World Academy of Science, Engineering and Technology 10 2005

On-line Handwritten Character Recognition: An

Implementation of Counterpropagation Neural Net
Muhammad Faisal Zafar, Dzulkifli Mohamad, and Razib M. Othman

line handwritten scripts are usually dealt with pen tip traces
Abstract—On-line handwritten scripts are usually dealt with pen from pen-down to pen-up positions.
tip traces from pen-down to pen-up positions. Time evaluation of the There is extensive work in the field of handwriting
pen coordinates is also considered along with trajectory information. recognition, and a number of reviews exist. General
However, the data obtained needs a lot of preprocessing including methodologies in pattern recognition and image analysis are
filtering, smoothing, slant removing and size normalization before presented in [25]. Character recognition is reviewed in [1, 6,
recognition process. Instead of doing such lengthy preprocessing, this
11, 19, 32] for off-line recognition, and in [28, 29] for on-line
paper presents a simple approach to extract the useful character
information. This work evaluates the use of the counter- propagation
recognition. Most of the researchers have chosen numeric
neural network (CPN) and presents feature extraction mechanism in characters for their experiment [2, 3, 10, 12, 15, 16]. So, some
full detail to work with on-line handwriting recognition. The maturity can be observed for isolated digit recognition.
obtained recognition rates were 60% to 94% using the CPN for However, when we talk about the recognition of alphabetic
different sets of character samples. This paper also describes a characters, the problem becomes more complicated. The most
performance study in which a recognition mechanism with multiple obvious difference is the number of classes that can be up to
thresholds is evaluated for counter-propagation architecture. The 52, depending if uppercase (A–Z) and lowercase (a–z)
results indicate that the application of multiple thresholds has characters are distinguished from each other. Consequently,
significant effect on recognition mechanism. The method is there is a larger number of ambiguous alphabetic characters
applicable for off-line character recognition as well. The technique is
other than numerals. Character recognition is further
tested for upper-case English alphabets for a number of different
styles from different peoples. complicated by other differences such as multiple patterns to
represent a single character, cursive representation of letters,
Keywords—On-line character recognition, character digitization, and the number of disconnected and multi–stroke characters
counter-propagation neural networks, extreme coordinates [18]. Few researches have addressed this complicated subject.
In fact, it can be said that character recognition still an open
I. INTRODUCTION problem [6].
Neural Nets (NN) and Hidden Markov Models (HMM) are
H ANDWRITING processing is a domain in great
expansion. The interest devoted to this field is not
explained only by the exciting challenges involved, but also
the popular, amongst the techniques which have been
investigated for handwriting recognition. It has been observed
that NNs in general obtained best results than HMMs, when a
the huge benefits that a system, designed in the context of a similar feature set is applied [17]. The most widely studied
commercial application, could bring [27]. Two classes of and used neural network is the Multi-Layer Perceptron (MLP)
recognition systems are usually distinguished: online systems [5]. Such an architecture trained with back-propagation [20] is
[4, 24, 34] for which handwriting data are captured during the among the most popular and versatile forms of neural network
writing process, which makes available the information on the classifiers and is also among the most frequently used
ordering of the strokes, and offline systems [33] for which traditional classifiers for handwriting recognition. See [37] for
recognition takes place on a static image captured once the a review. Other architectures include Convolutional Network
writing process is over. The field of personal computing has (CN) [21], Self-Organized Maps (SOM)[38], Radial Basis
begun to make a transition from the desktop to handheld Function (RBF) [5], Space Displacement Neural Network
devices, thereby requiring input paradigms that are more (SDNN) [26], Time Delay Neural Network (TDNN)[22],
suited for single hand entry than a keyboard. Online Quantum Neural Network (QNN) [39], and Hopfield Neural
handwriting recognition allows for such input modalities. On- Network (HNN) [23].
Few attempts have been found in the literature in which
counter-propagation (CPN) architecture has been used for the
Muhammad Faisal Zafar is doctoral candidate in the Faculty of Computer recognition of handwritten characters. Ahmed et al., [2] made
Science nd Information System at the Universiti Teknologi Malaysia (phone: an attempt but only for digit recognition. The main objective
+60 7 5532339 Fax: +607 5565044; e-mail: faisal@ gmm.fsksm.utm.my). of this work is the implementation of the CPN for the
Dzulkifli Mohamad is Associate Professor at Faculty of Computer Science
nd Information System at the Universiti Teknologi Malaysia (e-mail: recognition of online upper case English alphabets and to
[email protected]). evaluate its performance. Although this study deals with a
Razib M. Othman He is currently working toward the Ph.D. at the Software limited number of 26 upper case character classes, there is a
Engineering Department, Universiti Teknologi Malaysia (e-mail: space to extend this work for all alphanumeric characters
[email protected]).

