs 299

FIGURE 11.2 A “bottleneck” building block for Resnet50

11.4 METHODOLOGY
This section describes the steps of the proposed method to classify color-skin
images. It is divided into three subsections. The augmentation process of the input
color images of skin is discussed in the frst subsection, which describes different
augmentation methods that have been applied here. The implementation of the trans-
fer learning methodology is described in the second subsection for Alex-net and
Resnet50. The last subsection is devoted to describing the process of dropping out
the last three layers for Alex-net and Resnet50 and how authors replaced these layers
in order to adapt the process to the required task here.

11.4.1 THE AUGMENTATION PROCESS

DCNN required a large number of trained images to achieve good performance.
Working with small datasets of medical images is the biggest challenge for the
DCNN. Unfortunately, the available datasets of skin color images, such as MED-
NODE [14], DermIS [36], and DermQuest [37], are small datasets. The MED-NODE
dataset consists of 70 melanoma and 100 nevus images from the digital image archive
300 Deep Learning in Computer Vision

FIGURE 11.3 A detailed part of Resnet50.

of the Department of Dermatology, University Medical Center Groningen (UMCG)

in the Netherlands. It is used for the development and testing of the system for skin
cancer detection from macroscopic images.
The second dataset consists of 171 images of skin lesions. The images of this
dataset were obtained using standard consumer-grade cameras in varying and
unconstrained environmental conditions. These images, collected from the publicly
available online databases DermIS [36] and DermQuest [37], are classifed into 119
images of melanoma and 87 images of nevus; each image contains a single lesion of
interest. The melanoma and nevus images are rotated with different angles ranging
from 0° to 180° with fxed steps of 5° for each rotation. The number of images in
Skin Melanoma Classifcation Using Deep CNNs 301

TABLE 11.3
Resnet50 Layers*
Layer
name Conv1 Conv2_x Conv3_x Conv4_x Conv5_x Others FLOPs

Building 7 × 7, 3 × 3, max
blocks stride 2 pool, stide 2
° 1 × 1, 64 ˙ ° 1 × 1, 128 ˙ ° 1 × 1, 256 ˙ ° 1 × 1, 512 ˙ average 3.8 × 109
˝ ˇ ˝ ˇ ˝ ˇ ˝ ˇ pool,
˝ 3 × 3, 64 ˇ × 3 ˝3 × 3, 128ˇ × 3 ˝ 3 × 3, 256 ˇ × 3 ˝ 3 × 3, 512 ˇ × 3
˝˛1 × 1, 256 ˇˆ ˝˛1 × 1, 512 ˇˆ ˝˛1 × 1, 1024 ˇˆ ˝˛1 × 1, 2048ˇˆ 1000-d fc,
SoftMax
Output 112 × 112 56 × 56 28 × 28 14 × 14 7×7 1×1 FLOPs
size

* Brackets Refer to Building Blocks with the Numbers of Stacked Blocks.

TABLE 11.4
Relationship between Traditional Machine Learning and Various Transfer
Learning Settings
Transfer learning
Traditional
machine learning Unsupervised/inductive Supervised/transudative
Domains of target & Same Different but related Different but related
source
Tasks of target & Same Different but related Same
source

the frst dataset becomes (180/5+1) × 70 = 2590 melanoma images and (180/5+1) ×
100 = 3700 nevus images.
Similarly, the number of images in the second dataset becomes (180/5+1) × 84 =
3108 melanoma images and (180/5+1) × 87 = 3219 nevus images. The original images
are augmented by rotation with random angles ranging from 0° to 355° and trans-
lation with different translation parameters for every image. All dataset images
were randomly divided into images for training (85% of the images) and images for
testing samples of original, segmented, rotated, and translated images (15% of the
images) as displayed in Figure 11.4. The three columns on the left are from DermIS
& DermQuest, while those on the right are from MED-NODE.

11.4.2 TRANSFER LEARNING

To train a new DCNN model from scratch, it needs a huge amount of processing
time in addition to a massive number of images like ImageNet. The weights of
Alex-net and Resnet50 are initialized and fne-tuned to adapt weights according to
302 Deep Learning in Computer Vision

FIGURE 11.4 The three columns on the left were selected randomly from the DermIS &
DermQuest datasets, and the three right columns are from MED-NODE. The frst, second,
third, and fourth rows show original, segmented, rotated, and translated images, respectively.

the datasets used in the proposed model. The weights will not change dramatically
because of using a very small learning rate. The learning rate is used to update the
weights of the convolutional layers. Stochastic gradient descent (SGD) algorithm
has been used to update convolutional layer weights because it is computationally
fast, in addition to processing a single training sample, which makes it easier to ft
into memory. Unlike the convolutional layers, the weights of fully connected layers
have been initialized randomly because we have changed the classifcation layer to
multiclass SVM in Alex-net and to SoftMax in Resnet50.

11.4.3 DROPOUT LAYER

Overftting is one of CNN’s problems, particularly when using a small dataset for the
learning model. A powerful tool used for CNN task regularization is the dropout of
layers. Alex-net contains eight fully connected layers; the last fully connected layer,
FC8, outputs 1000 class scores for different tasks. In this chapter, the last fully con-
nected layer (FC8), the SoftMax layer, and the classifcation layer have been dropped
out and replaced with a new fully connected layer (for two classes only: melanoma
and nevus) and a multiclass SVM [38]. This process, clearly displayed in Figure 11.5,
Skin Melanoma Classifcation Using Deep CNNs 303

FIGURE 11.5 The architecture of Alex-net after transfer learning.

signifcantly increased the training speed and reduced the computational complexity
of the DCNN. Also, it led to learning extra new robust features and fnally reduced
the negative effect of overftting.
The same process of dropout layers has been applied with Resnet50 but with some
differences. The last three layers, called FC100, FC100_softmax, and Classifcation
layer_fc100, have been replaced with new layers to be ft with the required task here.
The new layers are FC2, SoftMax, and Class Output. These layers are able to clas-
sify images into 2 classes (melanoma and nevus) instead of the 1000 classes in the
original Resnet50. Figure 11.6 explains the modifed Resnet50.

FIGURE 11.6 Resnet50 architecture after transfer learning.

