Online Handwriting Recognition

for Ethiopic Characters

By
ABNET SHIMELES

A THESIS SUBMITTED TO THE SCHOOL OF GRADUATE

STUDIES OF ADDIS ABABA UNIVERSITY IN PARTIAL
FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE
OF MASTER OF SCIENCE IN COMPUTER SCIENCE

June, 2005
 ADDIS ABABA UNIVERSITY
SCHOOL OF GRADUATE STUDIES
FACULTY OF INFORMATICS
DEPARTMENT OF COMPUTER SCIENCE

Online Handwriting Recognition

for Ethiopic Characters

By
Abnet Shimeles

Name and Signature of Members of the Examining Board

1. Dr. Solomon Atnafu, Advisor __________________

2. __________________

3. __________________
Acknowledgement

First of all, I thank God for being with me all the time, not only during this research, but also in
my whole life. It is all His kindness that brings this happiness to my life.

What would have happened to my thesis paper without my advisor’s valuable ideas and
family-like support? I really thank Dr. Solomon Atnafu, my advisor, for his patience, friendly
treatment, valuable ideas, and supportive advices. I, not only gained academic knowledge, but
also learned humanity from him. I say, from the bottom of my heart, thank you, while I at the
same time feeling that the words are not still expressing my feelings.

‘Netsi’, starting from pushing me to join the M.Sc. program, he was always besides me in all
my activities in the last two years. I hope he will be with me forever. I know that you want me
to grow more. I promise, I will!

My parents all assisted me to stand here. Without their support full of love, I would have never
been ‘Me’. Mom and Dad were worrying about me while I was sitting in front of the PC at
home for long time. My sisters and brothers all follow every of my progress and sometimes
they joked at me which make me laugh. Thank you so much for your prayer, concern and
support.

My colleagues in Unity University College helped me a lot by sharing my burdens so that I can
finish this research. I don’t have words to thank them.
Table of Contents
1. Introduction----------------------------------------------------- 1
1.1. Background of the Study -------------------------------------------------------------------3
1.2. Problem Statement --------------------------------------------------------------------------4
1.3. Objectives--------------------------------------------------------------------------------------4
1.4. Justification of the study -------------------------------------------------------------------5
1.5. Scope and Limitation of the Study -------------------------------------------------------5
1.6. Methods of the Study------------------------------------------------------------------------6
1.6.1. Literature Review --------------------------------------------------------------------6
1.6.2. Design of the Online Handwriting Recognition System -----------------------6
1.6.3. Development of the Ethiopic Online Handwriting Recognition Engine-----7
1.6.4. Organization of the Thesis ----------------------------------------------------------7

2. Online Handwriting Recognition --------------------------- 8

2.1. Introduction to Pattern Recognition-----------------------------------------------------8
2.2. Character Recognition----------------------------------------------------------------------9
2.2.1. Handwriting Recognition -----------------------------------------------------------9
2.2.2. Online Vs Offline Handwriting Recognition Systems--------------------------10
2.3. Online Handwriting Recognition ---------------------------------------------------------12
2.3.1. Categories of Online Handwriting Recognition ---------------------------------12
2.3.1.1. Constrained Vs Unconstrained Systems-------------------------------12
2.3.1.2. Writer-dependent Vs Writer-independent Systems------------------13
2.3.2. Issues in Online Handwriting Recognition ---------------------------------------13
2.3.3. Major Steps in Online Handwriting Recognition -------------------------------15
2.3.3.1. Preprocessing --------------------------------------------------------------16
2.3.3.1.1. Noise Elimination -------------------------------------------16
2.3.3.1.2. Normalization------------------------------------------------17
2.3.3.2. Segmentation --------------------------------------------------------------18
2.3.3.3. Feature Extraction --------------------------------------------------------19
2.3.3.4. Classification --------------------------------------------------------------20

3. Review of the State of the Art ------------------------------- 21

3.1. Major Classification Approaches for Character Recognition ---------------------21
3.1.1. Pattern Representation---------------------------------------------------------------21
 3.1.2. Pattern Matching Techniques ------------------------------------------------------23
3.2. Review of Relevant Papers On Online Handwriting Recognition Systems -----24
3.2.1. Latin Online Handwriting Recognition Systems --------------------------------25
3.2.1.1. Kam-Fai Chab et al.: Flexible Structure Matching ------------------25
3.2.1.2. G. Louden et al.: Discrete Hidden Markov Model ------------------27
3.2.1.3. H.Shu: Discrete Hidden Markov Model-------------------------------28
3.2.1.4. E. Gomez et al.: Fuzzy Neural Network ------------------------------30
3.2.2. Japanese/Chinese Online Handwriting Recognition Systems ----------------31
3.2.2.1. S. Jaeger et al.: The State-of-the-art in Japanese Online Handwriting
Recognition ----------------------------------------------------------------31
3.2.2.2. C.L.Liu et al.: The State-of-the-art in Chinese Online Handwriting
Recognition ----------------------------------------------------------------35
3.2.3. Arabic Online Handwriting Recognition Systems ------------------------------36
3.2.4. Handwriting Recognition Systems for Ethiopic Characters -------------------40
3.2.4.1. Amharic Handwriting Recognition for Processing of Bank Checks
-------------------------------------------------------------------------------40
3.2.4.2. Amharic Handwriting Recognition for Postal Address Interpretation
-------------------------------------------------------------------------------41
3.2.4.3. Traditional Amharic Handwriting (“Yekum Tsifet”) Recognition 42
3.3. Summary---------------------------------------------------------------------------------------43

4. The Ethiopic Characters ------------------------------------- 45

4.1. Introduction -----------------------------------------------------------------------------------45
4.2. The Ethiopic Character Set----------------------------------------------------------------46
4.3. Analysis of the Ethiopic Character Set--------------------------------------------------48
4.3.1. First Order (Basic) Characters -----------------------------------------------------48
4.3.2. Second Order Characters------------------------------------------------------------48
4.3.3. Third Order Characters --------------------------------------------------------------49
4.3.4. Fourth Order Characters-------------------------------------------------------------49
4.3.5. Fifth Order Characters---------------------------------------------------------------49
4.3.6. Sixth Order Characters --------------------------------------------------------------50
4.3.7. Seventh Order Characters -----------------------------------------------------------50
4.4. Conclusion -------------------------------------------------------------------------------------51
5. Online Handwriting Recognition System for Ethiopic
Basic Characters --------------------------------------------------------------- 53
5.1. Introduction -----------------------------------------------------------------------------------53
5.2. Structural Approach for Recognition of Ethiopic Characters ---------------------54
5.3. Appropriateness of the Approach --------------------------------------------------------54
5.4. Design of Online Handwriting Recognition System for the Basic Ethiopic
Characters-----------------------------------------------------------------------------------55
5.4.1. Data Collection -----------------------------------------------------------------------57
5.4.2. Data Preparation----------------------------------------------------------------------58
5.4.3. Preprocessing -------------------------------------------------------------------------60
5.4.3.1. Extra Pen Up Data Points Noise Elimination------------------------60
5.4.3.2. Size Normalization -------------------------------------------------------61
5.4.3.3. Filtering --------------------------------------------------------------------63
5.4.3.4. 63Resampling -------------------------------------------------------------64
5.4.4. Feature Extraction--------------------------------------------------------------------66
5.4.4.1. Pattern Representation/Observation Coding--------------------------68
5.4.5. Training--------------------------------------------------------------------------------76
5.4.6. Classification -------------------------------------------------------------------------82
5.4.6.1. Coarse Classification -----------------------------------------------------84
5.4.6.1.1. The Computation in the Coarse Classification Layer--86
5.4.6.1.2. Inter-Stroke Distance Computation ----------------------89
5.4.6.2. Detailed Matching --------------------------------------------------------94
5.4.6.3. Matching By Superimposing--------------------------------------------96
5.4.6.3.1. Superimposing -----------------------------------------------96
5.4.6.3.2. Matching Distance Computation -------------------------98
5.5. Summary----------------------------------------------------------------------------------------98

6. Experimentation and Discussion---------------------------- 99

6.1. Experimentation -------------------------------------------------------------------------------99
6.2. Performance of the Classifier’s Layers ------------------------------------------------- 100
6.3. Discussion ----------------------------------------------------------------------------------- 101

7. Conclusion and Future Works ---------------------------- 104

References --------------------------------------------------------- 106
List Of Figures
Fig. 3.1a. A handwritten Ethiopic character called ‘ ’ (Collected by a digitizer software named
‘MovAlyzer’ using mouse as a writing device) ------------------------------------------------------ 22

Fig. 3.1b. Collected Data -------------------------------------------------------------------------------- 22

Fig. 4.1. A Two-level Character Entry Interface for the Non-basic characters-------------- 52

Fig. 4.2. Segmentation applied on (extract basic character and the stroke) ------------52

Fig. 5.0. Design for online handwriting Recognition system for Ethiopic characters ------ 56

Fig. 5.1. MovAlyzer Recording Window -------------------------------------------------------------- 57

Fig. 5.2. MovAlyzer Character Writing pad (after the character ” “ is written) ------------- 58
Fig. 5.3a The Character ‘ ‘ as it is written on the MovAlyzer writing pad ------------------ 59
Fig. 5.3b The set of data points during the formation of ‘‘ ‘ the character --------------- 59
Fig. 5.4. Extra Data point Noise Elimination Algorithm ------------------------------------------- 60
Fig. 5.5a. The letter " " with its original size (173×116)--------------------------------------61

Fig. 5.5b. The same letter " " after size normalization (40×60). ---------------------------61
Fig. 5.6. Size Normalization Algorithm --------------------------------------------------------------- 62
Fig. 5.7a. Original data for letter " " ---------------------------------------------------------------- 63
Fig. 5.7b. Data points after normalization applied to the Original data shown
in Fig. 5.7a--------------------------------------------------------------------------------------------------- 63
Fig. 5.8. Filtering (Data Reduction) Algorithm ------------------------------------------------------ 64
Fig. 5.9. Resampling Algorithm ------------------------------------------------------------------------ 64
Fig. 5.10a. Magnified view of " " after size normalization and before resampling ------ 65

Fig. 5.10b. Magnified view of " " after resampling ---------------------------------------------- 65

Fig. 5.11a. Preprocessed data for the letter " " (instance 1)---------------------------------66

Fig. 5.11b. Preprocessed data for the letter " " (instance 2)---------------------------------66
Fig. 5.12a. Letter " " (Writer B)-------------------------------------------------------------------67
Fig. 5.12b. Letter " " (Writer B) ----------------------------------------------------------------6 7
Fig. 5.13. Observation Code Sequence Extraction Algorithm ---------------------------------- 70
Fig. 5.14a-h ------------------------------------------------------------------------------------------71
Fig. 5.15a-d -------------------------------------------------------------------------------------------72
Fig. 5.16. X Observation Code Sequence for different instances of " " (Writer A)------73
Fig. 5.17. Sequence Refinement ---------------------------------------------------------------------- 74
Fig. 5.18. Observation Code Sequence Extraction Process ------------------------------------ 76
Fig. 5.19. A Reference File
[Generated for X observation sequence of the letter " " (Writer B)]-------------------------- 78

Fig. 5.20. Training conducted for the letter " " (Writer B) ------------------------------------- 79
Fig. 5.21. Training conducted for the letter " " (Writer B) ------------------------------------- 81
Fig. 5.24a. The Structure of the recognizer --------------------------------------------------------- 82
Fig. 5.24b. Flow in the recognizer --------------------------------------------------------------------- 83
Fig. 5.25. Inter-stroke distance------------------------------------------------------------------------- 85
Fig. 5.26. Algorithm for Coarse Classification Procedures -------------------------------------- 87
Fig. 5.27. Algorithm to compute inter-stroke distance -------------------------------------------- 89
Fig. 5.28. The Sequence Expansion Algorithm ---------------------------------------------------- 90
Fig. 5.29. An example Inter-stroke distance computation --------------------------------------- 91
Fig. 5.30a. Coarse Classification for the character " " ------------------------------------------ 95
Fig. 5.30b. Detailed Classification for the character " " ---------------------------------------- 95
Fig. 5.31. Superimposing Algorithm ------------------------------------------------------------------ 97
Fig. 5.32a. Characters after superimposing--------------------------------------------------------- 97
Fig. 5.32b. Characters after superimposing -------------------------------------------------------- 97
List Of Tables
Table 3.1. Improvement of accuracy by adding features ---------------------------------------- 30
Table 3.2: Previous works on Arabic Online Handwriting (adapted from [14]) ------------- 37
Table 3.3. Strength and Weakness of Approaches for Arabic online
handwriting system -----------------------------------------------------------------------------------38
Table 3.4 Recognition Accuracy reported in [14]-----------------------------------------------39
Table 3.5: Summary of result of [6]---------------------------------------------------------------41
Table 4.1. Derived Characters in the Amharic Character Set---------------------------------- 46
Table 4.2. The Amharic Core Character Set-------------------------------------------------------- 47
Table 4.3. The Ethiopic Numerals --------------------------------------------------------------------- 48
Table 4.4. Group1 (Basic character) ------------------------------------------------------------------------------ 48
Table 4.5. Group2 (Basic character) ------------------------------------------------------------------------------ 48
Table 4.6. Grouping the Sixth order character ------------------------------------------------------------------ 50
Table 5.1. Codification of observation ---------------------------------------------------------------- 68
Table. 5.2. Coarse Classification Results for three sample characters from the testing data
----------------------------------------------------------------------------------------------------------------- 93
Table 6.1. Experiment Results for Writer A --------------------------------------------------------- 99
Table 6.2. Experiment Results for Writer B --------------------------------------------------------- 99
Table 6.3. Performance of the layers of the recognizer --------------------------------------- 100
 Online Handwriting Recognition for Ethiopic Characters
MSc. Thesis by: Abnet Shimeles
Advisor: Dr. Solomon Atnafu

ABSTRACT
A new computing scheme, pen computing, which includes mobile devices and applications in
which electronic pen along with pen sensitive writing pad is used as the main input tool has
been emerging. To implement pen-computing applications, online handwriting recognition
system should be used. Online handwriting recognition engines have been developed for
various character sets. Despite that, no attempt has ever been made to build an online
handwriting recognition engine for Ethiopic character set. Pen-based inputting incorporated
with online handwriting recognition feature allows people to write texts and enter input data in
their own natural way of handwriting on an electronic pad.

This thesis then is the first attempt to develop an online handwriting character recognition
engine for Ethiopic characters. The pen-based devices are evidently unusual in Ethiopia and
one reason for that is the absence of localized applications. Bringing an online handwriting
recognition engine for Ethiopic character set to such devices would play an important role in
making these devices available and usable for the Ethiopian society.

In this study, a model for Ethiopic online handwriting character recognition is proposed and a
writer-dependent online handwriting character recognition engine for the 33+1 basic Ethiopic
characters is designed. The designed engine integrates five modules: the data collection and
preparation module, the preprocessing module, the feature extraction module, the training
module and the classification module. Data collection is done with the aid of digitizer
software named Neuroscript MovAlyzer, which samples data points along the trajectory of an
input device (electronic pen or mouse) while the character is drawn. Various algorithms are
designed for the preprocessing activities. In the feature extraction module, a new online
handwriting data representation scheme that makes use of the X and Y coordinate observation
code sequences is proposed. A training algorithm and most importantly a three-layered
recognizer is designed. We are able to show that a reasonably good accuracy is obtained by
implementing the proposed algorithms. On the average, a recognition accuracy of up to 99.4%
is achieved for the sampled two writers. Recognition accuracy 93.4%, 99%, 99.8% are also
obtained for each of the layers of the recognizer respectively.

Keywords: Online handwriting recognition, Online handwriting recognition for Ethiopic

Characters, algorithms for Ethiopic online handwriting recognition, Model for
Ethiopic character online handwriting recognition.
 CHAPTER ONE

INTRODUCTION

1.1. Background of the Study

The development of technology is changing how people live and work. Situations, which
were unthinkable, are becoming possible and practical. The powerful computers available
today capable of processing a large amount of data with a great speed were scientific
fictions before a century. The graceful giant machines that appeared in the 1980’s had been
a miracle. Out of imagination, computers became more and more powerful and at the same
time more and more compact.

Today, computers have become as compact as handheld devices that fit on human hands.
Not only their small size, but also their computing power to implement various
applications makes them very attractive. Historically, the whole thing was started when a
company called Palm produced PDAs (Personal Digital Assistant), whose main function
was to replace personal organizers[30]. Small applications such as calendar, and address
book were run on the PDAs. These devices were called palmtop since they were one-
handed and fit in one’s palm. More companies also involved in the production of such
devices. On the other hand, Microsoft produced minimized laptop-style computers with
screen of size 20cm wide and only 8 cm high and with an attached keyboard. These
computers used a version of Windows, named Windows CE and called handheld rather
than palmtops to indicate that they don’t fit in palm[30]. PocketPcs were the other devices
using the modified version of WindowsCE which are sometimes called handheld. This
time, these terms that emerged at different times have been used interchangeably. In
general, handheld computers are defined to be any small devices that provides computing
and information storage and retrieval and that can be easily carried and used[30].

Though these devices turn out to be common in an increasing number of countries around
the world, they are currently almost ignored in Ethiopia. Even people with full knowledge
of operating personal computers do not even know their existence or are not using them
due to various reasons. Localizing the applications of the devices could play an important

1
role in making them prevalent to the Ethiopian society. We have envisaged that devices
like handheld computers and smart phones that incorporate the wireless technology would
be in widespread usage in the future as the trend of people who use mobile phones is
highly rising. If this happens, then it would be compulsory to have localized applications
that run on such devices for the purpose of exploiting the opportunity of mobile computing
in this country.

When these devices are used, data and commands are needed to be entered while users are
moving. The keyboard as well as the mouse that are common in personal computers for
this purpose do not go with the mobility of these devices. Thus, pen-like devices
commonly called stylus will replace them. In relation to this situation, pen computing has
been an emerging trend of computing.

Pen computing includes computers and applications in which electronic pen is the main
input device[26]. This includes pen-based mobile computing devices such as Personal
Digital Assistants (PDA), and other Palmtop devices. These sorts of devices are becoming
affordable, and applicable in variety of domains. Consequently, nowadays they are
becoming popular and a very large number of customers are carrying and using them[27].
As it is stated earlier, the distinctive characteristic of handheld computing devices is the
use of electronic pen (or stylus)[27] to input data and commands to their system. Pen-based
inputting incorporated with online handwriting recognition feature allows people to write
in a natural way to input data, and provide a pen-paper like interface[8]. These kinds of
systems are very useful in terms of mobility, convenience, cost, capability of accessing
information at anytime and anywhere. Therefore, the above situations and many more
other reasons initiate many researchers to put their effort towards developing systems that
could mimic the pen-paper interface for various specific languages.

It has already been asserted that the pen-paper interface could possibly be available if
handwriting recognition is implemented. Handwriting recognition is the task of
transforming a language represented in its spatial form graphical marks into symbolic
representation[3]. Handwriting recognition systems are broadly divided into two: namely
offline and online handwriting recognition systems[3,4,24]. In offline handwriting
recognition systems, the whole data will be collected and provided for the recognizer as a

2
bitmap. On the other hand online handwriting recognition systems run and receive the data
as the user writes and they are expected to process the data and recognize in a real
time[3,24]. Both online and offline handwriting recognition systems for Western languages
have matured right now. These kinds of systems are also developed for other non-Latin
languages such as Chinese, Japanese, Thai and Arabic[28,13,29,12]. Handwriting
recognitions systems are language specific. This calls the need for localization of such
system to make it available for the Ethiopian society.

Designing and developing handwriting recognition systems is far from easy[4]. In spite of
the fact that a recognition system with 100% recognition rate is not still attained, a better
recognition rate has been exhibited from time to time by the use of various approaches. An
approach selection is another non-trivial task for which extensive study of the nature of the
language and other parameters should be investigated

1.2. Problem Statement

The Ethiopic character set is unique to Ethiopia and Eritrea and it is used to write different
languages including Amharic in Ethiopia. This character set contains more than 300
characters, of which the 231+7 are alphabetic letters. The remaining characters include the
Ethiopian numerals, punctuation symbols and some infrequently used special characters.
The Ethiopic character set, specifically the alphabetic letters set, is known by its systematic
arrangement. The letters are arranged in table form with 7 columns and 33+1 rows. All
characters in the first column are basic (core) characters and the remaining columns consist
of the non-basic characters. The entire arrangement and this classification are based on the
fact that the shape of the non-basic characters in each row is generally derived from the
basic character in the same row.

To the best of our knowledge, there is no research that has ever been conducted on online
handwriting recognition for Ethiopic characters. We have investigated that there are three
researches around offline handwriting recognition of Ethiopic characters [6,7,8]. Thus, this
area has not been even touched and this is a pioneer work, which shall open the way for
other researchers to involve in the area. This thesis then addresses the problem of online
handwriting recognition for Ethiopic characters.

3
1.3. Objectives
The general and specific objectives of this research work are outlined below:
General Objective:
The main objective of the study is to explore, analyze and design an appropriate algorithm
for the purpose of online Ethiopic handwriting recognition for Ethiopic characters on a
PDA Environment.
Specific Objectives:
To meet the mentioned general objective, the following specific objectives are
accomplished in this research work.
• Assess the different techniques for preprocessing, segmentation, feature extraction,
training and classification for online handwriting recognition task.
• Study the techniques in relation to their appropriateness for Ethiopic online
handwriting recognition and propose a technique.
• Design and implement algorithms for handwriting pattern representation scheme,
training, and classifications.
• Conduct experiments to evaluate the proposed algorithms for their accuracy.

1.4. Justification of the Study

Writing is a natural way of putting information. The emergence of very sophisticated
digital computers with facilitated input methods do not cover up the importance of the
handwriting activity. Particularly, in places in which taking short notes is necessary,
keyboards or keypads are illogical. For most people, a pen and a notebook are much
better[3]. On the other hand, learning and adapting to the unnatural way of feeding data to
computers had been a trouble for a significant number of people (this is even worse in
Ethiopia as the implementation of the available software is based on English language). To
see the idea in a different view, even computer literates are not perfectly happy by the
method employed to enter data in desktop computers, which doesn’t generally allow
mobility. However, devices that offer a pen-paper based interface mostly are very small in
size that can fit to a palm.
In summary, handheld devices particularly PDAs have the following useful features:
Convenience

4
 Portability
Affordability
Easy Information access any time/anywhere
Wireless communication
Some of the application domains of PDAs are:
In Schools for students and teachers.
In hospitals for doctors
Field data collection
For short note taking in meetings
Currently, in Ethiopia, handheld devices for instance PDAs are not common. One of the
technical reasons is that they are not still suited for local applications. The goal of this
research is to localize the online handwriting recognition system feature of handheld
devices so that Ethiopians can benefit from this technology.