232
 World Academy of Science, Engineering and Technology 10 2005

including lower case characters. This work also suggests a

Online input
Online from
input
simple approach to extract the useful character information. tablet
The performance CPN architecture is evaluated based on f bl
extracted character information by applying this simple
feature extraction approach. Extreme coordinates
Feature Extraction measurement
The recognition system applies a global decision module
which decides either to accept the recognition result or reject
it. In classification step, a pattern is considered ambiguous if it Grabbing character into grid
cannot be reliably assigned to a class, whereas a pattern Training Data set
assigned low confidence for all hypothesized classes can be
treated as an outlier. Three different criteria (thresholds) of Character digitization
decision making have been applied.
This paper is organized as follows. The Section 2 gives an
overview of the proposed system. Section 3 describes the Classifier
Trained data weight (Neural Network)
feature extraction steps from the handwritten character. File
Section 4 introduces CPN architecture. In Section 5 the
experimental results are provided with some analyses and Classification Result
discussions. Section 6 presents the concluding remarks and
future work.
Fig. 1 Block diagram of the system.

II. SYSTEM OVERVIEW

This section describes the simple technique involved in our
B. Character Detection
proposed online handwriting recognition system. This is a
writer-independent system based on the neural net method. As stated earlier that character is written with white digital
Conventionally, the data obtained needs a lot of preprocessing ink, so the algorithm for character detection is quite simple. It
including filtering, smoothing, slant removing and size searches from left to right for white pixels starting from left-
normalization before recognition process. Instead of doing top corner of the area specified for writing. A trace of a white
such lengthy preprocessing, we present a simple approach to pixel is the indication of the presence of a character.
extract the useful character information. The whole process
requires no preprocessing and size normalization. The method C. Calculating the Number of Rows
is applicable for off-line character recognition as well. A The algorithm searches for the presence and absence of
block diagram of the proposed online recognition system of white pixels going from top to bottom. The continuous
isolated roman characters is shown in Fig 1. The flow of data absence of white pixels (or presence of black pixels) could be
during training is shown by the dashed line arrows, while the a gap between two rows. To make sure whether it is a gap,
data flow during recognition is shown by solid line arrows. algorithm searches from left to right against every black pixel,
The input to the system is a sequence of handwritten character if there is no trace of white pixel for the entire row, the gap is
patterns. After receiving input from tablet the extreme confirmed. In this way, all the horizontal gaps, in the image,
coordinates i.e. left, right, top, and bottom are calculated. are traced out and from this, number of rows are calculated.
Then character is captured in a grid as shown in Fig. 3 and Fig. 2 shows a sequential algorithm for finding the number of
sensing the character pixels in grid boxes, the character is rows in the document.
digitized in a binary string. This binary string is applied at the
input of CPN for training and recognition. Grid size of i. Find the text boundary of the whole image by scanning from top to
bottom for upper border Y1, left to right for left border X1, right to
14x8(i.e. 14 rows and 8 columns) were used in the left for right borderX2 and bottom to top for lower borderY2.
experiments.
ii. Scan the binary image from Y1 towards Y2,
A. Data Acqusition iii. If there is a black pixel, then scan from X1 to X2 for that particular
Tablet SummaSketch III has been used to take the samples row to detect any white pixel.
from different subjects. Upper case alphabets characters have iv. If no white pixel found, there is a row gap.
been used. Each subject was asked to write on tablet board
v. Repeat steps ii - iii for the whole image to find total number of row
(writing area). No restriction was imposed on the content or gaps.
style of writing; the only exception was the stipulation on the
isolation of characters. The writers consisted of university Fig. 2 Sequential algorithm for finding the number of rows
students (from different countries), professors, and employees
in private companies. The simulation of each written character
could be seen on computer screen as white digital ink with
black background. Thus one can make use of black and white
colours for some useful processing of written characters.