304 Deep Learning in Computer Vision

11.5 EXPERIMENTAL RESULTS

The experiments are performed with an IBM-compatible computer equipped with a
Core i5 processor. Two datasets with color images defned in RGB space are used.
The datasets are divided into two classes, melanoma and nevus. Parallelization of
the implemented code in CUDA and running it using the GPU allowed the use of
a larger number of examples for the training process, which consequently provided
a large gain in the generalization error of the models. The classifcation layer has
been replaced by the multiclass SVM, where a fast-linear solver is compatible with
the two classes of skin images: melanoma and nevus. Two types of experiments
were performed. The frst type was performed with high-quality microscopic images
from the dataset [14], while the second type was performed with low-quality images
acquired by the camera with unconstrained and varying environmental factors,
which include noise such as hair and other factors.
Each experiment consisted of a number of runs where the batch size, the number
of training epochs (a measure of the number of times to use all training vectors
once to update the weights), and the initial learning rate were 10, 40, and 0.001,
respectively; these values were fxed for all experiments. The images were randomly
divided: 85% of the images were for training and 15% of the images were for test-
ing. All images from the two datasets were preprocessed by segmenting the region
of interest using the segmentation method [39]. The proposed model was evaluated
using three metrics, namely, accuracy, sensitivity, and specifcity [40], as follows:

t p + tn
Accuracy = (11.2)
t p + f p + fn + t n

tp
Sensitivity = (11.3)
t p + fn

tn
Specificity = (11.4)
f p + tn

Where, tp, fp, f n, and tn refer to true positive, false positive, false negative, and true
negative, respectively. The false-positive rate should be relatively small and the true
negative rate should be relatively large, resulting in most points falling in the left part
of the receiver operating characteristic (ROC) curve [41].

11.5.1 PERFORMANCE OF THE ALEX-NET

The modifed architecture of Alex-net has been tested using two datasets: MED-
NODE and DermIS [36] & DermQuest [37]. These datasets have been randomly
divided into 85% for training and the remaining 15% for testing. A total of six exper-
iments were conducted with proposed Alex-net using the frst dataset, MED-NODE.
The frst one was performed using the original skin images of the dataset without
Skin Melanoma Classifcation Using Deep CNNs 305

any preprocessing. Because of noise like hair, background, and other factors, a seg-
mentation process was carried out to extract ROI, keeping images in the same RGB
color space. The segmented ROI images were used in the second experiments to
measure the impact of noise reduction with Alex-net.
The segmented ROI images were augmented by rotating each image in the dataset
by an angle ranging from 0° to 90° with a fxed step of 5°. The numbers of melanoma
and nevus images increased to (90/5+1) × 70 = 1330 and (90/5+1) × 100 = 1900,
respectively. The third experiment was carried out to measure the performance of
the modifed Alex-net after the augmentation of the segmented images. The results
of all experiments are summarized in Table 11.5. Comparing these results with the
results of the frst and second experiments clearly shows that increasing the number
of training images results in a signifcant improvement in the evaluation metrics,
accuracy, sensitivity, and specifcity. This observation motivates us to conduct an
additional experiment with an approximately doubled number of training images.
The fourth experiment was performed with the same classes and the same condi-
tions except for the data augmentation. In this experiment, the segmented images
were augmented by rotating each image in the dataset by an angle ranging from
0° to 180° with a fxed step of 5°. Therefore, the numbers of melanoma and nevus
images increased to (180/5+1) × 70 = 2590 and (180/5+1) ×100 = 3700, respectively.
As shown in Table 11.5, the performance metrics were increased compared with the
previous experiments. The ffth experiment was performed where each segmented
image in the dataset was augmented by a rotation angle ranging from 0° to 355° with
a fxed step of 5°. The obtained result was improved compared with those from the
third and fourth experiments. We used the same augmentation process by rotating
each image with a fxed step angle of 5° from 0° to 90°, 0° to 180°, and 0° to 355° for
the third, fourth, and ffth experiments, respectively.
That last experiment was done by using a combination of different augmentation
approaches. Each segmented color image was translated with different translation
parameters and rotated with a random rotation angle in the range between 0° and
355°. The results of these six experiments clearly show that having more training
images increases the ability of the proposed system to identify skin lesions of the
type “melanoma” correctly (sensitivity) and its ability to identify skin lesions of type
“nevus” correctly (specifcity). To ensure this observation and the credibility of the
proposed system and its ability to identify melanoma and nevus skin lesions cor-
rectly, another group of experiments was performed with more complicated color-
skin images.
The second group of experiments was performed using noisy low-quality color-
skin images. The dataset used in these experiments was created by a random selec-
tion of images from DermIS [36] and DermQuest [37]. As mentioned before, the
images of the created dataset were segmented to isolate the ROI from the back-
ground to reduce the noise. The previous six experiments were repeated again using
the DermIS & DermQuest dataset. The acquired results have proved that increasing
the number of training images improved the evaluation metrics, accuracy, sensitivity,
and specifcity. Table 11.5 summarizes all obtained results of Alex-net using MED-
NODE and the DermIS & DermQuest datasets.
 306

TABLE 11.5
The Accuracy of the Proposed Model Using Alex-net
MED-NODE dataset DermIS & DermQuest dataset

Experiment Accuracy (%) Sensitivity (%) Specifcity (%) Accuracy (%) Sensitivity (%) Specifcity (%)

First 78.11 97.85 58.37 73.08 61.54 84.62

Second 81.05 95.53 66.56 76.93 84.62 69.23
Third 91.65 97.59 85.71 85.95 80.83 91.06
Fourth 94.62 92.41 96.83 90.79 94.86 86.71
Fifth 95.42 95.42 96.03 92.37 92.29 92.46
Sixth 96.82 95.27 98.47 93.35 93.48 93.21
Deep Learning in Computer Vision
Skin Melanoma Classifcation Using Deep CNNs 307

11.5.2 PERFORMANCE OF THE RESNET50

The modifed Resnet50 architecture was tested using the MED-NODE dataset.
The dataset was randomly divided into 85% for training and the remaining 15%
for testing. Because Resnet50 is deeper than Alex-net, it has the ability to classify
images better than Alex-net with a smaller number of images. Only three experi-
ments were done with modifed Resnet50 architecture. The frst experiments were
performed using the modifed Resnet50 architecture and the original images from
MED-NODE without segmentation and without augmentation. Images in the dataset
were segmented to extract foreground containing important features (ROI) and dis-
carding background, keeping segmented images in the same RGB color space. The
segmented color images were used in the second experiment. The performance mea-
sures were greatly improved compared with using a dataset without segmentation.
Finally, the segmented images in the dataset were augmented by rotating images
from 0° to 90° with a fxed step angle of 5°. To measure the impact of augmentation
in performance measures, a third experiment was performed. Table 11.6 shows the
performance of modifed Resnet50 with the three experiments.
The same experiments have been repeated using noisy low-quality color-skin
images from DermIS [36] and DermQuest [37]. As mentioned before to reduce
the noise, the images of the dataset were segmented to isolate the ROI from the
background. The previous three experiments were repeated using the DermIS &
DermQuest dataset by dividing it into 85% for training and the remaining 15% for
testing randomly. The acquired results proved that the modifed Resnet50 architec-
ture is better than the modifed Alex-net architecture.