1.5. Scope and Limitations of the Study

First attempts are always demanding. On top of this, handwriting recognition researches
are generally challenging mainly due to the variability in handwriting styles. Additionally,
standardized online handwriting recognition data set is not available at all. The number of
characters in the Ethiopic character set is also large relative to the Latin characters for
which researches have been conducted for decades to come up with improved approaches
that can bring about a better recognition accuracy. Thus, we put some constraints on the
recognition engine after examining these situations. Accordingly, the scope and limitations
of this research are that:
• Only character-level recognition is considered.
• The online handwriting recognition system developed in this work is writer-
dependent.
• Only the 33+1 basic Ethiopic characters are considered.

1.6. Methods of the Study

Online handwriting recognition engine is not a one-sep process. Thus, different techniques
are used in each stage as described below:

5
1.6.1. Literature Review
Research papers on online handwriting recognition systems for Latin, Chinese, Japanese
and Arabic character sets have been reviewed to see the state-of-the-art status and to
identify the various approaches used for this problem. Moreover, the existing offline
handwriting research papers have also been reviewed in order to get ideas on how they
treat the Ethiopic characters.

1.6.2. Design of the Online Handwriting Recognition System

Data Collection Technique
The online handwriting data of the 33+1 basic Ethiopic characters have been collected by
using the MovAlyzer digitizer software that samples data points when a character is drawn
by the mouse. This data serves for training and testing the recognition engine.

Preprocessing Algorithms
Various preprocessing activities involve in any handwriting recognition system. The
recognition system developed in this project employ algorithms for noise elimination, size
normalization, filtering, and resampling.

Pattern Representation Method

The online handwriting data as it cannot be directly used for recognition, an intermediary
pattern representation method can ease the recognition of the unknown character in many
ways. For this, a new pattern recognition scheme which attempts to represent patterns in
terms of X and Y observation sequences was designed.

Training and Testing

Training is a crucial action before any recognition system becomes capable of recognizing
entities, which are supposed to. Thus, a training algorithm that goes with the pattern
representation scheme has been designed. Experiments have also been conducted to test
the system for its accuracy.

6
1.6.3. Development of the Ethiopic Online Handwriting
Recognition Engine
The language used to implement the various algorithms is C/C++. Text files are used to
store the various data processed by the system. The original handwriting data is provided
by the digitizer software in the form of text files. The preprocessing module takes these
text files as an input and produces other text files that store the preprocessed data. The
feature extraction module, in turn, receives the preprocessed data and produces observation
codes sequences and put them in individual text files during training and classification. The
recognizer also takes the observation code sequences to predict what character is
represented by that specific input. To draw characters from a given data, the graphics tool
of the Turbo C++ compiler has been used.

1.6.4. Organization of the Thesis

This document contains a total of seven chapters. The second chapter presents general
concepts about online handwriting recognition. In chapter three, related works are
reviewed. The fourth chapter presents the analysis done on the shapes of Ethiopic
characters for the purpose of designing the recognition engine. Chapter five is the chapter in
which we present the design of the online handwriting recognition system for the basic
Ethiopic characters. All algorithms designed for the various activities that are involved in
the recognition system are detailed in this chapter. Chapter six presents the experimental
results by incorporating discussion about the results. Finally, the conclusion and future
works are presented in chapter seven.

7
 CHAPTER TWO
ONLINE HANDWRITING RECOGNITION

The chapter presents a short introduction to pattern recognition that is the parent research
stream of handwriting recognition. Then, the works done related to character recognition in
general and handwriting recognition in particular followed by the classification of online
handwriting systems are explained. Some issues, which deserve consideration during the
development of online handwriting recognition systems, are presented next. Finally, the
steps in online handwriting recognition task are detailed.

2.1. Introduction to Pattern Recognition

A pattern generally describes the way in which something is arranged to represent a given
identified and classified entity. Examples of what we call pattern are a fingerprint image, a
human face, a printed or handwritten character [1]. Pattern recognition is, therefore, the
study of how machines can be made capable of learning and distinguishing patterns to
classify them under predefined classes like humans are able to do [1,2]. Human beings are
naturally gifted to learn and identify patterns easily and intelligently. For instance, it is not
difficult for all of us to distinguish people we know by looking at their faces under normal
circumstances. However, it is not a trivial task to teach machines to do the same thing .

Researchers have been focusing on pattern recognition for it has applications in variety of
domains. It is also true that more applications are being generated due to the increasing
power of data processors used to implement the pattern recognition algorithms that are
known to be computationally complex[2]. Personal identification using attributes such as
fingerprint or face in biometric recognition, semantic document classification and
handwriting recognition are a few examples of these applications.

Various classification algorithms have been suggested and studied in relation to the
particular applications. Although more than 50 years have passed in conducting researches

8
around pattern recognition[1] to come up with well-defined and generic methodologies, it
is still a research topic having gaps here and there.
In summary, a pattern recognition system can be viewed as a black box receiving input and
producing output[2]. Inputs to such systems are input patterns and the expected output is a
classified and named entity.

2.2. Character Recognition

The term ‘recognition’ is a typical word used in many pattern recognition systems.
Recognition systems are in general devoted to classify input patterns to corresponding
entities. These entities vary from one system to another. Recognition systems, in which
characters are expected to be classified, are referred as character recognition systems.

One major distinguishing feature of character recognition systems is the type of characters,
which are supposed to be recognized. There are systems for recognition of printed
characters, typewritten characters or written characters. Due to the varieties occurring to
handwritten characters, development of the handwriting recognition systems is the most
challenging.

2.2.1. Handwriting Recognition

Handwriting recognition is the task of transforming a language represented in its spatial
form of graphical marks into its symbolic representation[3]. Symbolic representation refers
to the digital representation of characters like in the case of 8-bit ASCII character set.
Handwriting has been a basic form of communication and still a good way of expressing
ones’ ideas. In relation to this fact, handwriting recognition systems are useful and are used
to realize ideal applications of computers such as pen computing.

A number of classification methodologies to categorize handwriting recognition systems

have been advised based on different factors. The subsequent sections explain some of
these classification approaches, which are under interest.

9
2.2.2. Online Vs Offline Handwriting Recognition System
The basic input to character-based handwriting recognition systems is a pattern that
represents a character. In fact, this pattern should be digitized before it is available to the
system. The way the input is provided along with the digitization technique are taken as the
two factors to classify handwriting recognition systems as online or offline.

The basic differences between these online and offline handwriting recognition systems as
stated in a number of literatures[3, 4, 10, 24] are:
• Input Method:
In online handwriting recognition systems, the handwriting data is obtained with the
help of a transducer such as an electronic or tablet digitizer. Systems that simulate a
pen-paper like arrangement for a writer are the sources of online handwriting data.
Such systems record the pen-tip information as a sequence of (x,y) coordinates of data
points sampled over time. On the other hand, in offline handwriting recognition
systems, the data is captured optically by a scanner in the form of image. This
necessitates the development of a method to identify the pixels which are part of the
handwriting data pattern from pixels lie on the background [9].
• Kind of available information:
For online system, the coordinates of successive points as a function of time in order is
available where as in the offline case, only the completed writing is available.
Additionally the speed of the writer, stroke number and order, and the pen-up/pen-
down states are detected during the collection of online handwriting data. A pen-down
state is detected when the pen touches the digitizer(writing pad) and when the pen is
lifted off a pen-up state is sensed. This makes online systems easier to build than
offline systems taking the abovementioned additional information as an advantage.
Moreover, recognition rates generally have been reported to be higher for online
systems than offline ones for the same reason.
• Time of processing:
Clearly seen fact is that offline systems run after their data have been collected. The
handwriting ought to be put on a media such as paper exhaustively and brought to the
scanner, which in turn digitizes it. The case for online handwriting recognition systems
is quite opposite. An online system receives handwriting data and recognizes it in real

10
 or nearly real time. This situation adds one useful feature to online systems that is
interactivity. Being interactive, online handwriting recognition systems allow users to
edit what they have written and recognition errors can be corrected immediately.
Online handwriting recognition systems are mainly used as an input to handheld or
PDA-style computers that might replace the keyboard-based personal computers in the
future[10]. Signature verification systems that are aimed at verifying identity of an
individual with his/her signature, and educational software developed to aid children to
learn handwriting are other applications of online handwriting recognition systems.

Offline systems could play a significant role in digitizing knowledge and information
in the form of handwriting[4]. A research conducted by Wondesson Mulugeta on
offline handwriting recognition for Amharic handwriting written in a special type of
traditional style which is called ‘Yekum Tsifet’, is a good model to indicate the
usefulness of offline systems by contributing towards the digitization of a number of
historical and religious documents[8]. Other applications of offline systems are postal
address recognition and check reading[6,7].
• Speed Determination:
The speed of online systems depends on the writing speed of the user since the
handwriting is analyzed and recognized in real or near real time. Conversely, in offline
systems, the speed depends on the specification of the system in words or character per
second.
• Adaptation:
Adaptation of the writer to machine and machine to the writer is possible in online
handwriting recognition systems. When the writer sees that some of her characters are
not being classified correctly, she will alter the characters’ shapes to improve accuracy
and adapt to the system from time to time. On the other hand, some recognition engines
are made to be capable of adapting to the user, basically by storing samples of the
writer’s characters for subsequent recognition. This two-side adaptation could never be
made to happen to offline systems.

11
2.3. Online Handwriting Recognition
As stated earlier, online handwriting recognition systems attract researchers for their
currently appreciated applications, which involve in pen computing. In the context of
comparing these systems with offline ones, they take the advantage of the time information
that is incorporated in online handwriting data. However, the task remains to be
challenging in spite of the fact that the valuable additionally available time information
stands besides them.

This section attempts in revising technical issues regarding online handwriting recognition
systems. After introducing further classification of online handwriting systems, the major
steps in an online handwriting recognizer will be detailed.

2.3.1. Categories of Online Handwriting Recognition

Recognition accuracy is the leading parameter to evaluate online handwriting recognition
systems to decide on their usability. Looking into the history of commercial online
handwriting systems, it is proven that developers started to put different constraints on the
usage of their systems to get a reasonable accuracy rate[4]. Due to this reason, online
systems started to fall in two types namely constrained and unconstrained systems.

2.3.1.1. Constrained vs. Unconstrained Systems

Constrained systems placed restrictions on writing styles. Some of them may want users to
write in a discrete manner and some others force users to write in a given order of
strokes[11]. Surprisingly enough, Graffiti, a product of Palm Computing that took the
market by storm, placed the most suppressing restrictions on users by only allowing them
to use only single strokes to write all the Latin characters [4]. The way users write in
Graffiti is even far from natural handwriting since it obliges users to write only in a
specified direction [4]. It also modifies the shapes of the characters

On the other extreme, unconstrained handwriting recognition systems allow users to enjoy
writing in their own natural handwriting. Although these systems relax users, their
recognition accuracy could be evidently lower than constrained systems [4]. That is the

12
reason why Graffiti has been more acceptable than CalliGrapher, which is stated as
unconstrained system [4].

The amount of training data and its source used to train these systems is also a factor to
come up with a different classification approach which introduces two types of online
handwriting recognition systems: writer-dependent and writer-independent.

2.3.1.2. Writer-dependent vs. Writer Independent Systems

The goal of a writer-independent online handwriting recognition system is to recognize
handwriting in a variety of writing styles, while writer-dependent systems are trained to
recognize handwriting of a single individual[9]. A requirement of writer-independent
systems is that they are expected to be able to recognize handwriting that the systems have
not seen during training. Writer dependent systems lack this ability but still remain to be
satisfactory for users of personal small devices. Yet, writer independent systems are
necessary for applications like online form filling using a pen-like device.

Constructing writer independent systems is obviously harder than writer dependent

systems. Moreover, writer dependent systems generally present a better accuracy rate. The
difficulty of development of writer independent systems arises from the fact that the
system is expected to handle much greater varieties of handwriting styles[9].

A distinguishable difference between writer dependent and independent systems is

essentially in the amount of training data. It is possible to obtain large amount of training
data when training a writer-independent system since many writers involve in the data
collection process. For writer-dependent systems, the amount of training data will be
limited, since the only provider is a single individual whose handwriting is being trained to
the system[9].

2.3.2. Issues in Online Handwriting Recognition

In this section, we will explain the factors that make the construction of online handwriting
recognition system challenging and the possible design goals that could be set when
designing an online handwriting recognition engine.

13
Some of the common factors are explained below:
• Handwriting styles: Cursive and block
Writers usually use different handwriting styles which could be one of the known
defined ones or styles introduced by themselves. Variations resulting from this
situation import challenge to the development of handwriting recognition systems.
Some alphabets or languages have avoided some handwriting styles and others allow
based on a rule. For example, Latin characters are written in a discrete way or cursive
manner. When characters are written separately in boxes or virtual boxes created as a
result of spacing between characters, the writing style is referred as discrete or block.
Conversely, it is possible for writers of Latin alphabets to write characters by
connecting them to each other which results in cursive writing style[10]. There is no
way to control users not to write in mixed styles. However, in Amharic all characters
are naturally written in discrete style.

Variation is a source of errors in online systems and variations could occur both in time
and space[10]. Variation in time refers to the variation in writing speed. Shape
variation is another common difficult. These variations happen in both discrete and
cursive handwriting styles. But, cursive style brings an additional problem that
demands additional effort to identify the beginning and end of a character. Doing so is
referred as segmenting the handwriting. Segmentation will be worse in handwriting of
a mixed style[9].

• Stroke number and order variation

A stroke is defined as the sequence of sample points occurring between consecutive
pen-down and pen-up transitions[9]. A character could possibly be a uni-stroke
character or formed from two or more strokes. The number and order of strokes is
available for online systems, which is always considered as an advantage. In contrast,
variation in stroke number and stroke order in a single character is a source of
complication in online handwriting recognition system. Stroke order and stroke number
variations are even terribly severe in particular systems such as Chinese/Japanese
systems[12]. As a result, the issue of handling stroke order and number variations is
one of the design goals when designing online handwriting recognition systems
particularly the writer independent ones.

14
• Limited Resources in small devices
It sounds good to remind that online handwriting recognition systems help in avoiding
keyboard-based data input method in small handheld devices. It is not optional to let
users of such devices input data in a pen-paper like environment by aiding the process
with hardware digitizers along with pen-like devices and incorporating online
handwriting recognition feature.

It is good news that the storage and processing capability of handheld devices is
improving from time to time. Limitation in resources has been a problematic
phenomenon of these devices for recognition engine developers. The headache would
be severe for large character sets such as the Chinese/Japanese cases [12,13]. Carefully
designed algorithms to optimize the amount of data and the processing requirements
are supposed to be attained for online handwriting recognition systems. A related
design goal could be to make the systems platform independent and capable of dealing
with different pen capture technologies that may reveal different characteristics such as
varying sampling rate[5].

• Character Set or Dictionary Size

Some character sets introduce additional problems to handwriting recognition system
designers and developers. For instance, for character sets having a large number of
characters, finding a match for a to be recognized character would be much more tough
because the number of misleading prototypes would be higher in number. Evidently,
more than 200 characters are included in the Ethiopic character set. This is a large
number compared to the 26 Latin characters. However, there are bigger characters sets
such as the Chinese character set which includes about 5000 characters when only the
frequently used characters are counted [13]. Thus, character set size or dictionary size
is another factor that contributes in the complicatedness of the development of online
handwriting recognition systems.

15
2.3.3. Major Steps in Online Handwriting Recognition
Handwriting recognition is by no means a single-step process. It involves a number of
steps in which a to be recognized entity (character, word or sentence) passes through
before a recognition attempt. These steps are essentially used to avoid variations and are a
way to identify only the important part of the data which is believably useful for
recognition.

In general, recognition of handwriting patterns involves the following steps: Preprocessing,

segmentation, feature extraction, classification and post processing. The following
subsections detail these major steps.

2.3.3.1. Preprocessing
Handwriting data is subject to noises, which creates discrepancies and result in
misclassification. The goal of preprocessing is to reduce or eliminate these variances in
order to decrease the dissimilarity between a to be recognized entity and its correct
prototype[9]. The noises introduced in the data are caused by the inaccuracy of the data
capturing device, the erratic movement of the hand or fingers while writing and the varying
writing speed of the writer [3,9,13].

There are preprocessing activities that have been implemented in various recognizers.
These preprocessing measures applied on the online handwriting data could be grouped
into two, based on their purposes: noise elimination (reduction) steps and normalization
steps[13].

2.3.3.1.1. Noise Elimination

The term “noise” in pattern recognition systems in its broader sense is explained as:
“anything which is part of the data that hinders the pattern recognition system from
fulfilling its task.”.[2]
Naturally, the type of noise as well as the method to eliminate it is specific to pattern
recognition systems. In this thesis, we are concerned about online handwriting recognition
task and we narrow down our view to possible noises that may happen in online
handwriting data. Noise in online handwriting data can be explained as a set of sampled

16
points which makes the trajectory of the pen tip somewhat jagged[9]. In another sense,
additionally sampled points which are not part of the written character also tend to produce
noisy data. To name a few of the common approaches to avoid these and other possibly
occurring noises, we can state dehooking, filtering and smoothing[13, 18, 24].

Smoothing refers to the process of avoiding sampled points which make the trajectory
rough. It is usually performed by using an averaging scheme over neighboring points[19].
Filtering means the elimination of duplicate points and hence reduction of data [24]. On
the other hand, dehooking is a technique to detect and remove hooks which occur at the
beginning and end of a stroke by a quick movement of the pen when it is raised or
lowered[9]. Evidently more techniques exist, but as the quality of the input devices
advances, only simple smoothing will be a sufficient preprocessing task [13].

2.3.3.1.2. Normalization
Variation between handwritten characters, words or sentences happen naturally. The
variations that occur would be higher between in handwritings of different writers than of
the same individual. In spite of that, since even a slight diversity sometimes result in a
wrong classification results, techniques aimed at avoiding the variation are often
implemented in any type of handwriting recognition systems. Some of these variations that
probably occur include size difference, writing speed difference and difference in degree of
inclination of say characters. Even in isolated character recognition which is much less
complicated than word or sentence recognition, these variations greatly affect recognition
accuracy.

Writing speed determines the number of data points that are captured or sampled while the
pen is dragged on the electronic tablet or any other digitizer used. When the writer is
writing slowly, more number of points will be samples than when the writer is writing
quickly. To get rid of such variations, time normalization or most commonly called
resampling is applied [9,14]. Equidistant resampling is the process of resampling the data
such that the distance between adjacent points is approximately equal [13]. Besides the
removal of the variance, this step essentially reduces the data which is a desirable effect.

17
Very noticeably, size difference between instances of the same character occur both in
those written by a single writer or multiple writers. Sometimes, the extent of size variation
depends on the way the writer is writing. For example, in the case of a boxed-input
character recognition system, size dissimilarity would be minimal compared to the
situation without this constraint. Whatever the case may be, a concern has to be given to
size normalization. Size normalization is a way to reduce or enlarge the character to a
predefined size[14].

Slant correction or rotation normalization is another typical normalization in which a

character is rotated with the aim of avoiding writer-specific slants[14]. Translation
normalization is also a normalization technique to translate all the data to the same spot
relative to the origin[14].

The necessity of these normalization steps totally depend on the quality of data capturing
devices, and the overall nature of the recognition engine which is supposed to be
developed. It is the responsibility of the system designer/developers to decide on the
relevant and important preprocessing steps.

2.3.3.2. Segmentation
At this stage, the preprocessed data is available. But, it is still a set of sampled points that
are not yet meaningful. The preprocessed data needs to be segmented to parts to create
meaningful entities for classification[2]. Segmentation is a way to find the representation
of basic units that the recognition engine will have to process[3]. Segmentation may occur
at different levels depending on the recognition algorithm used. The handwriting input may
be segmented to individual characters or even into sub-character units[3].

For characters written in predefined boxes, most of the segmentation will be done by the
writer. On the other side, the segmentation task would be harder and more error-prone for
cursively written handwriting data[3] since it is not an easy task to determine the beginning
and ending of individual characters[3]. For such cases, it is necessary to apply internal
segmentation, which requires some recognition to aid the isolation of the writing units[24].
Conversely, external segmentation is a type of segmentation that is entirely done prior to
recognition[24].

18
Various segmentation techniques suited for online handwriting recognition have been
advised. Some segmentation techniques exploit temporal information for the purpose of
separating writing units such as by tracing the time difference between two writing
units[24]. Others use spatial information[24].

In some occasions, the segmentation step might not be explicit and are tightly coupled with
the previous preprocessing steps or the following steps[2]. Hidden Markov Models
(HMMs) are good examples of handwriting modeling methods without an explicit
segmentation step[10]. HMMs are tools to represent the continuous pattern of the
handwriting data of any kind (character, word or sentence) as a single unit. It is possible to
create character models, word models or sentence models as required. This makes HMMs
to be a good choice to model cursively written words eliminating the difficult problem of
segmentation of the words to individual characters. However, many researchers choose to
put a segmentation stage before the HMM models are created.

2.3.3.3. Feature Extraction

Once the data has been preprocessed and appropriate segmentation has been done on it (if
any), the data will be ready enough to apply classification. But, it is sensible that all the
sampled data points are not equally useful for recognition while the possibility to do
classification using the whole raw data is still possible[12,14]. Systems that rely on raw
data generally have lesser recognition rate than those that rely on only the important part of
the data or attribute(s) derived from the raw data[12]. Hence, the appropriateness of using
selected features.

Features are high-level attributes extracted from normalized raw data[12] that describe the
property of the handwriting data. For instance, number of strokes in a character is a very
simple example of a feature that could be used (in combination with others) to classify
characters. Many types of features are used in various recognition engines. Feature
extraction, is therefore, the process of extracting only the relevant features from the online
raw data with the aim of reducing within-class pattern variability and enhancing between-
class variability[2].

19
The purpose of feature extraction is not only deriving high-level attributes for classification
purpose but also reducing data. Data reduction in general has a positive effect in easing the
processing of the computationally complex classification algorithms.

There are no standardized and widely recognized features for describing online
handwriting pattern [12]. Due to this reason, features are usually selected manually[14].
And choosing appropriate features is a very critical issue for developers since this decision
affects the performance of the system greatly.

In general, there are two types of features namely online and offline features[12]. There are
advantages and drawbacks associated to each type of the features. To name a few, offline
features lack the dynamic part of the data and online features are not robust to stroke order
and number variations. Researchers have been incorporating offline features with online
features in online handwriting recognition process and the vice versa[12]. This generally
results in a better recognition accuracy rate. However, it is not straightforward to derive
online features from offline data and in turn produce offline features from online data. For
instance, it is required to extract the time information from an offline data to derive online
features from offline data.