233
 World Academy of Science, Engineering and Technology 10 2005

D. Character Boundary Calculation IV. NEURAL NET APPROACH

After detecting the character, the next step is to calculate its Neural network classifiers exhibit powerful discriminative
boundary. The algorithm checks from left to right and top to properties and they have been used in handwriting recognition
bottom for left, top, right and bottom boundaries of the particularly with digits, isolated characters, and words in small
character. While going from left to right, the first white pixel vocabularies(Alessandro et al., 2002). In this research work,
is the left boundary and last white pixel is the right boundary Counterpropagation Neural network (CPN) have been
of the character. Similarly from top to bottom, first white pixel implemented and its performances has been evaluated. This
is the top boundary and last white pixel is the bottom section presents a brief introduction of the architecture.
boundary of the character. If there is a vertical gap between A. Counterpropagation Neural Network
two portion of a same character, e.g., ‘ H’ then the algorithm
also check from top to bottom for that particular area. The Hecht-Nielsen [13] proposed CPN as an alternate function
presence of white pixel will eliminate the doubt of a true gap. approximator which can be developed on the available input-
Similar checks are employed for horizontal gaps between two output data. The first counter-propagation network consisted
portion of a same character like ‘S’ and ‘F’. In this way of a bi-directional mapping between the input and output
boundaries of the characters in a row are calculated and stored layers. In essence, while data is presented to the input layer to
in a data base. After calculating the total number of characters generate a classification pattern on the output layer, the output
in a row, the individual width and height of each character is layer in turn would accept an additional input vector and
measured. generate an output classification on the network's input layer.
The network got its name from this counter-posing flow of
III. FEATURE EXTRACTION information through its structure. Most developers use a uni-
flow variant of this formal representation of counter-
In the proposed online handwriting recognition system, propagation. In other words, there is only one feed-forward
feature extraction consists of three steps: extreme coordinates path from input layer to output layer. The forward-only
measurement, grabbing character into grid, and character counterpropagation network architecture, consists of three
digitization. The handwritten character is captured by its slabs: an input layer (layer 1) containing n fan out units that
extreme coordinates from left /right and top/bottom and is multiplex the input signals x1, x2,….., xn, (and m units that
subdivided into a rectangular grid of specific rows and supply the correct output signal values y1,y2 ,.., ym to the
columns. The algorithm automatically adjusts the size of grid output layer), a middle layer (layer 2 or Kohonen layer) with
and its constituents according to the dimensions of the N processing elements that have output signals zI, z2,..., zN,
character. Then it searches the presence of character pixels in and a final layer (layer 3) within processing elements having
every box of the grid. The boxes found with character pixels output signals y1’, y2’,…, ym’. The outputs of layer 3
are considered “on” and the rest are marked “off”. A binary represent approximations to the components y1, y2,…, ym of y
string of each character is formed locating the “on” and “off” = f(x) [14].
boxes (named as character digitization) and presented to the The underlying principle for CPN is simple: for a given
neural network input for training and recognition purposes. independent variable vector I not present in the available data
The total number of grid boxes represented the number of set, find the independent variable vector in the data set closest
binary inputs. A 14x8 grid thus resulted in 112 inputs to the to I. The criterion of closeness in the n-dimensional Euclidean
recognition model. An equivalent statement would be that a space can either be distance based (minimum Euclidean
14x8 grid provided a 112 dimensional input feature vector. distance) or can be angle based (minimum angle between
The developed software contains a display of this vectors of normalized lengths). If Xk is the vector found
phenomenon by filling up the intersected squares. The effect closest in the data set , then the value of f(I) can be
has been produced in Fig. 3. approximated as the dependent variable value corresponding
to Xk. This technique runs into problems when the data set
becomes very large [8]. During training, these "correct" values
are supplied to the units of the final layer from, the input units
of layer 1.
During training the network is exposed to examples of the
mapping f. After each xk is selected, yk = f(xk) is determined
and both xk and yk are input to the network. An important
component of training in the CPN is reduction of the data set
into a respective data set of lesser, specified size. This is
achieved and the estimates of the dependent variable values
corresponding to the new (and reduced in number)
independent variable vectors can also be calculated [14]. Thus
CPN actually operates as a closest-match lookup table and
training a CPN is an attempt to appropriately reduce the size
Fig. 3 Steps in Feature Extraction of the lookup table [2].
For the network to operate properly, the input vector must