11.6 DISCUSSION
It is noted that the results obtained from the second dataset are smaller than the
results obtained from the frst dataset. This is normal and predictable according to
the nature of the images in each dataset. The images of the frst dataset are high-
quality microscopic images acquired by specialized devices. On the other side, the
images of the second dataset are low-quality images acquired by using a common
camera in a noisy environment.
By comparing the performance measures of proposed Alex-net against proposed
Resnet50 using different datasets, we found that the best performance measures
were for the modifed Resnet50. The performance of Resnet50 is the best because it
contains many more layers than Alex-net. As discussed before, Resnet50 is enforced
to learn how to map some of the input to some of the output and input while Alex-net
tried to map some of the input to the output. Based on “skip connection”, Resnet50
passes the input from one layer to a far layer by escaping two to fve layers.
We could say that DCNN works well with any dataset of different resolutions.
The performance increased with high-quality images, high resolution, and low levels
of noise. The performance of the proposed methods was compared with the well-
known existing methods [11–21, 42] and the comparison is summarized in Tables 11.7
and 11.8. The accuracy and the ROC curves were used to measure the performance
of the proposed methods against the performance of the existing methods. For the
 308

TABLE 11.6
The Accuracy of the Proposed Model Using Resnet50
MED-NODE dataset DermIS & DermQuest dataset

Experiment Accuracy (%) Sensitivity (%) Specifcity (%) Accuracy (%) Sensitivity (%) Specifcity (%)

First 80 85.71 72.72 76.92 70.59 88.89

Second 92 100 83.33 92.31 86.67 100
Third 99.79 99.65 100 98.56 99.59 97.54
Deep Learning in Computer Vision
Skin Melanoma Classifcation Using Deep CNNs 309

TABLE 11.7
Comparative Study Using the DermIS & DermQuest Dataset
Preprocessing Image Classifcation Accuracy
Method (enhancement) Segmentation type method (%)

Amelard [11] Yes Yes RGB SVM 87.38

Karabulut [12] Yes Yes RGB SVM 71.4
Black & white CNN 71.4
Almaraz [13] Yes Yes RGB SVM 75.1
Hosny [42] No Yes RGB DCNN 96.86
Proposed Alex-net No Yes RGB DCNN 93.35
Proposed Resnet50 No Yes RGB DRNN 98.56

DermIS & DermQuest dataset, the performance of the proposed method was com-
pared with the performance of the existing methods [11–13, 42]. The obtained results
are listed in Table 11.7. All the mentioned methods in Table 11.7 used the same
DermIS & DermQuest dataset. Amelard et al. [11] enhanced the images and detected
the borders of skin lesions to segment important regions, and then they used the
segmented images in the classifcation process and achieved an accuracy percentage
of 87.38%. A similar approach was utilized in Karabulut and Ibrikci [12], where the
input images were enhanced to reduce the noise and then the regions of interest were
segmented in two ways. The frst way is to keep the segmented images in RGB color
space and the second way is to convert the segmented images to black-and-white
images. The SVM and CNN were used and achieved the same accuracy, 71.4%, for
the two methods. In Almaraz-Damian et al. [13], similar preprocessing steps were
used. The obtained accuracy was marginally increased to 75.1%. In Hosny et al.
[42], an automated process for skin cancer classifcation was proposed, where the
authors applied transfer learning with Alex-net and replaced the classifcation layer
with SoftMax. The achieved accuracy rate was 96.86%.
In order to improve the classifcation accuracy, the authors applied transfer learn-
ing with Alex-net and replaced the classifcation layer with the multiclass SVM.
Unfortunately, this method achieved an accuracy rate of only 93.35%. It is clear that
the achieved accuracy is lower than the achieved accuracy of the SoftMax method
[42]. Therefore, utilizing an alternative CNN could be a fruitful solution for this
problem. The authors applied transfer learning with Resnet50 to classify skin lesions
using the same dataset where the proposed Resnet50-based classifcation method
achieved a 98.56% accuracy rate and outperformed the Alex-net-based classifcation
methods. The ROC curves of the proposed methods and the existing methods [11–13,
42] are plotted and displayed in Figure 11.7.
Another comparison was performed using the MED-NODE dataset; all meth-
ods mentioned in Table 11.8 use this dataset. The performance of the proposed
methods was compared with that of the existing methods [14, 15, 19–21, 42]. Table
11.8 shows an overview of the conditions, image types, and classifers used in the
existing methods. It was noted that all of these methods applied two preprocessing
310 Deep Learning in Computer Vision

TABLE 11.8
Comparative Study for MED-NODE Dataset
Preprocessing Accuracy
Method (enhancement) Segmentation Image type Classifer (%)

Giotis [14] Yes Yes RGB Majority vote 81

M.H. Jafari [15] Yes Yes RGB SVM 79
Nasr-Esfahan [19] Yes Yes RGB CNN 81
Premaladha [20] Yes Yes Black & white ANN 90.12
DLNN 92.89
Kostopoulos [21] Yes Yes RGB PNN 76.2
Hosny [42] No Yes RGB DCNN 97.70
Proposed Alex-net No Yes RGB DCNN 96.82
Proposed No Yes RGB DRNN 99.79
Resnet50

steps. The frst step was enhancement, while the second was segmentation to extract
the region of interest. These methods used RGB color images, while the method of
[20] converted these color images to black and white. Table 11.8 clearly shows that,
on one hand, the performance of the proposed Alex-net method failed in terms
of its classifcation rates compared with other methods [14, 15, 19–21, 42]. On
the other hand, the proposed Resnet50-based classifcation method outperformed

FIGURE 11.7 The ROC curves for the different methods with the DermIS & DermQuest
dataset.
Skin Melanoma Classifcation Using Deep CNNs 311

FIGURE 11.8 The ROC curves for the different methods with the MED-NODE dataset.

the existing methods [14, 15, 19–21, 42]. The ROC curves of the different meth-
ods are plotted and displayed in Figure 11.8. All results are consistent and clearly
show that the proposed methods are superior to all the existing methods [11–21].
From the results displayed in Tables 11.7 and 11.8 in addition to the ROC curves
in Figures 11.7 and 11.8, it was proved that the proposed method is superior to the
state of the art.

