a study on handwritten marathi word recognition

“A STUDY ON HANDWRITTEN MARATHI WORD RECOGNITION”

A THESIS SUBMITTED TO

BHARATI VIDYAPEETH UNIVERSITY, PUNE

FOR AWARD OF DEGREE OF

DOCTOR OF PHILOSOPHY

IN

COMPUTER SCIENCE UNDER THE

FACULTY OF SCIENCE

SUBMITTED BY

CHANDRASHEKHAR HIMMATRAO PATIL

UNDER THE GUIDANCE OF

PROF. DR. M. S. PRASAD

DEPARTMENT OF COMPUTER SCIENCE,

YASHWANTRAO MOHITE COLLEGE OF ARTS, SCIENCE AND COMMERCE,

BHARATI VIDYAPEETH DEEMED UNIVERSITY PUNE.

JULY 2015

CERTIFICATE

This is to certify that the work incorporated in the thesis entitled

“A STUDY ON HANDWRITTEN MARATHI WORD RECOGNITION” for

the degree of ‘Doctor of Philosophy’ in the subject of Computer Science

under the faculty of Science has been carried out by

Mr. Chandrashekhar Himmatrao Patil in the Department of Computer

Science at Bharati Vidyapeeth Deemed University, Yashwantrao Mohite College of Arts, Science and Commerce, Pune during the

period from August 2010 to July 2015 under the guidance of Dr. M. S. Prasad.

Place: Pune (Signature of Head of the Institute with seal)

Date : Principal / Director

Seal

CERTIFICATION OF GUIDE

This is to certify that the work incorporated in the thesis entitled

“A STUDY ON HANDWRITTEN MARATHI WORD RECOGNITION”

Submitted by Mr. Chandrashekhar Himmatrao Patil for the degree of

‘Doctor of Philosophy’ in the subject of Computer Science under the

faculty of Science has been carried out in the Department of Computer

Science, Bharati Vidyapeeth’s Yashwantrao Mohite College of Arts, Science and Commerce, Pune during the period from August 2010 to

July 2015, under my direct supervision/ guidance.

Place : Pune (Research Guide)

Date : (Prof. Dr. M. S. Prasad)

DECLARATION BY THE CANDIDATE

I hereby declare that the thesis entitled “A STUDY ON HANDWRITTEN MARATHI WORD RECOGNITION” submitted by me to the Bharati

Vidyapeeth University, Pune for the degree of Doctor of Philosophy (Ph.D.) in Computer Science under the Faculty of Science is original piece of work

carried out by me under the supervision of Dr. M. S. Prasad. I further

declare that it has not been submitted to this or any other university or

Institution for the award of any degree or Diploma.

I also confirm that all the material which I have borrowed from other sources

and incorporated in this thesis is duly acknowledged. If any material is not

duly acknowledged and found incorporated in this thesis, it is entirely my

responsibility. I am fully aware of the implications of any such act which

might have been committed by me advertently or inadvertently.

Place : Pune Date : /07/2015 Research Student

Chandrashekhar H. Patil

ACKNOWLEDGMENT

I wish to express my sincere gratitude to Dr. M. S. Prasad, Research guide, Bharati Vidyapeeth Deemed University, Pune for his continuous support, encouragement and valuable guidance during my research work. I benefited a lot from his constructive suggestions, dedication and efforts to accomplish timely completion of my research. I shall always remain indebted to him.

My heartfelt thanks to Principal, Dr. K. D. Jadhav, Yashwantrao Mohite college and Prof. S. S. Shukla, Head, Dept. of Computer Science extending necessary facilities to carry forward my research.

I also thank Prof. Dr. M.G. Bodhankar, Prof. Dr. S. R. Patil for their continuous support and encouragement to complete my work.

My heartfelt thanks to my colleague Prof. Dr. S. M. Mali, for his valuable guidance, suggestions and critical comments helped me to carry forward my research.

I am beholden to Principal Dr. T. N. More of MAEER’S Arts, Commerce and Science College, Pune and Management of MAEER for supporting me to pursue for Ph. D. degree in computer Science. I also thank Mr. Vilas Shinde and other colleagues in the college for their constant support in completing my work.

I thank Ms. Manisha Bharambe and all fellow research scholars for their active participation in the technical discussions and providing a lively atmosphere during the course of research.

Lastly and most importantly, words cannot express my deepest gratitude to my beloved parents, my wife Kirti, daughter Rutu, my family members, relatives and friends for their love, support, patience and being source of a inspiration during the course of work.

Chandrashekhar H. Patil

Dedicated

To

My

Late Mother Pramila Patil

CONTENTS List of Figures

List of Tables

List of Abbreviations

1. INTRODUCTION ... 1

1.1 Optical Character Recognition (OCR) ... 2

1.1.1 Types of OCR ... 2

1.1.2 Data collection ... 3

1.1.3 Pre-processing ... 4

1.1.4 Segmentation ... 4

1.1.5 Feature extraction ... 5

1.1.6 Classification ... 6

1.2 Literature Review ... 7

1.3 Motivation for the present work, Problem statement ... 14

1.4 Organization of the Thesis ... 15

2. OBJECTIVES AND PROPOSED SYSTEM … 17

2.1 Objectives … 17

2.2 Description of the proposed system for Handwritten

Marathi word recognition … 18

3. DEVELOPMENT OF A DATABASE OF

HANDWRITTEN MARATHI WORDS, ISOLATED

CHARACTERS AND PREPROCESSING

... 22

3.1 Introduction ... 23

3.2 Marathi characters ... 24

3.3 Formation of Marathi words ... 26

3.4 Database Development ... 28

3.4.1 Database development for handwritten

Marathi simple words ... 28

3.4.2 Database development for handwritten ... 31

Marathi compound words

3.4.3 Database development for isolated

handwritten Marathi characters ... 33

3.5 Pre-processing ... 34

4. SEGMENTATION ... 38


4.2 Segmentation and Difficulties in Segmentation ... 39

4.3 Segmentation Methodology for simple words ... 42

4.4 Segmentation methodology for compound words ... 44

4.5 Discussion of Results ... 48

4.6 Analysis Of Results ... 58

5. MULTILEVEL CLASSIFICATION ... 60


5.2 Multilevel Classification ... 61


6. FEATURE EXTRACTION ... 71


6.2 Zone based symmetric density feature ... 72

6.3 Diagonal, Horizontal and Vertical features ... 76

6.4 Normalized chain code feature ... 81

6.5 Invariant moment feature ... 84

6.6 Zernike moment feature ... 87

6.7 Discrete wavelet transform ... 90

7. CLASSIFICATION AND RESULTS ... 94


7.2 Support Vector Machine Classifier ... 95

7.3 k-NN Classifier ... 99


8. SUMMARY AND CONCLUSIONS ... 123

8.1 Conclusions ... 123

8.2 Scope for further research ... 133

PUBLICATIONS

BIBLIOGRAPHY

List of Figures

1.1 Steps in offline OCR ………………………………………………… 3

2.1 Proposed system for offline handwritten Marathi word Recognition... 19

3.1 First Group of Marathi Vowels………………………………………. 24

3.2 Second Group of Marathi Vowels…………………………………… 24

3.3 Three words contains remaining two vowels………………………... 25

3.4 Group of Marathi consonants……………………………………….. 25

3.5 Marathi vowels and consonant ……………………………………… 26

3.6 Consonants and its corresponding half consonant…………………… 27

3.7 Dataset of handwritten Marathi simple words……………………….. 29

3.8 Sample of sheets for collection of handwritten Marathi simple words 30

3.9 Sample handwritten Marathi simple words………………………….. 30

3.10 Dataset of handwritten Marathi compound words…………………… 31

3.11 Sample A4 sheet for Handwritten Compound words………………... 32

3.12 Sample handwritten Marathi compound words……………………… 32

3.13 Sample A4 sheet for isolated handwritten Marathi characters………. 33

3.14 Example of Median filtering……………………………………........ 34

4.1 Words where no ‘shirorekha’ written……………………………….. 40

4.2 Words having touching characters…………………………………… 41

4.3 Words having slanted characters…………………………………….. 41

4.4 Words having broken characters……………………………………... 41

4.5 Words having overlapping characters………………………………... 42

4.6 Marathi vowels………………………………………………………. 53

4.7 Five base characters for Marathi vowels…………………………...... 54

4.8 Marathi consonants…………………………………………………... 54

5.1 Phases in Multilevel classification…………………………………… 62

5.2 Bar character…………………………………………………………. 63

5.3 No bar character……………………………………………………… 64

5.4 Enclosed region character……………………………………………. 65

5.5 Not enclosed region character………………………………………... 65

5.6 Two component character……………………………………………. 66

5.7 One component character……………………………………………. 66

5.8 80% row contains at least one black pixels character………………... 67

5.9 less than 80% row contains at least one black pixels character……… 67

5.10 Consonants having bar and enclosed region…………………………. 68

5.11 Consonants having bar, enclosed region and having two components 68

5.12 Consonants having bar, enclosed region, one component and black

pixels…………………………………………………………………. 68

5.13 Consonants having bar, enclosed region, one component and not

black pixels…………………………………………………………... 68

5.14 Consonants does not have bar and having enclosed region………….. 68

5.15 Consonants does not have bar and enclosed region………………….. 69

6.1 Character image divided into n zones and feature value for

corresponding zone…………………………………………………... 75

6.2 Diagonal Features……………………………………………………. 78

6.3 Horizontal Features…………………………………………………... 79

6.4 Vertical Features……………………………………………………... 80

6.5 Eight directional Chain code………………………………………… 83

7.1 Hyperplanes separating two classes correctly……………………….. 96

7.2 Soft margin training allows some training examples to remain on the

wrong side of the separating hyperplane…………………………….. 98

7.3 Linear and non linear classification………………………………….. 99

7.4 Test sample for k=3 and k=5………………………………………… 100

7.5 Data flow diagram of the system…………………………………….. 103

8.1 Recognition rate (%) of Marathi handwritten Characters using SVM

and k-NN classifiers………………………………………………… 126

LIST OF TABLES

4.1 Segmentation result for handwritten Marathi simple words………… 49

4.2 Segmentation result for handwritten Marathi compound words……. 51

4.3 Marathi vowels grouped depending on their base character………… 53

4.4 Isolated full characters after applying segmentation algorithm on

handwritten Marathi words………………………………………….. 55

4.5 Half characters after applying segmentation algorithm on handwritten

Marathi words………………………………………………………… 56

4.6 Modifiers after applying segmentation algorithm on handwritten

Marathi words……………………………………………………….. 57

4.7 Segmentation result comparison with other researchers…………….. 58

5.1 Multilevel classification result from Phase I to Phase IV…………… 69

5.2 Outcome of Multilevel classification………………………………... 70

6.1 First eight order Zernike moments…………………………………... 89

7.1 Results for Subclass I using SVM Classifier………………………... 107

7.2 Results for Subclass I using k-NN Classifier………………………... 107

7.3 Results for Subclass II using SVM Classifier……………………….. 108

7.4 Results for Subclass II using k-NN Classifier………………………. 108

7.5 Results for Subclass III using SVM Classifier……………………… 109

7.5 Results for Subclass III using k-NN Classifier……………………… 109

7.7 Results for Subclass IV using SVM Classifier………………………. 110

7.8 Results for Subclass IV using k-NN Classifier……………………… 110

7.9 Results for Subclass V using SVM Classifier……………………….. 111

7.10 Results for Subclass V using k-NN Classifier………………………. 111

7.11 Results for Subclass VI using SVM Classifier………………………. 112

7.12 Results for Subclass VI using k-NN Classifier……………………… 112

7.13 Confusion Matrix for fold I Subclass I using Density and Normalized

chain code feature SVM classifier……………………………………. 113

7.14 Confusion Matrix for fold I Subclass I using Density and Normalized 113

chain code feature k-NN classifier……………………………………

7.15 Confusion Matrix for fold I Subclass II using Density and

Normalized chain code feature SVM classifier………………………. 112

7.16 Confusion Matrix for fold I Subclass II using Density and

Normalized chain code feature k-NN classifier……………………..... 114

7.17 Confusion Matrix for fold I Subclass III using Density and


7.18 Confusion Matrix for fold I Subclass III using Density and

Normalized chain code feature k-NN classifier……………………..... 114

7.19 Confusion Matrix for fold I Subclass IV using Density and


7.20 Confusion Matrix for fold I Subclass IV using Density and

Normalized chain code feature k-NN classifier……………………... 115

7.21 Confusion Matrix for fold I Subclass V using Density and

Normalized chain code feature SVM classifier……………………… 116

7.22 Confusion Matrix for fold I Subclass V using Density and


7.23 Confusion Matrix for fold I Subclass VI using Density and

Normalized chain code feature SVM classifier……………………… 117

7.24 Confusion Matrix for fold I Subclass VI using Density and


7.25 Highest recognition rate for 41 Marathi characters using SVM and k-

NN Classifier……………………………………………………… 118

7.26 Handwritten Marathi simple words Recognition using SVM

classifier………………………………………………………………. 118

7.27 Handwritten Marathi compound words recognition using SVM

classifier………………………………………………………………. 120

8.1 Comparison of recognition rates of proposed methods for subclass I. 128

8.2 Comparison of recognition rates of proposed methods for subclass II 128

8.3 Comparison of recognition rates of proposed methods for subclass

III…………………………………………………………………… 128


IV……………………………………………………………………. 129

8.5 Comparison of recognition rates of proposed methods for subclass V 129


VI…………………………………………………………………….. 129

8.7 Comparison of recognition rates for handwritten marathi characters

with other methods in literature……………………………………... 130

8.8 Comparison of recognition rates for handwritten Marathi words with

other methods in literature…………………………………………… 133

LIST OF ABBREVIATIONS

OCR Optical Character Recognition

HMWR Handwritten Marathi Word Recognition

SVM Support Vector Machine

k-NN k-Nearest Neighbor

SC Sub-class

NCC Normalized Chain Code

IM Invariant Moment

DWT Discrete Wavelet Transform

CM Confusion Matrix

RR Recognition Rate

NM Not Mentioned

PM Proposed Method

Chapter 1

Introduction

---------------------------------------------------------------------------------------------------

1.1 Optical Character Recognition (OCR) 1.1.1 Types of OCR 1.1.2 Data collection 1.1.3 Pre-processing 1.1.4 Segmentation 1.1.5 Feature extraction 1.1.6 Classification

1.2 Literature Review 1.3 Motivation for the present work, Problem Statement 1.4 Organization of Thesis

---------------------------------------------------------------------------------------------------

The thesis entitled 'A Study on Handwritten Marathi Word Recognition' presented

here is OCR for handwritten Marathi words. OCR is acronym for Optical

Character Recognition in which text images are converted into digital text

without human intervention. This technology converts read only documents into

digitized formats that can easily be retrieved, searched, and archived. Document

analysis and recognition are two challenging research areas in pattern

recognition. Although sufficient amount of research work is reported for printed

offline OCR, little research work exists for offline handwritten OCR due to the

diversified nature in handwritings. Handwritten Marathi word recognition is a

challenging task because the total number of characters present in Marathi large.

Also Marathi consists of various modifiers and different forms of compound

characters which complicate the design of OCR procedures. In this chapter, we

2 Chapter 1: Introduction

give a brief description of OCR, literature review, motivation for the present work

and problem statement.

---------------------------------------------------------------------------------------------------

1.1 Optical Character Recognition (OCR):

Optical Character Recognition (OCR) converts text images into digital

text without user intervention. Since OCR has numerous applications like postal

automation, automatic form processing, historical document preservation and

many more, OCR is an area of interest for researchers working in document

analysis and recognition. OCR can be broadly classified into two types: Online

OCR and Offline OCR.

1.1.1 Types of OCR:

Online OCR: Online OCR converts input text to digital text as it is entered on

the device. Device used for input text can be a mobile, PDA or any special

digitizer. Pen movement, strokes and pen up and downs are recorded by the

sensors which can be used for recognition purpose.

Offline OCR: Offline OCR converts printed/handwritten text images into digital

text. Printed/handwritten texts documents are scanned using a scanner and

converted into digital text so that computer understands and processes that text.

The important steps in offline OCR are shown in the Figure 1.1.

Offline OCR can be broadly classified into two types: Printed OCR and

Handwritten OCR.

Printed OCR Vs Handwritten OCR: Input for offline OCR is either printed

documents or handwritten documents. Offline OCR converts printed/handwritten

documents into digital text. Printed OCR is used to digitize historical documents,


Figure 1.1 Steps offline OCR

books and printed forms. Handwritten OCR is used to digitize handwritten

documents. Designing and developing handwritten OCR is more complicated and

challenging task than printed OCR. Printed text has specific font type and

specific size where as in handwritten text considerable variation exists as each

person has different writing styles. Also segmentation process is difficult in

handwritten OCR as compared to printed OCR. In handwritten OCR

segmentation of text into characters is complicated task and which further reduces

recognition accuracy. Major steps in offline OCR are discussed below:

1.1.2 Data Collection:

In order to develop offline handwritten OCR, database of handwritten

sample images is to be created. Database has to be large in vocabulary and

variations. There are standard databases such as CEDAR, NIST and CENPARMI

which are used for experimentation of offline handwritten OCR. But all these

databases are only for isolated English and Devanagari characters, but not for

words, that is, strings of characters.


1.1.3 Pre-processing:

Pre-processing and image enhancement operations on images are carried

out to improve the quality of image data and to remove distortions. We have to

analyze information in the image so as to improve the quality and reduce

distortions. First, image data is transformed to gray scale using Ostu’s threshold

technique then gray scale image is converted to black and white image using

binarization. Conversion from a gray-scale image to a black-and-white image

may cause some loss of information. Dilation and erosion operations can reduce

this loss of information. We can remove noise by using appropriate structuring

elements. Slant corrections can be made to improve recognition rate. Also, we

have to normalize images to a specified standard plane. Normalization is carried

out to reduce the interclass variation of the shapes of the images. To carry out

feature extraction and classification process, preprocessing and image

enhancement is to be executed correctly; otherwise it may degrade the quality of

image and important information may be lost.

1.1.4 Segmentation:

Segmentation divides an image into meaningful components called

segments. Segmentation is of two types: contextual segmentation and non

contextual segmentation. Contextual segmentation is more useful to differentiate

objects using the pixels belongs to that object. There are two types of contextual

segmentation depending on signal discontinuity and signal similarity. Cluster,

compression based methods, histograms, edge detection are widely used in

contextual segmentation. Non contextual segmentation differentiates the pixels

irrespective of their location. A simple method of non contextual segmentation is

thresholding. The accuracy of offline handwritten OCR recognition largely

depends on the success of the segmentation phase.


1.1.5 Feature extraction:

Feature extraction is an important phase in OCR which occurs prior to

classification. Recognition accuracy of OCR is largely depends on the extracted

features. In this phase unique characteristics (features) are stored in a feature

vector for all input images. Features are broadly classified into two types (i)

global features, and (ii) geometric and topological features based on their

characteristics:

Global Features:

Global features are also known as statistical features. Global features are

not affected by noise or distortions and can be detected easily. Some commonly

used global features are moments, zoning, projection, histogram, n-tuples,

crossing and distances.

Geometric and Topological Features:

Geometric and topological features may represent both global and local

properties, but are not affected by distortions or style variations. Object

components, structure of objects and their properties can be represented using

geometric and topological features. Geometric and topological representations

can be broadly grouped into four categories:

i) Topological structures like lines, curves, loops, end points, branch

points, T- point etc.

ii) Approximating geometric properties like aspect ratio, difference

between x and y coordinates etc are a kind of representation.

iii) Codings like freeman chain code, normalized chain code, regular

expressions are forms of another representation of geometric features.

iv) Graphs and trees are another type of representations, in which first

topological features are extracted and those features are represented in

graph or tree formats.


1.1.6 Classification:

Image classification assigns a label to an unknown object. Classification

is broadly categorized into two types: supervised classification and unsupervised

classification

Supervised classification:

In supervised classification training data is used where available

predefined class labels and features are used to assign labels to unknown objects.

Supervised classification is appropriate when sufficient amount of training data is

available.

Unsupervised classification:

Unsupervised classification is more appropriate when there is less

information for classification. In unsupervised classification classes or groups are

formed according to randomly sampled data called clusters and unknown objects

are classified into these clusters.

Using various decision rules, unknown objects are classified into

respective classes. Some commonly used classification techniques and decision

rules are discussed below:

Multilevel slice classifier:

Multilevel slice classifier decision rule is defined on the basis of lowest

and highest values of classes. This classifier is also known as parallelpiped

classifier is very simple and easy to understand. In this case classification

accuracy depends on the lowest and highest values of classes chosen.

Minimum distance classifier:

In minimum distance classifier unknown image is classified into a class

that minimizes the distance between the image and the class. Decision rule is


based on the distances of image from the classes. Generally, Euclidian distance

or Mahalanobis distance are used in minimum distance classifiers.

Maximum likelihood classifier:

Maximum likelihood classifier decision rule is based on the posterior

probability of a pixel belonging to the class.

Fuzzy set theory and expert system:

Fuzzy set theory uses a ‘membership function’. It is difficult to define an

appropriate membership function and boundaries of different classes for

classification. Fuzzy set theory based classifiers are useful for qualitative data.

Expert system classifiers use knowledge based on experiences.

For the present work support vector machine and k-NN classifiers are

considered.

