CHARACTER RECOGNITION USING NEURAL NETWORK

SYAHFINASH BINTI SHAFII

This Report Is Submitted In Partial Fulfillment Of Requirement For The

Bachelor Degree of Electonic Engineering (Computer Engineering)

Fakulti Kejuruteraan Elektronik dan Kejuruteraan Komputer

Universiti Teknikal Malaysia Melaka

June 2016
iii
iv
 v

ACKNOWLEDGEMENT

I would like to express my greatest gratitude to ALLAH S.W.T for the mercy,
blessing and strength given for me to complete this thesis. With HIS blessing and
bestowed, I have able to complete this thesis.

My special thanks goes to my dedicated supervisor, En. Rosman bin Abd

Rahim, who always provides good supervision, encouragement and critics. I am truly
indebt with all the helps that he has done for me in completing this final project from
the beginning until the end of this thesis.

My thanks also go to my family for their support and endless encouragement

throughout my life. They have been a wonderful source of support, inspiration, and
encouragement throughout my education, and they deserve much credit for where I
am today.

Lastly, my deepest appreciation to all the persons that involved directly and
indirectly in this project with full willingness in contributing their efforts, time, energy
and idea in helping me completing this thesis. There are no other words that would
able to express my feeling of gratitude towards them except thank you.
 vi

ABSTRACT

Humans have the ability to recognize characters. For example, human can
distinguish between different characters and recognize them easily as an A or a B and
so on. Therefore, project is intended to develop a neural network system that is able to
perform character recognition, particularly English alphabets. Neural network is a
system inspired by human brain function; consists of neurons connected in parallel
with the ability to learn. A basic design of neural network has input layer, hidden layer,
and output layer. The use of neural network can improve the quality of recognition
while achieving good performance. A total of 650 samples characters are used with 25
samples of each character. The performance of evaluation is divided to 80% of training
and 20% for testing. Scaled conjugate gradient training function is used as this function
can perform faster in pattern recognition as well as its small memory requirement. Two
training methods are used. The first one is the Gradient Technique with 39 neurons in
hidden layer. The second training method is Geometric Feature Extraction with 35
neurons in hidden layer. Gradient Technique and Geometric Feature Extraction; both
show an excellent recognition rate of 94.6% and 94.3% respectively. The output of
recognized characters is shown in a .txt file.
 vii

ABSTRAK

Manusia mempunyai keupayaan untuk mengenal pasti abjad. Sebagai contoh,

manusia boleh membezakan antara abjad-abjad yang berbeza dan dengan mudah dapat
mengenali sama ada abjad tersebut adalah A atau B dan sebagainya. Oleh itu, projek
ini bertujuan untuk membina satu sistem rangkaian neural dengan keupayaan untuk
melaksanakan pengecaman abjad, terutamanya huruf bahasa Inggeris. Rangkaian
neural adalah satu sistem yang diilhamkan oleh fungsi otak manusia; yang terdiri
daripada neuron-neuron yang disambung secara selari yang berkebolehan untuk
belajar. Reka bentuk rangkaian neural yang asas mempunyai lapisan input, lapisan
tersembunyi, dan lapisan output. Penggunaan rangkaian neural boleh meningkatkan
kualiti pengecaman abjad disamping mencapai prestasi pengecaman yang baik.
Sejumlah 650 sampel abjad yang telah digunakan bersamaan dengan 25 sampel bagi
setiap abjad. Penilaian prestasi dibahagikan kepada 80% latihan dan 20% untuk ujian.
Skala konjugat kecerunan digunakan sebagai fungsi latihan kerana fungsi ini
mempunyai keupayaan yang lebih cepat dalam pengecaman corak serta memerlukan
memori yang kecil. Terdapat dua kaedah latihan yang digunakan. Yang pertama adalah
teknik kecerunan yang mempunyai 39 neuron di lapisan tersembunyi. Manakala
kaedah latihan yang kedua adalah pengekstrakan ciri geometri yang mengandungi 35
neuron di lapisan tersembunyi. Teknik kecerunan dan pengekstrakan ciri geometri ini
masing-masing menunjukkan kadar pengenalpastian abjad yang sangat baik iaitu
94.6% dan 94.3%. Output bagi abjad yang dapat dikenalpasti dipaparkan dalam fail
.txt.
 viii