11.7 CONCLUSION
This chapter proposed two models using the theory of transfer learning for the Alex-
net and Resnet50 architecture to classify tumors as melanoma or not-melanoma.
The last three layers (fully connected, SoftMax, and classifcation layers) have been
dropped out and replaced with three new layers. In Alex-net, the new layers were
fully connected for two classes only, namely, a multiclass SVM and an output layer;
in Resnet50, these layers were fully connected for two different classes, namely,
SoftMax and a classifcation layer. This modifcation was done to create two classif-
cations of skin images in addition to overcoming the overftting problem. Two differ-
ent datasets were used, one with high-quality images and the other with low-quality
images. These datasets were used to investigate the performance of the two proposed
models Alex-net and Resnet50 in a situation of using images of high and low quality.
Different ways of augmentation have been used to overcome the limitations of image
number in datasets. The proposed models have outperformed other models using
312 Deep Learning in Computer Vision

the same datasets. The sensitivity and specifcity of the proposed models proved the
credibility of the proposed model, because the increased sensitivity and specifcity
proved the truthfulness of the system.

BIBLIOGRAPHY
1. I. Maglogiannis, and C. N. Doukas, “Overview of advanced computer vision sys-
tems for skin lesions characterization”, Transactions on Information Technology in
Biomedicine, vol. 13, no. 5, pp. 721–733, 2009.
2. M. S. Arifn, M. G. Kibria, A. Firoze, M. A. Amini, and H. Yan, “Dermatological
disease diagnosis using color-skin images”, International Conference on Machine
Learning and Cybernetics, vol. 5, pp. 1675–1680, 2012.
3. J. M. Gálvez, D. Castillo, L. J. Herrera, B. S. Román, O. Valenzuela, F. M. Ortuño, and
I. Rojas, “Multiclass classifcation for skin cancer profling based on the integration of
heterogeneous gene expression series”, PLoS ONE, vol. 13, no. 5, pp. 1–26, 2018.
4. A. Masood, A. A. Al-Jumaily, and T. Adnan, “Development of automated diagnostic sys-
tem for skin cancer: Performance analysis of neural network learning algorithms for clas-
sifcation”, International Conference on Artifcial Neural Networks, pp. 837–844, 2014.
5. P. G. Cavalcanti, and J. Scharcanski, “Macroscopic pigmented skin lesion segmentation
and its infuence on lesion classifcation and diagnosis”, Color Medical Image Analysis,
vol. 6, pp. 15–39, 2013.
6. M. J. M. Vasconcelos, and L. Rosado, “No-reference blur assessment of dermatologi-
cal images acquired via mobile devices”, Image and Signal Processing, vol. 8509, pp.
350–357, 2014.
7. K. Bunte, M. Biehl, M. F. Jonkman, and N. Petkov, “Learning effective color features
for content-based image retrieval in dermatology”, Pattern Recognition, vol. 44, pp.
1892–1902, 2011.
8. S. V. Patwardhan, A. P. Dhawan, and P. A. Relue, “Classifcation of melanoma using
tree structured wavelet transforms”, Computer Methods and Programs in Biomedicine,
vol. 72, no. 3, pp. 223–239, 2003.
9. W. Y. Chang, A. Huang, C. Y. Yang, C. H. Lee, Y. C. Chen, T. Y. Wu, and G. S. Chen,
“Computer-aided diagnosis of skin lesions using conventional digital photography: A
reliability and feasibility study”, PLoS ONE, vol. 8, no. 11, pp. 1–9, 2013.
10. S. Kundu, N. Das, and M. Nasipuri, “Automatic detection of ringworm using Local
Binary Pattern (LBP)”, 2011, https://arxiv.org/abs/1103.0120.
11. R. Amelard, A. Wong, and D. A. Clausi, “Extracting morphological high-level intuitive
features (HLIF) for enhancing skin lesion classifcation”, International Conference of
the IEEE Engineering in Medicine and Biology Society, pp. 4458–4461, 2012.
12. E. M. Karabulut, and T. Ibrikci, “Texture analysis of melanoma images for computer-
aided diagnosis”, International Conference on Intelligent Computing, Computer
Science and Information Systems (ICCSIS), vol. 2, pp. 26–29, 2016.
13. J. A. Almaraz-Damian, V. Ponomaryov, and E. R. Gonzalez, “Melanoma CADe based
on ABCD rule and Haralick texture features”, International Kharkiv Symposium
on Physics and Engineering of Microwaves, Millimeter and Submillimeter Waves
(MSMW), pp. 1–4, 2016.
14. I. Giotis, N. Molders, S. Land, M. Biehl, M. F. Jonkman, and N. Petkov, “MED-NODE:
A computer-assisted melanoma diagnosis system using non-dermoscopic images”,
Expert Systems with Applications, vol. 42, no. 19, pp. 6578–6585, 2015.
15. M. H. Jafari, S. Samavi, N. Karimi, S. M. R. Soroushmehr, K. Ward, and K. Najarian,
“Automatic detection of melanoma using broad extraction of features from digital
images”, International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC), pp. 1357–1360, 2016.
Skin Melanoma Classifcation Using Deep CNNs 313

16. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, 2016.
17. P. Dubal, S. Bhatt, C. Joglekar, and S. Patil, “Skin cancer detection and classifca-
tion”, International Conference on Electrical Engineering and Informatics (ICEEI),
pp. 1–6, 2017.
18. D. Kumar, M. J. Shafee, A. Chung, F. Khalvati, M. Haider, and A. Wong, “Discovery
radiomics for computed tomography cancer detection”, 2015, https://arxiv.org/
abs/1509.00117.
19. E. Nasr-Esfahan, S. Samavi, N. Karimi, S. M. R. Soroushmehr, M. H. Jafari, K. Ward,
and K. Najarian, “Melanoma detection by analysis of clinical images using convolu-
tional neural network”, International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBC), pp. 1373–1376, 2016.
20. J. Premaladha, and K. S. Ravichandran, “Novel approaches for diagnosing melanoma
skin lesions through supervised and deep learning algorithms”, Journal of Medical
Systems, vol. 40, no. 96, pp. 1–12, 2016.
21. S. A. Kostopoulos et al., “Adaptable pattern recognition system for discriminating
Melanocytic Nevi from Malignant Melanomas using plain photography images from
different image databases”, International Journal of Medical Informatics, vol. 105, pp.
1–10, 2017.
22. A. Esteva, B. Kuprel, H. M. Blau, S. M. Swetter, J. Ko, R. A. Novoa, and S. Thrun,
“Dermatologist-level classifcation of skin cancer with deep neural networks”, Nature,
vol. 542, no. 7639, pp. 115–118, 2017.
23. S. Jain, V. Jagtap, and N. Pise, “Computer aided melanoma skin cancer detection using
image processing”, Procedia Computer Science, vol. 48, pp. 735–740, 2015.
24. D. Gautam, and M. Ahmed, “Melanoma detection and classifcation using SVM based
decision support system”, IEEE India Conference (INDICON), pp. 1–6, 2015.
25. Convolutional Neural Networks (CNNs / ConvNets), the Stanford CS class notes,
Assignments, Spring 2017, http://cs231n.github.io/convolutional-networks/, Accessed:
18 August 2017.
26. R. D. Azulay, D. R. Azulay, and L. Azulay-Abulafa, Dermatologia. 6th edition. Rio de
Janeiro: Guanabara Koogan, 2013.
27. Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to
document recognition”, Proceedings of the IEEE, vol. 86, no. 11, pp. 101–118, 1998.
28. S. Srinivas, R. K. Sarvadevabhatl, K. R. Mopur, N. Prabhu, S. S. S. Kruthiventi, and R.
V. Babu, “A taxonomy of deep convolutional neural nets for computer vision”, Frontiers
in Robotics and AI, vol. 2, pp. 1–18, 2016.
29. A. Krizhevsky, I. Sutskever, and G. Hinton, “ImageNet classifcation with deep convo-
lutional neural networks”, Advances in Neural Information Processing Systems, vol.
25, no. 2, pp. 1097–1105, 2012.
30. O. Russakovsky et al., “ImageNet large scale visual recognition challenge”,
International Journal of Computer Vision, vol. 115, no. 3, pp. 211–252, 2015.
31. J. D. Prusa, and T. M. Khoshgoftaar, “Improving deep neural network design with new
text data representations”, Journal of Big Data, Springer, vol. 4, pp. 1–16, 2017.
32. T. A. Martin, A. J. Sanders, L. Ye, J. Lane, and W. G. jiang, “Cancer invasion and
metastasis: Molecular and cellular perspective”, Landes Bioscience, pp. 135–168, 2013.
33. J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are features in deep
neural networks?” Advances in Neural Information Processing Systems, vol. 2, pp.
3320–3328, 2014.
34. S. J. Pan, and Q. Yang, “A survey on transfer learning”, Transactions on Knowledge
and Data Engineering, vol. 22, no. 10, pp. 1345–1359, 2010.
35. K. Weiss, T. M. Khoshgoftaar, and D. Wang, “A survey of transfer learning”, Journal
of Big Data, Springer, vol. 3, no 9, pp. 1–15, 2016.
36. Dermatology Information System, 2012, http://www.dermis.net, Accessed: 16 August 2017.
314 Deep Learning in Computer Vision

37. DermQuest, 2012, http://www.dermquest.com, Accessed: 16 August 2017.

38. C. Chih-Chung, and L. Chih-Jen, “LIBSVM: A library for support vector machines”,
Transactions on Intelligent Systems and Technology, vol. 2, pp. 1–27, 2013.
39. B. Basavaprasad, and R. S. Hegad, “Color image segmentation using adaptive Growcut
method”, Procedia Computer Science, vol. 45, pp. 328–335, 2015.
40. M. Stojanovi et al., “Understanding sensitivity, specifcity and predictive values”,
Vojnosanitetski pregled, vol. 71, no. 11, pp. 1062–1065, 2014.
41. T. Fawcett, “An introduction to ROC analysis”, Pattern Recognition Letter, vol. 27, no.
8, pp. 861–874, 2006.
42. K. M. Hosny, M. A. Kassem, and M. M. Foaud, “Classifcation of skin lesions using
transfer learning and augmentation with Alex-net”, PLoS ONE, vol. 14, no. 5, pp. 1–17,
2019.
43. K. M. Hosny, M. A. Kassem, and M. M. Foaud, “Skin cancer classifcation using
deep learning and transfer learning”, Cairo International Biomedical Engineering
Conference (CIBEC), pp. 90–93, 2018.
44. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition”,
Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.
Index
A B
abBRIEF, 184 BA, see Balanced accuracy
ABDC rule, 292, 293 Back-propagation algorithm, 3
Above Varosha, 196 Balanced accuracy (BA), 251
Accuracy (ACC), 251, 274, 304 Balanced classifcation rate (BCR), 251
ACs, see Autonomous cars Batch normalization (BN), 240–241
ACS, see American Cancer Society Batch parallelism, 8
Activation layers, 5 BCR, see Balanced classifcation rate
Active contour methods, 264 Bias, 4
AD, see Autonomous driving Bidirectional LSTM, 131–132
Adaptive Moment Estimator (ADAM), 190 Binary networks, 29–30
AlexNet, 235, 241–243, 296–298, 302, 303, Binary neural networks (BNNs), 6
307, 311 Biometric system, 90
performance, 304–306 BN, see Batch normalization
American Cancer Society (ACS), 262 BN layer, 6
ANNs, see Artifcial neural networks BNNs, see Binary neural networks
Approximate arithmetic, CNNs, 23–30 Brain tumor detection, DCNN
binary networks, 29–30 design, 249
dynamic fxed point, 28–29 simulation experiments and evaluation
extreme quantization, 29–30 criteria
fxed-point arithmetic, 23, 27 environment setup, 250
FPGA implementations, 29 evaluation criteria, 250–252
Artifcial neural networks (ANNs), 294 experimental results, 252
Augmentation process, 299–301
Automatic differentiation (AutoDiff), 244 C
Autonomous cars (ACs), 152
Autonomous driving (AD) C++, 244
ACs, 152 CAD, see Computer-aided detection; Computer-
analysis aided diagnosis
depth experiments, 167 Cancer Statistics Review (CSR) report, 262
optical fow experiments, 167–172 CD, see Contrastive divergence
autonomy levels, 152–153 CLAHE, see Contrast-limited adaptive histogram
CNNs (see Convolutional neural networks) equalization technique
decision-making system, 153 Classifcation performance, 11
experiments CNNs, see Convolutional neural networks
datasets, 166 CNNs inference accelerators, FPGAs
settings, 166 approximate computing, 23–32
fusion architectures approximate arithmetic, 23–30
early-fusion network, 165 reduced computations, 30–32
mid-fusion network, 166 computational transforms, 12–16
single-input network, 164 fast Fourier transform, 16, 17
history, 152 im2col transformation, 13–14
image segmentation (see Semantic Winograd transform, 14–15
segmentation) data-path optimizations, 16–23
motion and depth cues design space exploration, 21–22
depth estimation, 162–163 FPGA implementations, 22–26
optical fow estimation, 163–164 loop optimization, spatial architectures,
navigation, 153–154 18–21
technical challenges, 153 systolic arrays, 16, 18