1.2 Literature Review:

Handwriting recognition is one of the important research problems in the

field of document analysis and recognition. Document analysis and recognition is

challenging area in pattern recognition due to its varied applications. Many

systems have been proposed for recognition of printed as well as handwritten

characters, for Devanagari (Karwankar and Bhalchandra (2010); Desai and

Malik(2011); Desai et.al.(2011); Raj et.al.(2013); Holambe et.al.; Aggarwal

et.al.(2012); Malaviya et.al.(1996); Dhandra et.al.(2010); Bharath and

Madhvanath(2010); Shaw et. al.(2008); Singh et. al.(2011); Chavan et. al.(2013);

Koshti and Govilkar; Agrawal et. al.; Gohil et. al. (2012); Holambe et. al.(2010);

Rajput and Mishra; Malik and Deshpande (2009); Jangid(2011); Shukla et.

al.(2011); Mapari et. al.(2011); Sharma et. al.(2006); Garg et. al.(2011); Pratap

and Arya(2012); Sahu et. al.(2012); Murthy and Hanmandlu(2011); Ramana et.

al.(2012); Murthy and Hanmandlu(2011); Deshpande et. al. (2007;2008);


Mukherji and Rege(2008;2009); Patil and Ansari(2014); Kakde and Raut(2012);

Singh and Tyagi(2011); Ramteke(2010); Ramteke and Melhotra(2008);

Jayadevan et. al.(2011); Kapoor et. al.(2002); Kumar et. al.; Rathi et. al.(2012);

Khobragade(2013); Bajaj et. al.(2002); Arora et. al.(2008; 2009; 2011); Kamble

and Kamble ((2011); Kumar(2009; 2010); Kumar et. al. (2010;2012); Asthana et.

al.(2011); Kompalli et.al.(2009); Shelke and Apte(2011); Vaidya and

Bombade(2013); Bhattacharya and Chaudhuri(2005); Ladwani and Malik(2010);

Agnihotri(2012); Bansal and Sinha(2000); Kumar and Sengar(2010); Dongre and

Mankar(2010); Rani and Kumar(2013)), for Bengali (Sarkar and Biswas(2010);

Majumdar(2007); Das and Yasmin(2006); Shukla et. al.(2011); Parui et.

al.(2008); Bag and Harit(2013); Bhattacharya and Chaudhari(2005)), for

English (Talele et. al.(2011); Koerch et. al.(2010); Choudhary et. al. (2010);

Dhandra et. al.(2006); Romero et. al.(2007); Patel et. al.(2012); Pradeep et.

al.(2010); Hull et. al.(1990); Vaid and Gupta(2002); Prema and Reddy(2002);

Biswas and Parekh(2012); Sharma et. al.(2012); Asthana et. al.(2011)), for

Marathi (Ajmire and Warkhede (2010); Jane and Pund(2012); Mahender and

Kale(2011); Rajput and Mali(2010); Kale et. al.(2014); Tapkir and Shelke(2012);

Patil et. al.(2011); Ajmire et. Al.(2012); Jayadevan et. Al.(2011); Mali(2012);

Shelke and Apte(2010;2011); Pawar and Gaikwad(2014)), for Guajarati (Desai

(2012); Baheti et. al.(2011)), for Gurumukhi (Kumar and Jindal(2012); Singh and

Budhiraja(2012); Singh and Dhir(2012); Kumar and Sengar(2010)), for Kannada

(Dhandra et. al.(2009;2010;2011); Acharya et. al. (2008); Niranjan et. al.(2009);

Sangame et. al.(2009); Vaidya and Bombade(2013)), for Telugu (Dhandra et. Al.

(2009;2010); Jawahar et.al. (2003); Rao et.al.(2013); Rajashekararadhya and

Ranjan (2008); Asthana et.al.(2011)), for Malayalam (Chacko and Anto(2010);

Rajashekararadhya and Ranjan(2008)), for Hindi (Jawahar et.al.(2003);

Hanmandlu et.al.(2007); Garg et.al.(2010;2011;2013)), for Arabic (Chun

et.al.(2009);Abd(2007)), for Chinese (Liu et.al.(2010)), for Tamil (Aparna

et.al.(2004); Gandhi and Iyakutti(2010); Kannan and Prabhakar(2008);


Rajashekararadhya and Ranjan(2009); Asthana et.al.(2011)), for Farsi (Reza

et.al.(2011)), for Urdu (Asthana et.al.(2011)) and for Oriya (Bhattacharya and

Chaudhari(2005)). Also many systems have been proposed for numeral

recognition of different script (Holambe et.al.; Ashoka et.al.(2012); Aggarwal

et.al.(2012); Dhandra et.al.(2010); Romero et.al.(2007); Das and Yasmin(2006);

Rajput and Mali(2010)).

Pal and Chaudhari (2001) presented in their brief survey on Indian script

recognition sufficient amount of work is reported for printed and handwritten

character recognition. Also reported the status of present research and presented

scope for future work which consists of OCR for poor quality documents, multi

font OCR, multi script OCR, handwritten OCR and OCR for the visually

handicapped.

Aarti Desai et. al. (2011) proposed a system for handwritten Devanagari

character recognition. They have used minimum edit distance classifier and

combination of chain code, branch point and end point features. Using the

combination of these features is reported 87 recognition accuracy for 150

characters.

Chavan S. V. et. al. (2013) presented a system for recognition of

handwritten compound Devanagari characters. Moment base feature extraction

techniques are used to extract geometric features and Zernike moment features.

MLP and k-NN classifiers are used for classification and recognition accuracies

of 98.78% and 95.65% using MLP and k-NN classifier respectively are achieved

on a database of 27000 basic and compound characters.

Karbhari V. Kale et. al.(2014) presented a Zernike moment based feature

extraction technique for handwritten Marathi compound characters. Database of

9600 basic characters, 9000 compound characters and 3000 split characters has

been developed. Local structural classification and zone based zernike moment

features are extracted. Recognition is carried out by using SVM and k-NN


classifiers, where 98.37% recognition accuracy is achieved by SVM classifier and

95.82% accuracy by k-NN classifier.

Malik and Deshpande (2009) presented a novel approach for printed and

handwritten Devanagari characters by using regular expressions in finite state

models. Recognition accuracy reported for printed Devanagari characters is 100.

Shelke and Apte (2010, 2011) have suggested novel approach for

handwritten Devanagari compound character recognition consisting of multi-

feature and multi-classifier scheme. Database of 35000 character samples has

been developed. Structural classification, random transform, wavelet transform,

density, Euclidean distance, modified wavelet transforms are used as feature

extraction techniques. MLP and Neural network are used for classification.

Recognition accuracies reported are 94.22% when wavelet transform is used;

96.23% when modified wavelet transform is used, while for a combination of

modified wavelet transform, density and Euclidean distance gives 97.95%

recognition accuracy.

Bhattacharya and Chaudhari (2005) presented a brief survey on databases

for research on recognition of handwritten characters of Indian script. Databases

of 22556 samples of Devanagari numerals, 12938 samples of Bangala numerals,

5970 samples of Oriya numerals have been developed. Database of Devanagari

numerals is collected from 1049 users. Also 556 users have written Bangala and

Oriya numerals.

Sandhya Arora et. al. (2009, 2011) reported multiple feature and multi

classification approach for handwritten Devanagari character recognition.

Shadow features, view based features, chain code and moments are used as

features for recognition. Neural network classifier is used sequentially for

classification using multiple features. Recognition accuracy reported is 90.74%

when shadow features and chain code features are used.

Naresh Kumar Garg et. al.(2010) presented a segmentation method using

vertical and horizontal projection. Databases of 200 lines and 1380 words of


Hindi text were developed and results of 91.50% for line segmentation, 98.10%

for word segmentation, 79.12% for consonants segmentation and for modifiers

86% were reported.

Pal and Chaudhari (2001) presented a segmentation method for printed

and text line identification and the segmentation accuracy achieved 98.60%

accuracy.

Ajay Talele et. al. (2011) reported a system for handwritten legal amounts

written in English. Cavity and closed loop features are used for the recognition

purpose. They also have reported 92.50% recognition accuracy.

Alessandro L. Koerich et. al.(2013) proposed a system for verification of

unconstrained handwritten English words at character level. A database of 85092

English handwritten words is used for the experiments and recognition accuracy

is improved by 3.9%.

Bikash Shaw et. al. (2008) made significant contributions towards offline

handwritten Devanagari word recognition. They have developed a database of

39700 word samples for offline handwritten Devanagari words, consisting of 100

words. Both Holistic and segmentation based approaches are used for recognition

purpose. Chain code, 8 scaler, histogram and zone based features are are

extracted and HMM classifier is used for classification purpose. Using holistic

based approach 80.2% recognition accuracy is reported, while 81.63% recognition

accuracy is reported for segmentation based approach.

Brijmohan Singh et. al. (2011) proposed a novel approach for handwritten

Devanagari word recognition using curvelet transform. Database of 28500

samples for handwritten Devanagari word from 30 classes and database of 31860

samples for handwritten Devanagari characters from 46 classes were developed.

Curvelet transform and character geometry is used to extract features and

recognition accuracy is compared using SVM and k-NN classifier. Recognition

accuracy for words is 85.60% using SVM classifier and 93.21% using k-NN

classifier.


Gang Liu et. al. (2010) reported a novel approach for handwritten Chinese

words. Database of 44208 samples of words has been developed. Holistic

approach for recognition is used. LDA and MQDF classifiers are used for

classification purposes. Recognition accuracy reported for Chinese words is

91.96%.

R. Jayadevan et. al. (2011) presented a database and a recognition

approach for handwritten Devanagari legal amount words. A database of 26720

word samples is developed which contains all Devanagari legal amount words.

Gradient, structural features and cavity binary vector matching (BVM) is used for

recognition and achieved 80.65% recognition accuracy. A second approach using

vertical projection and dynamic time wraping (DTW) is reported with recognition

accuracy 76.69%.

Tapkir and Shelke (2012) reported OCR for handwritten Marathi script.

Projection methodology is used for line segmentation and word segmentation.

Density feature and Euclidean minimum edit distance classifier is used for

recognition. Reported result for line segmentation is 100% and for word

segmentation 98%. Recognition accuracy achieved is 92.77% for handwritten

Marathi script.

Veena Bansal and R.M.K. Sinha (2000) presented a complete Devanagari

OCR system and tested it with real-life printed documents of varying size and

font. Most of the documents used were photocopies of the originals. Recognition

accuracy reported is 90%.

Neha Avhad et. al. (2015) elaborated system for handwritten Devanagari

character recognition. The system addresses the segmentation of handwritten

Devnagari text document, the most popular script of Indian sub – continent into

lines, words and characters. They have used artificial neural network technique to

design to pre-process, segment and recognize Devanagari characters.


Priyanka Kulkarni (2015) et. al. presented brief review on Marathi and

Sanskrit word recognition using genetic algorithm. They have used dictionary

based approach and curvelet transform features are used for recognition purpose.

Kapil Bamne and Neha Sharma (2015) presented a system for offline

classifier for handwritten Devanagari script recognition. They have focused on

the recognition of offline handwritten Hindi characters that can be used in

common applications like commercial forms, bill processing systems ,bank

cheques, passport readers, offline document recognition generated by the

expanding technological society.

Snehal S.Patwardhan and R. R. Deshmukh (2015) reported a brief review

on offline handwritten recognition of Devanagari script. They have elaborated

detailed overview of different feature extraction and classification techniques for

recognition process Devanagari script by the researchers over the past few

decades.

From literature it has been observed that, due to non availability

benchmark database of handwritten words, experiments are performed on varied

number of samples. Very few experiments were performed on large databases.

Many researchers are considering holistic approach for word recognition, in

which dataset is limited. Analytical approach for word recognition is

segmentation based approach. There are many hazards in segmentation based

approach which reduces recognition accuracy. Also many characters are similar

in shape and presence of compound characters in some scripts complicates the

process of word recognition. It may be concluded that, development of

handwritten OCR is most challenging and fascinating task for researchers

working in pattern recognition.


1.3 Motivation for the present work, Problem statement:

Marathi is a well known language spoken by the people of Maharashtra. It

is written in Devanagari script which is third most widely used script in the world.

There are around 100 million speakers of Marathi language which is the fourth

largest number of native speakers in India.

Handwritten Marathi OCR has numerous applications like the reading

machines for blind and visually impaired, number plate recognition, for reading

invoices, postal automation, automated processing of bank cheque and bank

statements, digitization of 7/12 documents and ration cards, automated evaluation

of answer sheets, automated processing of admission forms and recommendation

forms.

Significant work has been reported for handwritten Devanagari/Marathi character

recognition and for printed Marathi OCR. However, handwritten Marathi word

OCR is not addressed satisfactorily in case of unconstrained handwritten Marathi

words. OCR for unconstrained handwritten Marathi word is very complex due to

many reasons as stated below:

1. Number of vowels and consonants in Marathi is large.

2. Word formation in Marathi is complex.

3. Vowels can be combined with consonants in forms.

4. Diacritic marks can be placed to the left or right or above or bottom of the

consonant.

5. Vocabulary is very large.

6. Marathi has fused characters also known as ‘Jodakshare’.

7. Number of ‘Jodakshare’s are more and are used frequently as compare to

other languages written in Devanagari script.

8. Some of the vowels and consonants are very similar in shapes and

structure.


9. Every consonant when it combined with consonant takes form of half

character.

10. Literature review shared that not much research is reported for

handwritten Marathi word recognition.

The goal of optical character recognition is to come up with a recognizer

which has best possible recognition accuracy. In this work we are designing such

type of recognizer. Hence the problem may be stated as: Given a character set

and a database of handwritten characters from the character set, design efficient

recognizer that recognizes all characters in the character set accurately.

Efficient recognizer is the recognizer which recognizes handwritten

characters using minimum number of features.

Accurate recognition can be defined as high recognition accuracy across

all handwritings.

For this problem we have chosen:

1. Character set consists of either Marathi character or Marathi words.

2. A Marathi word set is infinite since meaning of the words is not

considered in the present work.

1.4 Organization of thesis:

This thesis is organized into eight chapters.

In chapter 2, we are presenting objectives of the research work and brief

description of the proposed system to recognize handwritten Marathi word.

In chapter 3, we presented a brief description about Marathi; characters

used and the formation of Marathi words. The method of development of

database for handwritten Marathi words is elaborated. Also development of

database for handwritten Marathi simple words, compound words and isolated

characters is presented. Preprocessing techniques that were used to improve

quality of word images and to reduce noise are elaborated.


In chapter 4, the methodology for segmentation of handwritten Marathi

words was described. In this chapter we have described difficulties in

segmentation of handwritten Marathi word. Segmentation algorithms are

presented for handwritten Marathi simple and compound words and results are

compared with earlier work.

In chapter 5, we are presenting a multilevel classification technique which

categorized Marathi characters into six different groups depending upon their

special properties.

In chapter 6, feature extraction techniques are presented for handwritten

Marathi characters such as zone based symmetric density, moment invariant,

zernike moment, discrete wavelet transformations, diagonal, horizontal and

vertical features and normalized chain code. Finally we have discussed how to

create a knowledge base which contains feature vectors for every image and

corresponding class labels.

In chapter 7, classification process is described in detail. Methods used

for classification such as k-NN and SVM are described. For rigorous testing and

validation a fivefold cross validation technique is presented. A comparative study

of two the classifiers namely, k-NN and SVM is elaborated.

Finally, the chapter 8 contains summary, conclusions and future directions

of work carried out in this thesis. The results of all the methods proposed in this

thesis are compared. Further, the comparative study of proposed method and other

methods in literature is also carried out. Lastly, future directions for research

based on the present work are presented.

Chapter 2 OBJECTIVES AND PROPOSED SYSTEM

Chapter 2

Objectives and Proposed System

---------------------------------------------------------------------------------------------------

2.1 Objectives 2.2 Description of the proposed system for Handwritten Marathi

Word Recognition

---------------------------------------------------------------------------------------------------

In this chapter, we are presenting objectives of the system. Also description of the

proposed system for handwritten Marathi word recognition is elaborated.

---------------------------------------------------------------------------------------------------

2.1 Objectives:

The varied applications and challenging tasks in developing handwritten

Marathi word OCR motivated us to design and develop an efficient and robust

system for recognizing handwritten Marathi word of any length written in Marathi

by any writer. Thus, the main objectives of the thesis may be states as:

To design and develop a benchmark database for handwritten Marathi

words.

To design and develop segmentation methodology for handwritten

Marathi words.

To design appropriate and efficient feature extraction algorithms for

handwritten Marathi word recognition.

To use appropriate classification methodology so as to achieve significant


18 Chapter 2: Objectives and Proposed System

The proposed system for isolated handwritten Marathi word OCR is

discussed in the next section.

2.2 Description of the proposed system for handwritten Marathi word

Recognition:

There are two approaches for Handwritten Marathi Word Recognition

(further handwritten Marathi word recognition will be abbreviated as

HMWR): Holistic approach and analytical approach.

In holistic approach the word is considered as a single entity for

recognition. Holistic word recognition is also known as segmentation free

approach. Holistic approach of word recognition is simple and widely used if

domain of the words is limited.

In analytical approach the word is divided into their indivisible isolated

characters. Analytical approach is also known as segmentation based handwritten

word recognition. If domain of words is very large then analytical approach is

preferred.

We have adopted analytical approach for HMWR. The proposed model

for HMWR is shown in Fig. 2.1.

The process of handwritten Marathi word recognition is broadly classified

into two phases: Training phase and testing phase as shown in Fig. 1.2. Tasks in

the training phase are as follows:

Handwritten word database development: Training phase begins with

development of a reasonably large database to carry out experiments.

Database of 50000 unconstrained handwritten Marathi words is developed

and stored in the database. In addition, database of 10000 isolated

handwritten Marathi characters is developed and used in training phase. Preprocessing: Second step in handwritten Marathi word recognition is

preprocessing. Preprocessing is performed on input images to improve the


Figure 2.1 Proposed system for offline handwritten Marathi word Recognition


visual appearance and quality of image. Preprocessing operations like

image filtering, noise removal, morphological processing is performed.

Incorrect preprocessing may cause loss of information.

Segmentation: Third step in handwritten word recognition is

segmentation. Segmentation phase divides a word into meaningful

indivisible isolated characters. In handwritten word recognition system,

success of recognition largely depends on the effectiveness of

segmentation phase.

Multilevel classification: The multilevel classification technique is

developed for handwritten character categorization. In this, the character

set is divided into groups depending on specific properties of the

characters. The set of all Marathi characters are grouped into six classes

depending on their special properties.

Feature extraction: Feature is the property of input data to distinguish

objects uniquely. Feature extraction phase extracts features from input

image and is stored in a feature vector. In the proposed system features

are extracted from isolated characters as well as segmented characters and

stored in a feature vector.

Development of Knowledge base: Using the feature vectors of isolated

handwritten characters and feature vectors of segmented characters

knowledge base for handwritten Marathi word recognition is developed.

The knowledge base contains feature vectors for each image with their

class labels. The knowledge base developed in the training phase is

further for classification in the testing phase.

Tasks in testing phase are as follows:

In this phase tasks are similar to those in training phase such as

preprocessing, segmentation, multilevel classification and feature extraction as

shown in Fig. 1.2, are performed. In this phase some selected samples from the


database are used as input. Handwritten Marathi word input images are first

preprocessed, segmented, grouped into six groups using multilevel classification

and features are extracted from the segmented characters. In testing phase, the

segmented character is classified and recognized and a label is assigned to the

character. In the present work, two systems (OCR) are proposed for character

recognition based on SVM and k-NN methods respectively.

The input for the handwritten Marathi word recognition system is

handwritten Marathi word and the outcome of the system will be class labels for

every isolated character, half character and modifier present in the input word.

All the algorithms stated in this thesis are implemented in MATLAB Version 7.

In the next chapter development of database and preprocessing techniques

applied in this work are discussed.

Chapter 3 DEVELOPMENT OF A DATABASE OF

HANDWRITTEN MARATHI WORDS,

ISOLATED CHARACTERS AND

PREPROCESSING

Chapter 3

Development of a database of the handwritten Marathi

words, isolated characters and pre-processing ------------------------------------------------------------------------------------

3.1 Introduction

3.2 Marathi characters

3.3 Formation of Marathi words

3.4 Database Development 3.4.1 Database development for handwritten Marathi simple words

3.4.2 Database development for handwritten Marathi compound words

3.4.3 Database development for isolated handwritten Marathi characters

3.5 Pre-processing

------------------------------------------------------------------------------------

In this chapter, we are presenting brief description about Marathi language,

characters used in Marathi and formation of Marathi words. Also we have

presented the method for development of a database of the handwritten Marathi

isolated characters, simple words and compound words. Further pre-processing

techniques used to improve quality of word images and to reduce noise are

elaborated. Normalization is carried out for handwritten Marathi words and

isolated characters without disturbing aspect ratio.

--------------------------------------------------------------------------------------------------- Part of this has been chapter published in the Proceedings of National Conference on Challenging Research Areas in Computer Science and Information Technology - 2014, ISBN 978-93-83777-00-6.

23 Chapter 3: Development of database for handwritten Marathi words, characters and preprocessing

3.1 Introduction:

An overview of Marathi language, characters used in Marathi and various

word formations rule complicate the process handwritten Marathi word

recognition. The method of preparing the database for isolated handwritten

Marathi characters and isolated handwritten Marathi words is discussed in detail.

Pre-processing techniques helps to improve the quality and reduce the noise in the

images. Sufficient amount of work had been carried out and reported on isolated

Devanagari characters in the literature discussed below.

Bikash Shaw et. al. (2008) have reported a database of 39700 samples

using 100 classes from 436 writers for handwritten Devanagari words. Brijmohan

Singh et. al. (2011) has developed a database of 28500 word samples for 30

classes for handwritten Devanagari word from 950 writers. Laurent Guichard et.

al. (2010) have reported a database of 2000 samples for 10 classes of Devanagari

numerals written in word form from one to ten numerals. This database consists

of 10 classes and for each class 200 samples are stored. Naresh Kumar Garg et.

al. (2010, 2011, 2013) developed a database for handwritten Hindi text consisting

of 200 lines and 1380 words. R. Jayadevan et.al. (2011) developed a database of

26720 word samples for handwritten Marathi legal amounts consisting of 114

classes. G.G.Rajput et. al. (2010) used 100 blocks of handwritten Hindi script.

Rajiv Kumar et. al. [113] developed a database of 2,000 constrained and 2,000

unconstrained handwritten Devanagari words. Sandip N. Kamble et. al.(2011)

developed a database of 100 handwritten Devanagari words. Vijaya Rahul Pawar

et. al. (2014) developed a database of 3000 handwritten Marathi word.

It is observed from literature that experiments by researchers were

performed on databases various sizes ranging from 100 to 39700 having different

datasets. The method for database development and pre-processing techniques

applied are discussed in the next sections.


Marathi language: Marathi is well known language spoken by people of

Maharashtra. Marathi belongs to the Indo-Aryan group of languages. Indo-

Aryan languages are originated from Sanskrit. Currently balbodh script is used

for Marathi language and is originated from Devanagari script. Marathi has

influence of other languages like Sanskrit, Kannada and Telugu. Also lots of

words are entered into Marathi from Persian, Turkish and Arabic as well as

Portuguese and the British have influenced Marathi through their words.

3.2 Marathi Characters:

Marathi consists of a total 53 characters out of which 16 are vowels and 37

are consonants.

Marathi vowels:

The 16 Marathi vowels are classified into two groups; the first group

contains 12 vowels as shown in Fig. 3.1 while the second group contains four

vowels as shown in Fig. 3.2.

Fig. 3.1: First Group of Marathi Vowels

Fig. 3.2: Second Group of Marathi Vowels

First group of vowels are commonly used where as second group of

vowels are very rarely used. Out of four vowels of the second group two vowels

( ) have never been used in Marathi and remaining two vowels are found

only in three words called 'kLîptee', ‘RîshI’, ‘RîtU’ as shown in Fig. 3.3.


Fig. 3.3: Three words contains remaining two vowels

Since second group of Marathi vowels are not commonly used, we have

considered only first group of vowels for our study.

Marathi consonants:

The 37 consonants are broadly classified into six groups as shown in Fig. 3.4

according to their pronunciation.