TABLE OF CONTENTS

CHAPTER CONTENT PAGE

PROJECT TITLE i

REPORT VERIFICATION STATUS FORM ii

DECLARATION iii

SUPERVISOR'S DECLARATION iv

ACKNOWLEDGEMENT v

ABSTRACT vi

ABSTRAK vii

TABLE OF CONTENT viii

LIST OF TABLES xi

LIST OF FIGURES xii

LIST OF ABBREVIATIONS xiv

1 INTRODUCTION

1.1 Project Background 1

1.2 Problem Statement 1
1.3 Objectives 2
1.4 Scope and Limitation 2
1.5 Project Methodology 3
 ix

1.6 Report Structure 4

2 LITERATURE REVIEW

2.1 MATLAB 6
2.2 Image Processing 7
2.2.1 Image processing in MATLAB 9
2.3 Character Recognition 9
2.3.1 Handwritten character recognition 10
2.4 Neural Network 10
2.5 Comparison of literature review 12

3 METHODOLOGY

3.1 Project Design 18

3.2 Project Process Flow 18
3.2.1 Literature Review 19
3.2.2 Research on Algorithm 19
3.2.3 Coding Development in MATLAB 20
3.2.4 Character Recognition 20
3.2.5 Result and Analysis 20
3.2.6 Discussion and Future Work 20
3.3 Project Methodology 21
3.4 Image Pre-processing 22
3.4.1 Image Conversion 22
3.4.2 Morphology 23
3.4.3 Edge Detection 25
3.5 Feature Extraction 26
3.5.1 Gradient Technique 26
3.5.2 Geometric Feature Extraction 27
3.6 Training and Classification 28
3.6.1 Supervised Learning 28
 x

3.6.2 Unsupervised Learning 29

4 RESULT AND ANALYSIS

4.1 Result 30
4.1.1 Gradient Technique 31
4.1.2 Geometric Feature Extraction 32
4.1.3 Performance comparison between
training methods 34
4.2 Graphical User Interface (GUI) 44
4.3 Discussion 48

5 CONCLUSION

5.1 Conclusion 50
5.2 Future Work 51

REFERENCES 52

APPENDIX 54
 xi

LIST OF TABLES

NO TITLE PAGE

2-1 Comparison of literature review 12

3-1 Rule for Dilation and Erosion 24
4-1 Characters recognition rate for Gradient Technique 31
4-2 Characters recognition rate for Geometric Feature Extraction 32
4-3 The network performance after training 35
4-4 The output of recognition 36
4-5 Summary of optimal result when employing different methods 44
 xii

LIST OF FIGURES

NO TITLE PAGE

1.1 Process flow for Final Year Project (FYP) 4

2.1 MATLAB symbol 6
2.2 MATLAB platform 7
2.3 Fundamental of digital image processing 8
2.4 The components of image processing system 8
2.5 The example of image processing using MATLAB
for counting circular objects in an image 9
2.6 Example of handwriting recognition 10
2.7 A drawing of biological neurons 11
2.8 Neural Network modelled layer 11
3.1 Project process flow 19
3.2 Flow of algorithm development 21
3.3 Coding for converting RGB to grayscale to binary image 23
3.4 Output of image conversion 23
3.5 Coding for morphology function 24
3.6 Output images for morphology 24
3.7 Coding for edge detection 25
3.8 Example of edge detection output using Sobel algorithm 25
4.1 Neural network training block and progress 30
4.2 Neural network training progress for (a) Gradient Technique
and (b) Geometric Feature Extraction 34
4.3 The sample input image 36
4.4 Validation performance plot when employing gradient
technique method 36
4.5 Validation performance plot when employing geometric
feature extraction method 37
4.6 Gradient and validation checks plots when employing
 xiii

gradient technique method 38

4.7 Gradient and validation checks plots when employing
geometric feature extraction method 39
4.8 Confusion matrix plot employing gradient technique method 40
4.9 Confusion matrix plot employing geometric feature
extraction method 41
4.10 Regression plot employing gradient technique 42
4.11 Regression plot employing geometric feature extraction 43
4.12 Home Page 45
4.13 Image uploaded 45
4.14 Neural Network Training 46
4.15 Extracting characters 47
4.16 Output, recognized characters 47
 xiv