315
316 Index

FPGA-based deep learning, 11–12 Convolutional layers, 4–5, 237–238, 249,

inference, layers, and CNN models, 3–6 269, 296
activation layers, 5 Convolutional neural networks (CNNs), 2,
BN layer, 6 99–100, 128, 234, 242, 292–293,
convolutional layers, 4–5 296–298; see also Long short-term
fully connected layers, 5 memory
pooling layers, 5 deep learning, 154–155
Softmax layer, 6 layers, 236–239
inference vs. training, 3 BN, 156
overview, 3 convolutional, 155, 237–238
workloads and computations depth output, 156
computational workload, 6–8 fully connected, 157, 239
frameworks, 10 input, 155
hardware, 10 max-pooling, 238
libraries, 10 nonlinearity, 237, 238–239
memory accesses, 9–10 output, 157
parallelism, CNNs, 8–9 pooling, 157
CNNs object detection receptive feld, 156
datasets rectifed linear unit, 237, 238–239
MS COCO, 55 stride, 156
PASCAL VOC, 54–55 zero-padding, 156
VisDrone-DET2018, 55–58 types
detector architectures fully convolutional networks, 245
ResNet, 52–53 hybrid learning networks, 245
RetinaNet, 53–54 region-based convolutional neural
development networks, 244–245
FPN, 44 CSC, see Compressed sparse column
performance, 43 CRF, see Conditional random feld
evolution Cross-entropy, 272
image classifcation, 42 CSR, see Cancer Statistics Review
one-stage models, 43 Curvelet Transform-Fractional Brownian Motion
two-stage models, 43 (CT-FBM), 109
one-stage model
SSD, 50–51 D
YOLO, 48–49
YOLO9000, 49–50 Data-path optimizations, 16–23
YOLOv2, 49–50 design space exploration, 21–22
YOLOv3, 50 FPGA implementations, 22–23
two-stage model loop optimization, spatial architectures,
Faster R-CNN, 45–48 18–21
Fast R-CNN, 45–46 systolic arrays, 16, 18
R-CNN, 44–45 Data preprocessing and augmentation, 267–269
COLOR, 292 Dataset, 267
Compressed sparse column (CSC), 32 DBC, see Differential box-counting
Computational throughput, 11 DB-LSTM, see Deep bidirectional LSTM
Computational transforms, 12–16 DBN, see Deep belief network
fast Fourier transform, 16 DCNN, see Deep convolutional neural networks
im2col transformation, 13–14 Deconvolutional single-shot detector (DSSD), 51
Winograd transform, 14–15 Deconvolution Network (DeconvNet), 266
Computer-aided detection (CAD), 247 Deep belief network (DBN), 98–99
Computer-aided diagnosis (CAD), 293, 294 Deep bidirectional LSTM (DB-LSTM), 132, 133
Conditional random feld (CRF), 266 Deep convolutional neural networks (DCNN)
Connected layer, 296 AlexNet, 235, 241–243
Contrastive divergence (CD), 104 brain tumor detection
Contrast-limited adaptive histogram equalization design, 249
technique (CLAHE), 294 simulation experiments and evaluation
Conventional segmentation methods, 263–264 criteria, 250–253
Index 317

CNNs layers, 236–239 Digital Surface Model (DSM), 195

convolutional layers, 237–238 Discriminant power (DP), 252
fully connected layers, 239 DLNNs, see Deep learning neural networks
max-pooling layer, 238 DNNs, see Deep neural networks
nonlinearity, 237, 238–239 DP, see Discriminant power
rectifed linear unit, 237, 238–239 DPM, see Deformable Parts Models
CNNs types DRBM, see Discriminative restricted Boltzmann
fully convolutional networks, 245 machines
hybrid learning networks, 245 Dropout layer, 302–303
region-based convolutional neural Dropout property, 240
networks, 244–245 DSM, see Digital Surface Model
deep learning applications, medical imaging, DSP, see Digital signal processor blocks
245–248 DSSD, see Deconvolutional single-shot detector
feature extraction, 247 Dynamic fxed point, 28–29
tumor detection, 247–248
tumor segmentation, 248 E
deep residual networks, 243
GoogleNet, 241, 243 Earth Observation (EO), 186
operational properties, 241–242 Energy consumption, 11
optimization approaches, 239–241 Extreme quantization, 29–30
BN, 240–241
dropout property, 240 F
gradient-based optimization, 240
network depth, 241 Face identifcation experiments
tools and software, 244 ability
VGG-19, 243 FaceDeepNet system, 114
DeepLab network, 266 FRGC V 2.0 and UFI, 114
Deep learning applications, medical imaging, performance, 115
245–248 SDUMLA-HMT dataset, 117
feature extraction, 247 state-of-the-art approaches, 114
tumor detection, 247–248 FDT-DRBM approach, parameter settings
tumor segmentation, 248 CD learning algorithm, 109
Deep learning neural networks (DLNNs), 294 CNN and DBN, 111
Deep neural networks (DNNs), 90, 98, 213 CT-FBM, 109
Deep Residual Learning (ResNet) DBN model, 112
degradation, 52 FaceConvNet-A system, 111
design, 53 FaceConvNet-B system, 112
desired mapping, 52 IrisConvNet system, 111
identity mappings, 52 layers, 112
optimal mappings, 53 RBM model, 112
skip connection, 53 SDUMLA-HMT, 111, 113
Deep residual networks (ResNet), 243 10-fold cross-validation, 109
Deformable Parts Models (DPM), 43 VAR, 109
DermIS databases, 300 Face Recognition Grand Challenge Version 2.0
DermIS dataset (FRGC V2.0), 108
vs. DermQuest dataset, 309 Face verifcation experiments
ROC curves, 310 FaceDeepNet system, 119
Dermoscopy, 262 FDT-DRBM approach, 118
DermQuest databases, 300 Image-Restricted, Label-Free Outside Data
DermQuest dataset protocol, 117
vs. DermIS, 309 LFW datasets, 117
ROC curves, 310 performance, 118
Design space exploration, 21–22 Unrestricted, Labeled Outside Data
Detection performance, 11 protocol, 118
Dice coeffcient, 274 False negative rate (FNR), 251, 252
Differential box-counting (DBC), 102 False positive rate (FPR), 251
Digital signal processor (DSP) blocks, 10, 16, 30 Faster R-CNN
318 Index