234
 World Academy of Science, Engineering and Technology 10 2005

be normalized. This means that for every combination of input vectors? Each component of the resultant averaged vector was
values, the total "length" of the input vector must add up to average of the corresponding components of the k vectors.
one. Normalization of the inputs is necessary to ensure that This somewhat simplistic approach is mentioned by Freeman
the Kohonen layer finds the correct class for the problem. & Skapura in their discussion of the CPN ([9], pp 238-258).
Without normalization, larger input vectors, bias many of the This technique is intuitively attractive if the k vectors lie close
Kohonen processing elements such that weaker value input to one another in the n-dimentional Euclidean space (where n
sets cannot be properly classified. Because of the competitive = no. of extracted image features). The under laying
nature of the this layer, the larger value input vectors assumption would be that the clusters of input vector samples
overpower the smaller vectors [36]. A three layered CPN corresponding to different characters do not overlap. The
implements the principles discussed above. For a CPN in its performance of such models discussed in the next section
final form, each PE in the hidden layer represents an indicates that the above assumption was reasonable.
independent variable entry in the (reduced) lookup table; the Seven different data sets: 5 samples/character, 11
weights to one such PE in the hidden layer from all the PEs in samples/character, 22 samples/character 33 samples/character,
the input layer represent the components of the corresponding 44 samples/character, 55 samples/character, and 66
independent variable vector. The hidden layer PE whose samples/character were being experimented to evaluate the
incoming weights are closest to an input vector “wins” the performance of both models with gradually increasing the
competition and provides an output value of +1; all other number of samples/character.
hidden layer PEs supply zero outputs. The weights from the
C. Recognition Performance
hidden layer PEs to the output layer PEs represent the
dependent variable values. A CPN with one linear PE in the As mentioned earlier, models were evaluated on samples
output layer thus behaves as estimating one function. With taken from individuals who did not participate in the initial
multiple PEs in the output layer, the CPN becomes an process of setting up the training data set. This was done
estimator of more than one functions. keeping in view the eventual aim of using the model in
practical online recognition system. The quality of an online
V. EXPERIMENTS handwriting recognizer is related to its ability to translate
drawn characters irrespective of writing styles.
A. Data Set and Model Parameters For developed CPN model, closeness was evaluated by
The data used in this work was collected using tablet measuring the angel between the normalized input and weight
SummaSketch III . It has an electric pen with sensing writing vectors. If I is the normalized input vector and Wi is the
board. An interface was developed to get the data from tablet. normalized weight vector from the input layer to the ith hidden
Anoop and Jain [4] pointed out that the actual device for data layer PE, then the cosine of the angle between the two can be
collection is not important as long as it can generate a found by evaluating the dot product. (Wi . I = | Wi | | I | Cos θi
temporal sequence of x and y positions of the pen tip. = Cos θi) [9]. All the angles between each of the feature
However, the writing styles of people may vary considerably vectors of the unknown character and their closest
on different writing surfaces and the script classifier may corresponding feature vectors in the reference character are
require training on different surfaces. summed and missing or extra feature points are penalized.
Upper case English alphabets were considered in case Identification is then a matter of finding the character in the
study. In the data set, the total number of handwritten look up table that is within a certain threshold angle of the
characters is about 2000 characters, collected from 40 unknown character. Table 1 present the statistics for CPN.
subjects. Experiments were examined with grid size of 14x8. CRs, FRs, and RFs are abbreviation for Correct
Every developed model was tested on characters drawn by Recognitions, False Recognitions, and Recognition Failures
individuals who did not participate in the sample collection respectively.
for data set. Each subject was asked to write on tablet board
TABLE I
(writing area). No restriction was imposed on the content or PERFORMANCE OF CPN MODELS WITH THREE DIFFERENT CRITERIA OF
style of writing; the only exception was the stipulation on the CLASSIFICATION
shape of ‘I’. The grid based character digitization proved
Samples/
improper for characters with negligible width. The shape for ‘Threshold’: NONE ‘Threshold’: 0.5 ‘Threshold’: 0.75
handwritten ‘I’ was thus standardized with horizontal lines at Character
the top and the base. The writers consisted of university CRs FRs RFs CRs FRs RFs CRs FRs RFs
students (from different countries), professors, and employees
5 Each 80% 20% 0% 60% 40% 0% 70% 7% 23%
in private companies.
11 Each 83% 17% 0% 79% 21% 0% 72% 6% 22%
B. Training
22 Each 88% 12% 0% 76% 23% 1% 80% 6% 14%
In the CPN model, the look-up table grows with increase in
training samples. Instead of using Kohonen’s learning 33 Each 92% 8% 0% 84% 15% 1% 83% 4% 13%
algorithm for reducing the size of the look-up table, a much 44 Each 93% 7% 0% 82% 17% 1% 76% 8% 16%
simpler technique was employed. Since there were k samples
55 Each 87% 13% 0% 88% 8% 4% 86% 3% 11%
for a character in a particular model, why not reduce the k
vectors to one vector by taking the average of the sample 66 Each 94% 6% 0% 93% 6% 1% 92% 1% 7%