2.3.3.4. Classification
Classification could be regarded as the last and the most important step that produces the
final output of the recognition engine unlike the previous steps that produce only
intermediate outputs. The effort done to carefully design the techniques for the previous
activities is to facilitate the success of the classification step[12].

The ultimate goal of this step is to carry out some form of comparison between a given
unknown handwriting pattern to reference handwriting patterns to assign one of the
references to the unknown one. In literature, several types of classification methods are
advised in spite of the fact that none of them are perfect. Some systems divide the
classification step into two: coarse classification and detailed classification [13]. Coarse
classification is to assign a pattern to a group or multiple groups and detailed classification
is to find out the particular entity associated to the unknown input from the group to which
it is assigned [13]. Classification will be more detailed in the next chapter.

20
 CHAPTER THREE
REVIEW OF THE STATE OF THE ART

This chapter reviews the state-of-the-art in online handwriting recognition briefly. A

generalized description about the well-known approaches that have been applied in
classification of handwriting data will be given. Then, a subsection reviews the relevant
research papers on the trend of online handwriting recognition in three selected character
sets, which are Latin, Japanese/Chinese and Arabic. We believe that this section presents
the effort made to evaluate approaches used in recognition systems that deal with different
characters sets and with diversified properties to come up with the better methodology for
Ethiopic characters. Finally, the works on offline handwriting recognition systems for
Ethiopic characters are reviewed.

3.1. Major Classification Approaches for Character

Recognition
The most crucial step in handwriting recognition systems is classification. Classification is
the task of classifying an unknown input pattern to one of the prototype patterns. It is
evidently true that the effectiveness of the steps that have been carried out on the data
before it is presented to the classifier have a direct impact on the performance of the
classifier and hence on the recognition engine. The choice of the classification approach
may also deteriorate or improve the recognition accuracy rate.

Classification approaches generally depend on the method used to represent the

handwriting pattern which is referred by the feature extraction step[13]. Two issues are
considered in any classification algorithm. These are pattern representation and matching
method, which are discussed in brief in the following subsections.
3.1.1. Pattern Representation
Online handwriting data is essentially a set of sampled points where each point is described
by its x and y coordinate values apart from the other information detected by the digitizer

21
such as pen up/pen down information and the pressure of the pen tip (See Fig. 3.1a & b).
The raw data even after preprocessing will not be in the appropriate state to be compared
with prototype or reference data and classified accordingly. This demands representation
methods for the handwriting pattern.
198 -228 99
197 -228 99
196 -228 99
195 -228 99
193 -227 99
191 -226 99
190 -225 99
188 -224 99
186 -223 99
185 -222 99
183 -220 99
182 -219 99
181 -218 99
178 -211 99
.
.
Fig. 3.1a) A handwritten Ethiopic character
called ‘ ’ (Collected by a digitizer software Fig 3.1b) Part of the collected data for the
named ‘MovAlyzer’ using mouse as a writing letter shown in Fig 3.1a. Each triple includes
device) the x and y coordinate values and a number
to indicate the pen up/pen down information
(99 for pen down and 0 for pen up)

According to C.L.Liu, S. Jaeger, and M.Nakagawa [9], the pattern representation schemes
could be divided into three groups: structural, statistical, and hybrid statistical-structural.
Structural representation deals with expressing complex patterns in hierarchical
perspective where a pattern is viewed as being composed of simpler sub patterns that are
themselves built from yet other simpler sub patterns [1,13]. For instance, line segments
could possibly be built from a set of sampled points and the line segments may combine to
form strokes. Strokes, in turn, combine in a defined rule that specifies their relation to
create characters. Hierarchical structure could also be created on top of the character model
to organize the available strokes to share them to construct the character models of all
categories [13]. This is quite a good example to show the level-based description of the
pattern in which the lower components are used as primitives to build the upper
components. The difficulty in representing patterns structurally is to detect primitives from

22
pattern data (which is prone to noise and variations) and to formulate the rules to combine
primitives from a given training data set [1].
The other method for pattern representation used statistical techniques. These techniques
attempt to represent the pattern in terms of d features or measurements[1]. This creates a
feature vector representing the pattern. The feature vector could be viewed as a point in a
d-dimensional space. Feature set selection is the main issue in this methodology, since it is
expected that the feature set should enable the well separation of patterns into their
corresponding classes in the feature space. Statistical representation schemes are becoming
common in many recognition engines. To demonstrate feature vector description of
patterns in statistical representation, let us pick one of the referred recognition engines
systems. In [11], feature vectors are generated for each data point (xi, yi) as (xi, yi, xi-xi-1,
yi-yi-1) where xi-xi-1 and yi-yi-1 will be set to 0 if i<2 and i is the index to refer to the points.
This is a feature vector having four features. Actually, the system is an Hidden Markov
Model based system and these feature vectors are later quantized to generate observation
symbols that are known to be an input for the HMM [11].

The third representation scheme is a hybrid one that takes the advantages of both methods
which are described above. A pattern model is represented by a string, using tree or graph
structure, with the primitives and/or their relationships measured probabilistically by
replacing the attributes of primitives and/or relationships with a probability density
functions [13]. A good example is the Hidden Markov Model, which is a graph with nodes
and the between-node transition, are measured probabilistically.

3.1.2. Pattern Matching Techniques

All recognition systems are equipped by a set of prototype patterns corresponding to
entities or collection of primitives to build entities that are to be recognized. These
prototype patterns are obtained mostly by learning or occasionally integrated to the
recognition system when they are first constructed like in the case of Graffiti in which all
letters are represented by a single predefined stroke[4]. In any case, there will be a
matching operation to compare the unknown pattern to the existing prototype patterns for
the purpose of assigning it to one of the classes. Once pattern representation scheme is
designed, the next decision would be the way the matching must be done.

23
Pattern representation and pattern matching are related in many ways and usually treated
together and even integrated and referred by one name in many literatures. However, we
prefer to discuss them separately to match with our own classification approach that is
developed in this work, and that separately considers them. Structural matching,
probabilistic matching and statistical matching are generalized names for the three types of
matching techniques[13].

Structural matching is the process of finding a match for unknown pattern represented in
structural representation scheme from set of prototype patterns, which are also represented
structurally. The match will be searched by computing matching distances between the
unknown pattern primitives and the primitives of the candidate patterns and take a class
with the minimum distance[13]. Since a grammar or rule is defined in structural
representation of patterns, the match is found according to the rule. Structural matching is
sometimes called syntactic matching.

Probabilistic matching is computing distance between models and unknown patterns,

which are, represented structurally where the attributes are probabilistic [9]. In statistical
matching, a character is recognized by choosing either the most probable character or
choosing the character with the least probability of wrong classification [1].

Specific recognition algorithms are discussed in the following section while looking into
existing online handwriting recognition systems designed for various character sets. These
include structural string matching, template matching, Hidden Markov Model and neural
network.

3.2. Review of Relevant Papers on Online Handwriting

Recognition Systems
Naturally, a difference exists between characters in different character set. This difference
is in number, shape, degree of similarity between characters of the same set, and many
more other behaviors. Consequently, selected approaches should take the nature of the
characters in the character set to which a recognition engine is going to be developed into
consideration. Basically, the difference in the usage of classification approaches is
inherited from the fact that characters in different character sets possess diversified nature.

24
In this section, we perform a comparison of recognition systems for three character sets
namely Latin, Chinese/Japanese and Arabic characters.

These three character sets are chosen for the reason that they exhibit a great difference that
will reveal which approach is convenient to which type of character set. The Latin
character set contains 26 alphabetic letters and 10 digits. Each of the alphabetic letters have
two forms namely capital and small letters. Without including punctuation and other
symbols that involve in writing, there are 62 characters in the Latin character set.

The Japanese writing system is well described in [8]. It is a mixture of 3 types of

characters: Kanji (Chinese traditional characters), Hiragana and Katagana. The total
number of characters is roughly 6000-7000. Moreover, Japanese characters are divided
into two classes namely JIS first level (most common characters) and JIS second level (less
common ones) for the purpose of computer processing. The paper also stated that the main
difference between Japanese and Latin handwriting recognition systems is caused by the
large number of Japanese characters. The Chinese character set is divided into three
subsets: Traditional Chinese characters, Simplified Chinese character and Japanese Kanji
characters[13]. Moreover, there are three writing styles: Regular script, Fluent script, and
Cursive script. The main problem in Chinese online handwriting recognition is stroke-
order and stroke-number variability. Additionally, recognizing fluent style handwriting has
been a difficult task.

Arabic characters are 28 in number and all of them are cursive [14]. As indicated in [14],
the shape of the 28 characters is constructed based on 18 distinct shapes that vary
according to their connection to the preceding or following letters or by putting dots and
other symbols above or below them. The key difference between Arabic and Latin
characters is the fact that many letters differ only by a ‘.’ while the primary stroke is the
same.

25
3.2.1. Latin Online Handwriting Recognition Systems
In this subsection, we present four papers relevant to our work and for which summarized
explanations are offered.

3.2.1.1. Kam-Fai Chan et al.: Flexible Structural Matching

The research goal in this paper[16] is to build a handwriting recognition system, which is
speedy, accurate and flexible. The system is expected to tolerate variations and maintain
reusability hence facilitates extensibility.

The main approach is a structural string matching technique. Characters are viewed as they
are made up of line segments and curves. The line segments and curves incorporated with
directional information help to define primitives that are building blocks of strokes. These
primitives are line, up (curve going counter-clockwise), down (curve going clockwise),
loop (curve joining itself at some point, and dot (a very short segment). A hierarchical
structured representation is employed for the purpose of the handwriting pattern
representation in which a character is represented by stroke set, stroke set is a collection of
strokes, a stroke is a collection of primitives which are listed earlier. To model this, a 4-
tuple grammar is also introduced.

Once a character is handwritten, what we get is a set of points from which the primitives
that make up the character are going to be extracted. The structure extraction and
reconstruction mechanism is described for each of the primitives. After extracting these
primitives, their description will be compared with the model set to find a match. If an
exact match is not obtained (due to the variability of handwriting style), flexible structural
matching will be performed to change the shape of the character slightly in order to find
the closest character. Thus, the matching process will be continued until a match is found
for a character. Four levels of structural deformation are included namely no deformation,
primitive type deformation, directional deformation, and simultaneous type and directional
deformation. However, some ambiguities may still occur in flexible structural matching
and an additional step post-processing to find the most probable character is also
implemented.

26
The online handwriting data set collected by the MIT Spoken Language Systems [31] is
used to test the system and 98.60% accuracy for digits, 98.49 for upper case letters, and
97.44% for lower case letters are reported. The overall accuracy of the system is 97.40%.

The authors also claimed that reasonable speed has been achieved which is 7.5 characters
per second when running in Prolog on Sun SPARC Unix workstation and the memory
requirement is also reasonable low.

3.2.1.2. G. Loudon et al.: Discrete Hidden Markov Model

This system[11] is developed to run on smart phones and the authors stated that the system
considers the limitations of the device and the parameters to make recognition systems
usable. Accordingly, the following requirements are outlined:
• The system should run on slow processors and low memory,
• The system should have high accuracy,
• Fast text entry has to be facilitated,
• An easy text correction mechanism should be incorporated.
The authors claim that to achieve these goals, some restrictions (constraints) are placed on
the users of the system. These are:
• Users should write using only small letters (can change to upper case letter later
when editing).
• Users must use a predefined style of writing even though it is not like Graffiti,
which forces users to write a letter by a single stroke. In this system, the shape of
the letters is kept as their natural shape. However, every letter is written in a
predefined way as indicated in the paper.
• Users are only allowed to write in a hand printed manner (not cursive).
The main approach used is based on discrete Hidden Markov Model. The Hidden Markov
Model (HMM) is a statistical-structural pattern representation method. Hidden Markov
models are used extensively and successfully in speech recognition and are good to
represent sequential temporal data.

The system has two modes: Training and Recognition. In the training mode, isolated
characters are provided as a training data set. Character pre-processing; which includes
character normalization, stroke joining, stroke re-sampling, feature extraction and finally

27
vector quantization is applied. Character normalization is also done to avoid shape
distortion of characters. Stroke joining is the process of joining two strokes, which are part
of the same character (for eg. f, the and −). Stroke re-sampling is necessary to make the
average speed of writing a character is a constant across all characters. Feature extraction
is another important step to pick four features on each data point which are (xi, yi, xi-xi-1,
yi-yi-1). Vector quantization is to convert a multi-dimensional vector to a discrete symbol
and it is essentially a data-compression technique that will reduce computation complexity
in the HMMs. After the pre-processing tasks are accomplished, each character is modeled
by two ten-state HMMs from which one is based on xi, yi and the other is based on xi-xi-1,
yi-yi-1.

In the recognition mode, sentence recognition is supported. Four steps are going to be
followed namely segmentation which segments the sentence to isolated characters,
character pre-processing (same as the training case), the Viterbi search in which
probability of match between the input data and all characters will be calculated and sorted
and then the first five characters will be stored, and finally re-scoring which uses simple
language knowledge to improve accuracy.

The system has also an additional feature that is text correction. The module provides a
very interesting method of correction by allowing the user to write on top of the already
written text or delete by crossing words.

Experiment results are also reported. The training data consists of 11984 different
characters written by 13 writers. The test data includes 5559 characters written by 7
writers. Accuracy rate is reported to be 98.3%.

3.2.1.3. H. Shu: Hidden Markov Model

The paper[10] emphasized the fact that HMMs are initially applied to speech recognition
tasks and it had also been proven that they could also be used to build an online
handwriting recognition system. The main goal of this research is to improve the accuracy
of a system named BYBLOS by adding features. The BYBLOS[32] online handwriting
recognition system was built by adapting the BYBLOS CSR[33] speech recognition
system. Moreover, there is an attempt to build a real time handwriting recognition system

28
of the baseline BYBLOS handwriting recognition system without sacrificing much
accuracy. The BYBLOS recognizer is a writer-independent mainly concerned with cursive
handwriting. Moreover it is sentence-based recognition. No constraints are imposed on the
users except the following regulations while collecting the BBN data corpus [35] by which
the research is based on.
• All small letters should be connected which means mixed style (discrete and
cursive) is not allowed.
• The dots above letters like i should be written after the word is completely written
i.e. it is not allowed to lift the pen without finishing a word.

The BYBLOS online handwriting recognition system has three modules: Front-end,
trainer and the decoder. The front-end module is responsible for generating observation
sequence for the HMM from the input handwriting data. It includes three steps: pre-
processing (constructing invisible stroke and padding filter), computing feature vector with
6 features (pen up/pen down bit, delta-x position, delta-y position, writing angle, delta
writing angle, and sgn(x-max(x))) and computing the observation sequence (by converting
the feature vector to the observation sequence which could be the input for the recognizer).
The trainer module is implemented for the purpose of training the HMM. The decoder,
employs the fast match search and the Viterbi beam search to produce the most-likely
sentence. It is reported that it has achieved a good accuracy having only a 13.8% error rate.

Then, the paper suggested that it is still possible to improve the recognition accuracy to an
acceptable and usable level by adding more features which were not included in the feature
vector. Accordingly, four features are added: the vertical height (will help to distinguish
characters like ‘ and ,), the space feature ( a new way of handling inter-word spaces by
which space features will be calculated only for invisible strokes), hat stroke (to handle the
misleading situation of putting dotes and crosses once the word is written like in i and t)
and sub-stroke features (strokes are further divided to sub-strokes to incorporate
information within a bigger neighborhood of the sample). After adding these features, the
error rate is reduced by 34% and it becomes 9.1%. The improvement after each features is
added is summarized in Table 3.1 below.

29
 Feature added Feature Previous Current Error-reduction
vector size error rate error rate percentage
Vertical height 7 13.8% 13.6% Not significant
Space feature 8 13.6% 11.3% Significant
Hat stroke 9 11.3% 10.7% Significant
Sub-stroke features 10 10.7% 9.1% Significant
Table 3.1. Improvement of accuracy by adding features

A real time online recognition system that run on an SGI Indy workstation and GRiD. PC
is also built.

3.2.1.4. E. Gomez et al. :Fuzzy Neural Network

The following research goals are set in the paper[15].
• Processing time should be low enough to run on a Pentium 120MHz and to
complete the tasks within specific time (276ms/char for the test process, 7ms/char
for segmentation, 5ms /char for feature extraction and 256ms/char for the actual
test).
• Training time should be low.
The main approach is applying the neuro-fuzzy system named FasArt to build the
recognizer.

The system is presented in the paper to have three stages: preprocessing (basically
character segmentation), feature extraction and classification. The character segmentation
is done in two different approaches one of which is based on geometric features and the
other is based on biological models. Both methods try to find the optimal segmentation
points from a given character. Then feature extraction is conducted in two ways: Shannon
entropy and clustering maps. The features that they use are length of stroke, two phase
points, total phase and a feature to tell whether it is the beginning or the end of a stroke.
Finally the classifier based on Fuzzy logic system is applied.

The test is conducted based on the UNIPEN data set. The system achieved an accuracy rate
of about 85.39% for digits, and 59.57% for lower case letters and 66.6% for upper case
letters when geometric segmentation is used. On the other hand, employing biological
model segmentation, an accuracy rate of 82.52% for digits, 76.39% for upper case letters
and 58.92% for lower case letters is achieved.

30
Processing time required by the proposed system when tested on a Pentium 120MHz PC,
measured with Linux time, is 276 ms/char for the whole test process, including 7 ms/char
for segmentation, 5ms/char for feature extraction, and 265 ms/char for the actual test. An
overall speed of 0.28 sec/char is achieved by the proposed system.

3.2.2. Japanese/Chinese Online Handwriting Recognition Systems

Two papers that compare and contrast the currently used techniques in Japanese and
Chinese online handwriting recognition systems with the methods applied in Western
(Latin) systems are summarized below.

3.2.2.1. S. Jaeger et at.: The state of the Art in Japanese Online

Handwriting Recognition
This paper compares the techniques used in Japanese online handwriting recognition
systems with those used in Western systems[12]. The recognition is divided into three
processing steps: preprocessing, classification and post processing. It is also stated that
many statements made in the paper are also valid for Chinese character recognition, or
Asian character recognition in general, since the Japanese character set comprises a subset
of the Chinese character set. The set of techniques for each step in Japanese and Western
systems are analyzed as follows:

Preprocessing Techniques:
In this section, the paper compares the techniques used for the preprocessing tasks
normalization, re-sampling, slant correction and data compression in the two recognition
systems.
• The most common normalization in JOHR1 is size normalization, which could be linear
or nonlinear. It is also reported that nonlinear normalization techniques (such as line
density equalization technique) are better that linear methods as they improve
recognition accuracy.
• It is usual to compare Japanese characters and Western words in terms of complexity.
But, normalization techniques used for Japanese characters are not well suited for
Western words. For western words, size normalization is done by first identifying lines
such as base line, corpus line, ascending line, descending line and etc. Then the corpus

1
This is to say Japanese online handwriting recognition system.

31
 height (height between baseline and corpus line) will be mapped to a fixed height to
normalize the words.
• Re-sampling is common in Western systems where as it is not usually implemented in
JOHR.
• Slant correction is also another normalization task which is common in Western
systems, but are not generally implemented in Japanese systems.
• Data compression is a common task in western as well as in Japanese systems.

Feature Computation:
Feature computation is the next task once the preprocessing is done to the raw handwriting
data. Features are high-level attributes derived from normalized raw data. There are two
types of features namely online and offline features. These two types of features have their
own positive and negative sides.

In general, offline systems (solely based on offline features) are less accurate than online
systems and this implies that online features have positive impact. Conversely, online
recognizers suffer from stroke number and order variations (even worse in Japanese and
Chinese systems). Thus, incorporating online features with offline features in offline
handwriting recognition systems or vice versa is a common practice to improve accuracy.
Regarding this, the western people focus on incorporating online features with offline ones
for offline handwriting recognition systems. But in Japanese, the reverse is true. As stated
above, Japanese online handwriting recognition systems try to use offline features to
handle stroke order and number variation found in online data. But extracting offline
features from online data is not straightforward. Histogram features are statistics describing
absolute and relative frequencies of feature values in handwriting trajectories. Histograms
help to extract offline data from online data by avoiding the dynamic part of the online
data. The trajectory in the handwriting will be divided into equal cells and the selected
feature values (such as directional feature, directional feature) will be counted within the
cells, leaving their order of occurrence i.e of course the dynamic part of the data.

Classification:
Obviously, the final target of recognizers is to assign a character to unknown pattern. The
paper reviews some classification methods in Japanese and western recognizers.

32
Template Matching techniques:
Most JOHR systems are based on nearest neighbors classifiers (template matching
techniques). This technique will compute the distance between the unknown pattern and a
template stored in a database and assign the template with shorter distance to be the
classified character. More elaborated template matching techniques like elastic matching
techniques are also used. Elastic matching (sometimes called Dynamic Programming
matching-DP matching) finds the matching template, which requires the least deformation
necessary for transforming it to the unknown input pattern. DP matching is mostly used in
post-processing in western systems, but in Japanese systems, it is a core technique for the
classification step.

Hidden Markov Models:

HMMs are becoming common approaches in Western systems, however there are only a
few number of Japanese recognizers that use HMMs. The reasons are:
• HMMs are good for implicit segmentation and are used for cursive Latin words, but
segmentation is not a main problem in JOHR.
• HMM modeling will be very complicated for Japanese characters due to stroke
order and number variations. Even those JOHR systems applying HMMs use
different modifications.

AI Methods:
AI methods such as neural networks have shown the same level of applicability in both
western and Japanese systems. However, it is reported that, these methods were very much
appreciated around the 1990s and became faded during the middle of the last decade.

Pattern Representation:
Application-specific dictionaries in western systems facilitate recognition. The dictionary
will be represented by a tree whose nodes are characters in the word. This will make
searching very fast because it avoids flat searching techniques. But in the case of Japanese,
such a tree will be prepared for characters (not words) where the leaves of the tree
represent the subparts of the characters usually called radicals. It is here fair to note that

33
Japanese characters are usually compared with Western words (not western characters) in
terms of complexity.

Recognition rates:
Talking about recognition accuracy of systems really needs a standardized data on which
training and testing could be conducted. Considering this fact, there are different projects
around collecting online data set. For western systems, we can state two of them: UNIPEN
and IRONOFF. Japanese are also doing the same thing and two of the benchmark data sets
are mentioned in the paper. These are Kuchibue (having around 1,435,440 handwritten
Japanese characters) and Nakayosi. The two of them together contain more than 3 million
characters written by 283 writers. Recognition rates based on Kuchibue is around 90%. For
western systems, the paper reported the recognition rate only for the published Npen++
system[34] as 96% (WI mode & 5000 word-dictionary), 93.4% (20,000 word-dictionary)
and 91.2%(50,000-word dictionary).