Fig. 3.4 (a): First group of Marathi consonants Fig. 3.4 (b): Second group of Marathi consonants Fig. 3.4 (c): Third group of Marathi consonants Fig. 3.4 (d): Fourth group of Marathi consonants Fig. 3.4 (e): Fifth group of Marathi consonants Fig. 3.4 (f): Sixth group of Marathi consonants

Out of all 37 consonants first 25 consonants are classified into five groups

where each group contains five consonants. First group of consonants is called

‘Kantha’ because they are pronounced from the throat. Second group of

consonants is called 'Murdhanya' because they are pronounced by touching the

tongue to 'Murdhani' which is a part of the upper jaw between the roof and the

teeth. Third group of consonants is called 'Taalavya' because they are pronounced

by touching the tongue to the palate. Fourth group of consonants is called

'Dantya' because the tongue touches the teeth while pronouncing these. Fifth

group of consonants is called 'Aushthya' because they are pronounced by touching

the lips together. Sixth group consists of twelve remaining consonants which are

pronounced using combination of usage of tongue.


In Marathi out of 37 consonants, 36 consonants are commonly used but

one consonant ( ) is rarely used hence only 36 consonants are considered in this

work.

Finally the total 48 characters in Marathi are considered in this work

which consists of 12 vowels and 36 consonants as shown in Fig. 3.5.

Fig. 3.5(a) Marathi vowels

Fig. 3.5(b) Marathi Consonants

3.3 Formation of Marathi Words:

In Marathi, word formation is a very complex system because number of

vowels and consonants are more than English. Also during word formation

vowels may be combined with consonants. Whenever vowels are combined with

consonants they will take different forms called diactric marks such as ‘Kana’,

’Matra’, ’Ukar’, ’Velanti’, ’Anuswar’ or ’Visarg’.

In addition to this Marathi has a complex system of compound or fused

characters where more than one consonant are combined called ‘Jodakshare’. In

‘Jodakshare’ first consonant converted into half form and second has its full

form. Also two different words are also combined when second word is starting

with vowel.

Following Fig. 3.6 shows full consonant and its corresponding half

consonant.


Fig. 3.6(a): Group one consonants and its corresponding half consonant.

Fig. 3.6(b): Group two consonants and its corresponding half consonant.

Fig. 3.6(c): Group three consonants and its corresponding half consonant.

Fig. 3.6(d): Group four consonants and its corresponding half consonant.

According to their form half consonants are classified into 5 groups.

Group one contains those consonants are having vertical bar at the end and we get

corresponding half consonant form by removing vertical bar. Group two is

consonants not having vertical bar and we get corresponding half consonant form

by adding slanting line below the consonant. Group three contains only one

consonant with small vertical line on right-top end ( ) and we get half consonant

form by removing that small vertical line ( ). Group four contains two

consonants ( and ) having curve on the right side and formation of half

a b

c d


consonant is obtained by removing half curve. Group five contains consonants

which take multiple forms depending on the next character ( and ).

3.4 Database development:

In order to develop a system for offline handwritten Marathi word

recognition sufficient amount of database is required. Database has to contain

large vocabulary and variations. No standard database is available for Marathi.

3.4.1 Database for Handwritten Marathi simple words:

It is observed from literature that experiments by researchers were

performed on databases various sizes ranging from 100 to 39700 having different

datasets. Also literature review indicates that benchmark database for handwritten

Marathi word is not available for carrying out experiments. Since a benchmark

database is not available [32] our first attempt was to develop a database for

handwritten Marathi words.

Marathi contains two types of words such as simple words and compound

words. Simple words do not have ‘Jodakshare’ while compound words have. To

develop a database for handwritten Marathi simple words, a dataset consisting of

50 commonly used Marathi words were selected as shown in Fig. 3.7. While

selecting simple words we have taken care that all possible combinations of vowel

modifiers and consonants will appear in the words.


Fig. 3.7: Dataset of handwritten Marathi simple words

To develop a database for handwritten Marathi words, three A4 size sheets

were specially designed as shown in Fig. 3.8. These sheets were distributed to 50

writers of different age groups and professions which include students, clerks,

teachers etc. There are no constraints imposed on writers, except that they have to

write words in the given boxes. Every writer has to write a word for 10 times.

Finally, a database of 20210 handwritten Marathi simple word samples of 50

classes written by 50 different users was ready to carry out the experiments.


Fig. 3.8: Sample of sheets for collection of handwritten Marathi simple words

The handwritten data sheets were then scanned using a flat bed scanner at

a resolution of 1200 dpi and stored as gray scale images. Handwritten words

from the scanned gray scale images were manually cropped and stored in

respective class folders. The Fig. 3.9 shows some handwritten simple words in

gray scale cropped from the scanned image of a datasheet.

Fig. 3.9: Sample handwritten Marathi simple words


3.4.2 Database for handwritten Marathi compound words:

As discussed in section 2.4.1 Marathi consists of compound words also

called as fused words. Compound word contains fused characters known as

‘Jodakshare’. Occurrences of compound characters in Marathi is more frequent

(11 to 12%) as compared to other languages written in Devanagari (5 to 6%)

Shelke and Apte (2010).

To develop a database for handwritten Marathi compound words, the

dataset consisting of 47 commonly used Marathi words are selected as shown in

Fig. 3.10. While selecting the compound words all possible combinations of

vowels, modifiers and consonant clusters are considered.

Fig. 3.10: Dataset of handwritten Marathi compound words

Also A4 size sheets were specially designed to collect handwritten

Marathi word from 50 different users as shown in Fig. 3.11.


Fig. 3.11: Sample A4 sheet for Handwritten Compound words

We have adopted the same procedure for scanning, manually cropping

and storing in the respective class-folders as described in section 2.4.1. Finally, a

database of 16073 handwritten Marathi compound word samples of 47 classes

written by 50 users is ready for experiments. Sample handwritten compound

words are shown in Fig. 3.12.

Fig. 3.12: Sample handwritten Marathi compound words


3.4.3 Database for isolated handwritten Marathi characters:

We have designed A4 size sheets for data collection of Marathi

handwritten Marathi characters and distributed the sheets amongst 20. We have

adopted same procedure for scanning, cropping and storing into respective class-

folders as described in Section 3.4.1. Finally, a database of 9600 isolated

handwritten Marathi characters of 48 classes, written by 20 persons is ready for

experiments.

Fig. 3.13: Sample A4 sheet for isolated handwritten Marathi characters


3.5 Pre-processing:

Pre-processing refers to a number of operations that may be performed on

the input intensity images to obtain outputs with good quality intensity images.

The main objective of pre-processing is to remove noise from images, to enhance

quality of input images and to represent word images in standard plane.

3.5.1 Noise Removal:

Digital images contain noise due to various reasons such as movement in

acquisition process or inaccuracy in instrument for digitization. Noise removal is

an important step in preprocessing. There are several techniques for noise

removal like low-pass, high-pass, band-pass, spatial filtering, mean filtering,

median filtering. To reduce the blurring of word edges, suppress noise and

improve some features, the median filter is used. Median filter preserves edges

and removes noise. An example of the median filtering process on raw input

image is shown in Fig. 3.14.

Fig. 3.14 Example of Median filtering


3.5.2 Binarization:

Binarization operation takes input as indexed, intensity or RGB images

and outputs binary images. Here gray scale image is converted into binary image

having values 0 and 1. Gray scale image values are converted to 0 and 1

depending upon a threshold. The threshold for the gray scale image is calculated

by using histogram-shape based image thresholding suggested by Otsu. Otsu’s

method reduces interclass variance.

Ostu’s method assumes two classes of pixels (foreground and background)

in input image and calculates the optimum threshold value for separating those

two classes. The output binary image contains 0 if values of pixels in input image

are less than the calculated threshold value and 1 for all other pixels.

Ostu’s Algorithm:

Input: Nandwritten Marathi word / character images.

Output: Pre-processed handwritten Marathi word / character

Procedure:

1. Compute the normalized histogram of the input image. Denote the

components of the histogram by Pi, i=0,1,2,…L-1.

2. Compute the cumulative sums, P1(k), for k=0,1,2,…,L-1, using

퐏ퟏ(퐤) = 퐏퐢

풌

풊 ퟎ

3. Compute the cumulative means, m(k), for k=0,1,2,…,L-1, using

퐦(퐤) = 퐢 ∗ 퐏퐢

풌

풊 ퟎ

4. Compute the global intensity mean, mG, using

퐦퐆 = 퐢 ∗ 퐏퐢

푳 ퟏ

풊 ퟎ


5. Compute the between-class variance, 훔푩ퟐ(푲), for k=0,1,2,…,L-1, using

훔푩ퟐ(푲) =

[풎푮푷ퟏ(풌)− 풎(풌)]ퟐ

푷ퟏ(풌)[ퟏ − 푷ퟏ(풌)]

6. Obtain the Otsu threshold, k*, as the value of k for which 훔푩ퟐ (푲) is

maximum. If the maximum is not unique, obtain k* by averaging the

values of k corresponding to the various maxima detected.

7. Obtain the separability measure, η*, by evaluating

. 휼(푲) = 훔푩ퟐ (퐤∗)훔푮ퟐ

3.5.3 Normalization:

Handwritten words are varying in size and shape. We need to map these

word images onto a standard plane (with predefined size) so as to give a

representation of fixed dimensionality for classification. Normalization is

performed on the image to reduce the inter-class and intra-class variations of the

shapes of the words. Normalization operation facilitates segmentation process and

improves their segmentation accuracy. Linear normalization method is used to

standardize the word images. The standard plane is considered as a square of size

60 pixels x 90 pixels. The width and height ratio of the word image is not

disturbed due to normalization.

3.5.4 Thinning:

A morphological operation known as thinning, is also performed on word

images. The goal of character thinning is to remove pixels so that an object

without holes shrinks to a minimally connected stroke, and an object with holes

shrinks to a ring halfway between the hole and outer boundary. Thinning Marathi

words is very difficult task due to the presence of loops. This thinning operation

preserves Euler number. Thinning operation is related to hit-or-miss transform


and is represented as follows. Thinning of a set A by a structuring element B, is

defined as follows in terms of hit-or-miss transforms

퐴퐵 = 퐴 − (퐴 퐵) = 퐴 ∩ (퐴 퐵)

In this process we have used a sequence of structuring elements as

follows:

In the next chapter algorithms for segmentation of words into isolated

characters are described and analyzed.

},...,,,{}{ 321 nBBBBB

Chapter 4 SEGMENTATION

Chapter 4

Segmentation -------------------------------------------------------------------------------

4.1 Introduction

4.2 Segmentation and Difficulties in Segmentation

4.3 Segmentation Methodology for simple words

4.4 Segmentation methodology for compound words

4.5 Discussion of Results

4.6 Analysis of Results --------------------------------------------------------------------------------------------------- In this chapter, we have presented a brief description about segmentation of

handwritten Marathi words and difficulties in segmentation. Segmentation

algorithms are proposed for handwritten Marathi simple words and compound

words. Proposed algorithms are rigorously tested on the database developed for

this research and results are reported.

---------------------------------------------------------------------------------------------------

4.1 Introduction:

There are two approaches for handwritten text recognition. The first one is

a holistic approach which is more useful if words are limited. In this approach

features are extracted directly from word samples and classified. But, since

Marathi consists of many words, this approach is not appropriate.

The second approach is a segmentation based approach in which

handwritten Marathi words are divided into isolated indivisible characters, and

then these indivisible characters are used for classification process.

39 Chapter 4: Segmentation

The problem of segmentation of words and difficulties in segmentation are

well studied and reported in the literature. The problem of segmenting the old

typewritten Gujarati documents' is considered by Apurva desai (2012) and has

achieved 65% segmentation accuracy by using vertical projection method.

Bikash Shaw et. al. (2008) and Brijmohan Singh et. al. (2011) also reported the

use of projection method for handwritten word recognition. Dipankar Das and

Rubaiyat Yasmin [47] reported best cut method for touching Bangala numerals

and achieved 89.7% segmentation accuracy. Naresh Kumar Garg et. al. (2010)

has considered the problem of segmentation of Hindi text and reported 79.12%

segmentation accuracy by using vertical projection method. Morphological

approach for segmentation of handwritten Devanagari text is reported by Sandip

N. Kamble and Megha Kamble (2011) and achieved 52% segmentation accuracy.

Suryaprakash Kompalli et. al. (2009) reported a graph based segmentation

approach, and achieved 85% segmentation accuracy.

It has been observed from the literature that experiments of segmentation

were performed on a number of samples. However, few experiments were

performed on large databases. There are many hazards in segmentation based

approach which reduces recognition accuracy. Also presence of compound

characters complicates the process of word segmentation. Hence, we conclude

that, there is a need to address the problem of segmentation of handwritten

Marathi words.

4.2 Segmentation and difficulties in segmentation:

Segmentation is the process that decomposes the image into multiple

meaningful subparts. Text segmentation process divides written text into words

and characters. Text segmentation is a non-trivial problem because some written

languages have explicit word boundary markers, such as in written English and

the distinctive initial, medial and final letter shapes in Arabic. Such delimiters are


sometimes ambiguous and not present in all languages. Many techniques were

developed for image segmentation. These general techniques have to be

combined with domain knowledge to solve domain specific segmentation

problems. The general purpose segmentation methods are based on thresholding,

clustering, compression, histogram, edge detection, dual clustering, region

growing, partial differential equations and graphs. Segmentation of handwritten

Marathi word is a very difficult and challenging task because of various reasons

described below.

4.2.1. Shirorekha:

Marathi has a most prominent characteristic in every word called header

cap known as ‘Shirorekha’ which is written from left to right on the top of

characters in the words. Sometimes writers don’t write ‘Shirorekha’ or is broken

or slanted or is mixed with characters. Detecting the location of ‘Shirorekha’ is

an important part for segmentation process. If location of ‘Shirorekha’ is not

detected correctly segmentation of word fails which ultimately results in

misclassification of the word. Sample words without ‘Shirorekha’ and broken

‘Shirorekha’ are shown in Fig. 4.1.

Fig. 4.1: Words where no ‘shirorekha’ written

4.2.2. Touching characters:

Due to irregular handwritings it may happen that characters touch to each

other or connected to the modifiers of other characters. Touching characters

create problems during segmentation of words into isolated indivisible characters

that may lead to misclassification of words. Sample words with touching

characters are shown in Fig. 4.2.


Fig. 4.2: Words having touching characters

4.2.3. Slanting characters:

Due to different handwriting styles or style of keeping paper it may

happen characters in the words are slanted. It is very difficult to detect the

location of ‘Shirorekha’ if slanted characters or slanted ‘shirorekha’ present in the

word, which hampers segmentation process. Sample words with slanted

characters are shown in Fig. 4.3.

Fig. 4.3: Words having slanted characters

4.2.4. Broken characters:

Due to various reasons such as pen not working properly, incorrect writing

style, poor quality paper or damaged paper may result in broken characters. If

characters are broken in the word, it may cause over segmentation of character.

Over segmentation reduces the recognition accuracy of a word. Sample words

with broken characters are shown in Fig. 4.4.

Fig. 4.4: Words having broken characters


4.2.5. Overlapping characters:

It may happen that characters are overwritten due to improper writing

style or if writer is in hurry. Also modifiers are overwritten on characters. Due to

overlapping characters segmentation fails to segment isolated characters and

modifiers which lead to misclassification of words. Sample words with

overlapping characters are shown in Fig. 4.5.

Fig. 4.5: Words having overlapping characters

4.3 Segmentation methodology for simple words: An algorithm to segment simple Marathi words into characters is

described below.

Algorithm 1: Segmentation of Marathi simple words.

Input: Handwritten Marathi Word Image

Output: Segmented isolated indivisible characters.

1. Read input handwritten Marathi word image.

2. Perform pre-processing on input image.

3. Calculate horizontal projection for the word image.

4. Find out the row number which contains maximum number of white

pixel in horizontal projection label it as header_line (‘Shirorekha’).

5. Convert all white pixels to black pixels of the header_line identified in

step 4.

6. Divide word image into two parts. First part above ‘Shirorekha’

cropped from the word image labeled as top_strip of that word

contains top modifiers if any and second part labeled as core_area of

the word.


7. Calculate vertical projection for the core_area of the word image.

8. Find number of segments present in the core_area of word by using

vertical projection label the number as no_of_isolated_characters.

9. Repeat steps from 10 to 15 for the no_of_isolated_characters.

10. Skip all zeros.

11. Find out first column location contains nonzero value labeled as

starting_point.

12. Skip all nonzero numbers till zero.

13. Assign column location - 1 to the end_point.

14. Crop the word image from starting_point to end_point column

numbers mark it as isolated_character.

15. Assign remaining image to the core_area of word image.

16. Segment top modifiers in top_strip and assign modifiers to respective

segmented character in core_area using Algorithm 2.

Algorithm 2: Segmentation of modifiers and assign to characters.

Input: Top strip image identified in step 6 Algorithm 1.

Output: Segmented top_modifiers and assign to core_area word segments.

1. Calculate vertical projection for the top_strip.

2. Find number of segments present in top_strip.

3. Repeat step 4 to 6 for number of segments.

4. Find end_point of the segment.

5. If end_point of segment is greater than starting_point and less than

ending point of any segment in core_area word image assign top_strip

segment to the segment of core_area.

6. Otherwise if end_point of segment is greater than core_area segments

end_point and less than starting_point of next segment then assign

top_strip segment to first segment of core_area.


4.4 Segmentation of compound words:

Marathi consists of compound words, also known as fused words. A

compound word contains fused characters known as ‘Jodakshare’. Occurrences

of compound character in Marathi are about 11 to 12% where as in other

languages written in Devanagari is just 5 to 6% [145]. According to their form

half consonants are classified into 5 groups. Group one contains those consonants

that have vertical bar at the end for which we get corresponding half consonant

form is obtained by removing the vertical bar. Group two contains consonants

that do not have a vertical bar. In this case, corresponding half consonant form is

obtained by adding a slanting line below the consonant. Group three contains

only one consonant with small vertical line on right-top end ( ) and we get half

consonant form by removing that small vertical line ( ). Group four contains

two consonants ( and ) having curve on the right side and formation of half

consonant is obtained by removing the half curve. Group five contains

consonants take multiple forms depending on the next character ( and ).

A segmentation algorithm that depends on the groups mentioned above is

defined for compound words.

4.4.1 Segmentation methodology for compound words:

An algorithm to segment compound Marathi words into characters is

described below.

Algorithm 3: Compound word segmentation

Input: Handwritten Marathi Word Image

Output: Segmented isolated indivisible characters.

1. Read a handwritten Marathi word image as input.

2. Perform pre-processing on input image.

3. Find horizontal projection of the word image and label it as


horizontal_projection.

4. Find row number from horizontal_projection which contains maximum

value and label it as header_line. Header_line is also known as

‘Shirorekha’ in Marathi which contains maximum number of white

pixel.

5. Convert all white pixels to black pixels of the header_line identified in

step 4.

6. Divide the input word image into two parts depending on header_line.

First part is above header_line cropped from the input word image

labeled as top_modifier of that word which contains top modifiers if

any. Second part is below header_line cropped from input word image

labeled as core_area of the word.

7. Calculate vertical projection for the core_area of the word image

labeled as ca_vertical_projection.

8. Find number of segments present in the ca_vertical_projection label it

as number_of_characters.

9. Repeat steps from I to VII for the number_of_characters.

I. Scan ca_vertical_projection and skip all zeros.

II. Find out first column location contains nonzero value labeled it as

starting_point.

III. Skip all nonzero numbers till zero in ca_vertical_projection.

IV. Assign column location - 1 to the end_point.

V. Crop the input image from column number starting_point to column

number end_point and label it as char.

VI. To check whether compound character is present or not in char call

Algorithm 4.

VII. Mark the char who contains lower modifier based on threshold_height

of character. Crop lower_modifier from core_area of the char and

assign lower_modifier to respective char using Algorithm 5.


VIII.Assign remaining image of core_area to core_area.

10. Segment top_modifier and assign modifiers to respective segmented

character in core_area using Algorithm 6.

Algorithm 4: Check for compound character present or not and set cut point.

Input: Character image from step 10(vi) of Algorithm 3.

Output: Isolated indivisible characters.

1. Check if compound character is present or not.

if height(char) > width(char)

then

no compound character present;

return char image;

exit;

else

goto step 2;

end;

2. Find vertical_projection of char. Labled it as vp_char.

3. Check for vertical bar present or not.

for i=1 : 1 : number_of_columns

if vp_char(i) >= 42

then

vertical_bar=vertical_bar+1;

end;

end;

4. Set cut point for compound character segmentation.

if vertical_bar >= 2

then

difference=location_second_bar - location_first_bar;


cut_point = location_first_bar + (difference * 0.25);

else

if lower modifier present

then

cut_point=(start_point+end_point)/2;

else

difference=end_point – start_point;

cut_point=start_point + diff *0.40;

end;

end;

half_char=crop(char,start_point,cut_point);

full_char=crop(char,cut_point,end_point);

5. Return (half_char, full_char)

Algorithm 5: Separate lower modifier from character.

Prerequisite: There are three possible relationships between character and lower

modifiers if present.

A. Weak joining: Lower modifier below a middle bar or end bar character has weak

joining. Also some non bar character having small vertical bar also forms weak

joining with lower modifier.

B. Thick Joining: lower modifiers may be connected to some characters at more than

one location.

C. Gap: Sometimes lower modifiers are not joined to characters.

Input: Suspected character that contains lower_modifier.

Output: Segment lower modifier and assign to character.

1. Calculate horizontal projection for the character image.

2. If C relationship is present then find the row which contains no white

pixel, use that row to separate lower modifiers from the character

image.

3. If A relationship then find the row which contains minimum number of


pixels. Check if height (lower_modifier) >= (threshold_height * 0.20)

then crop char image from located row to bottom boundary. Otherwise

lower modifier is not present.

4. If B relationship then find row which contains minimum number of

pixels assigns to min_row_num below threshold_height, then set

cut_point = (min_row_num - threshold_height)/2. Crop char image

from cut_point to bottom boundary of character.

Algorithm 6: Segment Top_modifier and assign to respective character.

Input: Top strip image identified in step 6 Algorithm 1.

Output: Segmented top modifiers and assign to core_area word segments.

1. Calculate vertical projection for the top_strip.

2. Find number of segments present in top_strip.

3. Repeat step 4 to 6 for number of segments.

4. Find end point of the segment.

5. If end point of segment is greater than starting_point and less than

end_point of any segment in core_area word image, assign top_strip

segment to the segment of core_area.

6. Otherwise if end_point of segment is greater than core_area segments

end_point and less than starting_point of next segment then assign

top_strip segment to first segment of core area.

4.5 Discussion of results:

Segmentation algorithms proposed for simple words and compound words

are tested on the database. Database contains 20210 Marathi simple words and

16000 Marathi compound words. Segmentation results for simple words are

elaborated in Table 4.1 and for compound words are elaborated in Table 4.2.