235
 World Academy of Science, Engineering and Technology 10 2005

D. Performance Analysis (2002). Genetical Engineering of Handwriting Representations.

Proc. of the International Workshop on Frontiers in
For developed CPN models, there was no need of training Handwriting Recognition (IWFHR), Niagara-On-The-Lake,
parameters nor it is an iterative method like back-propagation August 6-8.
architecture which took a long time for learning. A general [4] Anoop M. Namboodiri, Anil K. Jain. (2004). Online
trend of increase in performance with increase in Handwritten Script Recognition. IEEE Trans. PAMI. 26(1):
124-130.
samples/character has also been observed in this case. The
[5] Bishop M. (1995). Neural Networks for Pattern Recognition.
difference in recognition rates with and without a threshold Oxford Univ. Press, Oxford-U.K.
for input classification is understandable (Table I). Though [6] Bortolozzi, F., Britto Jr,. A., Oliveira, L. S. and Morita, M.,
threshold reduces the correct recognitions but at the same time (2005). Recent Advances in Handwriting Recognition. In
it prevents the system to go for more false recognitions. False Umapada Pal et al editors, Document Analysis, 1-31.
Recognition (FRs) is another important factor in any [7] Dzulkifli Mohamad, and Zafar, M. F., (2004a). Recognition of
recognition system, lower the false recognition rate, more Complex Patterns Using a Novel Feature Vector by
Backpropagation Neural-Network. Pattern Recognition And
reliable the system [7, 35]. Instead of FRs, system goes for Image Analysis. 14(3): 479-487.
recognition failure (RFs) which is less dangerous than FRs. [8] Dzulkifli Mohamad, and Zafar, M. F., (2004b). Comparative
On the other hand, performance of the system increases Study of Two Novel Feature Vectors For Complex Image
without threshold but at cost of more FRs. Fig. 4 presents a Matching Using Counterpropagation Neural Network. Journal
graphical overview of CPN performances with three different of Information Technology, vol.16, No.1, FSKSM, UTM.
decision criteria of Recognition. [9] Freeman, J. A., and Skapura, D. M., (1991). Neural Networks:
Algrithms, Applications and Programming Tecniques, Addison-
5 Each 11 Each 22 Each 33 Each 44 Each 55 Each 66 Each Wesley Publishing Company.
100% [10] Gader P. D., Forester B., Ganzberger M., A. Billies, B.
90% Mitchell, M. Whalen, T. Youcum (1991). Recognition of
80% handwritten digits using template and model matching. Pattern
70%
Recognition, 5(24):421-431.
60% [11] Govindan, V. K.. and Shivaprasad, A. P. (1990), Character
50%
Recognition-a Review, Pattern Recognition, 23: 671-683.
40%
[12] Hebert Jean-Francois, Parizeau Marc and Nadia Ghazali.
30%
(1998). A new fuzzy geometric representation for on-line
isolated character recognition. Proc. of the 14th International
20%
Conference on Pattern Recognition, Brisbane: 1121-1123.
10%
[13] Hecht-Nielsen, R. (1987). Counterpropagation networks. Proc.
0%
CRs FRs RFs CRs FRs RFs CRs FRs RFs Of IEEE First Int’l Conference on Neural Networks. 1987 II:
‘Threshold’: NONE ‘Threshold’: 0.5 ‘Threshold’: 0.75
19-32.
[14] Hecht-Nielsen Robert.(1990). Neurocomputing. Addison-
Wesley Publishing Company.
Fig 7 Graphical Presentation of CPN performances with three [15] Hiroto Mitoma, Seiichi Uchida, and Hiroaki Sakoe. (2005).
different criteria of Recognition Online character recognition based on elastic matching and
quadratic discrimination. Proceedings of 8th International
Conference on Document Analysis and Recognition (ICDAR
VI. CONCLUSION 2005, Seoul, Korea).1(2):36—40.
From the results it can be concluded that CPN is a good [16] Iqbal, A., Zafar, M. Faisal., (1998). Pc Based Optical
Objective Test Marking System Using Neural Networks.
promise in terms of recognition capability which has not been International Workshop on Recent Advances in Computer
put on trial in the field of handwriting recognition.. More over Vision proc. SZABIST, Karachi, Pakistan,: 41-50.
CPN is more economical than convergence of other NN [17] Kapp, M.N., Freitas, C. and Sabourin, R. (2004).Handwritten
architectures e.g. back-propagation where the training time Brazilian Month Recognition: An Analysis of Two NN
can take long time. The experiments provided the authors an Architectures and a Rejection Mechanism. 9th International
Workshop on Frontiers in Handwriting Recognition (IWFHR-
opportunity to explore this pattern recognition methodology; 9), Tokyo, Japan,
the exercise provided a theoretical base for further [18] Koerich L. (2002). Large Vocabulary Off-Line Handwritten
investigations and impetus for development work in this Word Recognition. PhD thesis, École de Technologie
discipline. The obtained results motivate the continuity of the Supérieure, Montreal-Canada.
system development considering a preprocessing mechanism [19] Koerich A. L., Sabourin R., Suen C. Y. (2003). Large
including normalization and slant removal. Other future work vocabulary off-line handwriting recognition:A survey. Pattern
Anal Applic, 6: 97–121.
might involve some new feature extraction approaches.
[20] LeCun Y., Bottou L., Orr G. B., Muller K. R. (1998a). Eficient
backprop. In Orr G. and K. Miller, editors, Neural Networks:
REFERENCES Tricks of the Trade. Springer.
[1] Alessandro Vinciarelli (2002). A survey on off-line Cursive [21] LeCun Y., Bottou L., Bengio Y., and Haffner P. (1998b).
Word Recognition. Pattern Recognition, 35: 1433. Gradient-Based Learning Applied to Document Recognition.
Proc. IEEE, 86(11): 2278-2324.
[2] Ahmed, S.M.,et.al, (Nov.1995). Experiments in character
recognition to develop tools for an optical character [22] Lethelier, Leroux M., and Gilloux M. (1995). An automatic
recognition system, IEEE Inc. 1 st National Multi Topic Conf. reading system for handwritten numeral amounts on french
proc. NUST, Rawalpindi, Pakistan, 61-67. checks. In Proc. 3th International Conference on Document
Analysis and Recognition, Montreal-Canada, pp. pages 92-97.
[3] Alexandre Lemieux, Christian Gagne and Marc Parizeau.
[23] Ling, M. Lizaraga, N. Gomes, and A. Koerich (1997). A