Post-processing:
Postprocessing is the process of improving recognizer output by means of additional
information sources such as vocabulary, spelling etc. Most western handwriting
recognition systems employ bigrams and trigrams. Bigrams and trigrams provide
statistical information on character level or word level. They are generally used to indicate
the probabilities of combinations of characters in a word or words in a sentence. In
Japanese systems, a similar technique of employing bigrams and trigrams and even n-
grams for postprocessing has been applied only sometimes. Another postprocessing
activity is combining classifiers such as combining online and offline systems.
Improvement by combining online/offline systems has already been reported for Asian
character recognition.

The following points are emphasized in the paper regarding major classification
approaches applied in Japanese/Chinese and Western systems:
• Most western systems adhere to HMM, but most Japanese handwriting recognition
systems use nearest neighbor technique.
• HMMs and neural networks are not commonly applied in Japanese handwriting
recognition systems due to the high number of Japanese characters and stroke order

34
 and number variations occurred during writing. But substroke HMMs could be good
options.

3.2.2.2. C.L.Liu et al.: The state of the Art in Chinese Online

Handwriting Recognition
According to C.L.Liu[13], A typical Online Chinese handwriting recognition system
passes through the following steps:
a) Character segmentation: The handwriting input is segmented into character patterns,
but the segmentation is not for sure 100% correct at this early step until contextual
processing is done. So, what we get from this step is a set of segmented (candidate)
patterns.

b) Preprocessing: It includes noise elimination (by smoothing, filtering, wild point

correction and stroke connection), data reduction (equidistance sampling and line
approximation [feature extraction]) and normalization (linear, moment, and nonlinear
where nonlinear is the efficient technique for Chinese characters).

c) Pattern description (Feature Extraction): This step is a very important one since the
classification is actually carried out by finding a match between the pattern description
of the input data and the reference model. Three representation schemes are used:
statistical, structural, and hybrid statistical-structural.

d) Character classification: The character classification is divided into two stages namely
coarse classification and fine classification. During the coarse classification, multiple
classes are assigned to the input pattern either by using class set partitioning or
dynamic candidate selection techniques. The techniques used in fine classification are
grouped into three: structural matching, probabilistic matching and statistical
classification.
Structural Matching: The input pattern is matched with the structural model of each
candidate class and the class with the minimum matching distance is taken as the
recognition result. Various approaches are used. DP matching (choosing minimum edit
distance), Stroke correspondence (the distance is computed by summing up the
between-stroke distances), relational matching (the correspondence is searched under
the constraint of relationships) and knowledge-based matching (the prior knowledge of
character structure and writing is used as a constraint) are mentioned.

35
 Probabilistic Matching: Matching distance between the input pattern and the
references will be computed and the character with lesser distance will be taken as
recognized characters.
Statistical Classification: The input pattern will be described in feature vector and
various statistical techniques are applied to find a match.

e) Model Learning and Adaptation: During classification (whether it is coarse or fine),

the model database is consulted. The model database can be built by model learning or
manually using prior knowledge. But model learning improves accuracy than manually
built models. Different learning approaches are described. These are mean prototype
learning, multi-prototype learning, structured learning and writer adaptation.

f) Contextual Processing: This activity helps to find the optimal match by using linguistic
knowledge. Moreover, it also helps in assuring segmentation.

Performance Evaluation:
The performance of recognition systems is measured in terms of recognition accuracy,
memory requirement, and recognition speed. The evaluation reported on the paper is based
on the Nakayosi and Kuchibue data set. And in summary,
• The authors believe (based on collected information) that it is currently possible to
achieve a recognition rate above 98% for regular scripts. But for fluent (or fluent-
regular) scripts, it is difficult to achieve a recognition rate above 90%.
• Statistical methods use large amount of memory (For eg. over 30MB) and offer
high speed where as structural methods are slower but use a small amount of
memory (for eg. 166KB).

3.2.3. Arabic Online Handwriting Recognition Systems

A paper by T.Klassen[14] that presents Arabic online handwriting recognition system
developed by using neural networks is summarized below. The good thing about the paper
is that it also reviews prior works.

Review of Previous Works:

The author of the paper reviewed six previous works. The following table summarizes the
review. (See Table 3.2)

36
 No. Developer(s) Technique used Accuracy

1 Al-Sheik & Hierarchical Rule-based 100% for isolated letters, but unclear
Al-Taweel whether this rate is obtained on the
training or test data
2 El-Emami Segmented Structural Tested on 10 writers. They instructed
& Analysis: one of the tester (with 86%) to change
Usher (Structural analysis for his style of writing and they get 100%
feature selection and a accuracy.
decision tree for
classification)
3 Bouslama Structural and Fuzzy: It is reported on the paper that no test
& result had been listed.
Structural analysis is for
Amin
classifying different
characters & fuzzy
approach handle the
variability within
people’s handwriting for
the same class.
4 Alimi & Template matching and Using 9 prototypes, 96% accuracy on
Ghorbel Dynamic Programming test data for one author.

5 El-Wakil & Hierarchical Template 84% by testing 7 writers on sets of 60

Shoukry: Matching and k-nearest characters.
Neighbor Classification
6 Alimi Evolutionary Neuro-Fuzzy 89% without dot or diacritical
information
Table 3.2: Previous works on Arabic Online Handwriting (adapted from [14])

Then the paper also identifies the strength and weakness of the reviewed works. Another
table is provided below to outline the pros and cons of the methods (Table 3.3).

37
No. Method Strength Weakness
1 Hierarchical Rule- Good divide-and-conquer Sensitive to noisy data
based strategy
2 Segmented Structural Automatic segmentation of Sensitive to rotation and
Analysis words less data was used.
3 Structural and Fuzzy Had perfect training results No test result is reported
4 Template matching Accuracy (96%) was very Only one test subject
and Dynamic good involved and difficult to
Programming generalize
5 Hierarchical Template Good accuracy Sensitive to noisy data and
Matching and k- difficult to generalize since
nearest the test set contains only 60
Neighbor characters
Classification
6 Evolutionary Neuro- Was more robust, and The test set was constrained
Fuzzy segment in a novel way to
only one word and a small
subset of 7 different letters
and the system is writer-
dependent.
Table 3.3. Strength and Weakness of Approaches for Arabic online handwriting system.

General Description of the System:

The phases of this system are data collection, file storage, segmentation, sampling,
normalization, feature extraction and classification.
Data collection: The data collection was done using Wacom Graphire Model ET-0405-U.
2769 samples were collected. The program used to collect data is WinTab during data
collection, noises were evident to appear and handled in different ways.

File Storage: Though the ultimate goal is to make the system work on real time, it is
difficult to train and test the system repeatedly by bringing people every time an
experiment needed to be conducted. So, it could be simulated by storing the data in a
persistent form like a file. The data was stored in standard format namely in UNIPEN.

Segmentation: Since isolated characters are considered, word segmentation to letters is

assumed to be done. But segmentation of letters into primary and secondary strokes is
implemented.

Sampling (Critical Point Extraction): To reduce the amount of data, only critical points
will be extracted that can characterize the character. The procedure for ChkLOS (check
line of sight) is used to choose the critical points.

38
Normalization: The main purpose of normalization is to make the size of the letters the
same and put in the same position. Scaling normalization (size normalization) is done to
handle the size variability of people’s handwriting by mapping all handwritten letters to a
pre-defined size. Translation normalization is another normalization activity to translate
the letter to the same position relative to the origin. Another type of normalization is time
normalization, which is used to normalize the temporal sequence data. Rotation
normalization is the other normalization tasks included and used to make the data set
rotation-invariant. Moreover skew normalization is considered to correct slants.

Feature Extraction: The two purposes of feature extraction are to see that all sampled
points are not useful for pattern recognition and to reduce the amount of input data for the
case of neural networks. The system uses a neural network called SOM (Self-Organizing
Maps). It is reported that SOM has a capability of extracting relevant features used to
classify characters avoiding the manual extraction of features. Justification to use this tool
and its design are presented in the paper in detail.

Classification: Multi-layer perceptron (a kind of neural network) are employed to carry out
the classification task. The paper stated that online recognition of Arabic characters is a
non-linear problem and these neural network are appropriate. Optimization methods such
as pruning and partitioning are also applied.

Experimental Results:
The data for training and test were collected as described above. The experiments had 3
trials on 3 disjoint data sets: training (1656 letters), validation (1113 letters) and test (692
letters).

Trial Training data set Testing data set Accuracy

Trial 1 Training set Validation set & test 72%
without partitioning
and pruning
Trial 2 Partitioned & pruned on Validation set & test 80%
Training set set
Trial 3 Trained with training set & Test set 85%
partitioned and pruned on
validation test
Average 79%
Table 3.4 Recognition Accuracy reported in [14]

39
3.2.4. Handwriting Recognition Systems For Ethiopic Characters
As far as we know, there are only three researches [6, 7, 8] done on handwriting
recognition for Ethiopic characters. All of them are offline handwriting recognition
systems. Though the previously mentioned systems are all offline, we presented them here
to see how Ethiopic characters are handled.

3.2.4.1. Amharic Handwriting Recognition for processing of Bank

Checks
Nigussie Tadesse[6] conducted a research on handwritten character recognition for the first
time. He attempted to adopt a recognition algorithm for handwritten Amharic text so as to
apply the result of the research to the processing of bank checks. After collecting data
(amounts of money written on various checks in Amharic words), a pre-processing step
had been taken to minimize the intra-class variability through underline removal, stroke
width normalization and slant normalization if any. After extensive and careful pre-
processing activities were carried out, the segmentation of characters had been done by
extracting connected components first, then examining the aspect ratio and proximity of
each connected component from its neighbor.

A multilayer feed forward neural network with error-back propagation is applied for
recognizing the character after normalizing each segmented image of the character to a
16×16-pixel matrix. Then, the neural network is trained by giving specific values for
parameters such as target error, learning rate and momentum and this training took about
12 hours (See Table 3.5). Unfortunately, the recognition accuracy of the network was very
poor when it was tested after the training. This situation initiated the researcher to increase
the hidden layer neurons to 20 and the number of sample characters for training by 7 in an
attempt to get a better result. But no significant improvement was exhibited. For the third
time, a neural network is created and trained with a different value of target error and with
more characters. However, in spite of all the trials, the neural networks didn’t give any
better achievement. The result of the work is summarized in Table 3.5 below.

The parameters in the table (Table 3.5) will affect the recognition capability of the neural
network and they are part of the architecture of the neural network. The researcher
increases the number of hidden neurons to improve the accuracy of the neural network.

40
 Parameters to determine characteristics of 1st neural 2nd neural 3rd neural
the neural networks network network network
Input neurons 256 256 256
Hidden neurons 7 20 20
Output neurons 8 8 8
Target error 0.05 0.05 0.18
Learning rate 0.6 0.6 0.6
Momentum 0.8 0.8 0.8
No. of characters for training 135 498 498
No. of hours taken for training 12 24 14
No. of characters for testing 38 38 38
*
No. of characters correctly recognized 2 16
Table 3.5: Summary of result of [6]

3.2.4.2. Amharic Handwriting Recognition for Postal Address

Interpretation
Messay Hailemariam[7] attempted to apply OCR technology to the domain of postal
addresses interpretation by exploring and testing the application of simple geometric
calculations and line fitting techniques for handwritten Amharic character recognition.

In this work, pre-processing activities are conducted in a limited way to remove noises
(undesired artifacts) caused by interferences of strokes of some characters into other
characters. After the image of the document is scanned and digitized, segmentation
technique, which is known as stage-by-stage segmentation in which line, word and
character segmentation are done in order to come up with the image of the individual
characters. A 32×32 pixel size is chosen as a standard size so as to normalize the
characters. Once the normalized characters are produced, the feature extraction is done by
local line fitting (LLF) technique which is based on simple geometric operations. The
method is known to be efficient and fast because it avoids some pre-processing steps made
on the input. Finally, the neural network is trained by taking the extracted features as an
input vector.

*
The number of characters correctly recognized is not clearly stated, but it is reported that a significant
improvement did not exhibited.

41
The research is completed by getting the following results:
1. Testing the neural network for the data that it is already trained with results in having a
success rate ranging from 73.25%-91.9%.
2. The neural network was able to recognize new handwritten characters which were not
in the training set with a success rate ranging between 33.25% and 88.6%.

3.2.4.3. Traditional Amharic Handwriting (“Yekum Tsifet”)

Recognition
In 2004, Wondwossen Mulugeta[8] extended the research on Amharic handwriting
recognition using OCR. He intended to apply the OCR technology to recognize special
type of handwritten Amharic text which is traditionally called ‘Yekum Tsifet’ using all the
231 basic character sets [Wondwossen, 2004]. Wondwossen studied the basic features of
the traditional handwriting style and applied the artificial neural network approach to
develop a recognition engine for this purpose. He also investigated the application of the
result of the research by claiming that converting the existing large number of handwritten
historical and religious document written in ‘Yekum Tsifet’ would help to access the
information digitally. In his work, Wondwossen used the global thresholding approach to
binarize the documents. Stage-by stage segmentation techniques were found to fit
remarkably for the extraction of ‘Yekum Tsifet’ characters since they are written at the
same level without ascent and descent features [Wondwossen, 2004]. Visual C++.Net built
in function is used to normalize the extracted characters to a standard size of 20×20 pixels
or alternatively 16×16 pixels. Though this process may deform some of the characters, it is
a vital step to create an appropriate standard sized input vector for the neural network
recognizer. This step is followed by the production of the binary representation of the
normalized characters in a text file.

Wondwossen designed and developed an artificial neural network character recognition

engine that accepts the binary representation of a segmented character as an input and
produces the corresponding recognized character (or ‘Fedel’). The architecture of the
neural network, which is a feed forward neural net with back propagation algorithm, is a
multiplayer perceptron that has three layers namely the input, hidden and output layer.

42
The result of the research is reported as follows:
1. The artificial neural network recognizes most of the patterns it is trained with in a
recognition rate of 95.24%-98.7%.
2. The artificial neural network recognizes new patterns with a recognition rate of 16.1%-
31%.
3. The artificial neural network is tested with original training patterns with different level
of random noise imposed on it and a recognition rate of ranging between 27.27% and
94.8% is achieved.

3.3. Summary
To wind up the survey, we have provided the following points to summarize what we
obtained from the survey:
• Due to the cursive nature of Latin handwritten words and sentences, discrete Hidden
Markov Models are increasingly used in Latin online handwriting recognition systems.
The reason is that HMMs facilitate implicit segmentation which is the most difficult
problem in online word-based and sentence-based Latin recognition systems.
Limitations that are related to HMMs are the necessity to create individual models for
all entities (character, words or sentences) and stroke order dependency. The number of
Latin characters is minimal relative to Japanese/Chinese or Ethiopic character sets and
Latin online systems benefit from the fact that their characters are not as many and as
complex to be modeled by HMMs. However, as a direct consequence, HMMs are not
commonly adopted in Japanese/Chinese systems due to the large number of characters
and the complexity of the characters. Another reason for not considering HMMs for
Japanese/Chinese systems is that Japanese/Chinese writing is subject to stroke number
and stroke order variations. In spite of this, some researchers claimed that substroke
HMM are good options for Japanese/Chinese systems.
• The complexity of Japanese/Chinese characters make them to be compared with Latin
words (not characters). Structural pattern representations and template matching are
common techniques in Japanese/Chinese recognition systems. Hierarchical
representation such as tree representation scheme is employed to represent a
Japanese/Chinese character where the leaves of the tree are the strokes in the character
unlike the case of Latin systems in which trees are sometimes used to hierarchically
model words where leaves are letters.

43
• The analysis done on online systems for Arabic characters show that structural methods
are commonly used to build Arabic online handwriting recognition systems. However,
it is also reported that good recognition accuracy is obtained by applying neural
networks.

The Ethiopic character set is not as simple as Latin character set. However, it is not also as
complex as Japanese/Chinese character set. This conclusion has been reached based on the
fact that the number of characters in the Ethiopic character set is more than 300 which is a
big number as compared to the 26 Latin characters. The Japanese character set contains
6000-7000 characters that makes it more complex than the Ethiopic character set.
Moreover, the complexity of Japanese/Chinese characters is expressed in terms of the large
number of strokes used in a single character. The main difference between Ethiopic and
Arabic characters is that Arabic characters are cursive in nature and Ethiopic characters are
not.

From the lesson learnt from the trend of building online handwriting recognition systems
for the three character sets (Latin, Japanese/Chinese, and Arabic), we have set the
following guideline as to continue the development of the recognition system.

The structural approach seems to be more appropriate for Ethiopic character set as HMMs
are good options for cases in which segmentation is a big problem and the number of
character is less in number. Since Ethiopic characters are naturally written in discrete
manner, segmentation would not be a big problem. Moreover, the number of characters is
above 300, which is not a small number. From experience, we observed that different
Ethiopic writers write similar characters with different stroke type, number and order.
Thus, we underlined that the structural approach should not be based on prior identification
of strokes in the Ethiopic characters. Rather the set of strokes must be obtained from the
writer’s style. This also makes the system to be writer-dependent and evidently increase
the recognition accuracy.

44
 CHAPTER FOUR
THE ETHIOPIC CHARACTERS

As the main concern of this work is the Ethiopic character set, in this chapter we discuss
about the Ethiopic character set as a whole and the analysis done on the shape of the
individual characters while conducting this research.

4.1. Introduction
People started to represent ideas in marked symbolic form long ago. Writing has been a
means of communication and a persistent tool to represent information. Writing systems, in
particular full writing systems, have been constructed for different languages for the
purpose of expressing unambiguously any concept that can be formulated in the
language[20]. Archaeological discoveries suggested that the Egyptian hieroglyphs might
be the oldest form of writing which was active from about 3300 or 3200 BC[20]. Writing
systems evolved from logogram systems in which complete words are represented by signs
called logograms. To overcome the disadvantages of logogram systems such as their
ambiguous nature of representing words, syllabic systems were developed which are
concerned about representing sound by using signs. More improved systems are alphabetic
systems in which consonant and vowels are separately written[20]. Alphabets are set of
symbols or letters to represent sounds of a language so that words in the language could be
written by combining the alphabets[21]. Alphabetic writing systems are the most common
writing systems, though some writing systems like Chinese and Japanese still do not adopt
them[21].

The Ge’ez script, which dates back to AD 300, is one of the native African writing systems
[22]. The original Ge’ez script has only 26 consonants and has no vowel indications until
around 350 A.D.[7]. But, later, vowels were incorporated by adding six non-basic
characters for each consonant symbol. These were created by undergoing structural
changes to the existing consonant symbols and this made the total number of characters in
Ge’ez script 182[7]. Eventually, the Ge’ez script gave birth to the Amharic writing system,

45
which now is used to write Amharic and other Semitic languages like Tigre, Harari and
Gurage in Ethiopia[23].

In addition to the 26 consonants and their six non-basic forms that exist in Ge’ez system,
the Amharic writing system later added 7 other consonants whose shape derived from
some of the existing Ge’ez characters[7] (See Table 4.1.). This increases the total number
of characters to 231 (33×7) in the newly born Amharic character set.
Derived Derived
Character From

Table 4.1. Derived Characters in the Amharic Character Set

4.2. The Ethiopic Character Set

The total number of characters in the Amharic character set is 306 if the Ethiopian
numerals (Table 4.3), punctuation symbols, and the extended character group ( , etc)
are counted with the alphabetic characters. From these, 231+7 of them are alphabetic
letters (See Table 4.2). Basically, the number of basic characters is 33 and the total number
of alphabetic characters will be 231. But, the character is included as a special character
to represent the sound ‘v’ from Latin-based languages and it has also six non-basic forms.
That is the reason why the number of alphabetic characters is referred as 231+7. The
Ethiopic numerals contain 20 symbols defined for each of the first 10 natural numbers (1 to
10) and for numbers that are multiples of 10 in the range 11-1000. Other numbers are
written by combining these 20 distinct digits. For example, the number 25, is written as
, which is obtained by writing 20 ( and 5( side by side. But, these digit characters
are not frequently used by individuals to write numbers. For this reason, the Ethiopic
numerals are not considered in our recognition system. The other extended group of
characters such as , , and etc are also not considered in our work.

This leaves us with the remaining 231+7 alphabetic characters. The beauty of the Ethiopic
character set, particularly the alphabetic letters, is not only in its appearance [23] but also

46
in its systematic organization. The characters are arranged in the form of table with 33+1
rows and 7 columns. The characters in the first column are called basic characters and
they are 33+1 in number. The other six columns include the non-basic characters whose
shape is derived from the basic characters. More importantly, the shape derivation is quite
interesting for this research as character recognition is the main issue. We have analyzed
the way the shapes for the non-basic characters are derived for the purpose of designing a
better strategy or method of character recognition. The next section presents the analysis.

! "
# $ % & ' ( )
* + , - . /
0 1 2 3 4 5
6 7 8 9 : ;
< = > ? @ A
B C D E F G
H I J K L M
N O P Q R S
T U V W X Y Z
[ \ ] ^ _ `
a b c d e f
g h i j k l m
n o p q r s
t u p v w x
y z { | } ~ •
€ • ‚ ƒ „ …
† ‡ ˆ ‰ Š ‹
Œ • Ž • • ‘ ’
“ ” • – — ˜ ™
š › œ • ž Ÿ
¡ ¢ £ ¤ ¥
¦ § ¨ © ª « ¬
- ® ¯ ° ± ² ³
´ µ ¶ · ¸ ¹
º » ¼ ½ ¾ ¿
À Á Â Ã Ä Å
Æ Ç È É Ê Ë
Ì Í Î Ï Ð Ñ Ò
Ó Ô Õ Ö × Ø Ù
Ú Û Ü Ý Þ ß à
á â ã ä å æ ç
è é ê ë ì í

î ï ð ñ ò ó
Table 4.2. The Amharic Core Character Set

47
 ô õ ö ÷ ø ù ú û ü
1 2 3 4 5 6 7 8 9 10
ý þ
20 30 40 50 60 70 80 90 100 1000
Table 4.3. The Ethiopic Numerals

4.3. Analysis of the Ethiopic Characters

The characters in the Ethiopic character set are broadly divided into two groups as basic
characters and non-basic characters. The number of the basic characters is 33+1 and they
are listed in the first column of Table 4.2 and are referred as first order characters
afterwards. The characters in all the other six columns in Table 4.2 are non-basic
characters and each of them are derived from their basic relatives. These characters are
referred as second-order, third-order etc depending on the order that they appear in Table
4.2.