A. Simple words:

Table 4.1: Segmentation result for handwritten Marathi simple words

SR. NO. WORD SAMPLES RESULT IN

PERCENTAGE 1 324 79.01 2 324 90.43 3 357 84.31 4 335 70.45 5 373 92.49

6 384 80.47

7

373 93.83

8 357 87.68 9 363 82.92 10 384 95.31

11 373 88.47

12 457 94.09

13 477 77.78

14

346 80.06

15 477 91.61

16

384 85.16

17

371 85.71

18 477 81.34 19 477 87.00

20 457 88.84 21 406 77.34 22 384 63.28

23

374 58.56

24 384 78.13

25 488 88.93


26 357 77.87 27 466 88.84 28 384 69.27

29 467 85.65

30 358 65.64

31 360 86.11 32 477 76.52 33 368 73.64

34 372 83.06

35

384 89.58

36 396 67.17

37 395 82.28

38

371 80.32

39

364 77.75

40 477 86.37 41 477 87.63

42

357 70.87

43

357 68.07

44 477 91.40

45 368 81.52 46 477 84.91

47

479 88.73

48 455 87.69

49

384 86.72

50 477 87.84

Total word

Samples 20210 82.17


B. Compound Words:

Table 4.2: Segmentation result for handwritten Marathi compound words

SR. NO. WORD SAMPLES RESULT IN

PERCENTAGE 1

324 83.64

2

270 78.52

3

570 85.09

4

300 83.67

5

322 83.23

6

321 85.98

7

270 82.22

8

300 77.33

9

440 83.41

10

322 79.50

11

311 83.60

12

321 82.55

13

420 85.48

14

760 82.11

15

300 79.33

16

271 81.55

17

440 83.41

18

270 83.33

19

271 80.07

20

320 75.00

21

270 80.37


22

270 82.96

23

430 83.26

24

270 82.59

25

321 78.50

26

272 82.72

27

300 74.33

28

321 80.06

29

270 80.37

30

430 85.12

31

430 81.86

32

272 81.62

33

440 86.14

34

320 81.88

35

322 81.99

36

743 83.98

37

271 82.66

38

320 78.44

39

320 82.81

40

270 81.85

41

321 76.01

42

322 79.50

43

275 81.45

44

250 85.60

45

300 84.00

46

320 82.50


47

300 83.00

Total 16073 81.80

The outcome of segmentation process is isolated characters, top modifiers,

bottom modifiers and half characters. Marathi has 12 vowels as shown in Fig.

4.6.

Fig. 4.6: Marathi vowels

After segmentation above 12 vowels are grouped in to five groups based

on their base character as given below.

Table 4.3 : Marathi vowels grouped depending on their base character Group

No. Base

Character Vowels in the group

1.

2.

3.

4.

5.

Finally 12 Marathi vowels are based on only five base characters.

Therefore after segmentation process we are considering only five base characters


instead of 12 vowels. These five base characters for Marathi vowels are shown in

Fig. 4.7.

Fig. 4.7: Five base characters for Marathi vowels

Also in addition to vowels Marathi consists of 36 consonants as shown in

Fig. 4.8.

Fig. 4.8: 36 Marathi consonants

Finally we have total 41 isolated characters after segmentation process,

which includes five base characters of vowels and 36 consonants. Table 4.4

shows isolated characters and number of samples we get after applying

segmentation algorithms on handwritten Marathi words. Table 4.5 shows half

characters and number of samples we get after applying segmentation algorithm

on handwritten Marathi words. Table 4.6 shows modifiers and number of

samples we get after applying segmentation algorithm on handwritten Marathi

words. Section 4.7 gives comparative study of segmentation reported by other

researchers.


Table 4.4 : Isolated full characters after applying segmentation algorithm on handwritten

Marathi words

Sr. No. Symbol Number of samples

1 689

2 1009

3 802

4 1792

5 414

6 429

7 592

8 3025

9 760

10 1880

11 2127

12 1591

13 410

14 1103

15 777

16 3305

17 1489

18 4027


19 3949

20 4501

21 3149

22 2334

23 249

24 4537

25 484

26 888

27 362

28 3203

29 6584

30 1358

31

249

32 400

33 367

Total 58835

Table 4.5: Half characters after applying segmentation algorithm on handwritten Marathi

words


1 1681

2 464


3 484

4 492

5 493

6 1318

7 475

8 623

9 1356

10 1015

11 1001

12 2027

13 939

14 438

15 356

16 443

17 223

Total 13828

Table 4.6: Modifiers after applying segmentation algorithm on handwritten Marathi words


1

11978

2 9403

3

933

4 464

5 2292

6 497


7 605

8 1356

9 626

Total 28154

4.6 Analysis of Results:

Segmentation algorithms for handwritten simple and compound words

proposed in this chapter are tested on the databases of simple and compound

Table 4.7: Segmentation result comparison with other researchers

PN Author Language Method Result in Per Samples

[1] Apurva A. Desai

Gujrathi Vertical Projection

65% NA

[2] Bikash Shaw et al

Devanagari Morphology opening

NA NA

[6] Dipankar Das

Bangla Best cut (touching char)

89.7% NA

[9] Naresh Kumar Garg

Hindi Vertical Projection

79.12% 1380

[14] Sandip N.Kamble et. al.

Devanagari Morphology 52% 100

[15] Suryaprakash Kompalli

Devanagari Graph 85% NA

PM Proposed method

Marathi Plain words

Statistical information and Vertical Projection

82.17% 20210

PM Proposed method

Marathi Compound words

Statistical information and Vertical Projection

81.80% 16073

*PM = Proposed method


words. It is observed from literature that database is quite large to carry out

segmentation results. The segmentation results are encouraging for simple as well

as compound handwritten Marathi words. Comparison of segmentation

methodology and results with other researchers is shown in Table 4.7.

In the next chapter we are elaborating novel multilevel classification

approach which groups 41 Marathi characters into six groups.

Chapter 5 MULTILEVEL CLASSIFICATION

Chapter 5

Multilevel classification

------------------------------------------------------------------------------------

5.1 Introduction

5.2 Multilevel Classification


------------------------------------------------------------------------------------ In this chapter, a multilevel classification approach is described for handwritten

Marathi character recognition. In this method we have divided the character set

into six groups depending on special properties of the characters. This process of

classification is carried out in four phases.

------------------------------------------------------------------------------------

5.1 Introduction:

Marathi characters have more interclass and intra-class similarities. By

experiment it has been observed that single feature is not sufficient for

classification because recognition rate using single feature is very low. Also

classifying and recognizing 41 characters is time consuming task. In multilevel

classification approach Marathi characters are sub-classified into six groups

depending on their special properties such as presence of bar, presence of

enclosed region, presence of one component etc. Marathi 41 characters are sub-

classified into six subclasses using four phases which takes tree structure and the

61 Chapter 5: Multilevel Classification

tree has four levels. Since the approach of sub-classification has four levels we

are calling it as multilevel classification.

Chavan S. V. et.al (2013) has reported pre-classification approach for

handwritten Devanagari character recognition based on location of vertical bar

and number of components present in the character. Kale K. V. et.al (2014) has

reported local structural sub-classification for Marathi compound character

recognition. M. Hanmandlu et. al.(2007) has reported coarse classification

approach for handwritten Hindi characters using presence of vertical bar, location

of vertical bar and character is open to right or left side. Sushma Shelke and

Shaila Apte (2011) has reported structural classification approach for handwritten

Marathi compound character recognition.

It has been observed from literature that using single feature for

classification of large number of classes is difficult. In the next section 4.2 we are

describing multilevel classification approach where total numbers of Marathi

characters are sub-classified into six subclasses using their special properties.

5.2 Multilevel classification:

The outcome of segmentation process is 41 isolated full characters, half

characters, top modifiers and lower modifiers. These 41 isolated characters are

further classified into six sub-classes in phase I to phase IV as shown in the Fig.

5.1.


Fig. 5.1: Phases in Multilevel classification

Following is the discussion on phase I to phase IV sub-classification

where 41 characters are classified into six classes:


5.2.1 Phase I Sub-classification:

Phase I sub-classification is based on bar characters and no bar characters.

Marathi consonants are broadly classified into two major categories bar characters

and no bar characters. Bar characters are those characters having presence of

vertical bar. In order to verify whether a bar is present in the character, vertical

projection of image was taken and if any column contains more than 70% of black

pixels, then label it as bar character otherwise label as no bar character. As shown

in the Fig. 5.2 column number 57 contains 86% black pixels hence labeled as bar

character and as shown in the Fig. 5.3 all columns contains less than 70% black

pixels hence labeled as no bar character. Thus 41 characters are classified into

two subclasses, first contains 28 bar characters and second contains 13 no bar

characters as shown in phase I of Fig 5.1.

Fig. 5.2: Bar character


Fig. 5.3: No bar character

5.2.2 Phase II Sub-classification:

Phase II sub-classification is based on presence of enclosed region. In

phase II, 28 bar characters are broadly classified into two major categories having

enclosed region or not. To verify whether enclosed region is present, the numbers

of holes are counted in the character using eight connectivity as shown in the Fig.

5.4. If one or more than one enclosed region exists in the character then label it as

enclosed region character otherwise not enclosed region character as shown in

Fig. 5.5. Now 28 bar characters are sub-classified into 18 enclosed region bar

characters and 10 not enclosed region bar characters as shown in phase II of Fig.

5.1.

Similarly 13 no bar characters are classified into six enclosed region no

bar characters and seven not enclosed region no bar characters as shown in phase

II of Fig. 5.1.


Fig. 5.4: Enclosed region character

Fig. 5.5: Not enclosed region character

5.2.3 Phase III Sub-classification:

Phase III sub-classification is based on number of components present in

character. In phase III, 18 bar enclosed region characters are classified into two

subclasses, depending on whether number of component is one or more as shown

in Fig. 5.6 and Fig. 5.7. Presence of a component can be verified using region

properties for each labeled region.


In all 18 bar enclosed region characters are sub-classified into 14

characters having a component and four characters having two components as

shown in phase III of fig. 5.1.

Fig. 5.6: Two component character.

Fig. 5.7: One component character

5.2.4 Phase IV Sub-classification:

Phase IV sub-classification is based on number of rows containing at least

one black pixel. We have 14 bar enclosed region characters having one


component and further classified into two subclasses, depending on whether

character's 80% rows contains at least one black pixel in first 75% columns as

shown in phase IV of Fig. 5.1. Out of 14 bar enclosed region characters having

one component we get 10 characters satisfying above condition as shown in Fig.

5.8 and four characters does not satisfy the condition as shown in Fig. 5.9.

Fig. 5.8: 80% row contains at least one black pixels character

Fig. 5.9: less than 80% row contains at least one black pixels character.

Using above sub-classification method problem of handwritten Marathi

character recognition is simplified into six sub-classes as follows:


Sub-class I: Bar not enclosed region (10 characters)

Fig. 5.10: Consonants having bar and enclosed region

Sub-class II: Bar enclosed region with two components (4 characters)

Fig. 5.11: Consonants having bar, enclosed region and having two components.

Sub-class III: Bar enclosed region with one component and having 80% rows

contains at least one black pixel in first 75% columns (10 characters)

Fig. 5.12: Consonants having bar, enclosed region, one component and black pixels.

Sub-class IV: Bar enclosed region with one component and less than 80% rows

contains at least one black pixel in first 75% columns (4 characters)

Fig. 5.13: Consonants having bar, enclosed region, one component and not black pixels.

Sub-class V: No bar enclosed region (6 characters)

Fig. 5.14: Consonants does not have bar and having enclosed region.


Sub-class VI: No bar not enclosed region (7 characters).

Fig. 5.15: Consonants does not have bar and enclosed region.

The problem of classification of 41 isolated Marathi characters is now

simplified to small problems. Total 41 Marathi characters are divided into 6

different sub-classes as discussed above. Feature extraction methods suitable for

different sub-classes are discussed in the next chapter 6.

5.3 Discussion of Results:

Table 5.1: Multilevel classification result from Phase I to Phase IV

Sr. No. Level Phase Special Property Result in Percentage

1. I Phase I Presence of BAR 100%

2. II Phase II Presence of Enclosed Region 100%

3. III Phase III Presence of Number of components 100%

4. IV Phase IV Presence of at least one black pixel in 80% rows 100%


Table 5.2: Outcome of Multilevel classification

Sr. No. Subclass Characters

1. Subclass I

2. Subclass II

3. Subclass III

4. Subclass IV

5. Subclass V

6. Subclass VI

Experimental results for multilevel classification are shown in Table 5.1.

We got 100% accuracy for experimental results for all four levels. The outcome

of multilevel classification approach described in this chapter is shown in Table

5.2. After application of multilevel classification approach total 41 handwritten

Marathi characters are sub-classified into six subclasses in four phases. Subclass

I contains 10 characters, Subclass II contains four characters, Subclass III

contains 10 characters, Subclass IV contains four characters, Subclass V contains

six characters and Subclass VI contains seven characters. Further on these

subclasses various feature extraction techniques will be applied for classification

purpose.

In the next chapter various feature extraction techniques and algorithms to

extract features are elaborated.

Chapter 6 FEATURE EXTRACTION

Chapter 6

Feature Extraction -------------------------------------------------------------------------------

6.1 Introduction

6.2 Zone based symmetric density feature

6.3 Diagonal, Horizontal and Vertical features

6.4 Normalized chain code feature

6.5 Invariant moment feature

6.6 Zernike moment feature

6.7 Discrete wavelet transform

------------------------------------------------------------------------------------ In this chapter different feature extraction techniques are elaborated. The feature

extraction techniques are based on density, chain code, invariant moment,

Zernike moments and wavelet transform. Also we have presented different

algorithms to extract these features.

------------------------------------------------------------------------------------

6.1 Introduction

Feature extraction is important phase in OCR prior to classification. A

feature is a unique property that can describe image. The main objective of

feature extraction is to reduce the size of image and represent the image object

effectively in terms of a compact feature vector. Feature extraction takes image Part of this chapter has been published in the International Journal of Computer Applications (0975 – 8887) Volume 108 – No. 4, December 2014, ISSN 0975-8887. International Journal of Engineering Research & Technology (IJERT), Vol. 3 Issue 11, November-2014, ISSN: 2278-0181.

72 Chapter 6: Feature Extraction

as input, builds initial data and finally gives feature values which are non-

redundant and informative. Recognition accuracy of OCR largely depends on

features extracted in this phase. In this phase unique characteristics (features) of

an image are stored into feature vector for all input images which are further used

for recognition purpose.

Assigning handwritten Marathi character to predefined classes is very

difficult and challenging task due to interclass and intra-class similarities.

Sufficient amount of work is reported for isolated handwritten Devanagari

character recognition. Various feature extraction techniques like zone based

symmetric density, zone based diagonal, horizontal, vertical, normalized chain

code, moment invariant, Zernike moment and discrete wavelet transform are

reported. The major advantage zone based symmetric density, diagonal,

horizontal and vertical feature approach is that it is robust to small variations, easy

to implement and yields relatively high recognition rate. Many authors have

presented zoning mechanisms or regional decomposition methods to investigate

the recognition rates of patterns. Normalized chain code has several advantages

like it has compact representation and also translation invariant. Moment invariant

and Zernike moments are very important features they are rotation invariant,

independent of variability involved in the writing style of different individuals

and also thinning free. In the next sections we will elaborate all the feature

extraction techniques used in the present work.

6.2 Zone based symmetric density feature:

In this feature extraction technique hybrid zone based symmetric density

features are extracted. For correct classification of handwritten characters

suitable features should be extracted which are invariant with respect to shape.

The objective of this hybrid approach came from its robustness to small variation,

easy implementation and promising recognition accuracy. Zone based feature


extraction method gives good recognition accuracy even when certain

preprocessing steps like filtering, smoothing and slant corrections are not

performed. In this section, we elaborate on this feature extraction technique and

the algorithm.

6.2.1 Review of earlier work:

Ashoka H. N. et. al.(2012) reported zone based feature extraction methods

for handwritten numeral recognition and achieved 100% recognition accuracy.

B.V. Dhandra et. al. (2011) reported zone based density feature for recognition of

handwritten and printed Kannada and English numerals, and reported recognition

accuracy for Kannada numerals 95.25% and for English numerals 97.05%. B.V.

Dhandra and M. Hangarge [26] reported density and density ratio features as one

of the features for identification of script at word level. B.V. Dhandra et. al.

(2009, 2010) reported direction density estimation feature for Kannada, Telugu

and Devanagari numeral recognition and achieved 99.40% recognition accuracy

for Kannada numerals, 99.60% recognition accuracy for Telugu numerals and

98.40% recognition accuracy for Devanagari numerals. B.V.Dhandra et. al.

(2011) reported directional density feature for Kannada numerals and achieved

98.04% recognition accuracy. Dinesh Acharya U. et. al. (2008) reported direction

code frequency for horizontal and vertical blocks and achieved 92.68%

recognition accuracy for printed Kannada characters. Mahesh Jangid (2011)

reported pixel density and zone density features for Devanagari character

recognition and achieved 94.89% recognition accuracy using SVM classifier.

Vinaya Tapkir and Sushma Shelke (2012) reported pixel density feature for four

zones and achieved a recognition accuracy of 92.77% for handwritten Marathi

script. O.V. Ramana Murthy and M. Hanmandlu (2011) reported pixel density

feature for recognition of Devanagari character recognition.


It is observed from literature that for handwritten character recognition,

density feature is largely used by researchers and obtained significant recognition

accuracies. Structural features reflect the character’s structure information.

Statistical feature is the most relevant information extracted from the raw data,

which minimizes the inter-class distance and maximizes the between-class

distance. Density statistical feature is commonly used for character recognition.

Character’s structure feature method has a strong adaptability of character font

changes, so it can easily differentiate between similar characters, but its

computational complexity is large and its ability of anti-interference is bad.

Character’s statistical feature has advantage of anti-interference and simple

algorithm of classification and matching.

Hence, we have chosen zone based symmetric density feature for

handwritten character recognition. The feature extraction method that was used to

extract features and an algorithm is described in the following sections.

6.2.2 Feature extraction method:

To extract zone based symmetric density feature, the binary image

representing the handwritten character is pre-processed and is normalized to a size

of 60 x 60 pixels. The size-normalized image is divided into n equal zones. The

input image is divided into 4, 9, 16, 25 and 36 equal zones. For 4 equal zones,

one zone has 30 x 30 pixels; for 9 equal zones, one zone has 20 x 20 pixels; for 16

equal zones, one zone has 15 x 15 pixels; for 25 equal zones, one zone has 12 x

12 pixels and for 36 equal zones, one zone has 10 x 10 pixels. Therefore, features

are identified for n=4, 9, 16, 25 and 36 equal zones and they were stored in

feature vector for each image.

The density of each zone is computed by taking the ratio of total number

of object pixels to total number of pixels in that zone. This is carried out for every

zone in the image. Finally, 90 features are extracted from the image and feature


vector stores these 90 features. Zone based symmetric density features were

calculated for n=4, 9, 16, 25, 36 are shown in Fig. 6.1 using Equation (1).

Density(Z) = h h

------------------------------------(1)

Fig. 6.1: Character image divided into n zones and feature value for corresponding zone

6.2.3. Algorithm:

Algorithm: Zone based symmetric density feature extraction algorithm. Input: Gray scale character image

Output: Feature vector of size 90.

1. Pre-process the input image and resize it to 60 x 60 standard plane.

2. Divide the input image into four equal zones; calculate the density of

each zone that will give four features as shown in Fig. 6.1(a) & (b).


3. Divide the input image into nine equal zones; calculate the density of

each zone that will give nine features as shown in Fig. 6.1(c) & (d).

4. Divide the input image into 16 equal zones; calculate the density of each

zone that will give 16 features as shown in Fig. 6.1(e) & (f).


zone that will give 25 features as shown in Fig. 6.1(g) & (h).


zone that will give 36 features as shown in Fig. 6.1(i) & (j).

7. Store all features extracted in Step 2, 3, 4, 5 and 6 in feature vector.

Finally feature vector containing 90 features for each image is ready for

experimentation.

6.3 Diagonal, Horizontal and Vertical Features:


J. Pradeep et.al.(2010) reported diagonal feature extraction method for

handwritten character recognition and reported 99% recognition accuracy using

69 features. Om Prakash Sharma et. al. (2012) reported zone based diagonal

features for handwritten Devanagari alphabets and obtained 98.50% recognition

accuracy. It is observed that diagonal, horizontal and vertical features are having

quite encouraging recognition results.

Zone based diagonal, horizontal and vertical features are statistical

features and gives most relevant information from the raw data, which minimize

the inner-class distance and maximize the between-class distance. Diagonal,

horizontal and vertical feature methods have a strong adaptability of character

font changes, so it can easily differentiate the similar characters and its


computational complexity is large. Character’s statistical feature has advantage

of anti-interference and simple algorithm of classification and matching.

Hence we have decided to use a combination of diagonal, horizontal and

vertical features. Feature extraction technique used to extract these features and

algorithm is given in the next section.


To extract diagonal, horizontal and vertical features the binary image

representing the handwritten character is pre-processed and is normalized to a size

of 50 x 50 pixels. The size-normalized image is divided into 25 equal zones where

one zone has 10 x 10 pixels. The procedure to find diagonal, horizontal and

vertical features is described below.

6.3.2.1 Diagonal Features:

To extract diagonal features from the binary image representing the

handwritten character is preprocessed and is normalized to a size of 50 x 50

pixels. The size-normalized image is divided into 25 equal zones each of size is

10 x 10 pixels as shown in Fig. 6.2(a). Each zone has 19 diagonal lines, each

diagonal line is summed to get a single sub-feature and thus 19 sub-features are

obtained from the each zone as shown in Fig. 6.2(b).

These 19 sub-features values are averaged to form a single feature value

and placed in the corresponding zone. This procedure is sequentially repeated for

the all the zones as shown in Fig. 6.2(c).

Finally, 25 features are extracted for each character. In addition, 10

features are obtained by averaging the values placed in zones row-wise and

column-wise, respectively. As a result; every character is represented by 25+10

features, that is, 35 features.


Fig. 6.2: Diagonal Features

6.3.2.2 Horizontal Features:

To extract horizontal feature of the binary image representing the

handwritten character is first preprocessed and is normalized to size of 50 x 50

pixels. The size-normalized image is divided into 25 equal zones, each zone is of

size 10 x 10 as shown in Fig. 6.3(a). Each zone has 10 horizontal lines, each

horizontal line is summed to get a single sub feature and thus 10 sub-features are



and assigned as horizontal feature to the corresponding zone. This procedure is

sequentially repeated for the all the zones. Finally, 25 features are extracted for 25

zones for each character as shown in Fig. 6.3(c). In addition, 10 features are

obtained by averaging the values placed in zones row-wise and column-wise

respectively. Finally every character is represented by 35 features, that is 25+10

features.