anchors, 46 GIoU, see Generalized Intersection-over-Union

lower-dimensional vector, 46 Global Navigation Satellite System (GNSS), 184
RPN, 46 GM, see Geometric mean
two-stage detectors, 46 GNSS, see Global Navigation Satellite System
Fast Fourier Transform (FFT), 16, 17 GoogleNet, 241, 243
Fast R-CNN, 45, 46 Google Satellite Images (GSI), 195, 196, 202
FCLs, see Fully connected layers GPUs, see Graphics processing units
FCNs, see Fully convolutional networks Gradient-based optimization, 240
FCRN, see Fully convolutional residual networks Gradient descent (GD) computation, 240
FDT, see Fractal dimension transformation Graphical representation-based HFR method
FDT-DRBM approach (G-HFR), 94
DRBM Graphics processing units (GPUs), 2, 94, 243,
CD algorithm, 104 246, 296, 297, 304
phase, 105 GRU, see Gated recurrent unit
FDT GSI, see Google Satellite Images
DBC algorithm, 102
kernel function, 102 H
open set identifcation, 102
sigmoid function, 101 Handcrafted-descriptor approaches
Viola-Jones face detector, 101 challenges, 93
Feature pyramid networks (FPN), 44 feature-based, 93
FFT, see Fast Fourier Transform groups, 93
Field-programmable gate array (FPGA), 2, 70 holistic-based, 93
implementations, 22–26, 29, 32 hybrid-based, 93
Fine-tuning, 3, 241 Heterogeneous face recognition (HFR), 94
Fixed-point arithmetic, 23, 27 High-level intuitive features (HLIFs), 293
FNR, see False negative rate High-rank flters, 32
FPGA, see Field-programmable gate array HLIFs, see High-level intuitive features
FPGA-based deep learning, 11–12 Hue-saturation-value (HSV), 268, 269
FPN, see Feature pyramid networks Hu moment, 191
FPR, see False positive rate Hybrid learning networks, 245
FrCN, see Full-resolution convolutional networks
FRGC V2.0, see Face Recognition Grand I
Challenge Version 2.0
Full-resolution convolutional networks (FrCN), IDBC, see Improved differential box counting
263, 267, 275, 277, 279, 282, 283 im2col transformation, 13–14
architecture, 270–272 ImageNet large-scale visual recognition
convolutional and subsampling layers, challenge (ILSVRC), 234, 296
269–270 Improved differential box counting (IDBC)
cross-entropy loss function, 272 algorithm, 95
upsampling layers, 270 IMU, see Inertial measurement unit
Fully connected layers (FCLs), 5, 239 Inception v3, 294
Fully convolutional networks (FCNs), 162, 245, Inertial measurement unit (IMU), 184
265, 267, 277–280 Inference, 3
Fully convolutional residual networks (FCRN), 266 Initial deep learning segmentation methods,
264–265
G Intel Core i7 processor, 250
Inter-convolution parallelism, 9
Gated recurrent unit (GRU), 132 Inter-FM parallelism, 8
GD, see Gradient descent computation Inter-layer pipeline parallelism, 8
GEMMs, see General matrix multiplications International Skin Imaging Collaboration (ISIC),
Generalized Intersection-over-Union (GIoU), 54 263, 267
General matrix multiplications (GEMMs), 13 International Symposium on Biomedical Imaging
Geometric mean (GM), 252 (ISBI), 267
G-HFR, see Graphical representation-based HFR Intra-convolution parallelism, 9
method Intra-FM parallelism, 9
Index 319

ISBI, see International Symposium on evaluation metrics, 274–275

Biomedical Imaging full-resolution convolutional networks,
ISIC, see International Skin Imaging 269–272
Collaboration key component optimization, 275–277
training and testing, 272–273
J skin lesion segmentation results
ISIC 2017 test dataset segmentation
Jaccard index (JAC), 274, 281 performance, 277–280
network computation, 277–280
L segmentators comparisons, 280–282
Low rank approximation, 31–32
Labelled Faces in the Wild (LFW), 108 Low-rank flters, 32
LBP, see Local binary pattern LSTM, see Long short-term memory
Learning-based approaches
advantage, 94 M
category, 94
DeepFace framework, 94 MAC, see Multiply accumulate operation
problem, 94 Magnetic resonance imaging (MRI), 248
Lesion indexing network (LIN), 266 Matthews correlation coeffcient (MCC), 252,
LFW, see Labelled Faces in the Wild 274, 278, 282
LIN, see Lesion indexing network Max-pooling layer, 238, 249
Line of sight (LOS), 184 MCC, see Matthews correlation coeffcient
Local binary pattern (LBP), 93, 293 MED-NODE dataset, 299, 310, 311
Long short-term memory (LSTM) Melanoma, 262, 292, 305
action and activity recognition methods Memory accesses, 9–10
C3D network, 132 Minibatch normalization, 241
components, 135 MRI, see Magnetic resonance imaging
DB-LSTM, 132 MS COCO dataset, 55
framework learning, 133 Multi-biometric face recognition system
motion-based attention, 133 advantages, 90
network architecture, 138 approach
outcomes, 139–143 CNN, 99–100
RACNN, 135 DBN, 98–99
3DCNNs and RGB, 137 fractal dimension, 95
bidirectional, 131–132 RBM, 96–98
defnition, 130 datasets
GRU, 132 FRGC V 2.0, 108
image models, 144–145 LFW, 108
multi-layer, 131 SDUMLA-HMT, 107–108
multiple actions and activity recognition, 145 UFI, 108
utilization, large-scale datasets, 145 experiments (see Face identifcation
Loop tiling, 20–21 experiments; Face verifcation
Loop unrolling, 19–20 experiments)
LOS, see Line of sight methodology
Loss layer, 296 FaceDeepNet system, 105–106
Lossless deep learning FrCN, for skin lesion FDT, 102–103
boundary segmentation RBM, 104–105
conventional segmentation methods, 263–264 types, 91
deep learning methods, 265–266 Multiple LSTM, 131
deep learning works, 266–267 Multiply accumulate (MAC) operation, 6
initial deep learning segmentation methods,
264–265 N
materials and methods
data preprocessing and augmentation, Negative predictive value (NPV), 251
267–269 Network depth, 241
dataset, 267 Nevus, 305
320 Index