4.3.1. First Order (Basic) Characters

The basic characters are almost distinct in shape except some of them. In fact, the 26
original characters from the Ge’ez script differ much in shape. Even though it gives sense
to say that few characters from this set are similar (such as and á), we prefer grouping
the 26 original characters in one group and the remaining derived characters and the
special character in another category as indicated below.

#
T g y
Œ “ ¦ -
Ú Ì Ó á
Table 4.4. Group1 (Basic character) Table 4.5. Group2 (Basic character)

4.3.2. Second Order Characters

The shape of these characters is generally formed by attaching a horizontal stroke to the
middle of the right side of the basic characters. Examples are (from ) [ (from ),
´ (from ), º (from ), Ô (from Ó), etc. The second-order characters are consistent in
following this single rule. Exceptional to this rule are < (derived from and â (derived
from á).

48
4.3.3. Third Order Characters
According to our observation, two rules are mainly applied to derive the shape of the third
order characters. These are:
Rule 1: If the basic characters have leg(s), then extend the right leg horizontally to the right.
Examples are %(from #), +(from ), I(from ), ¡(from ).
Rule 2: If the basic characters have no legs, attach a leg and extend it in the similar way
stated in Rule 1.
Letters following this rule are (from ), 1 (from ), 7 (from ), Ž (from Œ),
• (from “), etc.
Only three characters from the third order character do not follow one of these rules. These
are =(from ), ¨(from ¦), ã(from á).

4.3.4. Fourth Order Characters

It is possible for us to see that the fourth order characters are derived by applying the next
rules:
Rule 1: If the base character has two or more legs, then shorten the left legs or all legs
except the right most one. For instance, &(from #), | (from y), ¶ (from ), È
(from ) are constructed using this rule.
Rule 2: For base characters with one leg, extend it to the left (or connect a horizontal stroke
running to the left to the leg). Characters like P (from ), ] (from ), c (from
), Ö (from Ó).
Rule 3: If the base character has no leg at all, reduce the size of the character by aligning to
the upper corner and attach a long vertical stroke to the right side of the base
character. Examples: (from ), 2 (from ), • (from Œ), © (from ¦).
However, the characters >, p, , ä seem to follow none of the abovementioned rules and
are considered exceptional in their group.

4.3.5. Fifth Order Characters

The fifth order characters mostly add loops to the base of the right legs of the characters.
Particularly, the following rules are applied to construct them.
Rule 1: For base characters with two or more legs with the right leg free from loop, attach a
loop to the right leg. Examples are ' (from #), ^ (from ), Ã (from ).

49
Rule 2: If the basic characters have no legs, reduce the size of the character by aligning to
the upper corner and attach a vertical stroke with a loop at its bottom end to the
right of the base characters. For instance, (from ), 3 (from ), • (from Œ) are
constructed this way.
Furthermore, the characters ª (from ¦), É (from ) are derived by adding a horizontal
stroke with a loop at its right end and ? (from ), å (from á) are formed by attaching a
loop to the horizontally lying end of the base characters.

4.3.6. Sixth Order Characters

Constructing rules for the sixth and seventh order characters is difficult as they are highly
irregular. Thus, we grouped the characters by putting the characters whose shapes are
constructed in a similar way in the same group. The following table (Table 4.5) shows the
categorization.
Group No. Characters
Group 1 ! „ Š ~
Group 2 (
Group 3 . R _ e Ä Ê ì
Group 4 Y
Group 5 4 :
Group 6 @ æ
Group 7 F L ˜ Ø
Group 8 l r w
Group 9 ‘
Group 10 ž ¤
Group 11 «
Group 12 ¸ ¾ ß Ñ
Group 13 ²
Table 4.6. Grouping the Sixth order character

4.3.7. Seventh Order Characters

The majority of the seventh order characters are constructed by shortening the right one or
two legs of the base characters with two or more legs. A few examples are G Z … Ÿ etc.
For characters with no leg, a straight (like in the case of 5) or left-declined vertical line (as
the case of ; ’ ™ Ù í ) is connected to the bottom of the characters. Other seventh
order characters such as " ` S … are formed by putting a loop at the top of the base
characters. Further, it is observed that the letters ) ¬ ³ are unique in derivation of their

50
shape than the others. In general, it can be said that the seventh order characters follow
different patterns and recognizers must consider these different patterns differently.

4.4. Conclusion
The shape and pattern formation of the Ethiopic characters is categorized and presented in
the subsections above. Using this information as a background, we are able to come up
with points of decisions discussed in the following paragraphs.

Though, some irregularities are seen in the non-basic characters particularly in the sixth
and seventh order characters, it is generally possible to say that the shapes of non-basic
characters is more or less similar to the basic ones. In other words, the non-basic characters
add some minor changes to the shapes of the basic characters. This fact lead us to a
strategic move to first consider the 33+1 basic characters whose shape is somehow distinct.
Then, an online Ethiopic handwriting recognition engine that recognizes the basic
characters can also easily recognize the associated non-basic characters. This strategy leads
to the approach of taking the basic characters as fundamental and primary set of characters
in a recognition system. This idea is strengthened by looking at the experiences of other
systems developed for Ethiopic characters. Though not exactly the same, identical strategy
was used for the construction of the manual Ethiopic/Amharic typewriter keyboard layout,
the traditional Ethiopic text entry for modern computer keyboards and the phonetic
Ethiopic text entry on computer keyboards. Then, the non-basic characters can be
considered as a secondary set of characters that can be recognized based on the basic
characters.

Ethiopic characters entering via keyboard in phonetic mode is one common example. Most
systems assign each key for English consonant to a Ge’ez basic character of similar sound.
In some cases, there will not be a single consonant for some Ethiopic characters with
similar sound. For example, the sound of the character (read as ‘che’) cannot be
associated to any of the consonants in English. Handling these cases differ from system to
system. In any case, all basic characters will be assigned to some consonant. The most
interesting case is the case of non-basic characters. Non-basic characters are usually
entered by using combination of keys, of which one is always the key assigned to their
basic character and the other corresponds to a vowel, which could probably create their

51
sound. For example, if k is assigned to (read as ‘ke’), then € (read as “ku”) can be
entered by pressing k and u together (k+u). This indicates that there is a habit of looking at
the basic characters and non-basic characters in different views. Our approach is no
different because we claim that it is also possible to create a two-level character entry for
non- basic characters by providing an interface that looks like the one shown in Fig. 4.1.

Six small boxes that

correspond to the written
Input area for basic character.
users to write
basic characters

Fig. 4.1. A Two-level Character Entry Interface

Users may write basic characters in the larger box as shown in Fig. 4.1. If the users are
interested to enter one of the non-basic characters, then they are expected to tap one of the
six boxes that correspond to the non-basic characters. Thus, the recognition of non-basic
characters will be based on the recognition of basic characters.

Another alternative to justify this strategy is by arguing that segmentation algorithms could
extract the shape of basic characters and the additional strokes used to form the non-basic
characters given the non-basic characters (See Fig. 4.2). This implies that an additional
effort is needed only to recognize the additional strokes and their position to recognize the
non-basic characters if the recognition of basic characters is done.

Segmentation

Fig. 4.2. Segmentation applied on (extract basic character and the stroke)
However, data entering via keyboard and using pen like devices should be handled
differently despite the similarity of the goals of the two systems which is interpreting the
input data to a character. In keyboard-based data entering scheme, once the keys are
assigned to the characters, the remaining task will be displaying characters by examining
the keys pressed. In online handwriting recognition system, a carefully designed algorithm
should be employed to handle the variability.

52
 CHAPTER FIVE
ONLINE HANDWRITING RECOGNITION
SYSTEM FOR ETHIOPIC BASIC
CHARACTERS

5.1. Introduction
The decreasing cost, the increasing capacity, the pervasive feature, and the increasing
availability of applications are the major factors that initiated us to think that an Ethiopic
online handwriting recognition system is a requirement of its usage in Ethiopia. The usage
of increasing applications and usage of handheld computers of demands an online
handwriting recognition system as one of the major input methods Taking this initiative,
we started our work by investigating the existing systems developed for other characters
such as Latin, Japanese/Chinese and Arabic. We have also studied the nature of the
Ethiopic characters in relation to these character sets and their existing online handwriting
recognition systems. This has given us the way to choose a general approach, and to design
appropriate algorithms that could help to realize the development of online handwriting
recognition systems for Ethiopic characters. The recognition system we designed is a
writer-dependent system. The reason why we give priority for writer-dependent online
handwriting recognition systems, is because the target devices usually belong to individuals
and we believe that attaining higher recognition accuracy is more important than writer-
independency.

Due to the variation in people’s handwriting styles, it is obviously true that it is a

challenging task to develop handwriting recognition systems. On top of this, other external
reasons made our research difficult. One of the main reasons is that no online handwriting
data has ever been collected and standardized. This restricted us to use a limited amount of
data to test our system for its usability. However, the system was tested for highly varying
handwriting styles and we strongly believe that it works for more number of writers.

Of the many and diversified approaches used in online handwriting recognition systems
that are discussed in chapter 2, we have chosen the structural approach for our recognition

53
engine. The structural approach is concerned about the various primitives that make up a
character and their relationship. It is the most natural way of treating characters, however
the approach has a disadvantage, which is the difficulty to extract the primitives and their
relationship from a given character [1].

5.2. Structural Approach for Recognition of Ethiopic

Characters
In spite of the fact that structural approach is usually criticized and regarded as an older
technique, it has got positive aspects, which would compensate its weak points. Some of
these are:
• If proper algorithms are developed for primitive extraction, this approach is the
most natural way of defining characters.
• It does not require huge amount of data to train the system [3].

5.3. Appropriateness of the Approach

The nature of Ethiopic characters itself guided us to use the structural approach where we
design a simple but a robust algorithm to represent primitives. Most of the Ethiopic
characters have defined shape. The set of the alphabetic characters in the Ethiopic character
set contains 33+1 basic characters and six non-basic characters with each basic character.
The shapes of these non-basic characters are derived from the related basic character by
adding a secondary stroke or by deforming their shape. This structural relationship is a
great advantage. After doing the analysis described in chapter four of this document, we
have concluded that if the recognition of the 33+1 basic characters is done successfully,
then the system could be easily extendible to recognize the remaining non-basic characters.

54
5.4. Design of Online Handwriting Recognition System
for the Basic Ethiopic Characters
The design of the online Ethiopic handwriting recognition engine developed in this work is
shown below (Fig. 5.0). The unknown pattern in general is acquired through the data points
collection process. This is achieved by using the digitizer software, MovAlyzer. The
collected data is brought to the preprocessing package to eliminate variations. The
preprocessed data is provided to the feature extraction module, which will produce the
observation code sequences from the given handwriting pattern. This code could be used to
train or test the system as required. In this case, training is the process of teaching the
recognition engine to make it aware which observation code sequences belong to which
strokes or characters. While testing, unknown handwriting recognition patterns represented
by observation code sequences are presented to the recognition engine and the recognizer is
expected to recognize what character is represented by the given observation code
sequences.

No explicitly defined segmentation is stated. The reason is that we have considered only
segmentation of characters into strokes which are already separated by pen-up points in the
collected online handwriting data. However, implicit segmentation is done in each step of
the online handwriting recognition system by looking for pen-up points to find the
beginning and end of strokes.

55
 Unknown
Data Collection Pattern
Raw Online
Handwriting data
for a character

Extra point Noise

Elimination Filtering

Size
Normalization Resampling

Preprocessed
Handwriting data
for a character

Observation
Coding
Observation Observation
Code Sequences Code Sequences
(Training data) (Unknown data)

Reference Sequence
Collection Coarse
Classification

Detailed
Matching
Reference
observation code Superimposing
sequences Matching

Recognized
Character

Fig. 5.0. Design for online handwriting

Recognition system for Ethiopic characters

56
5.4.1. Data Collection
Under normal circumstances, online handwriting data for this kind of research is collected
using tablet digitizers. However, for the purpose of our experiment, handwriting data is
collected using digitizer software named Neuroscript MovAlyzer that is developed to
sample points on the trajectory of the mouse or an electronic pen (digitizer) if any. When
digitizer is used to collect data, the common sampling rate is 100-200Hz and it is 50-60Hz
when the mouse is employed. The capability to use the mouse as an input device to draw
characters enables us to collect data using mouse in place of an electronic pen.

This software includes many features including segmentation and word extraction.
However, we used only the most important feature of the software, which is collecting
online handwriting data using mouse. Whenever data is needed to be collected, MovAlyzer
is started and an experiment is run. Then, MovAlyzer presents an interface (Fig. 5.1) that is
used to write characters by dragging the mouse over the mouse pad.

Fig. 5.1. MovAlyzer Recording Window

The writer then starts writing anywhere on the given blank screen and the system records
points along the path of the drawing. For a single sampled point, MovAlyzer records three
values, which are the x coordinate, y coordinate and the pen-pressure. The pen pressure is
equivalent to the pen-up or pen-down information where it has only two possible values
while using the mouse as an input device. These values are 99 and 0 which correspond to

57
pen-down and pen-up respectively. If a digitizer/tablet were the input device, the pen
pressure would vary to reflect the amount of pressure applied by the pen.

The difficulty during data collection is the inappropriateness of the mouse to write. In spite
of the fact that the mouse is simulating the light pen, it can by no means replace the pen
and is unnatural. Due to this reason, the data could be noisy and the trajectory is more or
less jagged. Related to this, writers had also difficulty in writing the same character
consistently using a mouse. The good thing is the system is supposed to be writer-
dependent and there is no need to collect a large amount of data to handle varieties. But, it
was tried to test the system for two writers whose handwriting vary much. These writers
are referred as writer A and writer B from now on.

Fig. 5.2. illustrates the appearance of the data recording window after a character is written.
Electronic ink marks all the pen-down sampled points and the pen-up points are not visible
though they are recorded. This capability makes the environment natural and convenient
for writers.

Fig. 5.2. MovAlyzer Character Writing pad (after the character ” “ is written)

5.4.2. Data Preparation

As stated earlier, MovAlyzer produces the raw (x,y) coordinates with a pen up/pen down
information of a data point. A data point is a point that is sampled along the path of the
mouse while the writer is writing the character. Thus, for a single character, number of data

58
points would be sampled. The software then presents the set of data points for a drawn
character collectively in a text file format.

136 -200 99
136 -201 99
136 -203 99
136 -204 99
136 -206 99
136 -208 99
136 -209 99
136 -211 99
136 -212 99
136 -214 99
136 -215 99
136 -216 99
136 -218 99
136 -220 99
Fig. 5.3a) The Character ‘ ‘ as it is written on the Fig. 5.3b) The set of data
MovAlyzer writing pad points during the formation of
‘‘ ‘ the character

Fig. 5.3a shows the character written by writer A and part of the data points collected for
this character is shown in Fig. 5.3b. For this character, 354 points are sampled by
MovAlyzer and each are recorded in terms of the x and y coordinate values and the pen
(mouse) pressure. This set of data points is saved in a text file. This is the original raw data
and used as an input to the recognition engine.

Even though online handwriting recognition systems are known to be real-time, the online
handwriting data collected in such researches is stored in a persistent way to carry out
repetitive experiments on the data during the development of the recognition engine.
Following the same trend, the sampled data points set collected for each character are kept
in a text file being accessible to the recognition engine.

The collected data should be grouped as training data and testing data. The training data is
supplied to the recognition engine while it is being trained and the test data serve to test the
system for its accuracy. The way the training is conducted as well as the categorization of
the data is explained in a subsequent section.

59
5.4.3. Preprocessing
The preprocessing steps play a very significant role to improve recognition accuracy by
avoiding the inter-character variation, that is variation that occurs between different
occurrences of the same character. In this system, we have found that the following
preprocessing stages are significantly necessary and are implemented.

5.4.3.1. Extra Pen Up Data Points Noise Elimination

At this stage, we eliminate only one type of noise that is caused by additional pen-up data
points recorded after the character is completely drawn. These points are not necessary and
more importantly create a series problem during the feature extraction process. Thus we
implemented the algorithm in Fig. 5.4 to eliminate them.

Input: File that contains the set of data points sampled for a character
//index: points to each data point
//NoOfPts: number of pen up and pen down data points collected for a given character
//OrigX, OrigY, OrigZ: set of x, y and z components of data points
//LastIndex: points to the last data point

index ←1
NoOfPts ← 0
Do
Read x, y, z of a data point from file
OrigX [index] ← x
OrigY [index] ← y
OrigZ [index] ← z
index ← index + 1
NoOfPts ← NoOfPts + 1
Until (end of file is reached)
LastIndex ← NoOfPts //Keep the last value of NoOfPts in LastIndex
While (OrigZ[LastIndex] = 0)
LastIndex ← LastIndex – 1 //Decrease LastIndex if the data point(s) is a
pen up.
index = 0
While (index < LastIndex)
Do
Write Origx [index] to file
Write Origy [index] to file
Write Origz [index] to file
index ← index+1
Until (index < LastIndex) //Unwanted data points are dropped

Fig. 5.4. Extra Data point Noise Elimination Algorithm

60
5.4.3.2. Size Normalization
The digitizer software MovAlyzer used for the data collection process presents a window,
which is evidently large enough to write lines of sentences. However, the writers are told to
write only a single character in an occasion for the reason that we are concerned about
character recognition and data of a single character is an input to the system. This made the
users free to write anywhere on the writing pad in any size.

This situation necessitates the application of size normalization to map all characters to a
box of standard size (40×60). It is observed that this size preserved the proportion between
the length and width of most of the characters hence do not alter their shape much. In spite
of this, few characters became abnormally longer in proportion to their width (See Fig. 5.5a
and Fig. 5.5b) after size normalization has been applied. But we didn’t do anything about it
because it doesn’t create any problem since our classification approach depends on
direction features.

Fig. 5.5b. The same letter " "

after size normalization
Fig. 5.5a. The letter " " with its (40×60).
original size (173×116)

Size normalization is done by translating each data point of a character to a box of a

standard size. This is accomplished by first finding the box that encloses the character with
its original size and then computing the width and length of the box. These values are later
used to find (x,y) coordinate values of newly produced data points in the standard sized box
that correspond to each data point from the original box. The complete algorithm is
presented below (Fig. 5.6).

61
 Input: File that contains the set of data points sampled for a character (noise
eliminated)
//MaxX, MinX: Maximum and minimum x values among the data points sampled for a
character
//MaxY, MinY: Maximum and minimum y values among the data points sampled for a
character

//Find the block that encloses the character.

Compute MaxX, MinX
Compute MaxY, MinY
Width ← MaxX-MinX //width of the box that encloses the character
Length ← MaxY-MinY //Length of the box that encloses the character

//Normalize character stroke-by-stroke

Do
Read x, y, z of a data point from file
if (z = 0) //pen-up data point
Write the triple 0, 0, 0 //Stroke separation (segmentation)
do
Read x, y, z of the next data point //To skip all other pen up points
Until (z ≠ 0)
else //pen-up data point
//Translate the data point to a box of size 40 × 60
newx ← (40 × x)/width
newy ← (60 × y)/length
Write newx, newy, z in a new file
Until (end of file is reached)

Fig. 5.6. Size Normalization Algorithm

This method results in having an equal number of pen-down data points, but replace
consecutive pen-up data points by the triple (0, 0, 0) to indicate separation of strokes. Note
that pen-up data points are recorded when the writer lift up the pen (mouse) to draw an
additional stroke in the character. This is clearly seen in Fig. 5.3a. It is also observed that
some data points are translated to the same point in the newly created point set. This
creates redundant data and further preprocessing, filtering is essentially implemented. For
instance, part of the collected data points for letter " " shown in Fig.5.5a and the
translated data points in its normalized version which is also shown in Fig. 5.5b are
provided below (Fig. 5.7a and Fig. 5.7b).

62
 220 -172 99 50 -88 99
219 -172 99 50 -88 99
217 -172 99 50 -88 99
214 -172 99 49 -88 99
210 -172 99 48 -88 99
206 -173 99 47 -88 99
204 -174 99 47 -89 99
201 -175 99 46 -89 99
199 -175 99 46 -89 99
197 -176 99 45 -90 99
196 -177 99 45 -90 99
194 -178 99 44 -91 99
44 -92 99
192 -180 99
43 -93 99
189 -182 99
43 -94 99
186 -185 99
42 -96 99
183 -189 99
. .
. .
. .
Fig. 5.7a. Original data for letter " " Fig. 5.7b. Data points after normalization
applied to the Original data
shown in Fig. 5.7a

From Fig. 5.7a and Fig. 5.7b, it is possible to see that the first three data points in the
original data set are mapped to the same data point which is (50, -88, 99). The
preprocessing step discussed in the next section avoids such points from the normalized
data.

5.4.3.3. Filtering
Filtering is another preprocessing technique that we have considered in this recognition
engine. It is the process of avoiding redundant data points that appear consecutively. These
data points are undesirable and have no use at all. Thus, eliminating them is essential.
Moreover, the action also contributes towards having a reduced data, which is one goal of
preprocessing.

Filtering the normalized data is accomplished by implementing the algorithm in Fig. 5.8 .

63
 Input: File that contains size normalized data for a character

Read x, y, z of the first data point from file

PreviousX ← x
PreviousY ← y
Do
Get x, y, z of next data point
if (x = PreviousX and y = PreviousY)
do nothing
else
Write x, y, z to a file
PreviousX ← x
Previousy ← y
Until (end of file is reached)
Fig. 5.8. Filtering (Data Reduction) Algorithm

5.4.3.4. Resampling
Relatively, the handwriting character data is to some extent normalized and reduced by the
previous two steps. But, the data points are not yet time normalized or equally spaced.
Furthermore, the trajectory is still jittered due to the improper arrangement of the data
points. This is corrected by the resampling procedure presented in Fig. 5.9.

Input: File that contains size normalized and filtered data for a character
Read x, y, z of the first data point from file
PreviousX ← x
PreviousY ← y
Write x, y, z to file //Accept the first point
Do
Read x, y, z of the next data point from file
//Compute d, distance between the previous data point and the current data point
d ← (x-PreviousX)2 + (y-PreviousY)2 .
if (d ≥ 25) //25 is a threshold value
Write x, y, z to a file //Accept the point
PreviousX ← x
PreviousY ← y
else
do nothing
Until (end of file is reached)
Fig. 5.9. Resampling Algorithm
The algorithm computes the distance between two consecutive data points and avoids one
of them (the second one) if the distance is less than a threshold value. After trying many
threshold values, it is concluded that 25 is a good threshold value. However, the sampled

64
data points will not be exactly equally spaced since points with distance greater than 25 are
accepted. This situation occurs rarely and if it occurs, the distance will not be much greater
than 25. Thus, the algorithm is left as it is. Fig. 5.10a and Fig. 5.10b show the appearance
of the letter “ ”, that we have considered in Fig. 5.5a and Fig. 5.5b, before resampling
and after resampling.