LIST OF ABBREVIATIONS

FYP Final Year Project

NN Neural Network

SVM Support Vector Machine

ANN Artificial Neural Network

MDA Multi-linear Discriminant Analysis

MSE Mean squared error

PSNR peak signal-to-noise ratio

RGB Red-Green-Blue (True color compound)

GUI Graphical User Interface

 CHAPTER 1

INTRODUCTION

This chapter discusses the project introduction, objectives of project, the

problem statement, and the scope of work and limitations of project, proceeding with
a brief explanation of project methodology and the report structure.

1.1 Project Background

Character recognition is described as an easy way for machine to read scripts

and digits. Neural network is a design of computational model based on human brain
to solve certain kinds of problems.

The idea is developing a system with neural network training that can helps
performing character recognition process. Due to the limitations of classical
computing, applying neural network serve for the emulation of human thinking in
computations.

This project aims to feed specifically handwritten English alphabets an input,

process the characters, train the characters using neural network algorithm, to
recognize the structure of characters, and matches the outputs with the desired targets.

To be more specific, this project is mainly focused on the development of a

neural network algorithm by using MATLAB software as the platform. The detailed
study of the performance of neural network through learning from its environment and
the accuracy of the system would be carried out.

1.2 Problem Statement

Conventional computers use an algorithmic approach which defines the

procedural term to solve problems. The algorithmic approach means that a computer
 2

follows a set of instructions in order to solve problems. The procedural term means; a
program starts at the first line of code, executes it, and goes to the next instruction in
linear pattern.

Unless the specific steps that the computer needs to follow are known the
computer cannot solve the problem. That restricts the problem solving capability of
conventional computers to problems that already been understand.

Neural networks, with its remarkable ability to derive meaning from

complicated or imprecise data, can be used to extract patterns and detect trends that
are too complex to be noticed by either humans or other computer techniques.
Therefore, neural network can be utilized to reduce humans’ errors while performing
mundane tasks like digitization and statistical analysis. A trained neural network can
be thought of as an "expert" in the category of information it has been given to analyze.
This expert can then be used to provide project and implementations given new
situations of interest and answer "what if" questions.

1.3 Objectives

The main requirement of this project is to develop a system that can recognize
character using the neural network method. Therefore, the following objectives need
to be achieved to satisfy the development of the project.

i. To study Neural Network algorithm and develop a system that is able to

recognize characters using MATLAB.
ii. To detect, extract and recognize characters, particularly English alphabets
using Neural Network.
iii. To analyze performance of the neural network.

1.4 Scope and Limitation

This project is only focusing in developing neural network algorithm for

character recognition application. For this project, characters are freely written in Paint
tools, saved the characters in any image format, and later is took as samples for training
 3

and testing purposes. The entire process of completing the project will be done in
MATLAB. The MATLAB is chosen because it is a user friendly tool and easy to
understand the coding implementation.

There are several limitations that have been considered while completing this
project. The first limitation is in data collecting. Handwritten data will not be included
to perform the process of recognizing characters. No scanned handwritten image
character will be used. There is also limitation in training approach method. Due to the
high difficulty and deep learning of unsupervised training, the supervised training is
chosen as the desired output has been provided for the network.

1.5 Project Methodology

There are three main components to be concerned to complete this project.

There are:

i. Pre-processing
ii. Feature extraction
iii. Classification and recognition
 4

Figure 1.1: Process flow for Final Year Project (FYP)

1.6 Report Structure

The report consists of five chapters that conclude a complete discussion of the
project. Chapter 1 describes an overview of the project. In Chapter 1, the objectives,
problem statement, and scope of work are explained. The objectives define the desired
and expected outcome to be achieved from the project. Problem statement discusses
the related issues to be solved while improving the current solved problems. Scopes of
work define the limitations and consideration taken throughout the project completion.