Fig. 6.3: Horizontal Features

6.3.2.3 Vertical Features

To extract vertical features from the binary image representing the

handwritten character is preprocessed and is normalized to a size of 50 x 50

pixels. The size-normalized image is divided into 25 equal zones; each zone of

size is 10 x 10 as shown in Fig. 6.4(a). Each zone has 10 vertical lines, each

vertical line is summed to get a single sub-feature and thus 10 sub-features are



and placed in the corresponding zone. This procedure is sequentially repeated for

the all the zones as shown in Fig. 6.4(c). Finally, 25 features are extracted for

each character. As a result every character is represented by 25 features. In

addition, 10 features are obtained by averaging the values placed in zones row-

wise and column-wise, respectively. Finally every character is represented by 35

features, that is 25+10 features.


Fig. 6.4: Vertical Features

6.3.3 Algorithm:

Algorithm: Diagonal, Horizontal and Vertical feature extraction algorithm

Input: Gray scale character Image

Output: Diagonal, Horizontal and Vertical Features.

1. Pre-process the input Image and resize to 50 x 50 standard plane.

2. Divide the input image into 25 zones, each zone is of size 10 x 10 pixels.

3. Calculate diagonal feature value for each zone, repeat the process to find

diagonal features for 25 zones.

4. Calculate average values for row-wise diagonal features and column-

wise diagonal features, in all five row-wise and five column-wise values.

5. Feature vector of 35 diagonal features is prepared for each image.

6. Calculate horizontal feature value for each zone, repeat the process to

find horizontal features for 25 zones.

7. Calculate average values for row-wise horizontal features and column-

wise horizontal features, in all five row-wise and five column-wise

values.


8. Feature vector of 35 horizontal features is prepared for each image.

9. Calculate vertical feature value for each zone, repeat the process to find

vertical features for 25 zones.

10. Calculate average values for row-wise vertical features and column-wise

vertical features, in all five row-wise and five column-wise values.

11. Feature vector of 35 vertical features is prepared for each image.

6.4 Normalized Chain Code:

Chain codes are the features which represents the boundary of a character.

There are several advantages of using normalized chain code feature extraction

listed below:

1. Compact representation of a character.

2. Feature values are not affected by translation of character.


Aarti Desai et.al. (2011) reported chain code features for Devanagari

character recognition. They have divided a character image into 25 blocks and for

each block 8 chain code features are extracted, finally they have used 200 chain

code features for recognition and achieved 87% recognition accuracy. Bikash

Shaw et.al. (2008) reported directional chain code feature for handwritten

Devanagari word recognition and achieved 80.2% recognition accuracy. G.G.

Rajput and S.M. Mali (2010) reported freeman chain code features in combination

with fourier descriptor for handwritten Marathi numeral recognition and achieved

98.1% recognition accuracy. Gunvantsinh Gohil et.al. (2012) reported chain code

and holistic features for printed Devanagari script and achieved 66.35% and

80.55% using ANN and SVM classifier respectively. N. Sharma et.al. (2006)


reported directional chain code feature extraction for 49 zones for handwritten

Devanagari characters and obtained 98.86% and 80.36% on Devanagari numerals

and characters respectively. Ravi Sheth et.al. (2011) reported normalized chain

code feature extraction technique for handwritten English character recognition

and obtained 92% recognition accuracy. S. Arora et.al. (2011) reported chain

code feature extraction method in combination with shadow and view based

features and obtained 98.61% recognition accuracy for handwritten Devanagari

characters. S. Arora et.al. [122] reported zone based chain code histogram feature

in combination with shadow features for recognition of non-compound

handwritten Devanagari character and obtained 90.74% recognition accuracy.

It is observed from literature review that chain code directional features

are having quite encouraging recognition results. Hence we have decided to use

normalized chain code features for recognition of handwritten characters. Feature

extraction technique used to extract these features and algorithm is elaborated in

next sections.


To extract freeman chain codes first locate any boundary pixel, called as

starting pixel, and then move along the boundary of character either clockwise or

anticlockwise direction, find out next boundary pixel and allocate this new pixel

a number depending upon its direction from the previous pixel is called code for

that pixel. The process is repeated till starting pixel is not encountered. The codes

may be 4-directional or 8-directional depending upon 4-connectivity or 8-

connectivity of a pixel to its neighboring contour pixel. An 8-directional chain

coded image is given in Fig. 6.5.


Fig. 6.5: Eight directional Chain code

The chain code extracted from above process is different for different

characters as length of each chain code depends on the size of the handwritten

characters.

Example shows Chain code extracted for the image shown in Fig. 5.5.

Chain code: [0 7 6 6 6 0 6 4 3 4 5 4 2 2 2 0 2 0 2]

V1= [0 7 6 6 6 0 6 4 3 4 5 4 2 2 2 0 2 0 2]

Compute the frequency of the codes 0, 1, 2, ….., 7. For vector V1 we have the

frequency vector V2 as below.

V2= [4 0 5 1 3 1 3 1]

The normalized frequency, represented by vector V3, is computed using the

formula

V3 =| |

, where |V1|=ΣV2

For the example considered above, we have

V3= [0.22 0 0.27 0.05 0.16 0.05 0.16 0.05]

Finally, V3 is the required feature vector of size 8.

6.4.3 Algorithm: Algorithm: Normalized chain code feature extraction algorithm

Input: Gray scale character image

Output: Normalized chain code feature vector for each image.



2. Extract the boundary of the character image.

3. Resample the boundary in order to obtain a uniform resampling along the running arc length of the boundary.

4. Trace the boundary in counterclockwise direction and generate 8 directional chain codes 0 to 7.

5. Compute the frequency of the codes 0 to 7.

6. Divide frequency of each code by sum of the frequencies.

7. Store eight features in feature vector.

8. Finally feature vector of 8 features is ready for each input image.

6.5 Moment Invariant

Moment invariant features are based on statistical moments of characters.

They are traditional and widely-used tool for character recognition. Classical

moment invariants were introduced by Hu (1962) and they were successfully used

in numerous applications not only for character recognition. Hu invariants are

invariant under translation, rotation and scaling. Moment invariants features are

extracted for the image which contributes to improve the overall recognition

accuracy.


Ajmire and Warkhede (2010) reported seven moment invariant features

for handwritten Marathi vowel recognition. They have computed mean and

standard deviation for each feature and these 14 features were used for

recognition using Gaussian distribution function. S.V.Chavan et.al.(2013)

reported geometric and Zernike moments for handwritten Devanagari compound

character recognition and achieved 98.78% recognition accuracy using MLP


classifier and 95.56% recognition accuracy using k-NN classifier. Nilima Patil

et.al.(2011) reported moment invariant and affine moment invariant for

handwritten Marathi vowel recognition and obtained 75% recognition accuracy.

R. J. Ramteke (2010) reported invariant moment based feature extraction

technique for handwritten Devanagari vowels recognition using 3 different feature

sets by dividing image into four or two zones. R. J. Ramteke and S. C. Mehrotra

(2008) reported invariant moment based feature extraction technique for

handwritten Devanagari numerals recognition using 3 different feature sets by

dividing image into four or two zones and achieved 92% recognition accuracy

using 78 features. Reena Bajaj et.al.(2002) reported density and moment feature

extraction technique for Devanagari numeral recognition and obtained 63.4%

recognition accuracy for Devanagari numerals. S. Arora et.al.(2009) reported

chain code histogram and moment based features for handwritten Devanagari

character recognition and reported 98.03% recognition accuracy. S. M. Mali

(2012) reported moment and density features for handwritten Marathi numeral

recognition and reported 97.69% recognition accuracy.

It is observed from literature review that moment invariant features are

having quite encouraging recognition results in case of handwritten characters.

Hence we have decided to use moment invariant features for recognition of

handwritten characters. Feature extraction technique used to extract these

features and algorithm is elaborated in next sections.

6.5.2 Feature extraction Technique:

The method to calculate invariant moment is described below:

The two Dimensional moment of order (pique) of image is calculated as

follows

푚 = 푥 푦 푓(푥, 푦)


Where p=0, 1,2,… and q=0,1,2….

Using 2D moments, central moment of order (p+q) can be calculated as follows

µ = (푥 − 푥̅) (푦 − 푦) 푓(푥, 푦)

For p=0,1,2,… and q=0,1,2…. Where 푥̅ = 푎푛푑 푦 =

Normalized central moments can be derived by using above central moments as

follows

휂 =µµ

Where

훾 =푝 + 푞

2 + 1

For p+q=2,3…

Set of seven invariant moments can be derived from second and third moments

ø = 휂 + 휂

ø = (휂 − 휂 ) + 4휂

ø = (휂 − 3휂 ) + (3휂 − 휂 )

ø = (휂 + 휂 ) + (휂 + 휂 )

ø = (휂 − 3휂 )(휂 + 휂 )[(휂 + 휂 ) − 3(휂 + 휂 ) ]

+ (3휂 − 휂 )(휂 + 휂 )[3(휂 + 휂 ) − (휂 + 휂 ) ]

ø = (휂 − 휂 )[(휂 + 휂 ) − (휂 + 휂 ) ] + 4휂 (휂 + 휂 )(휂 + 휂 ) ]

ø = (3휂 − 휂 )(휂 + 휂 )[(휂 + 휂 ) − 3(휂 + 휂 ) ]

+ (3휂 − 휂 )(휂 + 휂 )[3(휂 + 휂 ) − (휂 + 휂 ) ]


6.5.3 Algorithm to compute moment invariant features: Algorithm: Moment Invariant feature extraction algorithm.

Input: Gray scale character image.

Output: Moment invariant features for each image.


2. Compute seven moment invariant feature for the whole image and store into feature vector.

3. Divide character image into four equal zones, each zone of size 25 x 25 pixels.

4. Compute moment invariant feature for each zone. Add these 28 features into feature vector.

5. Feature vector of size 35 is ready for each image.

6.6 Zernike Moment:


K. V. Kale et. al. (2014) reported Zernike moment feature extraction

technique for handwritten Marathi compound character recognition. They had

extraction zone based first 8 order Zernike moments and achieved 98.37% and

95.82% recognition accuracy using SVM and k-NN classifier.

The Zernike moment were first proposed in 1934 by Zernike. Zernike

moments are complex numbers by which an image is mapped on to a set of two-

dimensional complex Zernike polynomials. The magnitude of Zernike moments is

used as a rotation invariant feature to represent a character image pattern. Zernike

moments are a class of orthogonal moments and have been shown effective in

terms of image representation. The orthogonal property of Zernike polynomials

enables the contribution of each moment to be unique and independent of


information in an image. A Zernike moment does the mapping of an image onto a

set of complex Zernike polynomials. These Zernike polynomials are orthogonal to

each other and have characteristics to represent data with no redundancy and able

to handle overlapping of information between the moments. Due to these

characteristics, Zernike moments have been utilized as feature sets in applications

such as pattern recognition and content-based image retrieval. These specific

aspects and properties of Zernike moment are supposed to found to extract the

features of handwritten characters. Feature extraction technique and algorithm to

extract Zernike moments is elaborated in next sections.


The Zernike moments introduce a set of complex polynomials which form

a complete orthogonal set over the interior of a unit circle, i.e., x2 + y2 ≤ 1.

Zernike moments are the projection of the image function on some orthogonal

basis functions. Let the set of these basis functions be denoted by Vn,m(x, y).

These polynomials are defined by Vn,m(x, y) = Vn,m(ρ, θ) = Rn,m(ρ)ejmρ (1)

where n is a non-negative integer, m is a non-zero integer subject to the following

constrain: n − |m| is even and |m| < n. Also, ρ is the length of the vector from

origin to the (x, y) pixel, θ is the angle between vector ρ and x axis in a counter-

clockwise direction, and Rn,m(ρ) is the Zernike radial polynomial. The Zernike

radial polynomials, Rn,m(ρ), are defined as :

Rn, m(ρ) =(−1) (n− s)!

s! n + |m|2 − s ! n − |m|

2 − s ! ρ

| |

Note that Rn,m(ρ) = Rn,−m(ρ). The Zernike moment of order n with repetition m

for a digital image is


푍 , =n + 1π 푓(푥, 푦)푉 ,

∗ (푥, 푦)∆x∆y

where V∗n,m(x, y) is the complex conjugate of Vn,m(x, y).

To compute the Zernike moments of a given image, the image center of mass is

taken as the origin.

Table 6.1: First eight order Zernike moments

Order Dimension Zernike moment

0 1 Z0,0

1 2 Z1,1

2 4 Z2,0,Z2,2

3 6 Z3,1,Z3,3

4 9 Z4,0,Z4,2,Z4,4

5 12 Z5,1,Z5,3,Z5,5

6 16 Z6,0,Z6,2,Z6,4,Z6,6

7 20 Z7,1,Z7,3,Z7,5,Z7,7

6.6.3 Algorithm:

Algorithm: Zernike moment feature extraction algorithm.


Output: First eight orders of Zernike moment features for each image.


2. Compute first eight orders Zernike moment feature for the whole image and store into feature vector.

3. Divide character image into four equal zones, each zone of size 25 x 25 pixels.


4. Compute first eight orders Zernike moment feature for each zone and append into feature vector.

5. Feature vector containing first eight orders Zernike moment is ready for experiments.

6.7 Discrete Wavelet Transform:


The Discrete Wavelet Transform (DWT) provides a decomposition of an

image into details having different resolutions and orientations; it is a bijection

from the image space onto the space of its coefficients. It has been mainly used

for image compression. Diego J. Romero et.al. (2007) has reported directional

continuous wavelet transformed for recognition of handwritten numerals.

G.G.Rajput and Anita H. B. (2010) has reported discrete cosine transform and

discrete wavelet transform for handwritten script recognition. Pritpal Singh and

Sumit Budhiraja (2012) has reported wavelet transformation for handwritten

Gurumukhi character recognition. Sushama Shelke and Shaila Apted (2010) has

reported discrete wavelet transform for the recognition of handwritten Marathi

compound character.


Discrete wavelet transforms (DWT) are applied to discrete data sets and

produce discrete outputs. Transforming signals and data vectors by DWT is

a process that resembles the fast Fourier transform (FFT), the Fourier method

applied to a set of discrete measurements. Discrete wavelet transforms map data

from the time domain (the original or input data vector) to the wavelet

domain. The result is a vector of the same size. Wavelet transforms are linear and

they can be defined by matrices of dimension if they are applied to inputs


of size . Depending on boundary conditions, such matrices can be either

orthogonal or ''close'' to orthogonal. When the matrix is orthogonal, the

corresponding transform is a rotation in in which the data (a -typle) is

a point in . The coordinates of the point in the rotated space comprise the

discrete wavelet transform of the original coordinates. The discrete wavelet

transform (DWT) has a large number of applications in computer science. It is

used for signal coding, to represent a discrete signal in a more redundant form,

often as a preconditioning for data compression. Practical applications can also be

found in signal processing of accelerations for gait analysis, in digital

communications and many others.

It is shown that discrete wavelet transform (DWT) is discrete in scale and

shift, and continuous in time. DWT is successfully implemented as analog filter

bank in biomedical signal processing for design of low-power pacemakers and

also in ultra-wideband (UWB) wireless communications. Wavelets are localized

basis functions which are translated and dilated versions of some fixed mother

wavelet. The decomposition of the image into different frequency bands is

obtained by successive low-pass and high-pass filtering of the signal and down-

sampling the coefficients after each filtering. Here various discrete wavelet

transforms Daubechies is used.

DWT Single-level discrete 1-D wavelet transform. Single-level one-

dimensional wavelet decomposition with respect to Daubechies wavelet transform

is used. It performs a multilevel one-dimensional wavelet analysis using

Daubechies wavelet and returns the wavelet decomposition of the signal. The

output decomposition structure contains the wavelet decomposition vector C and

the bookkeeping vector L. Compute the approximation coefficients using the

wavelet decomposition structure [C, L].


The wavelet transform exhibits the features like separability, scalability,

translatability, orthogonality and multiresolution capability. The discrete wavelet

transform of an image f(x, y) of size MxN is

푊 (푗 ,푚,푛) =1

√푚푛푓(푥, 푦)휑 , , (푥, 푦)

푊 (푗,푚,푛) =1

√푚푛푓(푥, 푦) 훹 , , (푥,푦)

Where,

휑 , , = 2 휑 2 푥 −푚, 2 푦 − 푛

And,

훹 , , = 2 훹 2 푥 −푚, 2 푦 − 푛

are the two dimensional scaling and wavelet functions respectively and the index i

identifies the directional wavelets that takes the values H, V and D i.e. horizontal,

vertical and diagonal details respectively. j0 an arbitrary starting scale and the

푊 (푗 ,푚, 푛) coefficients define an approximation of f(x, y) at scale j0. The

푊 (푗,푚, 푛) coefficients add horizontal, vertical and diagonal details for scales j≥

j0. Normally, j0 = 0 N = M = 2J so that j = 0, 1, 2…, J-1 and m, n = 0, 1, 2,…, 2j-

1. The discrete wavelet transform can be implemented using digital filters and

down samplers. The high pass or detail component characterizes the image’s

high-frequency information with vertical orientation; the low-pass, approximation

component contains its low-frequency, vertical information. Both sub images are

then filtered column wise and down sampled to yield four quarter size output

images.


6.7.3 Algorithm:

Algorithm: Discrete wavelet transforms feature extraction algorithm.


Output: Eight discrete wavelet transform features for each image.

1. Pre-process the input Image and resize to 50 x 50 standard planes.

2. Number of black pixels along each row of the binarized image has been

counted to form a 50 sized vector.

3. The 1D discrete wavelet transform on row count vector at level 3 using

Daubechies db1 wavelet has been applied.

4. Compute approximation coefficients and add to these four values to

feature vector.

5. Number of black pixels along each column of the binarized image has

been counted to form a 50 sized vector.

6. The 1D discrete wavelet transform on column count vector at level 3

using Daubechies-db1 wavelet has been applied.

7. Compute approximation coefficients and add to these four values to

feature vector.

8. Feature vector containing eight discrete wavelet transformations is ready

for experiments.

All these feature extraction methods are used to extract the features and

the extracted features are further used for classification. In the next chapter SVM

and k-NN classifiers are discussed. Also results are presented for SVM and k-NN

classifier.

Chapter 7 CLASSIFICATION AND RESULTS

Chapter 7

Classification and Results ---------------------------------------------------------------------------------

7.1 Introduction

7.2 Support Vector Machine Classifier

7.3 k-NN Classifier


--------------------------------------------------------------------------------- In this chapter classification is elaborated. SVM and k-NN classifiers are

discussed and used for handwritten Marathi words recognition. Fivefold cross

validation technique is used to compute the results. Recognition rates for

handwritten Marathi simple and compound words are reported using SVM

classifier using combination of density and normalized chain code features. Also

recognition accuracy for isolated handwritten Marathi character is reported

using SVM and k-NN classifier. Recognition accuracy using SVM and k-NN

classifier is compared also results are compared with other researchers.

------------------------------------------------------------------------------------

7.1 Introduction:

Classification is a process in OCR that groups the individual items

depending on the similarity of item and groups properties. Different types of item

are distinguished in classification process. Image classification assigns label to

the unknown object. Classification broadly categorized into two types:

supervised classification and unsupervised classification

95 Chapter 7: Classification and Results

Supervised classification:

Supervised classification first applies knowledge and then classifies. In

supervised classification we use training data where predefined class labels and

features are available which are used to assign labels to unknown objects.

Supervised classification is useful when sufficient amount of training data is

available.

Unsupervised classification:

Unsupervised classification process first classifies and then applies

knowledge. Unsupervised classification is more useful where there is less

information is available for classification. In unsupervised classification classes

or groups are formed according to randomly sampled data called clusters and

unknown objects are classified into that clusters. Using various decision rules,

unknown objects are classified to respective class.

In the present work two supervised classifiers are used for classification

purpose namely support Vector Machine (SVM) and k-Nearest Neighbor (k-NN).

We are elaborating SVM and k-NN classifiers in the next sections.

7.2 Support vector machines:

Support vector machines are introduced in COLT-92 by Boser, Guyon &

Vapnik. Support vector machines are supervised classification method used for

classification. SVM has successful applications in the fields of bioinformatics,

text, image recognition etc. SVM is effective in high dimensional spaces; also

effective even if number of dimensions is greater than the number of samples. In

SVM different kernel functions can be specified for decision functions, also they

are memory efficient. SVM creates a hyper plane or set of hyper planes for

classification and uses it for classification. SVM correct classification can be


achieved by hyper plane that has the largest distance to the nearest training

features of any class called functional margin shown in Fig. 7.1.

Support vector machines are frequently used for classification in statistical

pattern recognition. Classification into separable two classes can be achieve by

maximizing the distance between two classes. The distance between two classes is

defined as the discrimination hyper-surface in n-dimensional feature space. The

closest training patterns are called as support vectors. The advantage of this

approach is that it identifies the optimal discriminating hyper-surface between two

classes when many such hyper-surface exist. Basically support vector machine

classifier was developed first for linear separation of two classes. This drawback

was overcome by introducing nonlinearly separable classes, nonseparable classes,

combining multiple 2-class classifiers which results into multi-class classification,

and other extensions.

Fig. 7.1: Hyperplanes separating two classes correctly

In support vector machines a data point is a ρ dimensional vector and is

separated in to ρ-1 dimension hyperplane called as linear classifier. Decision

hyperplane is b and decision hyperplane normal vector is perpendicular to

hyperplane is also called as weight vector. Since hyperplane is perpendicular


to normal vector, all the points on hyperplane → will satisfy 푤̇⃗ 푥⃗ = −푏̇ . Training

data set is 픻 = {(푥⃗ , 푦 )}, where 푥⃗ is a pair of points on hyperplane and 푦 are

class labels.

Then linear classifier푓(푥⃗) = 푠푖푔푛(푤̇⃗ 푥⃗ + 푏̇ ), returns -1 if 푓(푥⃗) < 0 and

+1 if 푓(푥⃗) ≥ 0 for different classes. Functional margin of 푥⃗ with respect to

hyperplane < 푤̇⃗ , b > is 푤̇⃗ 푥⃗ + 푏̇ . Functional margin can be increased by scaling

푤̇⃗ and 푏.

The discrimination function in terms of support vectors and multipliers is

given below:

푓(푥) = 훼 휔 (푥є

.푥) + 푏

Each of the Lagrange multipliers 훼 shares a corresponding training vector푥 .

Those vectors that contribute to the maximized margin have non zero 훼 are the

support vectors. Since remaining training vectors do not contribute to the final

discrimination function, summation is performed only for support vectors푥 .

The decision whether test vector belongs to which class +1 or -1 totally

depends on the support vectors associated with the maximum margin as identified

in the training phase. Hence the discrimination hyperplane in the feature space

can be obtained from the vectors in the input space and the dot products in the

feature space. Also training can be based on a small set of support vectors, even

in large training sets, thus limiting the computational complexity of the training

with explicitly represented feature vectors.