Fig. 5.10a. Magnified view of " " after Fig. 5.10b. Magnified view of " "
size normalization and after resampling
before resampling

After resampling, the data is significantly reduced and the irregularly placed data points
that create the jitter on the trajectory are somehow removed. This makes the resampling
step very useful in noise elimination as well as data reduction. The number of the original
sample points in the raw data set for the character " ", for example, is 189. By chance, the
data for this character do not include unnecessary pen-up points which are sampled after
the character is written completely and hence no change has been exhibited in the original
data after applying the noise elimination procedure discussed in section 5.6.1.1. After size
normalization, the number of points remains the same since the character is written by a
single stroke. Note that the size normalization step plays its own role in reducing the data
by replacing the pen-up points between strokes by a single (0,0,0). Filtering which avoids
repeated data points reduced the number of data points to 160. Finally, the resampling
action greatly reduced the data size to be only 45 data points.

65
5.4.4. Feature Extraction
The original data has passed through the preprocessing steps and refined to an extent
described in the previous sections. But this doesn’t mean that the data points for different
instances of the same character become similar or identical. The following figures (Fig.
5.11a and Fig.5.11b) illustrate the preprocessed data set for two instances of the same
character " ".

91 -108 99 72 -79 99
100 -109 99 77 -80 99
107 -110 99 82 -82 99
114 -110 99 88 -82 99
120 -110 99 94 -83 99
125 -110 99 99 -84 99
130 -110 99 104 -85 99
0 0 0 109 -85 99
114 -112 99 0 0 0
114 -117 99 90 -82 99
114 -125 99 90 -87 99
114 -132 99 90 -93 99
114 -137 99 90 -98 99
114 -142 99 89 -103 99
114 -147 99 89 -108 99
114 -152 99 89 -113 99
114 -158 99 89 -119 99
114 -163 99 89 -124 99
113 -168 99 89 -129 99
0 0 0 89 -134 99
97 -136 99 89 -139 99
103 -135 99 0 0 0
108 -135 99 77 -110 99
113 -135 99 82 -110 99
118 -135 99 87 -111 99
123 -134 99 92 -111 99
128 -134 99 97 -111 99
102 -111 99
Fig. 5.11a. Preprocessed data for the 107 -111 99
letter " " (instance 1)
Fig. 5.11b. Preprocessed data for the
letter " " (instance 2)

Three strokes are used by the writer (Writer A) to write this letter as it is indicated by the
triple (0,0,0) that separate the various strokes. As can be seen in the data collected from
Writer A, she wrote many other letters such as " " with three strokes. Thus, the number of
strokes alone cannot be a distinguishing factor to measure the similarity between two or

66
more characters. This implies that there is a need to examine and compare the
corresponding strokes and identify their order of occurrence to quantify the similarity
between characters.

In any classification task, an unknown pattern must be compared with the existing
reference patterns to pick one, which is believed to be the most similar. Comparing patterns
in their original form is error prone and computationally complex. Therefore, formulating
pattern representation method that reflect similarity between identical characters and
difference between characters from different classes is a crucial matter to make the
comparison efficient and effective.

We have observed that the similarity between strokes could be measured by looking how
the x values and y values of the data points change as moving from one data point to the
next one. We have identified that the change can be expressed by three possible conditions.
These are increasing, decreasing and constant (or nearly constant). Despite the variations,
this property will be more or less maintained between strokes that involve in different
instances of the same letter. This idea is illustrated by the samples given in Fig. 5.11a and
5.11b. The first stroke in the instance 1 and instance 2 of the letter " " have seven and
eight sampled data points. The x values of the data points in the strokes that appear in both
instances increase as one pass from one data point to another until the last data point. The y
values in the data points generally do not change much and it is possible to conclude that
the y values are not changing (nearly constant). A similar analysis could be done to the
other strokes and each of them would be compared to the prototype strokes on this basis.
The order of the strokes occurrence will also be taken into account to determine what
character they form when combined. But, the classification result from this approach
shouldn’t be taken for granted, as it doesn’t consider the relationship between strokes. For
example, the characters in Fig. 5.12a and Fig. 5.12b are often confused by this approach.

Fig. 5.12a. Letter " " (Writer B) Fig. 5.12b. Letter " " (Writer B)

67
The reason is because writer B wrote the two characters by combining two similar strokes
in the same order. In both cases, she drew the vertical line and then drew the horizontal line
as a secondary stroke. The second stroke crosses the vertical line in letter " " but is placed
at the top of the vertical line in letter " ". By following the approach explained
previously, there is no way to identify this distinction.

Therefore, examining the consecutive data points to see whether their x and y values
increase, decrease or remain constant is a good approach to derive a common
representation for the strokes, while more features need to be considered to refine the
classification result of the character. Note that similar strokes are commonly found in
different characters in the same order or different order. Thus, we have used this feature
only for coarse classification, which produces a number of characters with higher degree of
similarity with the unknown pattern from the whole prototype set. After getting the ‘Top
characters’, detailed matching is performed to pick the finally classified character.

The next section presents the way to derive a common representation method for the
strokes involved in character patterns to make them ready for comparison and hence for
coarse classification. We called this process observation coding. Then, the technique of
training and the classification of characters are presented in subsequent sections.

5.4.4.1. Pattern Representation /Observation Coding

The three possibilities that occur between x and y values of consecutive data points are
encoded as shown in Table 5.1 below.
Number
Condition
Code
Beginning of a new stroke 1
Increasing 2
Decreasing 3
Constant
4
(Nearly constant)
Stroke separator 5
Table 5.1. Codification of observation

68
All handwriting patterns will be represented by sequences of these number codes appearing
according to what happens between the x and y values of a pair of consecutive data points.
Two observation code sequences are created for a single stroke representing the way the x
and y values vary or remain constant. While extracting the sequences, a number code will
not be assigned to each individual pair of data points. If series of data points exhibit the
same characteristics in the nature how their x or y values vary, a single number code is
enough to represent them. The algorithm to extract the sequences is shown in Fig. 5.13.
This algorithm is particularly for obtaining an observation sequence code to represent the
characteristics of the x values. The same algorithm is also applied to extract observation
code sequence for the y values.

The algorithm in Fig. 5.13 starts by assigning 1 as the first element of the code sequence.
This is to indicate the beginning of the first stroke (See Table 5.1). Then, the first data point
is considered and its x value is kept. The next procedure is to pass to the next data point
and compare its x value with the previously kept x value to determine whether a value that
is less than, greater than or equal (nearly equal) is encountered. Accordingly, the values 2,
3, or 4 are assigned to indicate the situations increasing, decreasing or constant (nearly
constant) respectively. If two x values in consecutive data points are different only by the
value 1, they will be regarded as if they are the same and as a result an observation code 4
(for nearly constant) is recorded. In each step, if a formerly recorded observation code and
the observation code that is currently to be assigned are the same, the assignment will be
canceled to avoid repetitive observation codes in the sequence. This means that for set of x
values having a uniform way of changing, only a single observation code represents the
whole set. However, if this were the only task of the algorithm, it would have been
impossible to properly extract observation code sequences from characters having more
than one stroke. Thus, additional observation codes 1 and 5 are included to show the
beginning of a stroke and as a separation mark between strokes respectively. That is the
reason why x is checked if it is 0 (to identify whether the stroke ends and the triple (0,0,0)
is came across). The value of PreviousX is also checked if in case it is 0. If it is 0, it implies
that the current data point is the first point in the new stroke. The algorithm handles these
two conditions by placing the values 1 and 5 in the sequence respectively.

69
 Input: File that contains preprocessed data for a character
//CntObsX represents the number of extracted code sequences and ObsX is the code sequence itself.

CntObsX ← 1
ObsX [CntObsX] ← 1 //Beginning of the first stroke
Get x value of the first data point from file
PreviousX ← x
Do
Get x value of the next data point
if (x = 0) //It indicates that the triple (0,0,0) is encountered
CntObsX ← CntObsX+1
ObsX[CntObsX] ← 5 //separation of strokes
else if (PreviousX = 0) //starting a new stroke
CntObsX ← CntObsX+1
ObsX[CntObsX] ← 1 //Beginning of a new stroke
else
if (x != PreviousX and |x-PreviousX|>1)
if (x > PreviousX and Obsx[CntObsX] != 2)
CntObsX ← CntObsX+1
ObsX[CntObsX] ← 2 // Increasing
else if (x < PreviousX and Obsx[CntObsX] != 3)
CntObsX ← CntObsX+1
ObsX[CntObsX] ← 3 //Decreasing
else
if (Obsx[CntObsX] != 4)
CntObsX ← CntObsX+1
ObsX[CntObsX] ← 4 //Constant or nearly constant
PreviousX = x
Until (end of file is reached)

Fig. 5.13. Observation Code Sequence Extraction Algorithm

In Fig. 5.14a-5.14d, two instances of the letter " " (written by Writer A) and the
preprocessed data for the letters are shown. The data points are sequentially numbered to
facilitate the discussion. The observation code sequences derived (from the two
preprocessed data set) for x and y values are shown in Fig. 5.14e-5.14h. The sequences in
Fig. 5.14e and 5.14f are analyzed in the following paragraphs.

70
Both sequences start with the observation code 1. The x values among the first few data
points (data point 1 – data point 9) are constant (or nearly constant) and hence observation
code 4 is used to represent them in the x sequence (See Fig. 5.14e). In the next few data
points (data point 10 – data point 18), x starts to increase. This makes the sequence
extractor to record the observation code 2. The x values in the last six points are constant or
nearly constant and hence the observation code 4 represents them in the observation code
sequence.

1. 102 -51 99
2. 102 -56 99 1. 42 -51 99
3. 102 -61 99 2. 42 -58 99
4. 101 -67 99 3. 42 -66 99
5. 101 -74 99 4. 42 -73 99
6. 101 -81 99 5. 42 -80 99
7. 101 -87 99 6. 42 -86 99
8. 101 -93 99 7. 42 -91 99
9. 102 -99 99 8. 42 -96 99
10. 105 -104 99 9. 44 -102 99
11. 109 -108 99 10. 47 -106 99
Fig. 5.14a. The letter 12. 114 -110 99 11. 52 -108 99
(instance 1) 13. 120 -111 99 12. 57 -109 99
14. 125 -110 99 13. 63 -110 99
15. 130 -106 99 14. 68 -111 99
16. 134 -102 99 15. 74 -110 99
17. 136 -97 99 16. 78 -106 99
18. 138 -91 99 17. 81 -100 99
19. 140 -85 99 18. 81 -94 99
20. 140 -79 99 19. 82 -89 99
21. 141 -73 99 20. 82 -83 99
22. 141 -67 99 21. 82 -78 99
23. 141 -62 99 22. 82 -73 99
23. 81 -67 99
24. 141 -57 99
24. 80 -62 99
Fig. 5.14b. The letter Fig. 5.14c. The 25. 80 -57 99
(instance 2) preprocessed data for Fig. 5.14d. The
the letter in Fig. 5.14a preprocessed data for
the letter in Fig. 5.14b

1 1 1 1
4 3 4 3
2 4 2 4
4 2 4 2
Fig. 5.14e. X Fig. 5.14f. Y Fig. 5.14g. X Fig. 5.14h. Y
observation code observation code observation code observation code
sequence for the sequence for the sequence for the sequence for the
data in Fig. 5.14c data in Fig. 5.14c data in Fig. 5.14d data in Fig. 5.14d

71
The y values between the data points in the set (Fig. 5.14c) can be analyzed similarly to
derive the sequence shown in Fig. 5.14f. In the y observation code sequence, 3 represent
the characteristics of the y values in the first 11 data points. The y values in the next three
data points are almost the same and the observation code 4 is recorded. The rest of the data
points contain y values, which increase from time to time and hence the last observation
code 2.

Identical sequences are being extracted from the second data set as illustrated in Fig. 5.14g
and 5.14h. This example has shown the possibility of coming up with identical
representations for different data sets. In general, these two sequences together characterize
these patterns and other similar patterns irrespective of the actual values of the coordinates.
Note that letter is written by a single stroke (as Writer A wrote) and the observation code
sequences do not include code 5, which is a stroke separator code.
Fig.5.15 demonstrates how the extractor algorithm treats characters with multiple strokes.
1. 91 -108 99
2. 100 -109 99
3. 107 -110 99
4. 114 -110 99
5. 120 -110 99
6. 125 -110 99
7. 130 -110 99
0 0 0
1. 114 -112 99
Fig. 5.15a. The letter " "
2. 114 -117 99
(Writer A)
3. 114 -125 99
4. 114 -132 99
5. 114 -137 99
6. 114 -142 99
1 1 7. 114 -147 99
2 4 8. 114 -152 99
9. 114 -158 99
5 5 10. 114 -163 99
1 1 11. 113 -168 99
4 3 0 0 0
1. 97 -136 99
5 5 2. 103 -135 99
1 1 3. 108 -135 99
2 4 4. 113 -135 99
5. 118 -135 99
Fig. 5.15c. X Fig. 5.15d. Y 6. 123 -134 99
observation code observation code 7. 128 -134 99
sequence for the sequence for the
data in Fig. 5.15b data in Fig. 5.15b Fig. 5.15b. The preprocessed
data for letter " " (Writer A)

72
In general, the algorithm in Fig.5.13 produces identical, if not similar sequences for
handwriting patterns that belong to the same character. However, it has been observed that
the algorithm in its original form produced very diversified sequences for characters with
rough trajectories. It is also very sensitive for slight variations that occur between instances
of the same character. The following sequences (Fig. 5.16) are X observation sequences
which are produced for different instances of the letter " ".
a b

1, 3, 2, 4, 3, 4 1, 3, 4, 2, 4, 2, 4, 3, 4, 3
c d

1, 3, 4, 2, 4, 2, 4, 3, 4 1, 3, 4, 2, 4, 3, 4
e f

1, 3, 4, 2, 4, 2, 4, 3, 4, 3
1, 3, 4, 2, 4, 2, 4, 3, 4, 2, 4
g h

1, 3, 4, 2, 4, 2, 4, 3, 4 1, 3, 4, 3, 4, 2, 4, 2, 4, 3, 4

Fig. 5.16. X Observation Code Sequence for different instances of " " (Writer A)

73
The eight sequences are produced for instances of the character " " which are written by
the same writer (Writer A). Contrary to the expectation, these sequences are different. This
is mainly due to the shape variations. To overcome this problem, the algorithm (Fig. 5.13)
is modified to reject observation codes, which last for too short time. To accomplish this, it
is required to keep the length of each observation code in the sequence, which refers to the
number of data points that cause the observation code to appear in the sequence. The length
of an observation code denoted by Length(ObsvCode) is computed based on the following
rule:
0 if ObsvCode = 1 or 5

Length (ObsvCode) =
Number of data points
that cause the code to
appear if ObsvCode = 2, 3 or 4

where Length(ObsvCode) is the length of the observation code,

ObsvCode is an observation code.

By incorporating this rule with the algorithm, the length of each observation code along
with the observation code sequence is obtained. Observation codes with length 1 are
considered to be codes with shorter length and they evidently increase inter-class variations
while reducing intra-class variations. Further, their role in defining the shape of the
character is minimal. Therefore, discarding such observation codes from the sequence is
advantageous in enhancing the matching between sequences that belong to similar
characters.

Likewise, after the observation code sequence is extracted, the sequence will be refined by
removing observation codes with length 1. By doing so, the sequence in Fig.5.16b, can be
refined as shown in Fig. 5.17.

Observation 1 3 4 2 4 2 4 3 4 3 1 3 2 2 3 4
Length 0 5 1 6 1 4 1 3 16 1 0 5 6 4 3 16

1 3 2 3 4
Fig. 5.17. Sequence Refinement 0 5 10 4 16

74
After this refinement step, observation code sequences may include identical consecutive
observation codes (See Fig. 5.17). Thus, the sequences are further refined to remove such
codes by merging them into one observation code, as they have no use and the action
simplifies the sequence. This also helps to reduce the variation between the sequences of
similar characters.

Even though, the length of observation codes in an observation code sequence is originally
computed for the purpose of avoiding shorter codes, we have observed that it is also a
useful parameter for matching sequences during classification. Basically, it measures and
quantifies the degree of similarity or dissimilarity between sequences, which is an
important input to decide on which stroke or character the sequence belongs to. To take this
advantage, the length of each observation code is considered while observation code
sequences are extracted from a given preprocessed data.

But, we have noticed that there is still an intolerable variation between sequences of the
same character, which might result in misclassification of characters. It has also been
observed that most variations between sequences of the same character are created by the
observation code 4. In the majority of the sequences, this code appears with relatively
shorter length except for horizontal and vertical line strokes. Horizontal strokes are
characterized by the x observation sequence (1,2) or (1,3) depending on the direction of
writing, and by the y observation sequence (1,4) since the y values will be constant (nearly
constant) among the data points along the horizontal line. Similarly, the x observation
sequences extracted for vertical line strokes will be (1,4) and the y observation sequences
will be (1,2) or (1,3) depending on whether the vertical line is drawn upward or downward.
It is possible to conclude that the observation code 4 has a considerable part in describing
such strokes. In spite of this fact, it was preferred to leave all occurrences of the
observation code 4 from all sequences that are used to classify strokes and characters for
the coarse classification task. These codes would be included while detailed matching is
later performed in the second-stage classification. The whole process to extract observation
code sequence, which is ready for training and classification, is summarized and illustrated
in Fig. 5.18. Note that the length of element observation codes is processed in each step
along with the observation code sequences.

75
 Step 1:
Input: Preprocessed handwriting data for a character
Procedure: Observation Code Extraction:
(Using the algorithm in Fig. 5.13)
Output: X and Y observation code sequences along with the length of each
element code (original form)
Step 2:
Input: Original X and Y observation code sequences
Procedure: First stage refinement:
Removal of shorter observation codes (codes with length 1)
Output: Refined X and Y observation code sequences
Step 3:
Input: Refined X and Y observation code sequences
Procedure: Second stage refinement:
Removal of identical consecutive codes in the sequence
Output: Further refined X and Y observation code sequences
Step 4:
Input: Further Refined X and Y observation code sequences
Procedure:
If the X and Y observation code sequences are needed for coarse
classification, then third stage Refinement will be done which
constitutes the removal of the observation code 4. Otherwise, the
sequence will be left as it is.
Output: X and Y observation code sequences which are ready for training or
classification.

Fig. 5.18. Observation Code Sequence Extraction Process

In conclusion, this section discusses the pattern representation scheme used to represent
handwriting patterns that involve in the online handwriting recognition engine developed in
this work. During training, the engine stores reference handwriting patterns that are
represented by observation code sequences. In the same way, unknown patterns are also
represented by this scheme before they can be compared with the reference patterns. The
next section explains how the training is conducted. The subsequent section explains the
classification.

5.4.5. Training
The main goal of the training process is to teach the recognition system an individual’s
handwriting so that it could be able to classify unknown handwriting patterns written by the

76
same writer. Particularly, training a writer-dependent system focuses on identifying a
writer’s handwriting style who is supposed to use the system. The recognition engine learns
the handwriting of an individual by observing and recording (persistently) how each
character is written. This requires to have set of characters written by the individual and the
data is usually referred as training data. The training data needs to be essentially complete
in the sense that it should include samples for each character. The training should also be
effective enough to bring about better recognition accuracy for the recognition engine.
However, it is also a desirable feature for a training process to use small amount of training
data as much as possible.

The recognition engine has a module named ‘Trainer’ which manages the training
activities. Observation code sequences, which are extracted from the training data by the
feature extraction component of the system, are the input to the trainer (Fig. 5.4). The
trainer then acts on the input observation code sequences to produce a set of reference
observation code sequences for each stroke of a character that is supposed to be recognized
by the recognition system.

The general procedure used to train the system is to identify and record unique observation
code sequences for the various strokes that constitute a character. A number of samples for
a character will be collected and observation code sequences for each stroke will be
extracted from each sample. Then, the sequences for similar strokes in the different
samples are compared to each other to identify unique observation code sequences. The
trainer then keeps these sequences as reference sequences that represent the strokes in the
character. It should be noted that the length of element observation codes in the sequences
are also stored as they are important during classification. Keeping the length of each code
that involve in the sequences is a way to know the chance of the code to appear in the
sequence in that particular order. The training is conducted both for the x and y observation
sequences of the strokes.

The output of the trainer module is a text file that stores all possible observation sequences
of strokes of a character that are obtained from the given samples. This file is refereed as a
reference file. After training, two reference files are generated for each character that
correspond to the X and Y observation sequences. There is a possibility that more than one

77
pair of reference files would be produced for a character if the writer claims that he/she
writes the same character by using different number of strokes and/or in different order.
The reference file mainly stores the possible observation sequences of strokes of a
character, which also includes the length of each code in the sequences. An example of a
reference file is shown in Fig. 5.19. When observation sequences are written in the
reference file, an observation code and its length are placed in the same line as a pair.
These pairs are listed in different lines until the last code of the sequence. At the end of
each unique sequence of a stroke, the pair (10,10) is written for the purpose of marking its
end.

4 Number of Strokes

1 Number of unique sequences for the first stroke

1 1
2 3 Unique sequence for the first stroke
10 10

3 Number of unique sequences for the second stroke

1 5
2 2
10 10
1 5 3 unique sequences for the second stroke
10 10 Each sequence is separated by the pair 10,10.
1 5
3 2
10 10

2 Number of unique sequences for the third stroke

1 5
10 10 2 unique sequences for the third stroke
1 5 Each sequence is separated by the pair 10,10.
3 4
10 10

1 Number of unique sequences for the fourth stroke

1 3
2 3 1 Unique sequence for the fourth stroke
10 10 Fig.

Fig. 5.19. A Reference File

[Generated for X observation sequence of the letter " "
(Writer B)]
In addition to the possible observation sequences of strokes of a character, the reference
file is made to contain other useful information such as the number of strokes. The number
in the first line in all reference files corresponds to the number of strokes of the character as

78
written by the writer for whom handwriting the training was conducted. A number in the
second line is the number of unique observation sequences for the first stroke. Then, the
sequences of the particular stroke will be written. When all the unique sequences are
exhaustively written for a stroke, another number is inserted before the listings of the
sequences for the next stroke. This number is the number of unique sequences for the next
stroke.