Nonlinearity, 238–239 ResNet, see Deep residual learning; Deep

Nonlinear rectifed linear unit layer, 296 residual networks
Normalized Otsu’s segmentation (NOS) Resnet50, 297–298, 300, 301, 303, 311
method, 294 performance, 307
NPV, see Negative predictive value Resource utilization, 11
Restricted Boltzmann machine (RBM), 94, 96–98
O RetinaNet
accuracy, 53
OP, see Optimization precision GIoU, 54
OpenStreetMap (OSM), 195, 196, 203, 205 RGB images, 268, 269
Optimization precision (OP), 252 RIDER, see Reference image database to
OSM, see OpenStreetMap evaluate response
OverFeat, 43 RNNs, see Recurrent neural networks
Robot vision
P category, 224
challenges
Parallel integration (PI) technique, 266 biological learning, 223
Parallelism, CNNs, 8–9 geometry-based methods, 224
PASCAL VOC dataset, 54, 55 open-world learning, 223
PCA, see Principal component analysis power and memory, 222
Per-pixel classifcation, 245 solving problems, 224
PEs, see Processing elements STM, 223
PI, see Parallel integration technique visual system, 223
Pooling layers, 5, 296 CNNs, 212
Positive prediction value (PPV), 251 GPUs, 212
Power consumption, 11 object detection and categorization
PPV, see Positive prediction value detectors, 214, 215
Principal component analysis (PCA), 293 image processing and DNNs, 216
Processing elements (PEs), 16, 18 LIDAR data, 214
Pseudo-binary networks, 30 RGBD data, 214
Python, 244 Selective Search/EdgeBoxes, 214
sensors, 214
R soccer, 216
state-of-the-art methods, 213
RACNN, see Recurrent attention CNN task-dependent methods, 213
RBM, see Restricted Boltzmann machine object grasping and manipulation
R-CNN, see Region-based convolutional neural graph-based/Jacobian-based
networks algorithms, 216
Rectifed linear units (ReLUs), 46, 164, 238–239, image-based reinforcement learning, 217
271, 272 multimodal data, 217
Recurrent attention CNN (RACNN), 135 network layers, 216
Recurrent neural networks (RNNs), problem, 216
129–130, 235 requirements, 212–213
Reduced computations, CNNs, 30–32 scene representation and classifcation
FPGA implementations, 32 place recognition, 218
low rank approximation, 31–32 semantic segmentation, 218
weight pruning, 31 3D geometry, 220
Reference image database to evaluate response visual navigation, 218–220
(RIDER), 250 spatiotemporal vision
Region-based convolutional neural networks human action recognition, 221
(R-CNN), 44, 45, 244–245 object recognition, 221
Region-based methods, 264 UCF-101 dataset, 221
Region of interest (ROI), 294, 305 video, computer vision techniques, 222
Region proposal network (RPN), 47 Tensorfow, 211
ReLU function, 5 Roofine method, 21
ReLUs, see Rectifed linear units RPN, see Region proposal network
Index 321

S Skin lesions, 261, 262

Skin lesion segmentation results
SAE, see Stacked autoencoder (SAE) ISIC 2017 test dataset segmentation
Scale-invariant feature transform (SIFT), 93, performance, 277–280
184, 187 network computation, 277–280
Scene parsing, 245 segmentators comparisons, 280–282
SEER, see Surveillance, Epidemiology, and End Skin melanoma classifcation, DCNN
Results CNNs
SegNet architecture, 265, 266, 277 Alex-net model, 296–297
Semantic segmentation, 245 Resnet50, 297–298, 308
automotive datasets experimental results
real datasets, 160 Alex-net performance, 304–306
synthetic datasets, 161 Resnet50 performance, 307
automotive scene literature review, 293–294
multi-camera network, 159 methodology
road pixels, 160 augmentation process, 299–301
zoom augmentation, 159 dropout layer, 302–303
data pre-processing, 189 transfer learning, 301–302
end-to-end training, 158 preprocessing, 294–295
experiment, 201 transfer knowledge, DNNs, 298–299
loss function, 190 Skip connection, 297
metrics, 201 Softmax classifer, 272
multiscale, 158 Softmax layer, 6
network, 189 Sørensen index, 190
optimization algorithm, 190 Spatial architectures, 18–21
outcomes, 201–205 loop optimization
pixel-wise annotation, 157 tiling, 20–21
regularization and normalization, 190 unrolling, 19–20
sequence frames Specifcity (SPE), 251, 274, 304
frame rate, 193 Speeded-Up Robust Features (SURF), 184
ORB, 193 SSD, see Single-shot detector
SURF and SIFT, 193 Stacked autoencoder (SAE), 247
shape matching State-of-the-art CNN models, 8
calibration step, 192 Stochastic gradient descent (SGD), 240,
computer vision techniques, 190 276, 302
contextual knowledge, 193 Subsampling layers, 269
Hausdorff distance, 191 Supervised segmentation approaches, 264
Hu moment, 191 Support vector machine (SVM), 248, 293, 309
outcomes, 205–206 SURF, see Speeded-Up Robust Features
Siamese Network, 185 Surveillance, Epidemiology, and End Results
SLAM, 184 (SEER), 262
training, 201 SVD, see Single-value decomposition
transfer learning, 190 SVM, see Support vector machine
Sensitivity, 251, 274, 304 Systolic arrays, 16, 18
SGD, see Stochastic gradient descent
Siamese Network, 185
SIFT, see Scale-invariant feature transform T
Simultaneous Localization And Mapping TensorFlow, 211, 244
(SLAM), 184 Theano, 244
Single-precision foating-point representation, 23 Thresholding methods, 263
Single-shot detector (SSD) Tiles, 14
architecture, 50, 51 TNNs, see Traditional neural networks
DSSD, 51 TNR, see True negative rate
one-stage detector, 50 Torch, 244
Single-value decomposition (SVD), 32 TPR, see True positive rate
Skin cancer, 262 Traditional neural networks (TNNs), 236, 237
322 Index

Training, 3 VGG-19, 243

Transfer learning, 298, 301–302 VGGnet, 241
True negative rate (TNR), 251 Viola-Jones face detector, 101
True positive rate (TPR), 251 VisDrone-DET2018 dataset, 55–58
2D Gabor Wavelet, 93
2D-separable flter, 31 W

U Weight pruning, 31
Winograd transform, 14–15
UAVs, see Unmanned aerial vehicles Workloads and computations, CNN models
Ubuntu, 194 computational workload, 6–8
Unconstrained Facial Images (UFI), 108 frameworks, 10
U-Net method, 265, 267, 277–280 hardware, 10
Unmanned aerial vehicles (UAVs) libraries, 10
applications, 184 memory accesses, 9–10
calibration, 187–188 parallelism, 8–9
datasets
ISPRS WG II/4 Potsdam, 195 X
videos, 196–197
defnition, 183 Xilinx ZCU102 device, 15
EO image registration
calibration component experiment, 198 Y
local feature registration results, 199–200
sequence experiment, 198–199 YI, see Youden’s index
framework architecture, 184–187 YOLO, see You Only Look Once
GNSS, 184 YOLO9000, 50
IMU, 184 YOLOv2, 49–50
LOS, 184 YOLOv3, 50
platform, 194 Youden’s index (YI), 252
semantic segmentation (see Semantic You Only Look Once (YOLO)
segmentation) architecture, 49
Unsupervised clustering techniques, 263–264 bounding boxes, 48
Unsupervised pretraining, 234 grid cell, 48
network design, 48
V
Z
Validation accuracy rate (VAR), 109
VGG16 classifcation model, 265 Zero skip scheduler, 32