Chapter 2 is the background study of the project. Previous studies and

researches that have been done are reviewed to help in processing data while
conducting the project. The relevant researches are taken from journals, technical
papers, and books.
 5

Chapter 3 consists of the procedures and methodology performed in order to

complete the project. This chapter will also describe the project flow in details.

The outputs from the project is discussed in Chapter 4. All the results will be
stated in this chapter in details explanation. The performance of proposed model will
be discussed briefly.

The last part, Chapter 5 concludes the project following some

recommendations to help improving the project in future. This chapter will also state
either the objectives of the project are effectively achieved or partially achieved with
some comments.
 CHAPTER 2

LITERATURE REVIEW

This chapter describes the background studies related to the development of

project. Some important ideas and findings from previous researches are presented in
order to provide better understanding on the subject. It also helps to formulate new
ideas, methods and act as the benchmark for this project. Also the basic theories of
character recognition and neural network will be review.

2.1 MATLAB

MATLAB or Matrix Laboratory is a fourth-generation programming language

and numerical computing analysis environment. MATLAB is used in many fields
include image processing, communications, control design, and robotics. MATLAB
allows matrix calculation, developing and running algorithms, creating user interfaces,
data visualization, and interfacing with programs developed in other languages
including Java and FORTRAN. MATLAB's built-in graphics make it easy to visualize
and gain insights from data. A vast library of prebuilt toolboxes lets users get started
right away with algorithms essential to required domain.

Figure 2.1: MATLAB symbol

 7

Figure 2.2: MATLAB platform

2.2 Image processing

A process that deals with manipulating digital images through digital computer
as processing tool is known as digital image processing. Digital Image Processing
focuses on developing a system that is able to perform processing on an image. The
input of that system is a digital image and the system process that image using efficient
algorithms, and gives an image as an output.

Two branches of digital image processing are pre-processing and post-

processing. Pre-processing includes image acquisition, binarization, noise filtering,
and many other methods. Post-processing includes edge detection, segmentation,
pattern recognition, and many other methods.
 8

Figure 2.3: Fundamental of digital image processing

Figure 2.4: The components of image processing system

Data preprocessing converts raw data and signals into data representation
suitable for application through a sequence of operations. The objectives of data
preprocessing include size reduction of the input space, smoother relationships, data
normalization, noise reduction, and feature extraction.
 9

Research in the field of pre-processing on character recognition using neural

network is an improvement of the image data that suppresses unwanted distortion.
Image pre-processing is the technique of enhancing data images prior for further
processing.

2.2.1 Image processing in MATLAB

Digital image processing is the use of computer algorithms to create, process,

communicate, and display digital images. Digital image processing algorithms can be
used to convert signals from image sensor into digital images, improve clarity, remove
noise, compress images for communication across a network, and extract the size,
scale, or number of objects in a scene. Effective techniques for processing digital
images include using algorithms and tools that provide a comprehensive environment
for data analysis, visualization, and algorithm development.

Figure 2.5: The example of image processing using MATLAB for counting
circular objects in an image

2.3 Character Recognition

Character recognition has been a topic of interest recent few years. The ability
to identify machine printed characters in an automated or semi-automated manner has
obvious applications in numerous fields. It is impossible to develop an algorithm with
perfect correct recognition rate, hence while designing algorithms it is important to
consider failures of a system so that it will at least be predictable when mistakes are
made. Different algorithms and approaches can be performed with these failures in
mind to create a very low error system.
 10

2.3.1 Handwritten character recognition

Handwritten character recognition takes handwritten input from sources such

as touch screens and paper documents to be interpreted by a computer. The written
text image may be sensed as offline from a piece of paper by intelligent word
recognition. For a pen-based computer screen surface, the movements of the pen tip
may be sensed online and the recognition task is easier.

Figure 2.6: Example of handwriting recognition

2.4 Neural Network

Today, neural network has been a very popular computer tool used for solving
lot of different practical problems. Neural network provides the best solutions to many
problems in image recognition, natural language processing, and speech recognition.
The basic idea of neural network is to simulate interconnected brain cells inside a
computer so it can learn things all by itself, recognize patterns, and make decisions
just like human way. One of the advantages of using neural network is the network
learning. Network learning provides an efficient ability in recognition.