If a separating hyperplane cannot be found to partition the feature space

into two classes that is linear inseparability of the training patterns. Support vector

machines trains classifier in such non-separable sets, soft margin training allows

some training examples to remain on the wrong side of the separating hyperplane.

Support vector classifier splits these two class patterns as accurate as possible

with minimum number of patterns on the wrong side as shown in Fig. 7.2. Hence


the equation is modified to allow some wrong-class patterns to remain within the

margin as follows: 휔(푤.푥 + 푏) ≥ 1 − 흃

Fig. 7.2: Soft margin training allows some training examples to remain on the wrong side of

the separating hyperplane

The minimization of ||w|| can be achieved using Langrange multipliers, or

by setting a dual optimization problem to eliminate ξ. The kernel trick facilitated

the extension to non-linearly separable problems. To determine the similarity of

two patterns xi and xj in a linear space, a kernel function k(xi, xj) may be

determined by calculating the dot product k(xi, xj) = (xi.xj). The dot product of

the linear support vector classifier can be replaced with non-linear kernel

functions k(xi, xj ) = φ(xi) . φ( xj). Due to the kernel trick, the support vector

classifier can locate the linearly separating hyperplane in the transformed space

by defining an appropriate kernel function.

A number of simple kernels include

Homogeneous dth order polynomials, k(xi, xj) = (xi. xj)d

Non-homogeneous dth order polynomials, k(xi, xj) = (xi. xj + 1)d

Radial basis functions, k(xi, xj) = exp ( - γ || xi – xj || 2 )

Gaussian radial basis functions, k(xi, xj) = exp || 퐱퐢 – 퐱퐣 || ퟐ , etc


Fig. 7.3: Linear and non linear classification

Multiple class classification can be achieved by combining N 2-class classifiers,

where each classifier will discriminate between a specific class and the rest of the

training set. During the classification stage, a pattern is assigned to the class with

the largest positive distance between the classified pattern and the individual

separating hyperplane for the N binary classifiers. Algorithm for support vector

machine learning and classification is given below.

1. Training:

1. Select an appropriate kernel function, k(xi, xj).

2. Minimize ||w|| subject to the constraint.

3. Store only the non zero αi‘s and the corresponding training vectors xi.

These are the support vectors.

2. Testing/Classification:

1. For each pattern x compute the discrimination function, using the

support vectors xi and the corresponding weights αi. The sign of the

function determines the classification of x.

7.3 K Nearest Neighbor classifier:

k-Nearest Neighbor (k-NN) algorithm is very simple to understand and has

numerous applications. k-NN is non parametric lazy algorithm, does not make

any assumptions on the data distribution. In k-NN algorithm no explicit training


phase or is minimum, but testing phase is costly in terms of time and memory. K-

NN algorithm can be used for classification, where if x is unlabeled data item then

find data closest to x if it is y then assign the label of y to x using nearest neighbor

algorithm. K-NN looks up for its k nearest points (for k an integer number) and

then label the new point according to which set contains the majority of its k

neighbours. The best value for k totally depends on the data. Large value

of k may reduce the effect of noise, but it creates very short boundaries between

distinct classes. A good k can be selected by various heuristic techniques. In case

where k = 1 is called the nearest neighbor algorithm.

Fig. 7.4: Test sample for k=3 and k=5

For example, in the above Fig. 7.4 there is a test sample shown by green

circle has to be classified into first class of blue squares or to the second class of

red triangles. Consider k = 3 indicated by solid line circle then unlabeled circle is

assigned to the second class because there are 2 triangles and only 1 square inside

the inner circle. If k = 5 shown by dashed line circle then unlabeled circle is

assigned to the first class because 3 squares vs. 2 triangles inside the outer circle.

An alternative to using the Euclidean distance are the city-block distance,

Minkowski distance, and the weighted Euclidean distance and the weighted city-

block distance. The k-nearest neighbor classifiers performs well when there are a

lot of training patterns. However, the more training patterns there are, the more


distance have to be calculated, and consequently the computation required

increases, thus slowing down the process of classification. But because of their

fairly high classification performance, they serve as good benchmarks for

evaluating other classifiers.

Algorithm:

1. Training set: (x1,y1), (x2,y2), …, (xn,yn)

2. Assume xi = (xi1, xi

2, …, xid) is a d-dimensional feature vector of real

numbers, for all i.

3. yi is a class label in {1…C}, for all i

4. Determine ynew for xnew

5. Find k closest training points to xnew w.r.t. Euclidean distance between x

and y defined as:

푑 = |푥 − 푦 |

= (푥 − 푦)(푥 − 푦)

6. Classify by yknn = majority vote among the k points.

7.4 Discussion of Results:

The input for handwritten Marathi word recognition system is handwritten

Marathi words and Marathi isolated characters. We have collected 9600 images

of isolated Marathi characters and 36283 handwritten Marathi words from

different users. In Marathi isolated characters we are considering five base

characters of vowels. Thus we obtained 8200 images for isolated Marathi

characters. The outcome of segmentation process is 58835 isolated Marathi

characters, along with 13828 half characters and 28154 modifiers. After

combining 8200 Marathi isolated characters and 58835 segmented isolated


characters, a database of 67035 images for 41 handwritten Marathi characters is

ready for experiments.

Now this database is fed to multilevel classification, where these 41

characters are grouped into six subclasses as discussed in chapter 5. The outcome

of multilevel classification is, sub-class I consist of 10 characters with 23621

images; sub-class II consist of four characters with 7648 images; sub-class III

consist of 10 characters with 11464 images; sub-class IV consist of four

characters with 8719 images; sub-class V consist of 6 characters with 6225

images and sub-class VI consist of 7 characters with 9358 images.

OCR for handwritten Marathi word recognition system is shown in

Fig.7.5.

The multilevel classification groups 41 Marathi characters into six sub-

classes. Now suitable feature extraction techniques discussed in chapter 5 are

used to extract features for all sub-classes. We found promising recognition

results for following features:

1. Zone based symmetric density (90 features)

2. Diagonal features (35 features)

3. Horizontal features (35 features)

4. Vertical features (35 features)

5. Normalized chain code (8 Features)

6. Invariant moment (35 features)

7. Zernike moment (16 features)

8. Discrete wavelet transformation (8 features)

9. Zone based symmetric density and Normalized chain code (98

features)

10. Diagonal and Zernike moment(51 features)

11. Horizontal and normalized chain code (43 features)

12. Horizontal and invariant moment (42 features)

13. Horizontal and Zernike moment (51 features)


Fig. 7.5 : Data flow Diagram of the system


After extracting features for each sub-class separately we have used SVM

and k-NN classifiers for recognition. The experiments were performed on the

67035 samples of handwritten Marathi characters. We have used 5-fold cross

validation technique. In this method, five test subsets were created. Each subset

contains disjoint image samples. The 5-fold cross-validation involves the

determination of classification accuracy for multiple partitions of the input

samples used in training. The 5-fold cross-validation partitions available data into

5 sets. In each run 4 sets are used for training and the remaining 5th set is used for

testing. Finally, an average accuracy over 5 runs is obtained.

The experiments were performed on the sub-class I contains 10 characters

and 23621 image samples. Table 7.1 and Table 7.2 show recognition results for

sub-class I using SVM and k-NN classifier respectively. The highest average

recognition rate 91.85% achieved for combination of features symmetric density

and normalized chain code using SVM classifier. The highest average


and normalized chain code using k-NN classifier.

The experiments were performed on the sub-class II contains four

characters and 7648 image samples. Table 7.3 and Table 7.4 show recognition

results for sub-class II using SVM and k-NN classifier respectively. The highest

average recognition rate 94.72% achieved for combination of features symmetric

density and normalized chain code using SVM classifier. The highest average



The experiments were performed on the sub-class III contains 10


results for sub-class III using SVM and k-NN classifier respectively. The highest






The experiments were performed on the sub-class IV contains four


results for sub-class IV using SVM and k-NN classifier respectively. The highest

average recognition rate 88.62% was achieved when symmetric density and

normalized chain code features were combined and SVM classifier was used.

The highest average recognition rate 82.62% achieved for combination of features

symmetric density and normalized chain code using k-NN classifier.

The experiments were performed on the sub-class V contains six


results for sub-class V using SVM and k-NN classifier respectively. The highest





The experiments were performed on the sub-class VI contains seven


results for sub-class VI using SVM and k-NN classifier respectively. The highest






Table 7.1: Results for Subclass I using SVM Classifier:

Sr. No. Feature No. of

Features

SUBCLASS I

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 91.00 92.25 92.50 91.50 92.00 91.85

2. HORIZONTAL (35) + NCC(8) 43 87.50 87.50 88.75 87.75 87.50 87.80

3. HORIZONTAL (35) + ZERNIKE

(16) 51 87.50 87.25 88.75 88.00 87.25 87.75

4. HORIZONTAL (35) 35 86.25 85.75 86.75 85.25 86.50 86.10

5. DENSITY(90) 90 84.00 81.25 84.75 83.50 83.75 83.45

Table 7.2: Results for Subclass I using k-NN Classifier:


Features

SUBCLASS I

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 87.50 90.25 89.75 88.50 89.50 89.10

2. DENSITY(90) 90 84.50 86.00 88.75 84.75 85.00 85.80

3. HORIZONTAL(35) + ZERNIKE 98 84.25 82.75 83.50 81.00 82.75 82.85

4. HORIZONTAL(35) + NCC(8) 98 84.25 83.25 83.50 80.50 82.75 82.85

5. HORIZONTAL (35) 35 82.00 81.00 80.75 79.75 81.50 81.00


Table 7.3: Results for Subclass II using SVM Classifier:

Sr. No

. Feature No. of

Features

SUBCLASS II

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 95.63 93.12 94.25 96.25 94.37 94.72


(16) 51 85.00 86.87 92.50 91.87 84.37 88.12

3. HORIZONTAL (35)+NCC(8) 43 85.00 87.50 92.50 90.62 84.37 87.99

4. DENSITY(90) 90 88.13 78.75 87.50 88.13 88.13 86.12

5. HORIZONTAL (35)+IM(7) 42 83.12 84.37 88.13 90.62 80.62 85.37

Table 7.4: Results for Subclass II using k-NN Classifier:

Sr. No.

Feature No .of Features

SUBCLASS II

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 88.75 86.25 91.25 93.75 90.00 90.00

2. DENSITY(90) 90 89.37 83.12 88.75 92.50 87.50 88.24

3. HORIZONTAL (35)+NCC(8) 43 83.12 79.38 83.75 88.13 81.88 83.25


(16) 51 82.50 79.38 84.37 88.13 80.62 83.00

5. HORIZONTAL (35)+IM(7) 42 83.75 80.00 83.75 86.87 80.62 82.99


Table 7.5: Results for Subclass III using SVM Classifier:

Sr. No

. Feature No. of

Features

SUBCLASS III

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 82.00 83.75 83.00 84.50 83.25 83.30


(16) 51 76.75 79.00 76.75 77.40 77.75 77.53

3. HORIZONTAL (35)+NCC(8) 43 77.00 79.00 74.75 77.50 77.5 77.15

4. HORIZONTAL (35)+IM(7) 42 74.75 74.75 76.00 77.25 74.75 75.50

5. HORIZONTAL (35) 35 73.25 76.50 72.75 76.25 73.5 74.45

Table 7.6: Results for Subclass III using k-NN Classifier:

Sr. No

. Feature No .of

Features

SUBCLASS III

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 75.75 78.75 77.75 78.00 73.5 76.75

2. DENSITY(90) 90 71.75 76.75 76.00 72.75 74.75 74.4

3. HORIZONTAL (35)+NCC(8) 43 70.75 72.25 72.25 73.75 70.25 71.85


(16) 51 70.50 72.25 72.25 73.75 70.50 71.85

5. HORIZONTAL (35)+IM(7) 42 70.50 69.25 69.25 74.50 69.25 70.55


Table 7.7: Results for Subclass IV using SVM Classifier:

Sr.

No.

Feature No. of Features

SUBCLASS IV

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 90.00 91.87 86.25 89.37 85.63 88.62

2. HORIZONTAL (35) 35 87.50 90.00 88.75 87.50 86.87 88.12

3. HORIZONTAL (35)+NCC(8) 43 83.75 90.62 88.13 87.50 87.50 87.50


(16) 51 83.12 90.00 88.13 87.50 87.50 87.25

5. DIAGONAL(35) 35 83.75 82.50 83.12 82.50 75.00 81.37

Table 7.8: Results for Subclass IV using k-NN Classifier:

Sr.

No.


SUBCLASS IV

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 85 82.5 85 87.5 83.12 84.62

2. DENSITY(90) 90 81.88 89.37 82.5 83.75 79.38 83.37

3. HORIZONTAL (35)+NCC(8) 43 80.62 85 82.5 84.37 81.25 82.74


(16) 51 80.62 84.37 82.5 83.75 82.5 82.74

5. HORIZONTAL (35) 35 80.62 82.5 81.62 80 79.38 80.82


Table 7.9: Results for Subclass V using SVM Classifier:


Features

SUBCLASS V

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 92.50 92.92 91.25 95.42 92.08 92.83

2. HORIZONTAL (35)+NCC(8) 43 88.75 87.92 88.33 89.58 90.00 88.91


(16) 51 88.37 87.92 88.33 89.58 90.00 88.84

4. HORIZONTAL (35) 35 85.42 89.17 86.77 87.50 86.25 87.02

5. DENSITY(90) 90 83.33 80.83 82.5 89.17 88.33 84.83

Table 7.10: Results for Subclass V using k-NN Classifier:

Sr. No

. Feature No. of

Features

SUBCLASS V

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 86.67 90.83 86.67 95.00 89.58 89.75

2. DENSITY(90) 90 85.00 87.92 87.80 91.77 86.25 87.74

3. HORIZONTAL(35)+NCC(8) 43 83.75 83.32 83.33 86.77 87.50 84.93

4. HORIZONTAL(3

5) + ZERNIKE (16)

51 83.33 83.33 83.33 87.08 87.08 84.83

5. HORIZONTAL(35) 35 83.75 85.00 84.17 81.77 82.98 83.53


Table 7.11: Results for Subclass VI using SVM Classifier:

Sr. No.

Feature No .of Features

SUBCLASS VI

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 92.14 94.64 95.71 94.29 94.29 94.73

2. HORIZONTAL (35) 35 91.07 87.50 89.29 97.21 91.79 91.37

3 HORIZONTAL (35)+NCC(8) 43 91.07 89.29 91.43 91.43 93.57 91.35


(16) 51 91.43 88.93 91.43 91.07 93.57 91.28

5. DENSITY(90) 90 82.07 85.71 86.79 86.07 86.79 85.48

Table 7.12: Results for Subclass VI using k-NN Classifier:

Sr. No.


SUBCLASS VI

FOLD I

FOLD II

FOLD III

FOLD IV

FOLD V

AVERAGE in %

1. DENSITY(90) + NCC(8) 98 92.50 93.21 94.29 92.86 93.57 93.28

2. DENSITY(90) 90 81.50 92.50 92.14 88.93 91.43 89.30


(16) 51 87.86 85.71 91.07 91.07 87.50 88.64

4. HORIZONTAL (35)+NCC(8) 43 87.50 85.71 90.71 91.07 87.86 88.57

5. HORIZONTAL (35) 35 83.93 86.43 84.64 86.79 86.79 85.71

Table 7.13 to Table 7.24 gives confusion matrix for subclass I to subclass

VI for fold I using both SVM and k-NN classifier.


Table 7.13: Confusion Matrix for fold I Subclass I using Density and Normalized

chain code feature SVM classifier

Confusion Matrix

40 0 0 0 0 0 0 0 0 0

1 38 0 0 0 0 0 1 0 0

1 0 36 0 1 0 0 0 1 1

0 0 1 36 0 0 0 0 3 0

0 0 0 0 37 0 0 0 2 1

0 0 0 0 0 35 1 4 0 0

0 0 0 0 0 0 40 0 0 0

1 2 0 0 1 3 0 33 0 0

2 0 0 1 2 0 0 1 33 1

0 0 1 0 1 0 0 0 2 36

Table 7.14: Confusion Matrix for fold I Subclass I using Density and Normalized chain code

feature k-NN classifier

Confusion Matrix

38 0 0 1 0 0 0 1 0 0

0 35 0 1 0 1 0 3 0 0

1 0 36 0 1 0 0 0 1 1

0 0 1 35 0 0 0 0 4 0

0 0 0 0 34 0 0 0 5 1

0 1 0 0 0 36 1 2 0 0

0 0 0 0 0 2 38 0 0 0

0 1 0 0 2 3 0 33 1 0

2 1 2 2 0 0 0 0 31 2

0 0 3 0 0 0 0 2 1 34


Table 7.15: Confusion Matrix for fold I Subclass II using Density and Normalized chain code

feature SVM classifier

Confusion Matrix

40 0 0 0

0 36 2 2

1 1 38 0

1 0 0 39

Table 7.16: Confusion Matrix for fold I Subclass II using Density and Normalized chain code


Confusion Matrix

35 2 3 0

1 34 4 1

0 0 38 2

1 1 3 35

Table 7.17: Confusion Matrix for fold I Subclass III using Density and Normalized chain

code feature SVM classifier

Confusion Matrix

40 0 0 0 0 0 0 0 0 0

1 34 0 2 2 0 1 0 0 0

0 0 36 1 0 0 0 1 2 0

0 3 0 24 7 2 0 3 1 0

1 5 0 3 28 0 1 0 2 0

0 0 1 2 1 30 5 0 1 0

0 1 1 0 0 2 36 0 0 0

0 0 3 1 0 2 3 30 1 0

1 1 1 0 0 2 1 0 34 0

0 0 0 2 0 0 1 1 0 36


Table 7.18: Confusion Matrix for fold I Subclass III using Density and Normalized chain

code feature k-NN classifier

Confusion Matrix

40 0 0 0 0 0 0 0 0 0

1 31 0 2 5 0 1 0 0 0

0 0 34 1 0 1 0 1 3 0

0 4 0 25 5 2 0 4 0 0

0 10 0 2 24 0 1 0 3 0

0 0 1 1 1 29 6 0 1 1

0 0 1 0 0 4 32 2 1 0

0 1 4 1 0 1 3 26 2 2

0 0 4 0 0 3 1 0 32 0

0 0 1 2 0 1 1 5 0 30

Table 7.19: Confusion Matrix for fold I Subclass IV using Density and Normalized chain


Confusion Matrix

37 1 1 1

0 32 8 0

2 1 37 0

0 0 2 38

Table 7.20: Confusion Matrix for fold I Subclass IV using Density and Normalized chain


Confusion Matrix

39 0 1 0

5 26 8 1

2 4 32 2

0 0 1 39


Table 7.21: Confusion Matrix for fold I Subclass V using Density and Normalized chain code

feature SVM classifier

Confusion Matrix

38 0 0 0 0 2

0 37 0 1 2 0

1 0 35 2 1 1

0 0 0 39 0 1

1 0 2 0 37 0

0 0 1 2 1 36

Table 7.22: Confusion Matrix for fold I Subclass V using Density and Normalized chain code


Confusion Matrix

32 0 0 3 4 1

0 37 0 1 2 0

1 0 35 3 0 1

0 0 0 38 1 1

1 2 4 2 30 1

0 0 1 2 1 36


Table 7.23: Confusion Matrix for fold I Subclass VI using Density and Normalized chain


Confusion Matrix

36 0 0 2 1 1 0

3 36 0 0 1 0 0

0 0 37 0 0 2 1

0 0 0 39 0 0 1

4 0 0 0 36 0 0

0 1 0 3 0 34 2

0 0 0 0 0 0 40

Table 7.24: Confusion Matrix for fold I Subclass VI using Density and Normalized chain


Confusion Matrix

38 0 0 0 1 1 0

1 38 0 0 1 0 0

1 0 36 0 0 1 2

0 0 0 38 0 1 1

2 1 0 0 37 0 0

0 1 2 2 0 34 1

0 0 0 1 0 1 38

It has been observed from the experimentation that we got highest

recognition accuracy using density and normalized chain code for 41 Marathi

characters for all sub-classes using SVM and k-NN classifiers. The highest

recognition rate for handwritten Marathi characters using SVM classifier is

91.01% and using k-NN classifier 87.25% as shown in Table 7.25.


Table 7.25: Highest recognition rate for 41 Marathi characters using SVM and k-NN

Classifier

Classifier Features SC I

SC II

SC III

SC IV

SC V

SC VI

AVG FOR 41 CHARS

SVM

Density and

Normalized chain code

( 98 features )

91.85 94.72 83.30 88.62 92.83 94.73 91.01

k-NN

Density and

Normalized chain code

( 98 features )

89.10 90.00 76.75 84.62 89.75 93.29 87.25

*SC=Subclass

The OCR described above is further extended for Handwritten Marathi

word recognition. The experiments were carried on 50 simple word and 47

compound words. As we got best results using density and normalized chain code

features, we have used same for handwritten Marathi word recognition. Also it

has been observed that SVM classifier performs better than k-NN classifier, hence

we have used SVM classifier for word recognition. Recognition results for

simple words are elaborated in Table 7.26 and for compound words in Table 7.27.

Table 7.26: Handwritten Marathi simple words Recognition using SVM classifier

SR. NO. WORD SAMPLES

RESULT IN PERCENTAGE

SVM 1. 100 88 2. 100 90 3. 100 91 4. 100 89


5. 100 90

6. 100 92

7.

100 94

8. 100 88 9. 100 90 10. 100 91

11. 100 89

12. 100 88

13. 100 90

14.

100 92

15. 100 94

16.

100 88

17.

100 90

18. 100 91 19. 100 91

20. 100 90 21. 100 89 22. 100 89

23.

100 90

24. 100 85

25. 100 88 26. 100 90 27. 100 92 28. 100 88

29. 100 92

30. 100 87

31. 100 85 32. 100 93


33. 100 92

34. 100 90

35.

100 92

36. 100 89

37. 100 90

38.

100 91

39.

100 87

40. 100 86 41. 100 88

42.

100 90

43.

100 91

44. 100 92

45. 100 90 46. 100 94

47.

100 94

48. 100 93

49.

100 89

50. 100 90 Average Recognition Rate 90

Table 7.27: Handwritten Marathi compound words recognition using SVM classifier

SR. NO. WORD SAMPLES

RESULT IN PERCENTAGE

SVM 1.

100 90

2. 100 91

3.

100 87

4. 100 88


5.

100 86

6.

100 88

7.

100 89

8.

100 83

9.

100 89

10.

100 90

11.

100 88

12.

100 89

13.

100 87

14.

100 88

15.

100 87

16. 100 86

17.

100 89

18.

100 87

19. 100 90

20.

100 90

21.

100 89

22.

100 89

23.