In our case, three samples are provided for the trainer for a single character as training data.
The number of samples can however be increased as the need may be. In Fig. 5.20 and Fig.
5.21, two examples are given to illustrate how training of a single character is conducted
from three given samples. The trainings shown in the figures were conducted for Y
observation code sequence of the character " " and X observation code sequence of the
letter " " respectively. All the samples in the figures are written by Writer B.

Y sequence for Y sequence for Y sequence for

" " (Sample 1) " " (Sample 2) " " (Sample 3)
T=5 T=5 T=5
1 0 1 0 1 0
3 11 3 10 3 11
2 10 2 10 2 11
3 9 3 11 3 10
2 11 2 11 2 11

TRAINER

Training Result
Reference file for Y sequence for
1
1
1 0
3 10
2 10 Fig. 5.20. Training conducted
3 10 for the letter " " (Writer B)
2 11
10 10

79
In Fig. 5.20, three different samples of the letter " " are provided and the indicated Y
observation code sequences are extracted from each sample by the feature extraction
module. These sequences are input to the trainer module and the trainer generates the
reference file. The trainer first determined the number of strokes in each sample and it finds
out that only a single stroke is used to write all the samples of the letter (by Writer B) and 1
is recorded as the number of strokes in the reference file. Then, the trainer compared the
sequences of the strokes from the three samples. The three sequences are identical in terms
of the number of code elements and the order that they appear in the sequence. This leads
the trainer to the decision of making the number of unique sequence for the stroke 1, hence
1 in the second line of the reference file. Then, the only unique sequence was recorded
right after this number. Another important action done while recording the sequence is that
the length of each code element is computed by finding the average of the length of the
corresponding code elements in the given three sequences.
The second example in Fig. 5.21 is quite different from the previous case. The writer used
three strokes to write the character and this can be seen from the sequences. Each stroke
sequence is separated by the pair (5,0). Moreover, the sequences of the first stroke differ in
the three samples. Though the gap between these sequences seems to be big, it is not in real
sense. The sequences in the first and second sample differ by the observation code 2 which
is found between 1 and 3 in the sequence of the second sample. However, the length of this
observation code is only 2 which is not a large number. Another interesting relation is
exhibited between the first stroke of the first and third samples. The existence of the two
observation codes 3 and 2 between the first code 1 and the last code 3 are the reasons to
create a difference between the two sequences. But, if 2 does not exist between the second
3 and the third 3, then it is possible to combine the two to have only a single 3 with length
which is equal to the sum of the length of the two identical codes. The length becomes 5
(3+2) which is near to 6 that is the length of the code 3 in the sequence of the stroke in the
first sample. The resulting sequence would be {1,3} with the corresponding length 0 and 5.
In any case, the trainer records all the three unique sequences in the reference file as the
possible sequences for the first stroke of the letter " ". A similar analysis is done to treat
the second and the third strokes (See Fig. 5.21).
The training is also done in the same way for sequences that contain the code 4, because
sequences including this code are sometimes needed during classification.
The section that follows discusses the classification procedures.

80
X sequence for X sequence for X sequence for
(Sample 1) (Sample 2) (Sample 3)
T=9 T=9 T=12
1 0 1 0 1 0
3 6 2 2 3 3
5 0 3 2 2 3
1 5 5 0 3 2
2 3 1 5 5 0
5 0 2 4 1 5
1 5 5 0 2 5
2 5 1 5 3 2
3 2 2 5 5 0
1 5
2 7
3 2
TRAINER

Training Result
Reference X sequence for
3
3
1 0
3 6
10 10
1 0
2 2
3 2
10 10
1 0
3 3
2 3
3 2
10 10
2
1 0
2 3
10 10
1 0
2 5
3 2
10 10
2
1 0
2 5
3 2
10 10
1 0
2 5
10 10
Fig. 5.21. Training conducted for the letter " " (Writer B)

81
5.4.6. Classification
A trained character handwriting recognition system is ready to receive character
handwriting data and attempt to classify it as one of the characters in the reference data set
by which it was trained. In the previous section, the training procedures are discussed. The
training activity aims at making the system familiar with one’s handwriting and capable of
recognizing characters written by the same writer.

A module named ‘Recognizer’ is available in the recognition system for the purpose of
classification. Similar to the trainer, the recognizer gets its input from the feature extraction
module, which produces the X and Y observation code sequences from the online
handwriting data of an unknown character. Thus, the job of the recognizer is to assign the
given observation code sequences to one of the characters in the reference set. A three-
stage classification method is designed and developed to implement the recognizer. This
layered structure of the recognizer is necessary to enhance the accuracy of the result.

Each classification stage is given a name. The first-stage classification is referred as coarse
classification. The second and the third stage classification steps are named as detailed
matching and superimposing matching respectively. The layered structure of the recognizer
imposes that it should be passed through an upper layer before getting to a lower layer.
This means that the top most layer of the recognizer is the coarse classification, the detailed
matching is the middle layer and the superimposing matching is the bottom layer. Output
of upper layers is an input for lower ones. Fig. 5.21a and Fig. 5.24b show the structure of
the recognizer and the control flow in the recognizer respectively.

Coarse Classification

Detailed Matching

Superimposing Matching

Fig. 5.24a. The Structure of the recognizer

82
 Observation
Code Sequences
(Unknown Data)

Stroke Matching for

coarse classification

Yes Satisfactory? No Detailed Matching

Yes Satisfactory?
Recognized
Character No

Superimposing
matching

Fig. 5.24b. Flow in the recognizer

As illustrated in Fig. 5.24b, the X and Y observation sequences coming from the feature
extraction module are first passed through the coarse classification layer. The layer in its
part examines the sequences and the characters in the reference set to produce the five most
likely characters in order. Of the five characters, the character in the first order is suggested
to be the character that belongs to the given sequences by the coarse classification layer.
However, the recognizer has a mechanism to judge on what the first layer claims. If the
recognizer finds that the result is not acceptably certain, it will pass the five characters to
the second layer so that the second layer does its best to choose one of the five characters
by conducting detailed matching. In some cases, it is still possible that the detailed
matching layer hesitates on choosing one of the five characters and the third layer will be
allowed to resolve this situation. Such steps help to guarantee the correctness of the output.
This implies that a character might not necessarily pass through the second and/or the third
layers of the recognizer.

Each layer utilizes different techniques to classify unknown handwriting patterns. The
simplest method is used by the first layer and its simplicity is expressed by the resources
needed to accomplish it. During detailed matching, more resources could be used. The

83
superimposing matching technique is computationally complex as compared to the
methods applied in the other two layers. The layered structure of the recognizer has an
advantage in minimizing the number of reference characters that are to be compared with
the unknown character when moving from one layer to another. This is useful to avoid
unnecessary efforts that would be made in the second and third stage classification. Hence,
the speed of the recognizer will be improved.

The following subsections explain the three classification stages in detail.

5.4.6.1. Coarse Classification

The X observation and Y observation code sequences of an unknown character are always
initially presented to the coarse classification layer. The coarse classification layer utilizes
the sequences after they are refined to remove shorter codes and the observation code 4 as
explained in section 5.6.2.1. Thus, all sequences presented for this layer do not include the
observation code 4 and consequently are less detailed.

In general, the coarse classification layer is expected to find closest matches for the
unknown sequences from the reference set. To realize this, it should compare the input
unknown sequences with the available sequences from the reference set. The reference
sequences are already prepared by the trainer and stored in the system persistently. The
comparison can possibly be carried out in different ways. In this system, the comparison is
aided by computing distance between sequences (inter-sequence distance). Distance in this
sense is a measure that indicates the degree of similarity of the sequences and it leads to a
conclusion of choosing one of the reference sequences as a classification result. Reference
sequences with smaller inter-stroke distances are generally accepted as the most likely
match of the unknown sequences.

As stated earlier, the coarse classification layer processes both the X and Y observation
code sequences for a given character. Thus, two types of inter-sequence distances are
calculated and a total inter-character distance is computed by adding up the two X and Y
inter-sequence distance values. From now on, we refer to these distances as the X inter-
sequence distance and the Y inter-sequence distance. Both types of inter-sequence
distances are computed by first finding distance between strokes and adding them to get the

84
total distance between the sequences. Note that we are dealing with three types of distances
where one is derived from the other. These are inter-stroke distance, inter-sequence
distance and inter-character distance. Inter-stroke distance is illustrated in Fig. 5.25.

3 3

1 T=7 1
1 1
1 1 1 1
10 10 5 0 3 4
10 10
1
1 0 1 0 1
2 7 2 8 1 5
10 10 10 10
5 0
1 1
1 0 1 5
2 5 1 0 2 6
10 10 2 6 10 10

X Reference file X Reference file

Unknown X
for the letter " " sequence for the letter " "
(Writer B) (Writer B)

Fig. 5.25. Inter-stroke distance

In Fig. 5.25, an unknown X sequence with three strokes is compared with two reference
file which are products of the trainer after the system has been trained for the letters and
. Note that the distance is calculated between corresponding strokes as shown by the
arrows in the figure. The inter-sequence distance is the sum of the distance between the
three strokes. The relationship between the abovementioned three distances is outlined
below:
n
dist sequence = dist stroke
1

dist character = Xdist sequence + Ydistsequence

where diststroke is inter-stroke distance,
distsequence is inter-sequence distance,
distcharacter is inter-character distance,
Xdistsequence is X inter-sequence distance,
Ydistsequence is Y inter-sequence distance, and
n: is the number of strokes in the given sequences.

85
5.4.6.1.1. The Computation in the Coarse Classification Layer
The set of the general procedures for classifying unknown characters coarsely is shown in
Fig. 5.26. Given the unknown sequence (say the X sequence) and the reference files (X
reference files in this case), the distance between the unknown sequence and the sequences
in the reference files (inter-sequence distance) is computed.

The process is started by first determining the number of strokes in the unknown sequence.
It is also possible to get the number of strokes in the reference sequence in charge by
reading the number in the first line in the reference text file (See Fig. 5.17 to see in what
format the information is structured in any reference file which is the product of the
trainer). Acquiring this information is quite important in order to avoid reference
sequences with different number of strokes than that of the unknown pattern. This is to say
that the distance will be computed if and only if the number of strokes is equal in the
unknown and reference sequences. If the opposite happens, the algorithm simply assigns
the value 500 as an inter-sequence distance. This number is large as compared to the
expected computed distances. Thus, it is an indication to show the large gap between the
unknown and reference sequence. As a result, this sequence would be one of the last
choices to be selected as a match.

86
Input: X Observation code Sequence OR Y Observation code sequence

Get NoOfStrokesFromInputSeq, no of strokes in the input sequence

n←1
Do
Get (open) the nth reference file.
Read NoOfStrokesFromRefFile, number of strokes in the reference file
If (NoOfStrokesFromInputSeq = NoOfStrokesFromRefFile)
Begin
i←1
Do
Get StrokeSeqFromInputi, ith stroke of the unknown input sequence
Read NoOfUniqueSeqi, number of unique sequences for the ith stroke in the
reference file
j←1
Do
Get StrokeSeqFromRefi,j, jth unique stroke sequence of the ith stroke in
the reference file
Compute InterStrokeDistancei,j, inter-stroke distance between
StrokeSeqFromInputi and StrokeSeqFromRefi,j (Algorithm in Fig 5.25)
Until (j <= NoOfUniqueSeqi)
Compute MinOfInterStrokeDist, minimum of inter-stroke distances
StrokeDistancei = MinOfInterStrokeDist
Increment i
Until (i <= NoOfStrokesFromInputSeq)
Compute TotalSeqDistance, the total distance between the input sequence and
the reference sequence by adding up all inter-stroke distances.
End
Else
TotalSeqDistance = 500 //This is to make this reference out of choice
Increment n
Until (n <= 34)
Get TopFive, five sequences with the smallest distance.
MinDist ← DistOfTopCand 1, distance computed for the top most character
k←2
TopCnt ← 1
Do
If (| DistOfTopCand k - DistOfTopCand 1| <=5)
TopCnt ← TopCnt + 1
k←k+1
Until (k > 5)
If (TopCnt > 1)
Pass to the next layer taking the nearest candidates to the first one.
else
Report that DistOfTopCand 1, top most character as the recognized character
Fig. 5.26. Algorithm for Coarse Classification Procedures

87
If the number of strokes in the unknown and reference sequences is equal, the distance
computation is started by retrieving the sub-sequence for the first (it could be the only one
if the number of strokes is 1) stroke from the unknown sequence.

This is followed by reading the number of unique sequences for the first stroke from the
reference file. This is done to make sure that the unknown sequence is compared with all
possible sequences of the corresponding stroke in the reference file. Accordingly, distance
between the first stroke in the input sequence and the unique sequence(s) listed for the first
stroke in the reference file are calculated. If the number of unique sequences for the strokes
is greater than one, then the stroke sequence with the minimum of the distances will be
associated to the input stroke. The Inter-stroke distance calculation is discussed in the next
section.

Similarly, all strokes are analyzed exhaustively in this way. At the end of this process, we
have the inter-stroke distances in hand and the total inter-sequence distance is equal to the
sum of all the inter-stroke distances. The coarse classification layer finally picks five
sequences with the smallest distances as the top five candidates ordered by their closeness
to the reference character. The top most from the five candidates is the recognized
character as proposed by the layer. Though, it seems that it is uncertain to take this
proposal for granted, we have proved through experimentation that 93.4% of the suggested
top most characters are the correct characters. To decide on whether accepting the top most
character as a final output or to further compare the five candidates, the recognizer attempts
to examine the distance difference between the top most character and each of the
remaining four characters. If the distance difference between the top most character and the
remaining four characters is greater than 3, the recognizer will accept it as the recognized
character. Otherwise, the top most character and its nearest candidates are passed to the
detailed matching layer. The distance difference value 3 is chosen based on experiments.
That means, we have seen that the possibility of one of the remaining four characters (other
than the top most) to be the correct one is much lesser in cases in which a distance
difference of 3 exists.

Next, the most important issue is the technique to compute the distance between stroke
sequences. We designed and implemented a simple but clever algorithm for this purpose. It
has been proven that the algorithm produces a good inter-stroke distance measure that

88
determines the degree of similarity between stroke sequences. This means that smaller
distance measures are obtained for sequences with higher level of similarity. An
explanation of this algorithm is presented in the subsection that follows.
5.4.6.1.2. Inter-Stroke Distance Computation
The algorithm for inter-stroke distance computation is presented in Fig. 5.27. In this
algorithm, the length of a stroke sequence is the number of elements in the observation
codes sequence. If the stroke sequences are of the same length, the algorithm compares
each corresponding observation code in the two sequences to see if they are identical. If
they are identical, then the codes might differ only in their length. Thus, the difference
between the lengths of the two code values is added to the distance (Fig. 5.27, line 9).
However, if the codes are different, the difference is in observation codes which is a much
bigger difference than the previous case. Therefore, the lengths of both observation codes
are added to the distance (Fig. 5.27, line11).
Input: Two stroke sequences

// UnknownStrokeSeq: Sub-Sequence for a stroke in the unknown sequence

// RefStrokeSeq: Sub-Sequence for a stroke in the reference sequence
//UnknownStrokeLen: set of lengths for the codes in InputStrokeSeq
// RefStrokeLen: set of lengths for the codes in RefStrokeSeq

1: Compute LengthOfUnknownStrokeSeq, no of element observation codes in -

UnknownStrokeSeq
2: Compute LengthOfRefStrokeSeq, no of element observation codes in -
RefStrokeSeq

3: dist ← 0 //dist is inter-stroke sequence

4: if (LengthOfUnknownStrokeSeq ≠ LengthOfRefStrokeSeq)
5: Expand the shorter sequence (Algorithm in Fig. 5.26)

6: i←0
7: Do
8: if (UnknownStrokeSeq i = RefStrokeSeq i)
9: dist ← dist + | UnknownStrokeLen i - RefStrokeLen i |
10: else
11: dist ← dist + UnknownStrokeLen i + RefStrokeLen i
12: i←i+1
13: Until (i > LengthOfUnknownStrokeSeq)

Output: distance between the two sequences

Fig. 5.27. Algorithm to compute inter-stroke distance

89
The difficult case is when the sequences are of different length. In such cases, it is
impossible to create a one-to-one correspondence between the elements observation code
sequences. To solve this problem, the sequence with the shorter length is expanded to be of
equal length with the longer one. The expansion is not straightforward and it needs careful
examination of the sequences. The reason is because there is a need to align observation
codes to their likely match in the longer sequence.

In Fig. 5.28, the expansion algorithm is shown. The algorithm, in general, finds a suitable
place of each observation codes in the shorter sequences as to align them to their match in
the longer sequences. The rest of the code places and their length in the shorter sequence
are filled with the value 0.
Input: Two stroke sequences (one shorter, the other longer)

// ShortSeq: The shorter sequence

// ShortSeqLen: Set of Lengths for the code in the shorter sequence
// LongtSeq: The longer sequence
// LongSeqLen: Set of Lengths for the code in the longer sequence

1: Compute LengthOfShortSeq, length of the ShortSeq

2: Compute LengthOfLongSeq, length of the LongSeq
3: i←1
4: min ← 50 //a large number is assigned
5: Do
6: j←1
7: Do
8: if (ShortSeq i = LongSeq j)
9: l ← |ShortSeqLen i – LongSeqLen j|
10: if (l < min)
11: min = l
12: index = j
13: j←j+1
14: Until ( j > LengthOfLongSeq
15: if (i ≠ index)
16: ShortSeq index = ShortSeq i
17: ShortSeqLen index = ShortSeqLen i
18: ShortSeq i = 0
19: ShortSeqLen i = 0
20: i←i+1
21: Until (i > LengthOfShortSeq)
22: Fill the rest of the places in the shorter sequence by 0.

Fig. 5.28. The Sequence Expansion Algorithm

90
An example is provided to illustrate the inter-stroke computation in Fig. 5.29. Two
sequences for the different instances of the same letter " " are compared to the X reference
file that was produced during training conducted for letter " " of Writer A. It can be seen
that the writer used only one stroke to draw the characters. In the second instance of the
character, the length of the observation code sequence and the length of the reference code
sequence are equal. Moreover, the sequences are identical in their element observation
codes. Thus we need to compute only the length difference of each code. In the first
instance, the lengths of the sequences are different and consequently an additional step,
expansion is needed before computing the distance. Accordingly, the X reference sequence
(i.e. the shorter sequence) is expanded to have the same length with the first instance
sequence. Then, the two sequences are compared to calculate the distance between them.
Following the steps of the algorithm, each corresponding code is compared. The first and
the second element codes are identical in both sequences. But, the third element is 0 in the
reference sequence and 3 in the other sequence. This difference implies that the length of
both will be added to the distance value (Fig. 5.27, Line 11).

X Reference file
trained for the
letter (Writer A)
X sequence stroke for X sequence stroke for
the letter 1 the letter
(Instance1, Writer A) 1 (Instance 2, Writer A)
1 0
T=3 2 9 T=2
1 0 10 10 1 0
2 11 2 10
3 2
computation
Distance
Expansion

d = 0 + |11-9| + 2 + 0
d=4 d = 0 + |10-9|
d =1

1 0
2 9
Distance 0 0
Computation
Fig. 5.29. An example Inter-stroke distance computation

91
Note that the X distance is larger for the first instance. This is due to the slightly deformed
shape of the character in its right side. The existence of the bend results in having the last
observation code 3 (decrease in x values) in its sequence. While computing the distance,
the existence of this code in the sequence has a major part in making the distance larger.
Therefore, the inter-stroke distance computation algorithm is effective in modeling the
level of similarity between strokes. This is important, as the classification result highly
depends on it.

Table 5.2 shows the coarse classification results for three sample characters (written by
Writer A). The table lists the X observation code distances and Y observation code
distances of the sample test characters against each of the reference characters. It also
includes the total distance. In all the three cases, the smallest distance belongs to the closest
characters. Despite this, the first layer'
s proposal in the second case is not taken as a final
classification result since the first two ("#" and " ") from the group are competing
candidates as the distance of "#" is less than only by 1 from that of " ". This situation
forces the recognizer to take a measure before reporting the classification result, which is
passing the two characters to the next layer. This implies that the next layer is supposed to
identify the correct one, if not pass them to the third layer.

92
 Reference Character Character # Character
No.
character X Y Tot X Y Tot X Y Tot
1 4 3 7 500 500 1000 500 500 1000
2 # 500 500 1000 5 2 7 500 500 1000
3 500 500 1000 19 33 52 500 500 1000
4 12 25 37 500 500 1000 500 500 1000
5 4 23 27 500 500 1000 500 500 1000
6 6 13 19 500 500 1000 500 500 1000
7 500 500 1000 17 30 47 500 500 1000
8 500 500 1000 500 500 1000 15 26 41
9 15 35 50 500 500 1000 500 500 1000
10 T 4 16 20 500 500 1000 500 500 1000
11 500 500 1000 7 1 8 500 500 1000
12 500 500 1000 500 500 1000 13 22 35
13 g 5 18 23 500 500 1000 500 500 1000
14 6 12 18 500 500 1000 500 500 1000
15 500 500 1000 11 3 14 500 500 1000
16 y 12 11 23 500 500 1000 500 500 1000
17 4 12 16 500 500 1000 500 500 1000
18 500 500 1000 14 25 39 500 500 1000
19 Π500 500 1000 500 500 1000 19 19 38
20 “ 17 2 19 500 500 1000 500 500 1000
21 500 500 1000 500 500 1000 1 1 2
22 500 500 1000 500 500 1000 500 500 1000
23 ¦ 500 500 1000 13 7 20 500 500 1000
24 14 15 29 500 500 1000 500 500 1000
25 500 500 1000 18 23 41 500 500 1000
26 - 6 13 19 500 500 1000 500 500 1000
27 4 32 36 500 500 1000 500 500 1000
28 14 41 55 500 500 1000 500 500 1000
29 Ú 500 500 1000 19 32 51 500 500 1000
30 Ì 19 17 36 500 500 1000 500 500 1000
31 Ó 500 500 1000 17 14 31 500 500 1000
32 á 500 500 1000 12 7 19 500 500 1000
33 500 500 1000 13 14 27 500 500 1000
34 500 500 1000 500 500 1000
, distance 7 #, distance 7 , distance 2
, distance 16 , distance 8 , distance 35
Top Five
, distance 18 , distance 14 Œ, distance 38
characters
, distance 19 á distance 19 , distance 41
“, distance 19 ¦ distance 20 , distance 1000
# and are
Recognized passed to the
Character
middle layer.