236
 World Academy of Science, Engineering and Technology 10 2005

prototype for brazilian bankcheck recognition. In S.Impedovo System (FSKSM), UTM, Malaysia. His research interests include pattern
et al, editor, International Journal of Pattern Recognition and recognition, neural networks, machine print and handwriting recognition.
Artificial Intelligence,World Scientific, pp. 549-569.
[24] Liu Cheng-Lin, Stefan Jaeger, and Masaki Nakagawa (2004). Dzulkifli Mohamad received the BSc degree in computer science and
Online Recognition of Chinese Characters:The State-of-the- statistics from the National University of Malaysia in 1978, the Postgraduate
Art. IEEE Trans. on Pattern Analysis and Machine Intelligence, Diploma in computing science from University of Glasgow in 1981, and the
26(2), 198-203. MSc and PhD degrees in computer science from the University of
[25] Mantas, J. (1986), An Overview of Character Recognition Technology, Malaysia in 1991 and 1997. He has been with the University of
Methodologies, Pattern Recognition, 19 (1986) 425-430. Technology, Malaysia since 1978, where he is currently an associate
professor of computer science. His research interests include image
[26] Matan O., Burges J. C., LeCun Y., Denker J. S. (1992). Multi- processing, pattern recognition, medical imaging, information theory, and
digit recognition using a space displacement neural network. In artificial intelligence.
J. E. Moody, S. J. Hanson, and 165 R. L. Lippmann, editors,
Advances in Neural Information Processing Systems, volume 4, Razib M. Othman received B.Sc. in computer science in 1999 and M.Sc. in
Morgan Kaufmann, pp. 488-495. computer science in 2003 both from the Universiti Teknologi Malaysia. He
[27] Morita M., Sabourin R., Bortolozzi F., Suen C. Y. (2003). A is currently working toward the Ph.D. at Universiti Teknologi Malaysia in
Recognition and Verification Strategy for Handwritten Word the field of computational biology. Since 1999 he has been working as a
Recognition. ICDAR'03), Edinburgh-Scotland: 482-486. lecturer at the Software Engineering Depart-ment, Universiti Teknologi
[28] Nouboud, F., and Plamondon,(1990). On-Line Recognition of Malaysia and researcher at the Artificial Intelligence and Bioinformatics
Handprinted Chara.cters: Survey and Beta Tests, Pattern Laboratory in the same university. His main research interests are genetic
Recognition, 23: 1031-1044. algorithms, software agents, and semantic web.
[29] Plamondon Rejean, and Sargur N. Srihari, (2000). On-Line and
Off-Line Handwriting Recognition: A Comprehensive Survey.
1EEE Transactions on PAMI. 22(1): 63-84.
[30] Rumelhart, D. E., Hinton, G. E., and Williams, R. J., (1986).
Learning internal representations by error propagation. in
Rumelhart, D. E. And McClelland, J. L. [Ed], Parallel
Distributed Processing: Explorations in the Microstructure of
Cognition, 1, 318-362, MIT Press, Cambrige.
[31] Shridhar M. and Badreldin A. (1986). Recognition of isolated
and simply connected handwritten numerals. Pattern
Recognition, 19(1):1-12.
[32] Suen, C. Y., Berthod, M., and Mori, S. (1980), Automatic
Recognition of Ha.nd printed Character-the State of the Art,
Proceedings of IEEE. 68 (1980) 469-487.
[33] Steinherz T., Rivlin E., and Intrator N. (1999). Offline Cursive
Script Word Recognition—A Survey. Int’l J. Document
Analysis and Recognition, vol. 2, 90-110.
[34] Tappert C.C., Suen C.Y., Wakahara T.(1990). The state of the
art in on-line handwriting recognition. IEEE Trans. on Pattern
Analysis and Machine Intelligence, 12(8), 787-808.
[35] Zafar, M. F., and Dzulkifli Mohamad. (2005). Comparison of
Two Different Proposed Feature Vectors For Classification of
Complex Image. Jurnal Teknologi, Universiti Teknologi
Malaysia, 42(D): 65-82.
[36] Zafar, M. F., and Dzulkifli Mohamad. (2002). A Comparison
Of Two Neural Network Techniques On Feature Based
Complex Image Recognition. 2 nd World Engineering Congress,
WEC2002, Kuching, Sarawak, Malaysia.
[37] Zhang G. P. (2000). Neural networks for classification: a
survey. IEEE Transactions on Systems, Man, and Cybernetics -
Part C: Applications and Reviews, 30(4):451-462.
[38] Zhang, Fu M., Yan H., and Fabri M. A. (1999). Handwritten
digit recognition by adaptativesubspace self organizing map.
IEEE Trans. on Neural Networks, 10:939-945.
[39] Zhou J (1999). Recognition and Verification of Unconstrained
Handwritten Numeral. PhD thesis, Concordia University,
Montreal-Canada.

Muhammad Faisal Zafar received the BSc degree in mathematics and

physics in 1990 from Govt, College Faisalabad, Pakistan, MSc degree in
physics in 1993 from Govt. College Lahore, Pakistan, , the Postgraduate
Diploma in computer training in 1995 from Computer Training Centre (CTC)
Islamabad, Pakistan. He has been with Informatics Complex (ICCC),
Islamabad, Pakistan, since 1995 as a scientific / senior scientific officer, where
he worked on pattern recognition tasks. He was offered fellowship from
Universiti Teknologi Malaysia (UTM), in 2001 and he completed his MSc
degree in computer science in 2003 from UTM. Currently he is a PhD
(computer science) student at Faculty of Computer Science and Information

237