100 88

24. 100 87

25. 100 84

26.

100 92

27.

100 91

28.

100 84

29.

100 94


30.

100 88

31.

100 86

32. 100 89

33.

100 91

34.

100 91

35.

100 89

36.

100 92

37.

100 93

38.

100 85

39.

100 83

40.

100 88

41.

100 86

42.

100 84

43.

100 88

44.

100 87

45.

100 90

46.

100 88

47.

100 89

Average Recognition Rate 88

In the next chapter we are summarizing results of the research and

concluding remarks discussed.

Chapter 8

Summary and Conclusions

---------------------------------------------------------------------------------------------------

8.1 Conclusions

8.2 Scope for further research

---------------------------------------------------------------------------------------------------

In this chapter we present contributions and summary of the present research

work. A comparative study of the present work with research work reported in

literature is made. Limitations of the present work and scope for the further

research are discussed.

---------------------------------------------------------------------------------------------------

8.1 Conclusions:

The work presented in this thesis has addressed the problem of

handwritten Marathi word recognition. Approaches to recognition largely depend

on the nature of the data to be recognized. Since handwritten Marathi words

could be of various shapes and sizes, the recognition process needs to be much

efficient and accurate to recognize the words written by different users. The

present work has addressed this problem by a novel multilevel classification

approach that groups Marathi characters into six groups. In additions to this,

suitable features for different subclasses are extracted. Recognition accuracy for

handwritten Marathi characters are cross validated using fivefold method and

tested using two classifiers, viz. k-Nearest Neighbor and Support Vector Machine

classifier. The different feature sets used are:

124

Chapter 8: Summary and Conclusions

1. Symmetric density features based upon the zoning approach. The feature

vector size is 90.

2. Diagonal features based upon the zoning approach. The feature vector

size is 35.

3. Horizontal features based upon the zoning approach. The feature vector

size is 35.

4. Vertical features based upon the zoning approach. The feature vector size

is 35.

5. Normalized chain code features. The feature vector size is 8.

6. Invariant moment features. The feature vector size is 35.

7. Zernike moment features. The feature vector size is 16.

8. Discrete wavelet transformation features. The feature vector size is 8.

9. Combination of symmetric density features based upon the zoning

approach and normalized chain code. The feature vector size is 98.

10. Combination of diagonal features based upon the zoning approach and

Zernike moments. The feature vector size is 51.

11. Combination of horizontal features based upon the zoning approach and

normalized chain code features. The feature vector size is 43.


moment invariant features. The feature vector size is 42.

125



Zernike moment features. The feature vector size is 51.

The effectiveness of the features proposed in the thesis is evaluated by

performing experiments on the database developed for the work. The database of

handwritten Marathi words contains 36283 images which are obtained from 100

writers belonging to different age groups and professions. In addition to this

database of isolated handwritten Marathi characters contains 9600 images which

are obtained from 20 writers belonging to different age groups and professions.

Since the data was collected in a preformatted paper the skew/slant was assumed

to be negligible and hence ignored in preprocessing stage. We adopted fivefold

cross validation technique, for performance evaluation of the classifier. With

fivefold cross-validation, all objects in the data set are used both as test objects as

well as training objects. This ensures that the classifier is tested on both rare and

common types of objects. The graphical presentation of recognition accuracy for

all subclasses using combination of zone based symmetric density and normalized

chain code feature is shown in the Figure 8.1 for both the classifiers SVM as well

as k-NN.

126


Figure 8.1: Recognition rate (%) of Marathi handwritten Characters using SVM and k-NN classifiers

From literature it is observed that some research work is reported for

handwritten word recognition, but dataset used for the research work is related to

specific domain such as city names, district names, legal amounts, numerals

written in characters. Also experiments were performed on databases of varying

sizes, ranging from 100 to 39700. All the work reported in the literature is based

on segmentation free approach. Marathi characters are very similar in shape and

0

10

20

30

40

50

60

70

80

90

100

1 2 3 4 5 6

RE

CO

GN

ITIO

N A

CC

UR

AC

Y IN

PE

RC

EN

TA

GE

SUBCLASSES

Recognition accuracy using SVM and k-NN classifiers

SVM k-NN

127


structure; hence large numbers of features are required to achieve good


All these limitations are considered and resolved in this work. The

contributions of the present work may be summarised as follows:

1. Any handwritten word of any size can be recognized using the system

proposed in this thesis.

2. Database developed for handwritten Marathi words and isolatd characters

is large enough to carry out experiments in further work.

3. Segmentation algorithms are developed and tested on the developed

database.

4. A novel multilevel classification approach is proposed to reduce the

number of features and hence improved the recognition efficiency of

character classification.

5. The usual claim by researcher that SVM clasifier performs better than k-

NN classifier is verified from this work.

The recognition accuracy from subclass I to subclass VI is computed using

listed features for both the classifiers. The recognition rates are higher for all

subclasses using combination of zone based symmetric density and normalized

chain code features. Top five features and their recognition accuracies for all

subclasses using SVM and k-NN classifiers is given in Table 8.1 to Table 8.6.

128


Table 8.1: Comparison of recognition rates of proposed methods for subclass I

Sr. No.

Feature extraction method

Number of features

Recognition Accuracy

k-NN SVM

1. Density(90) + NCC(8) 98 89.10 91.85

2. Horizontal(35) + NCC(8) 43 82.85 87.80

3. Horizontal(35) + Zernike (16) 51 82.85 87.75

4. Horizontal(35) 35 81.00 86.10

5. Density(90) 90 85.80 83.45

Table 8.2: Comparison of recognition rates of proposed methods for subclass II

Sr. No.


Number of features


k-NN SVM

1. Density(90) + NCC(8) 98 90.00 94.72


3. Horizontal(35)+ NCC(8) 43 83.25 87.99

4. Density(90) 90 88.24 86.12

5. Horizontal(35) +IM(7) 42 82.99 85.37

Table 8.3 Comparison of recognition rates of proposed methods for subclass III

Sr. No.


Number of features


k-NN SVM

1. Density(90) + NCC(8) 98 76.75 83.30


3. Horizontal(35)+NCC(8) 43 71.85 77.15

4. Horizontal(35)+IM(7) 42 70.55 75.50

5. Horizontal(35) 35 68.65 74.45

129


Table 8.4 Comparison of recognition rates of proposed methods for subclass IV

Sr. No. Feature extraction method Number of

features Recognition Accuracy

k-NN SVM

1. Density(90) + NCC(8) 98 84.62 88.62

2. Horizontal(35) 35 80.82 88.12

3. Horizontal(35)+NCC(8) 43 82.74 87.50


5. Diagonal(35) 35 70.09 81.37

Table 8.5 Comparison of recognition rates of proposed methods for subclass V

Sr. No. Feature extraction method Number of

features Recognition Accuracy

k-NN SVM

1. Density(90) + NCC(8) 98 89.75 92.83

2. Horizontal(35)+NCC(8) 43 84.93 88.91


4. Horizontal(35) 35 83.53 87.02

5. Density(90) 90 87.74 84.83

Table 8.6 Comparison of recognition rates of proposed methods for subclass VI

Sr. No.


Number of features


k-NN SVM

1. Density(90) + NCC(8) 98 93.28 94.73

2. Horizontal(35) 35 85.71 91.37

3. Horizontal(35)+NCC(8) 43 88.57 91.35


5. Density(90) 90 89.30 85.48

130


The performance of the method based upon combination of ‘Zone

Symmetric Density and Normalized Chain Code’, presented in Chapter 6, is

compared with the work reported in literature. The recognition rate and the

performance comparison are given in Table 8.7. The proposed method performs

well and appears promising as compared to other methods in the literature. Also

the database used for experiments is quite large and recognition accuracy is

encouraging by reducing number of features. Work reported in literature is

performed on isolated characters where as in the proposed system experiments are

carried on both characters which are segmented from words and isolated

characters.

Table 8.7 Comparison of recognition rates for handwritten marathi characters with other methods in literature

Sr. No.

Method Database Size

Features Feature

Vector Size

Classifier Recognition Rate

(%)

1. Aarti Desai et. al. [4] 150 End Pt., Branch

Pt., Chain Code 200 NM 87

2. Ajmire P.E.

and Warkhede S.E. [8]

120

Standard Deviation and

Mean of Moment Invariant

14 NM 62

3. Anilkumar N. Holambe [12] 5000 Gradient

Features 32 SVM 97

4. Archana P.Jane

and Mukesh. A.Pund [14]

1020 Smoothing and Fuzzy Pattern NM NM 90

5. Ashutosh

Aggarwal et. al. [19]

7200 Gradient Features 200 SVM 94

6. Brijmohan Singh et. al.

[35] 31860

Curvelet Transform and

Geometric Features

1024 SVM and

k-NN 93.8

131


7. Chavan S.V. et. al. [40] 27000 Moment

Features 36

MLP and

k-NN 98.78

8. Holambe A.N. et. al. [57] 20000 Gradient

Features 400 k-NN 96

9. J.Pradeep et. al. [58] NM Diagonal

Features 69 Neural Network 99

10. Karbhari V. Kale et. al. [64] 27000 Zernike

Moments NM SVM and

k-NN 98.37

11. Latesh Malik

and P.S. Deshpande

NM Regular Expression NM NM 100

12. M. Hanmandlu et. al. [67] 4750 NM NM NM 90.64

13. Mahesh Jangid [70] 12240 Statistical

Features 314 SVM 94.89

14.

Vinaya. S. Tapkir and Sushma. D.

Shelke

NM Density Features 16

Euclidean

Minimum

Distance Classifier

92.77

15. N. Sharma et. al. [80] 11270 64 Directional

Features 64 Quadrati

c Classifier

80.36

16. Nilima P. Patil et. al. [87] 1500

Moment Invariant,

Affine Moment Invariant

NM

Fuzzy Members

hip Classific

ation

89.09

17.

O. V. Ramana Murthy and M. Hanmandlu[89]

4713 Zone Based Features 64 SVM 88.9

18. P. S.

Deshpande et. al. [94]

5000 Regular Expression NM

Minimum Edit

Distance Classifier

82

19. Prachi

Mukherji and Priti P. Rege

NM

Average Compressed

Direction Codes

45 NN Classifier 92.8

132


20. R. J. Ramteke 250 Invariant Moment NM

Fuzzy Gaussian Members

hip

94.56

21.

Rakesh Rathi et. al. [114]

9191

Recursive Subdivision

Feature Extraction Technique

NM k-NN 96.14

22. S. Arora et. al. [119] 1500

Chain Code And Shadow

Features 200

Multilayer

Perceptron

89.58

23. Sandhya Arora et. al. [122] 7154

Chain Code And Shadow

Features 200

Multilayer

Perceptron

90.74

24. Proposed Method 67035

Density And Normalized Chain Code

98 SVM 91.01


Density And Normalized Chain Code

98 k-NN 87.25

*NM= Not mentioned

The proposed method performs well and appears promising as compared

to other methods in the literature for handwritten Marathi word recognition. Also

the database used for experiments is quite large and recognition accuracy is

encouraging by reducing number of features. Work reported in literature is on

limited dataset and on specific domain like city name, district names and legal

amounts written in Marathi where as in the proposed system experiments are

performed on commonly used 97 Marathi words. Also this method for

handwritten Marathi word recognition is applicable to any word having any

number of characters. The performance comparison in terms of recognition

accuracy of handwritten Marathi word recognition is given in Table 8.8.

133


Table 8.8 Comparison of recognition rates for handwritten Marathi words with other

methods in literature Sr. No.

Method Database size

Features Feature vector

size

Classifier Recognition rate (%)

1. Bikash Shaw et. al. [32] 39700 Stroke based

features NM HMM 84.31

2. Bikash Shaw et. al. [33] 39700

Directional chain code

feature NM HMM 80.20

3. Brijmohan

Singh et. al. [36]

28500 Curvelet

transormed based features

200 SVM and k-NN 93.21

4.

C.Namrata Mahender and

K. V. Kale [39]

2800 Structure based features NM

Rule based

classification

approach

85.00

5. Naresh Kumar

Garg et. al. [83]

2016 Shape based features 59 SVM 76.40


Density and normalized chain code

98 SVM 90.00

8.2 Scope for further research:

Though we have successfully attempted the problem of recognition of

handwritten Marathi words and presented encouraging results in terms of

recognition accuracy, still there is considerable scope for further research.

1. In the proposed handwritten Marathi word recognition system the words

are not compared with lexicon. The outcome of the system is class labels

134


for every isolated character, half character and modifier. The performance

of the system may be improved by including lexicon.

2. The proposed feature extraction methods have been tested for handwritten

words collected in preformatted sheets. In real world situations, the words

are segmented out from a handwritten document and are input to the OCR

for recognition. Hence, there is a need to consider such words to test the

robustness of the proposed OCR system.

3. Though the data is collected in preformatted sheets, there is always some

slant in the written words. Including slant correction algorithms in

preprocessing stage may definitely increase the recognition rate of

handwritten Marathi words.

4. The characters in words obtained from old handwritten documents are

often disconnected. The proposed system takes care of disconnected

characters, where the disconnectivity is one or two pixels. Beyond that,

there is a need to consider other methods to take care of disconnected

characters.

5. For similar character symbols, the recognition rate can be improved by

combining multiple classifiers.

6. The proposed methods can be extended for recognition of words written in

other Indian scripts.

Publications

International Journals:

1. “Recognition of Handwritten Marathi Vowels using Zone based Symmetric

Density Features”, International Journal of Computer Applications (0975 –

8887) Volume 108 – No. 4, December 2014, ISSN 0975-8887. Impact factor-

0.715.

2. “Recognition of Handwritten Marathi Vowels using Combination of

Topological and Statistical Features”, International Journal of Engineering

Research & Technology (IJERT), Vol. 3 Issue 11, November-2014, ISSN: 2278-

0181. Impact factor-1.76

National:

1. “Isolated Handwritten Marathi Character Recognition”, Proceedings of

National Conference on Challenging Research Areas in Computer Science and

Information Technology - 2014, ISBN 978-93-83777-00-6.

Bibliography

1. A. L. Koerich, R. Sabourin, C. Y. Suen; 2003; Large vocabulary off-line

handwriting recognition:A survey; Springer Pattern Anal Applic (2003) 6: 97–121 DOI 10.1007/s10044-002-0169-3.

2. A. R. Karwankar, A.S. Bhalchandra; 2010; Stroke Pattern Identification in Devanagari Character Recognition; Published in International Journal of Advanced Engineering & Application, Jan. 2010.

3. Aarti Desai, Latesh Malik; 2011; A Novel Approach To Thinning Of Devnagari Characters; International Journal of Engineering Science and Technology (IJEST); ISSN : 0975-5462.

4. Aarti Desai, Latesh Malik, Rashmi Welekar; 2011; A New Methodology For Devnagari Character Recognition; JM academy of It & Management Volume -1 Issue 1; ISSN: 2229-6115.

5. Aditya Raj, Ranjeet Srivastava, Tushar Patnaik, Bhupendra Kumar; 2013; A Survey of Feature Extraction and Classification Techniques Used In Character Recognition for Indian Scripts; International Journal of Engineering and Advanced Technology (IJEAT) , Volume-2, Issue-3, February 2013; ISSN: 2249 – 8958.

6. Aisharjya Sarkar, Arindam Biswas; 2010; Word Segmentation and Baseline Detection in Handwritten Documents Using Isothetic Covers; 12th International Conference on Frontiers in Handwriting Recognition 978-0-7695-4221-8/10, 2010 IEEE.

7. Ajay K. Talele, Sanjay L. Nalbalwar, Milind E. Rane; 2011; Automatic Recognition and Verification of Handwritten Legal and Courtesy Amounts in English Language Present on Bank Cheques; International Journal of Computer Applications(0975–8887) Volume 21– No.8, May 2011.

8. Ajmire P. E., Warkhede S.E.; 2010; Handwritten Marathi character (vowel) recognition; Advances in Information Mining, ISSN: 0975–3265, Volume 2, Issue 2, 2010, PP-11-13.

9. Alessandro L. Koerich Alceu de S. Britto Jr. Luiz Eduardo S. de Oliveira; 2010; Verification of Unconstrained Handwritten Words at Character Level; 12th International Conference on Frontiers in Handwriting Recognition 978-0-7695-4221-8/10, 2010 IEEE.

10. Amit Choudhary, Rahul Rishi, Savita Ahlawat, Vijaypal Singh Dhaka; 2010; Totally Unconstrained Handwritten Character Recognition using Improved BP Neural Network; International Journal of Information Technology and Knowledge Management July-December 2010, Volume 2, No. 2, pp. 645-649.

11. Angshul Majumdar; 2007; Bangla Basic Character Recognition Using Digital Curvelet Transform; Journal of Pattern Recognition Research 1 (2007) 17-26.

137 Bibliography

12. Anilkumar N. Holambe, Ravinder C.Thool, Sushilkumar N. Holambe; Printed and Handwritten Character & Number Recognition of Devanagari Script using Gradient Features; Advances in Multimedia - An International Journal (AMIJ).

13. Apurva A. Desai; 2012; Segmentation of Characters from Old Typewritten Documents using Radon Transform; International Journal of Computer APPlications (0975–8887) Volume 37– No.9, January 2012.

14. Archana P. Jane, Mukesh A. Pund; 2012; Recognition of Similar Shaped Handwritten Marathi Characters Using Artificial Neural Network; Global Journal of Computer Science and Technology Neural & Artificial Intelligence Volume 12 Issue 11 Version 1.0 Year 2012; Online ISSN: 0975-4172 & Print ISSN: 0975-4350.

15. Ashish Ghosh, B. Uma Shankar, Saroj K. Meher; 2009; A novel approach to neuro-fuzzy classification; Elsevier Neural Networks 22 (2009) 100 109.

16. Ashoka H.N., Manjaiah D.H., Rabindranath Bera; 2012; Feature Extraction Technique for Neural Network Based Pattern Recognition; International Journal on Computer Science and Engineering (IJCSE) Vol. 4 No. 03; ISSN : 0975-3397.

17. Ashu Kumar, Simpel Rani Jindal; 2012; Segmentation of handwritten Gurmukhi text into lines; International Conference on Recent Advances and Future Trends in Information Technology (iRAFIT2012) Proceedings published in International Journal of Computer Applications (IJCA).

18. Ashutosh Aggarwal, Rajneesh Rani, Renu Dhir; 2012; Recognition of Devanagari Handwritten Numeralsusing Gradient Features and SVM; International Journal of Computer Applications (0975–888). Volume 48– No.8, June 2012.

19. Ashutosh Aggarwal, Rajneesh Rani, RenuDhir; 2012; Handwritten Devanagari Character Recognition Using Gradient Features; International Journal of Advanced Research in Computer Science and Software Engineering Volume 2, Issue 5; ISSN: 2277 128X.

20. Ashutosh Malaviya, Christoph Leja, Liliane Peters; 1996; Multi-Script Handwriting Recognition with FOHDEL; Published in the Proceedings of the Biennial Conference of North American Information Proecessing Society (NAFIPS’96), Berkeley, IEEE, PP. 147-151, 1996.

21. B. V. Dhandra, V. S. Malemath, Mallikarjun H. , Ravindra Hegadi; 2006; Skew Detection in Binary Image Documents Based on Image Dilation and Region labeling Approach; 0-7695-2521-0/06/2006 IEEE.

22. B. V. Dhandra, V. S. Malemath, Mallikarjun H., Ravindra Hegadi; 2006; Multi-font Numeral Recognition without Thinning based on Directional Density of Pixels; 1-4244-0682-X/06/2006 IEEE.

23. B.V. Dhandra, Gururaj Mukarambi, Mallikarjun Hangarge; 2011; Mixture Of Printed Kannada Numerals And Vowels Recognition System; World Journal of Science and Technology 2011,1(8):109-114 ISSN: 2231 – 2587.

138 Bibliography

24. B.V.Dhandra, Gururaj Mukarambi, Mallikarjun Hangarge; 2011; Kannada and English Numeral Recognition System; International Journal of Computer APPlications (0975–8887) Volume 26– No.9, July 2011.

25. B.V.Dhandra, Mallikarjun Hangarge; 2007; On Separation of English Numerals from Multilingual Document Images; Journal Of Multimedia, Vol. 2, No. 6.

26. B.V.Dhandra, Mallikarjun Hangarge; Morphological Reconstruction for Word Level Script Identification; International Journal of Computer Science and Security, Volume (1) : Issue (1).

27. B.V.Dhandra, P. Nagabhushan, Mallikarjun Hangarge, Ravindra Hegadi, V.S. Malemath; 2006; Script Identification Based on Morphological Reconstruction in Document Images; 0-7695-2521-0/06/2006 IEEE.

28. B.V.Dhandra, R.G.Benne, Mallikarjun Hangarge; 2010; Kannada, Telugu and Devanagari Handwritten Numeral Recognition with Probabilistic Neural Network: A Novel APProach; IJCA Special Issue on “Recent Trends in Image Processing and Pattern Recognition” RTIPPR, 2010.

29. Baheti M. J., Kale K.V., Jadhav M.E.; 2011; Comparison Of Classifiers For Gujarati Numeral Recognition; International Journal of Machine Intelligence ISSN: 0975–2927 & E-ISSN: 0975–9166, Volume 3, Issue 3, 2011, PP-160-163.

30. Benne R. G., Dhandra B. V., Mallikarjun Hangarge; 2009; Tri-scripts Handwritten Numeral Recognition: A Novel Approach; Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009, PP-47-51.

31. Bharath A., Sriganesh Madhvanath; 2010; On the Significance of Stroke Size and Position for Online Handwritten Devanagari Word Recognition: An Empirical Study; International Conference on Pattern Recognition; 1051-4651/10 2010 IEEE.

32. Bikash Shaw, Swapan Kr. Parui, Malayappan Shridhar; 2008; Offline Handwritten Devanagari Word Recognition: A Segmentation Based Approach; 978-1-4244-2175-6/08/2008 IEEE.

33. Bikash Shaw, Swapan Kr. Parui, Malayappan Shridhar; 2008; Offline Handwritten Devanagari Word Recognition: A holistic approach based on directional chain code feature and HMM; International Conference on Information Technology; IEEE ICIT 978-0-7695-3513-5/08 2008.

34. Binu P Chacko, Babu Anto P; 2010; Pre and Post Processing Approaches in Edge Detection for Character Recognition; 12th International Conference on Frontiers in Handwriting Recognition 978-0-7695-4221-8/10 2010 IEEE.

35. Brijmohan Singh, Ankush Mittal, Debashis Ghosh; 2011; An Evaluation of Different Feature Extractors and Classifiers for Offline Handwritten Devnagari Character Recognition; Journal of Pattern Recognition Research 2 (2011) 269-277.

139 Bibliography

36. Brijmohan Singh, Ankush Mittal, M. A. Ansari, Debashis Ghosh; 2011; Handwritten Devanagari Word Recognition: A Curvelet Transform Based Approach; International Journal on Computer Science and Engineering (IJCSE) Vol. 3 No. 4 ; ISSN : 0975-3397.