Table. 5.2. Coarse Classification Results for three sample characters from the testing data

93
5.4.6.2. Detailed Matching
Characters with equal or nearly equal (with difference less than 3) distance are passed to
the second layer of the recognizer so that detailed matching could be done to identify the
most likely character that would be considered as the classification result. This requires
using a more detailed sequence to compare the characters. We, purposely, took out all
occurrences of the observation code 4 from all sequences before they were provided to the
coarse classification layer. Now, it is time to include this code in the sequences, in order to
make them more detailed than the previous sequences that have been used in the first layer.

The main difference between the coarse classification layer and the detailed matching layer
is the type of the sequences they use to classify. Moreover, the competition would be
among fewer numbers of characters in the detailed matching layer than the coarse
classification. As we did in the coarse classification layer, distances are computed between
the detailed sequences in a similar manner by applying the algorithms discussed in the
preceding section.

An example is provided below (Fig. 5.30a and Fig. 5.30b) to show how the inclusion of the
observation code 4 makes a difference. In the first layer, the distance between the character
" " (writer A) and the reference files trained for the characters " " and "-" is 0. This is
because the reference files are identical. This means that the first layer often confuses the
two characters. To resolve this, the detailed matching (Fig. 5.30b) compares the original
form of the sequences that include the observation code 4. Distance values that can clearly
distinguish the two characters are obtained. Thus, the detailed matching layer produced the
character " " as the final classification result.

94
 X observation Y observation
code sequence code sequence
T=2 T=2
1 0 1 0
2 7 3 10

X Reference Y Reference
files files

- -
1 1 1 1
1 1 1 1
1 0 1 0 1 0 1 0
2 7 2 7 3 10 3 10
10 10 10 10 10 10 10 10

d=0 d=0 d=0 d=0

Fig. 5.30a. Coarse Classification for the character " "

X observation Y observation
code sequence code sequence
T=2 T=2
1 0 1 0
4 9 3 10
2 7 4 6

X Reference Y Reference
files files

- -
1 1 1 1
1 1 1 1
1 0 1 0 1 0 1 0
4 9 2 7 3 10 4 6
2 7 4 10 4 5 3 10
10 10 10 10 10 10 10 10

d=0 d = 33 d=1 d = 32

Fig. 5.30b. Detailed Classification for the character " "

Yet, this is not always the case. Some character are also confused by the middle layer
which in that case passes them to the last layer, namely the superimposing matching.

95
5.4.6.3. Matching by Superimposing
As stated earlier, characters that have been passed to this layer are those that cannot be
identified by the first as well as the second layers of the recognizer. For this reason, the
classification here must be handled in a special way. This means that observation code
sequences are no more useful in distinguishing these characters. This necessitates
employing a different approach to distinguish one character from the other.

The methodology employed in the third layer classification scheme constitutes

superimposing the unknown character on top of the reference characters and measure the
degree of matching, hence superimposing matching. To measure the matching, the distance
between the corresponding data points (inter-point distance) that form the reference
characters and the superimposed character is computed. After that, the inter-point distances
are added to get the matching distance. A reference character that results in smaller
matching distance is considered to have high degree of matching and it will be the result of
the recognizer.

The superimposing matching process incorporates two major steps namely overlying the
unknown character on top of each of a reference character and computing the matching
distance.

5.4.6.1.3. Superimposing
Superimposing is the action of taking the unknown character and putting it on top of the
reference character. An algorithm was designed to perform this task and it is presented in
Fig.5.31. The algorithm first identifies the boxes in which the reference and the unknown
character are drawn turn by turn. This is achieved by computing the maximum and
minimum values of the x and y coordinate for both characters. To put the unknown
character on top of the reference character, it is necessary to translate all the data points that
make up the unknown character. The translation is done by adding values to the x and y
coordinate values of the points. Thus, the difference between the maximum x value of the
reference character and that of the unknown character, and the difference between the
minimum x value of the reference character and that of the unknown character are averaged
to get the change in x. The same thing will be done for obtaining the change in Y. Then,

96
 Input: Preprocessed data of a reference character and an unknown character
//Each data point of the unknown character is a triple (x, y, z)

1: Get UnknownMaxX, UnknownMinX, UnknownMaxY, UnknownMinY

//to get the box in which the unknown character is drawn
2: Get RefMaxX, RefMinX, RefMaxY, RefMinY
//to get the box in which the reference character is drawn
3: ChangeInX ← [(RefMaxX – UnknownMaxX)+( RefMinX- UnknownMinX)]/2
4: ChangeInY ← [(RefMaxY – UnknownMaxY)+( RefMinY- UnknownMinY)]/2
5: i←1
6: Do
7: if (x i ≠ 0)
8: newx i ← x i + ChangeInX
9: newy i ← y i + ChangeInY
10: newz i ← 0
11: else
12: newx i ← 0
13: newy i ← 0
14: newz i ← 0
15: i←i+1
16: Until (i > n) //n is the number of data points that make up the unknown
character
each data point of the unknown character is translated to a relative data point inside the box
in which the reference character has been drawn.

Fig. 5.32a. Superimposing Algorithm

Fig. 5.32b. Characters after superimposing

97
The character "#" and "T" (Writer B) are often confused by the second layer and they are
sent to the third layer for further examination. Fig. 5.32a and Fig. 5.32b show how
superimposing is done on them. It can be clearly seen that the first set of characters (Fig.
5.32a) are well matched than the second set (Fig. 5.32b). This difference could be
quantified by calculating the distance between each corresponding data point of the two
overlapped characters.

5.4.6.1.4. Matching Distance Computation

The matching distance is equal to the sum of the inter-point distances. The inter-point
distance is calculated as:

d i = x − newx + y − newy where (x,y) is a data point on the reference

character and (newx, newy) is the corresponding data point on the character that is
superimposed on the reference character.
The matching distance is:
n
D= d i where d i is the inter-point distance
1

In the example given in Fig. 5.32a and Fig. 5.32b, after superimposing the characters, the
matching distance between each pair is computed. It was obtained that the matching
distance for the characters in Fig. 5.32a is 465 and it is 875 for the characters in Fig. 5.32b.
This led to the conclusion that the tested character is recognized to be " #".

5.5. Summary
In this chapter, the online handwriting recognition engine designed and developed for the
basic Ethiopic characters has been discussed. The design of the recognition engine with the
detailed explanation of its components has been presented. We have also described the
Movalyzer digitizer software that was used for data collection. Algorithms that were
designed for accomplishing the preprocessing activities have also been presented. We dealt
with the feature extraction process widely. Under this, we covered the observation code
sequence extraction process, which is the proposed handwriting pattern representation
technique. We have claimed that we developed a recognition engine based on this pattern
representation scheme. Along this, the new algorithms designed for the purpose of training
and classification have been presented.

98
 CHAPTER SIX
EXPERIMENTATION AND DISCUSSION

Experiments have been conducted to test the system for its accuracy. In this chapter, we
present the results of these experiments to evaluate the performance of the recognizer and
the performance of each of the layers in the recognizer.

6.1. Experimentation
Two major experiments were conducted to test the recognition engine for its accuracy and
ability to handle variations. It was already mentioned that two writers (writer A and writer
B) involved in the data collection process and the experiments were carried out on the
handwriting of each writer. Nine instances of each of the 33+1 character written by these
writers were collected. This constitutes the whole available data. The first three instances
of each character from the collected data were used to train the system. Now, the data was
divided into two groups namely the training data and the testing data. The training data
contained the first three instances of each of the 33+1 characters and the rest of the
instances are made to be part of the testing data.

Two types of experiments were conducted for each writer. The first experiment type was
performed by utilizing the training data itself as testing data. The second experiment was
conducted by using the real testing data. Note that the two experiments are different in the
type of data used. In the first experiment, the engine was provided data that it has seen
previously. The results of the experiments are presented in Tables 6.1 and 6.2.

Experiment No. Writer Data Used Accuracy

1 Writer A Training data 99.7%
2 Writer A Testing Data 99.4%
Average 99.55%
Table 6.1. Experiment Results for Writer A

Experiment No. Writer Data Used Accuracy

1 Writer B Training data 99.5%
2 Writer B Testing Data 99%
Average 99.25%
Table 6.2. Experiment Results for Writer B

99
6.2. Performance of the Classifier’s Layers
The final outputs of the recognizer are tested by the experiments discussed in the last
section. From the same experiments, the performance of the three layers of the recognizer
is also analyzed as follows.

On the average, only about 16% of the characters are passed to the second layer. However,
this doesn’t necessarily mean that all of these characters were not recognized correctly by
the first layer. But, it means that the recognizer is not sure about the results proposed by the
first layer. For writer A, from the characters that are not passed to the second layer (the
result is reported based on the proposal of only the first layer), 99.8% of them are correctly
recognized. Conversely, from the characters that are passed to the second layer, 61.3% of
them were correctly recognized. In summary, the recognition accuracy of the first layer
(coarse classification layer) is up to 93.9% for writer A. A similar analysis has shown that
the recognition accuracy of the same layer for writer B is 91.9%.

To determine the accuracy of the second layer (detailed matching), it suffices to see how
many of the characters passed to it are correctly recognized. We have noticed that 99% of
them are correctly recognized for both writers.

Characters are passed to the third layer only in rare cases. For example, only 1% of the
total characters are made to pass to the third layer for writer A. Actually, the number of
characters that has been given to the third layer are greater in number for writer B. This is
because writer B writes different characters with the same number of strokes in the same
order. This situation increases the probability of passing the characters written by Writer B
to the third layer. Whatsoever, characters passing to the third layer are fewer in number as
compared to the total number of characters that are tested. After the computationally
complex transformation and distance computation, the third layer is able to identify 99.8%
of the characters correctly (see Table 6.3).

Layer Writer A Writer B Average Accuracy

Coarse Classification 93.9% 91.9% 93.4%
Detailed Matching 99% 99% 99%
Superimposing 99.8% 99.7% 99.8%
Matching
Table 6.3. Performance of the layers of the recognizer.

100
6.3. Discussion
It is reminded that this is the first attempt to solve the problem of online handwriting
recognition for Ethiopic characters. This necessitates conducting a study on the various
approaches used to solve this problem for other character sets in relation to the nature of
the Ethiopic characters. To this end, we have also studied the shape of Ethiopic characters.

Online handwriting data collection is the primary step and it is accomplished by digitizer
software named MovAlyzer using the mouse as an input device. This software collects pen
down and pen up data points along the path of the mouth. A stroke is the sequence of data
points between a pen down and a pen up. Thus, the character can be defined as group of
strokes in the order they appear while the character is written where the strokes in the are
separated by set of pen up points which are sampled by MovAlyzer while the user lift up the
mouse to draw another stroke. This data is presented to the recognition engine where part
of it is utilized for training and the rest for testing purpose.

Regardless of the usage of data (for training or testing), it is first presented to the
preprocessing module in which different preprocessing activities that help in reduction of
data and reduction of variations are performed. The feature extraction module takes the
preprocessed data to represent each stroke is in terms of X and Y observation code
sequences. Then, the set of strokes in a character represented in terms of X and Y
observation code sequences are stored in a text file in which their order is maintained.

While training, the training algorithm collects X and Y observation code sequences of
strokes in a character from the sample data and creates a reference file. The major content
of this file is the collected X and Y observation code sequences from the sample data.

Classification is accomplished by using the three-layered recognizer module. The first layer
is the coarse classification layer in which inter-strokes distances are used to match
characters. In the second layer, detailed matching is performed by using detailed
observation code sequences. The last layer, namely the superimposing layer, uses a
mechanism of comparing two characters by superimposing one on another and finding
inter-point distances to measure the matching

101
The experimental results show that the pattern representation scheme, the algorithms that
we have designed for the various steps in the recognition process and the three-layered
structured recognizer together make most of the recognition tasks successful. Another great
discovery in this research is only small amount of data is needed to train the system and
this is an advantage to users. On top of this, the layered structure of the recognizer
facilitates use of resources efficiently by allowing processing computationally complex
algorithms implemented in lower layers for only ambiguous results of upper layers.

According to an analysis that has been done on Ethiopic character set, it is concluded that
the majority of non-basic characters are formed by adding secondary strokes on the basic
characters. Some others are formed by modifying the shapes of the basic characters based
on different rules except some exceptional cases. Thus, we proposed a strategy to approach
the problem by first solving the online character handwriting problem for the basic
characters. Accordingly, we narrowed down our work to consider only the 33+1 basic
characters. We strongly believe that the system can be extended to include the non-basic
characters though we didn’t do that due to time constraint. Though it seems that this is a
series limitation, we justified that the non-basic characters could be recognized based on
the basic characters as this strategy has been applied for the construction of other systems
such as the Ethiopic/Amharic typewriter keyboard layout and the phonetic Ethiopic text
entry system on computer keyboards.

In its easy form, extending the system will not be different from adding model characters
and train the system with them. This means that the system could be trained with instances
of the 231+7 Ethiopic alphabetic letters to incorporate the remaining non-basic characters.
However, this approach should not be directly pursued as it may make the system
inefficient both in memory requirement and speed. Thus, additional techniques to avoid
such consequences must be considered. Some of these are consideration of the structural
relationship between strokes, formation of an efficient structure to store the available
strokes and their order to form characters (so as to minimize the number of set of reference
sequences stored for similar strokes that occur in different characters).

Another limitation of the whole approach is that the approach is stroke number and stroke
order dependent. For writer-dependent systems, this is tolerable. In the case of writer-
independent systems, this causes to have many more reference sequences of characters

102
which might widen the search space and consequently reduce the efficiency of the
recognition system. This situation might also increase the amount of training data needed to
train the system. But for our consideration, we believe that it is not a big problem.

103
 CHAPTER SEVEN
CONCLUSION AND FUTURE WORKS

In broader sense, online handwriting character recognition is the process of finding out
what character is represented by a given online handwritten character data. Online
handwriting character data is, in general, a set of the coordinate values of data points that
are sampled along the trajectory of the character as it is written. The difficulty of
identifying what character is actually represented by the data set start from the fact that
identical set of data points will not be obtained for different occurrences of the same
character. Thus, pattern representation scheme that produces similar representations for
different occurrences of the same character solve the problem of character classification
greatly.

We have introduced a new online handwriting character data representation scheme, which
we have proved that it models the pattern in a manner that creates similarity between
different occurrences of the same character. Major contributions of this work are outlined
below:
• The nature of the shapes of Ethiopic characters is analyzed and a strategy to
approach the problem of online handwriting recognition for Ethiopic characters is
proposed. This strategy is to first concentrate only on the basic characters since it
would be reasonably possible to extend the system to incorporate the non-basic
characters. This is evidently logical, as we have shown that the shapes of most of
the non-basic characters are derived by applying different modifications on the
shapes of the basic characters.
• An online handwriting recognition engine for Ethiopic basic characters is designed.
• Algorithms for preprocessing activities are designed and implemented. These
include extra point noise elimination, size normalization, filtering and resampling
algorithms.
• A new method of online handwriting pattern representation scheme, which makes
use of X and Y observation code sequences derived from the data is introduced and
algorithms for generating these sequences are designed and implemented. This

104
 constitutes the feature extraction process, which is considered as the most important
factor in determining the classification result.
• Training algorithm by which the engine is trained is also designed and
implemented. This algorithm takes the observation code sequences extracted from
the sample characters in the training data and produces a reference file that
systematically stores the reference observation code sequences. The training
algorithms needs relatively small amount of data to train the writer-dependent
recognition system.
• Most importantly, a three-layered recognizer is designed and implemented. The
recognizer has three layers namely coarse classification, detailed matching and
superimposing matching, through which unknown character passed if needed. The
logical structure employed by the recognizer attempts to pass a to be recognized
character to a lower layer if and only if the character is not surely identified by the
current layer. We have done this to minimize the inaccuracy of the engine.

As it is only the beginning, future works are expected. We suggest the following:
• Improve the approach to make it stroke number and order independent.
• Extend the recognition engine to incorporate the non-basic Ethiopic characters.
• Extend the system to incorporate word-level and/or sentence-level recognition.
• Extend the recognition engine to be writer-independent.

105
References
1. Anil K., Jain, Robert P.W., Duin, Jianchang Mao: “Statistical Pattern Recognition: A
Review”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, no.
1, January 2000.

2. Jorma Laaksonen, “Subspace Classifiers in Recognition of Handwritten Digits”,

(Thesis for the degree of Doctor of Technology), Helsinki University of Technology,
Department of Computer Science and Engineering, Laboratory of Computer and
Information Science, May 1997.

3. Palmondon, S.N. Srihari: “Online and Offline Handwriting Recognition: A

Comprehensive Survey”, IEEE Transactions on pattern Analysis and Machine
Intelligence, Vol. 22, No-1, 2000.

4. A. Drissman, “Handwriting Recognition Systems: An Overview”, available from

www.drissman.com/avi/school/HandwritingRecognition.pdf

5. “Pen-based Solutions and Handwriting Recognition”, HP Labs, Hewlett- Packard

Development Company, L.P., 2004. Available in site:
www.hpl.hp.com/india/research/lt-penhw-interfaces.html

6. Nigussie Tadesse: “Handwritten Amharic Text Recognition Applied to the Processing

of Bank checks”, (Masters Thesis). Addis Ababa University, School of Information
Studies for Africa, Addis Ababa University, 2000.

7. Messay Hailemariam: “Handwritten Amharic Character Recognition:The Case of

Postal Addresses ”, (Masters Thesis). Addis Ababa University, School of Information
Studies for Africa, Addis Ababa University, 2003.

8. Wondwossen Mulugeta, ““OCR for Special Type of Handwritten Amharic Text

(“Yekum Tsifet”), Neural Network Approach (Masters Thesis). Addis Ababa
University, School of Information Studies for Africa, Addis Ababa University, 2003.

9. Scott D.Connell, “Online Handwritten Recognition Using Multiple Pattern Class

Models”, (Doctor of Philosophy Dissertation), Department of Computer Science and
Engineering, Michigan State University, 2000.

10. H. Shu, “On-line Handwriting Recognition Using Hidden Markov Models”, (Master
Thesis), Department of Electrical Engineering and Computer Science, Massachusetts
Institute of Technology, 1997.

106
11. G. Loudon, O. Pellijeff, Li zhong-wei, “A Method for Handwriting Input and
Correction on Smartphones”, , Ericsson Research, Cyberlab Singapore, 2000.

12. S. Jaeger, C.-L. Liu, M. Nakagawa, “The state of the art in Japanese Online
Handwriting Recognition Compared to Techniques in Western Handwriting
Recognition”, International Journal on Document Analysis and Recognition, 2003.

13. C.L.Liu, S. Jaeger, M.Nakagawa, “Online Recognition of Chinese Characters, The

State-of-the-Art”, IEEE Transactions on Pattern Analysis and Machine Intelligence,
Vol. 26, No. 26, 2004.

14. T.Klassen, “Towards Neural Network Recognition of Handwritten Arabic

Letters”(Masters Thesis). DalHousie University, Faculty of Computer Science, 2001.

15. E. Gómez Sánchez, Y.A. Dimitriadis, M. Sánchez-Reyes Más, Sánchez García, J.M.
Cano Izquierdo, J. López Coronado, “Online Character Analysis and Recognition
With Fuzzy Neural Network”, Intelligent Automation and Soft Computing, Vol. 7, No.
3, pp. 161-162, 1998.

16. Kam-Fai Chan, Dit-Yan Yeung “Recognizing online handwritten alphanumeric

characters through flexible Structural Matching”, The Journal of Pattern Recognition
Society, October, 1998

17. W. Thimbleby, “A Better Calculator: Processing Handwritten Mathematical

Expressions to Solve Problems”, May, 2004

18. N. Joshi, “Tutorial On Handwriting Character Recognition”, Biomedical Lab,

Department of Electrical Engineering, Nov.2003

19. M.Aksella, “Handwritten Character Recognition: A Palmtop Implementation and

Adaptive Commmitte Experiments”(Masters Thesis). Helsinki University of
Technology, Department of Engineering Physics and Mathematics, May 2000.

20. Ignace Jay Gelb, R. M. Whiting, “Writing”, Microsoft Encarta Reference Library 2004,
Microsoft Corporation.

21. Robert A. Fradkin, “Alphabets”, Microsoft Encarta Reference Library 2004, Microsoft
Corporation.

22. Alwiya S. Omar, “African Languages”, Microsoft Encarta Reference Library 2004,
Microsoft Corporation.

107
23. “Ethiopic Script and Language”, Microsoft Encarta Reference Library 2004,
Microsoft Corporation.

24. C.C. Tappert, C.Y. Suen, T.Wakahara, “The State of the Art in Online Handwriting
Recognition”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.
12, No. 8, August, 1990.

25. A.Lenaghan, R. Malyan and G.A Jones, “Matching Structural Descriptions of

Handwritten Characters Using Heuristic Graph Search”, Kingston University,
School of Computer Science and Electronic Systems.

26. “Pen Technologies”, IBM Pen Technologies, Available from

“www.research.ibm.com/electricInk/”

27. K. Chan, D. Yeung, “PenCalc: A Novel Application of Online Mathematical

Expression Recognition Technology”, Hong Kong University of Science and
Technology, Department of Computer Science.

28. S.S.Tang, I.Methaste, “Thai Online Handwritten Character Recognition Using

Windowing Backpropagation Neural Networks”, Information Research and
Development Division, National Electronics and Computer Technology Center,
National Science an Technology Department Agency, Rachathewi, Bangkok, Thailand,
available from www.links.nectec.or.th/ uploads/upload/040336/onlineBNN.pdf

29. P. Burrow, “Arabic Handwriitng Recognition”, (Masters Thesis), School of

Informatics, University of Edinburgh, 2004.

30. D. Perry, “Handheld Computers (PDAs) in Schools”, Becta ICT Reseach, March
2003.

31. R. Kassel, “A comparison of approaches to on-line handwritten character

recognition”. Ph.D. Thesis, MIT Department of Electrical Engineering and Computer
Science, June 1995.

32. T. Starner, J. Makhoul, R. Schwartz, and G. Chou, “On-Line Cursive Handwriting

Recognition Using Speech Recognition Methods,” presented at International
Conference on Acoustics, Speech, and Signal Processing, 1994.

33. F. Kubala, A. Anastasakos, J. Makhoul, L. Nguyen, R. Schwartz, and G.Zavaliagkos,

“Comparative Experiments on Large Vocabulary Speech Recognition,” presented at
International Conference on Acoustics, Speech, and Signal Processing, 1994.

108
34. Jaeger S, Manke S, Reichert J, Waibel A, “Online Handwriting Recognition: the
Npen++ Recognizer”, International Journal on Document Analysis and Recognition,
2001.

35. D. Paul, “The Design for the Wall Street Journal-based CSR Corpus,” presented at
Proceedings of the DARPA Speech and Natural language Workshop, 1992.

109