37. Brijmohan Singh, Nitin Gupta, Rashi Tyagi, Ankush Mittal, Debashish Ghosh; 2011; Parallel Implementation of Devanagari Text Line and Word Segmentation Approach on GPU; International Journal of Computer Applications (0975–8887) Volume 24– No.9, June 2011.

38. C. V. Jawahar, Pavan Kumar, S. S. Ravi Kiran; 2003; A Bilingual OCR for Hindi-Telugu Documents and its Applications; Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 2003 IEEE.

39. C. Namrata Mahender, K.V.Kale; 2011; Structured based Feature Extraction of Handwritten Marathi Word; International Journal of Computer Applications (0975–8887) Volume 16– No.6, February 2011.

40. Chavan S. V., Kale K. V., Kazi M. M., Rode Y. S.; 2013; Recognition Of Handwritten Devanagari Compound Character A Moment Feature Based Approach; International Journal of Machine Intelligence ISSN: 0975-2927 & E-ISSN: 0975-9166, Volume 5, Issue 1, 2013, pp.-421-425.

41. Chun Lei He Louisa Lam Ching Y. Suen; 2009; A Novel Rejection Measurement in Handwritten Numeral Recognition Based on Linear Discriminant Analysis; 10th International Conference on Document Analysis and Recognition 978-0-7695-3725-2/09 IEEE.

42. Dhandra B. V., Benne R. G., Mallikarjun Hangarge; 2011; Printed And Handwritten Kannada Numerals Recognition using Directional Stroke And Directional Density With Knn; International Journal of Machine Intelligence ISSN: 0975–2927 & E-ISSN: 0975–9166, Volume 3, Issue 3, 2011, pp-121-125.

43. Diego J. Romero, Leticia M. Seijas, Ana M. Ruedin; 2007; Directional Continuous Wavelet Transform Applied to Handwritten Numerals Recognition Using Neural Networks; JCS&T Vol. 7 No. 1.

44. Dileep Kumar Patel, Tanmoy Som, Sushil Kumar Yadav, Manoj Kumar Singh; 2012; Handwritten Character Recognition Using Multiresolution Technique and Euclidean Distance Metric; Journal of Signal and Information Processing, 2012, 3, 208-214 doi:10.4236/jsip.2012.32028 Published Online May 2012.

45. Dinesh Achaya U., N. V. Subba Reddy and Krishnamoorthi; 2008; Hierarchical Recognition System for Machine Printed Kannada Characters; IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.11.

46. Dipak V. Koshti, Sharvari Govilkar; Segmentation of Touching Characters in Handwritten Devanagari Script; UACEE International Journal of Computer Science and its Applications - Volume 2: Issue 2 [ISSN 2250 - 3765].

140 Bibliography

47. Dipankar Das, Rubaiyat Yasmin; 2006; Segmentation and Recognition of Unconstrained Bangla Numeral; Asian Journal of Information Technology 5(2).

48. Dipti Deodhare, NNR Ranga Suri R. Amit; 2005; Preprocessing and Image Enhancement Algorithms for a Form-based Intelligent Character Recognition System; International Journal of Computer Science & Applications Vol. II, No. II, PP. 131 - 144 2005.

49. Dipti Pawar; 2012; Extended Fuzzy Hyperline Segment Neural Network for Handwritten Character Recognition; Proceedings of the International MultiConference of Engineers and Computer Scientists 2012 Vol I;ISBN 978-988-19251-1-4 ISSN 2078-0958.

50. Edgar Erdfelder, Carolina E. Küpper-Tetzel, Sandra D. Mattern; 2011; Threshold models of recognition and the recognition heuristic; Judgment and Decision Making, Vol. 6, No. 1, February 2011, PP. 7–22.

51. G. G. Rajput, S. M. Mali; 2010; Fourier Descriptor based Isolated Marathi Handwritten Numeral Recognition; International Journal of Computer Applications (0975 – 8887) Volume 3 – No.4.

52. G. G. Rajput, Anita H. B.; 2010; Handwritten Script Recognition using DCT and Wavelet Features at Block Level; IJCA Special Issue on “Recent Trends in Image Processing and Pattern Recognition” RTIPPR, 2010.

53. G. G. Rajput, S. M. Mali; 2010; Isolated Handwritten Marathi Numerals Recognition Based upon Fourier Descriptors and Freeman Chain Code; International Journal of Computational Intelligence Research ISSN 0973-1873 Volume 6, Number 2 (2010), PP. 289–298.

54. Gang Liu, Lianwen Jin, Kai Ding, Hanyu Yan; 2010; A new approach for synthesis and recognition of large scale handwritten Chinese words; 12th International Conference on Frontiers in Handwriting Recognition 978-0-7695-4221-8/10 2010 IEEE.

55. Gaurav Agrawal, Kshitij, Amitabha Mukerjee, Nimit Kumar; Handwritten Devanagari Script Segmentation using Support Vector Machines.

56. Gunvantsinh Gohil, Rekha Teraiya, Mahesh Goyani; 2012; Chain Code And Holistic Features Based OCR System For Printed Devanagari Script Using ANN And SVM; International Journal of Artificial Intelligence & Applications (IJAIA), Vol.3, No.1, January 2012.

57. Holambe A. N., Thool R. C., Shinde U. B. and Holambe S. N.; 2010; Brief review of research on Devanagari script; International Journal of Computational Intelligence Techniques, ISSN: 0976–0466, Volume 1, Issue 2, 2010, PP-06-09.

58. J.Pradeep, E.Srinivasan, S.Himavathi; 2010; Diagonal Feature Extraction Based Handwritten Character System Using Neural Network; International Journal of Computer Applications(0975–8887) Volume 8– No.9, October 2010.

59. Jonathan J. Hull, Alan Commike, Tin-Kam HO; 1990; Multiple Algorithms for Handwritten Character Recognition; Int. Workshop on Frontiers in Handwriting Recognition, Montreal, Canada, April 2-3, 1990.

141 Bibliography

60. Jyotsna Vaid, Ashum Gupta; 2002; Exploring Word Recognition in a Semi-Alphabetic Script: The Case of Devanagari; 2002 Elsevier Science (USA) Brain and Language 81, 679–690 (2002).

61. K. V. Prema, N. V. Subba Reddy; 2002; Two-tier architecture for unconstrained handwritten character recognition; Sadhana Vol. 27, Part 5, PP. 585–594.

62. K. Y. Rajput, Sangeeta Mishra; Recognition and Editing of Devnagari Handwriting Using Neural Network; Proceedings of SPIT-IEEE Colloquium and International Conference, Mumbai, India Vol. 1.

63. K. H. Aparna, Vidhya Subramanian, M. Kasirajan, G. Vijay Prakash, V. S. Chakravarthy; 2004; Online Handwriting Recognition for Tamil; Proceedings of the 9th Int’l Workshop on Frontiers in Handwriting Recognition (IWFHR-9 2004) 0-7695-2187-8/04 2004 IEEE;

64. Kapil Bamne, Neha Sharma; 2015; Offline Classifier For Handwritten Devanagari Script Recognition; International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE) Volume 4, Issue 2, February 2015; ISSN: 2278 – 909X.

65. Karbhari V. Kale, Prapti D. Deshmukh, Shriniwas V. Chavan, Majharoddin M. Kazi, Yogesh S. Rode; 2014; Zernike Moment Feature Extraction for Handwritten Devanagari (Marathi) Compound Character Recognition; (IJARAI) International Journal of Advanced Research in Artificial Intelligence,Vol. 3, No.1, 2014.

66. Latesh Malik, P.S. Deshpande; 2009; Recognition of Printed and Handwritten Devanagari Characters With Regular Expression in Finite State Models; Digital Technology Journal 2009, Vol. 2, PP. 1{7, ISSN 1802-5811 (print), ISSN 1802-582X (online).

67. Laurent Guichard Alejandro H. Toselli Bertrand Couasnon; 2010; Handwritten word verification by SVM-based hypotheses re-scoring and multiple thresholds rejection; 12th International Conference on Frontiers in Handwriting Recognition 978-0-7695-4221-8/10 2010 IEEE.

68. M. Hanmandlu, O.V. Ramana Murthy, Vamsi Krishna Madasu; 2007; Fuzzy Model based recognition of handwritten Hindi characters; Digital Image Computing Techniques and Applications 0-7695-3067-2/07 2007 IEEE DOI 10.1109/DICTA.2007.82 IEEE.

69. M. Omidyeganeh, S. Ghaemmaghami, S. Shirmohammadi; 2012; Application of 3D-wavelet statistics to video analysis; Multimed Tools Appl DOI 10.1007/ s11042-012-1012-5 Springer.

70. M.K. Jindal, R.K. Sharma, G.S. Lehal; 2007; Segmentation of Horizontally Overlapping Lines in Printed Indian Scripts;International Journal of computational Intelligence Research. ISSN 0973-1873 Vol.3, No.4 (2007), PP. 277–286.

142 Bibliography

71. Mahesh Jangid; 2011; Devanagari Isolated Character Recognition by using Statistical features; International Journal on Computer Science and Engineering (IJCSE) Vol. 3 No. 6 ;ISSN : 0975-3397.

72. Manoj Kumar Shukla, Tushar Patnaik, Shrikant Tiwari, Sanjay Kumar Singh; 2011; Script Segmentation of Printed Devnagari and Bangla Languages Document Images OCR; International Journal of Computer Sci ence and Technology Vol. 2, Issue 2,;ISSN:2229-4333(Print)|ISSN:0976-8491(Online).

73. Mansi Shah, Gordhan B Jethava; 2013; A Literature Review On Hand Written Character Recognition; Indian Streams Research Journal Vol -3 , ISSUE –2, March.2013 ISSN:-2230-7850.

74. Mehmmood Abdulla Abd; 2007; Effective Arabic Character Recognition using Support Vector Machines; Springer Innovations and Advanced Techniques in Computer and Information Sciences and Engineering, 7–11.

75. Mithun Biswas, Ranjan Parekh; 2012; Character Recognition using Dynamic Windows; International Journal of Computer Applications (0975 – 8887) Volume 41– No.15, March 2012.

76. Mohammad Reza, Jenabzade Reza Azmi , Boshra Pishgoo, Samanesadat Shirazi; 2011; Two Methods for Recognition of Hand Written Farsi Characters; International Journal of Image Processing (IJIP), Volume (5) : Issue (4) : 2011.

77. Mohit Mehta, Rupesh Sanchati, Ajay Marchya; 2010; Automatic Cheque Processing System; International Journal of Computer and Electrical Engineering, Vol. 2, No. 4, August, 2010 1793-8163;

78. Vinaya. S. Tapkir, Sushma. D. Shelke; 2012; OCR For Handwritten Marathi Script; International Journal of Scientific & Engineering Research Volume 3, Issue 8, August-2012 1; ISSN 2229-5518.

79. Mudit Agrawal, Ajay S Bhaskarabhatla and Sriganesh Madhvanath;Data Collection For Handwriting Corpus Creation In Indic Scripts.

80. N. B. Mapari, A. L. Telang, R. K. Rajbhure; 2011; A Study Of Devnagri Handwritten Character Recognition System; International Journal Of Computer Science And Applications Vol. 4, No. 2, June-July 2011 ISSN: 0974-1003.

81. N. Sharma, U. Pal, F. Kimura, S. Pal; 2006; Recognition of Off-Line Handwritten Devnagari Characters Using Quadratic Classifier; Springer ICVGIP 2006, LNCS 4338, PP. 805 – 816, 2006.

82. Naresh Kumar Garg, Lakhwinder Kaur, M. K. Jindal; 2010; Segmentation of Handwritten Hindi Text; 2010 International Journal of Computer Applications (0975 – 8887) Volume 1 – No. 4.

83. Naresh Kumar Garg, Lakhwinder Kaur, M. K. Jindal; 2011; The Hazards in Segmentation of Handwritten Hindi Text; International Journal of Computer Applications (0975 – 8887) Volume 29– No.2, September 2011.

84. Naresh Kumar Garg, Lakhwinder Kaur, M. K. Jindal; 2013; Recognition of Offline Handwritten Hindi Text Using SVM; International Journal of Image Processing (IJIP), Volume (7) : Issue (4) : 2013.

143 Bibliography

85. Nazih Ouwayed, Abdel Belaïd; 2011; A general approach for multi-oriented text line extraction of handwritten documents;Springer IJDAR DOI 10.1007/s10032-011-0172-6.

86. Neeraj Pratap, Shwetank Arya; 2012; A Review of Devnagari Character Recognition from Past to Future; International Journal of Computer Science and Telecommunications [Volume 3, Issue 6, June 2012].

87. Neha Avhad, Shraddha Darade,Neha Gawali, Apurva Matsagar; 2015; Handwritten Devnagari Character Recognition System; International Journal Of Engineering And Computer Science Volume 4 Issue 4 April 2015, Page No. 11233-11236; ISSN:2319-7242.

88. Neha Sahu, R. K. Rathy, Indu Kashyap; 2012; Survey and Analysis of Devnagari Character Recognition Techniques using Neural Networks; International Journal of Computer Applications (0975 – 888) Volume 47– No.15, June 2012.

89. Nilima P. Patil, K. P. Adhiya, Surendra P. Ramteke; 2011; A Structured Analytical Approach to Handwritten Marathi vowels Recognition; International Journal of Computer Applications (0975 – 8887) Volume 31– No.3, October 2011.

90. Niranjan S.K, Vijaya Kumar, Hemantha Kumar G, Manjunath Aradhya V N; 2009; FLD based Unconstrained Handwritten Kannada Character Recognition; International Journal of Database Theory and Application Vol. 2, No. 3, September 2009.

91. O. V. Ramana Murthy, M. Hanmandlu; 2011; Zoning based Devanagari Character Recognition; International Journal of Computer Applications (0975 – 8887) Volume 27– No.4, August 2011.

92. O. V. Ramana, Sujoy Roy, Vipin Narang, M. Hanmandlu; 2012; Devanagari Character Recognition in the Wild; International Journal of Computer Applications (0975 – 8887) Volume 38– No.4, January 2012.

93. O.V. Ramana Murthy, M. Hanmandlu; 2011; A Study on the Effect of Outliers in Devanagari Character Recognition; International Journal of Computer Applications (0975 – 8887) Volume 32– No.10, October 2011.

94. Om Prakash Sharma, M. K. Ghose, Krishna Bikram Shah; 2012; An Improved Zone Based Hybrid Feature Extraction Model for Handwritten Alphabets Recognition Using Euler Number; International Journal of Soft Computing and Engineering (IJSCE) ISSN: 2231-2307, Volume-2, Issue-2, May 2012.

95. P. Bhaskara Rao D.Vara Prasad Ch.Pavan Kumar; 2013; Feature Extraction Using Zernike Moments; International Journal of Latest Trends in Engineering and Technology (IJLTET) Vol. 2 Issue 2 March 2013; ISSN: 2278-621X.

96. P. S. Deshpande, Latesh Malik, Sandhya Arora; 2008; Fine Classification & Recognition of Hand Written Devnagari Characters with Regular Expressions & Minimum Edit Distance Method; Journal Of Computers, Vol. 3, No. 5.

144 Bibliography

97. P. S. Deshpande, Latesh Malik, Sandhya Arora; 2007; Recognition of Hand Written Devnagari Characters with Percentage Component Regular Expression Matching and Classification Tree; IEEE 1-4244-1272-2/07 2007.

98. P.E.Ajmire, R.V. Dharaskar, V. M. Thakare; 2012; A Comparative Study of Handwritten Marathi Character Recognition; National Conference on Innovative Paradigms in Engineering & Technology (NCIPET-2012) Proceedings published by International Journal of Computer Applications (IJCA).

99. Partha Pratim Roy, Umapada Pal, Josep Lladós; 2011; Text line extraction in graphical documents using background and foreground information;Springer IJDAR DOI 10.1007/s10032-011-0167-3.

100. Prachi Mukherji, Priti P. Rege; 2008; Fuzzy Stroke Analysis of Devnagari Handwritten Characters; WSEAS TRANSACTIONS on COMPUTERS Issue 5, Volume 7; ISSN: 1109-2750.

101. Prachi Mukherji, Priti P. Rege; 2009; Shape Feature and Fuzzy Logic Based Offline Devnagari Handwritten Optical Character Recognition; Journal of Pattern Recognition Research 4 (2009) 52-68.

102. Prachi Mukherji, Priti P. Rege; Stroke Analysis Of Devnagari Handwritten Characters; 6th Wseas International Conference On Circuits, Systems, Electronics, Control & Signal Processing, Cairo, Egypt, Dec 29-31, 2007.

103. Prachi Patil, Saniya Ansari; 2014; Online Handwritten Devnagari Word Recognition using HMM based Technique; International Journal of Computer Applications (0975 – 8887) Volume 95– No.17, June 2014.

104. Prashant M. Kakde, Vivek R. Raut; 2012; Performance Analysis Of Handwritten Devnagri Characters Recognition Through Machine Intelligence; International Journal of Emerging Technology and Advanced Engineering www.ijetae.com (ISSN 2250-2459, Volume 2, Issue 7, July 2012).

105. Prerna Singh, Nidhi Tyagi; 2011; Radial Basis Function For Handwritten Devanagari Numeral Recognition; (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No. 5, 2011.

106. Pritpal Singh, Sumit Budhiraja; 2012; Handwritten Gurmukhi Character Recognition Using Wavelet Transforms;International Journal of Electronics, Communication & Instrumentation Engineering Research and Development (IJECIERD) ISSN 2249-684X Vol.2, Issue 3 Sep 2012 27-37.

107. Priyanka Kulkarni, Sonal Patil, Ganesh Dhanokar; 2015;Review On Marathi And Sanskrit Word Recognition Using Genetic Algorithm; International Journal of Informative & Futuristic Research ISSN (Online): 2347-1697, Volume 2 Issue 7 March 2015; ISSN (Online): 2347-1697.

108. R. C. Tripathi, Vijay Kumar; 2012; Character Recognition: A Neural Network Approach; National Conference on Advancement of Technologies – Information Systems & Computer Networks (ISCON – 2012) Proceedings published in International Journal of Computer Applications (IJCA).

145 Bibliography

109. R. Indra Gandhi, K.Iyakutti; 2010; A Technique for Segmentation over Overlapping Line of Uniform Sized Text on Non-Headline Based Distorted Tamil Scripts; Int. J. of Advanced Networking and Applications Volume: 02, Issue: 02, Pages: 491-495 (2010).

110. R. J. Ramteke ; 2010; Invariant Moments Based Feature Extraction for Handwritten Devanagari Vowels Recognition; International Journal of Computer Applications (0975 - 8887) Volume 1 – No. 18.

111. R. J. Ramteke, S. C. Mehrotra; 2008; Recognition Of Handwritten Devanagari Numerals; International Journal of Computer Processing of Oriental Languages;

112. R. Jagadeesh Kannan, R. Prabhakar; 2008; An Improved Handwritten Tamil Character Recognition System using Octal Graph; Journal of Computer Science 4 (7): 509-516, 2008;ISSN 1549-3636.

113. R. Jayadevan, S. R. Kolhe, P. M. Patil, U. Pal; 2011; Automatic Processing Of Handwritten Bank Cheque Images: A Survey; Springer IJDAR DOI 10.1007/ s10032-011-0170-8.

114. R. Jayadevan, S. R. Kolhe, P. M. Patil, Umapada Pal; 2011; Database Development and Recognition of Handwritten Devanagari Legal Amount Words; International Conference on Document Analysis and Recognition 1520-5363/11 2011 IEEE.

115. Rajiv Kapoor, Deepak Bagai, T. S. Kamal; 2002; Skew Angle Detection Of A Cursive Handwritten Devanagari Script Character Image; J. Indian Inst. Sci., May-Aug. 2002, 82, 161–175.

116. Rajiv Kumar, Amresh Kumar, Pervez Ahmed; A Benchmark Dataset for Devnagari Document Recognition Research; Recent Advances in Telecommunications, Signals and Systems; ISBN: 978-1-61804-169-2.

117. Rakesh Rathi, Ravi Krishan Pandey, Mahesh Jangid; 2012; Offline Handwritten Devanagari Vowels Recognition using KNN Classifier; International Journal of Computer Applications (0975 – 8887) Volume 49– No.23, July 2012.

118. Ratnashil N Khobragade, Nitin A. Koli, Mahendra S Makesar; 2013; A Survey on Recognition of Devnagari Script; International Journal of Computer Applications & Information Technology Vol. II, Issue I, January 2013 (ISSN: 2278-7720);

119. Ravi Sheth, N C Chauhan, Mahesh M Goyani, Kinjal A Mehta; 2011; Handwritten Character Recognition System using Chain code and Correlation Coefficient; International Conference on Recent Trends in Information Technology and Computer Science (IRCTITCS) 2011 Proceedings published in International Journal of Computer Applications (IJCA).

120. Reena Bajaj, Lipika Dey , Santanu Chaudhury; 2002; Devnagari numeral recognition by combining decision of multiple connectionist classifiers; Sadhana Vol. 27, Part 1, PP. 59–72.

146 Bibliography

121. Rohit Verma, Jahid Ali; 2012; A-Survey of Feature Extraction and Classification Techniques in OCR Systems; International Journal of Computer Applications & Information Technology Vol. I, Issue III, November 2012 (ISSN: 2278-7720).

122. S. Arora D. Bhattacharjee, M. Nasipuri, D.K. Basu, M.Kundu; 2011; Complementary Features Combined in a MLP-based System to Recognize Handwritten Devnagari Character;Journal of Information Hiding and Multimedia Signal Processing 2010 ISSN 2073-4212 Ubiquitous International Volume 2, Number 1, January 2011.

123. S. Arora, D. Bhattacharjee, M. Nasipuri, D.K. Basu, M.Kundu; 2009; Application of Statistical Features in Handwritten Devnagari Character Recognition; International Journal of Recent Trends in Engineering, Vol 2, No. 2, November 2009

124. S. Arora, D. Bhattacharjee, M. Nasipuri, D.K. Basu, M.Kundu, L.Malik; 2009; Study of Different Features on Handwritten Devnagari Character; Second International Conference on Emerging Trends in Engineering and Technology, ICETET-09.

125. S. Arora, D. Bhattacharjee, M. Nasipuri, D. K. Basu, M. Kundu; Recognition of Non-Compound Handwritten Devnagari Characters using a Combination of MLP and Minimum Edit Distance; International Journal of Computer Science and Security (IJCSS),Volume (4) : Issue (1).

126. S. K. Parui, K. Guin, U. Bhattacharya, B. B. Chaudhuri; 2008; Online Handwritten Bangla Character Recognition Using HMM; IEEE 978-1-4244-2175-6/08.

127. S. M. Mali; 2012; Moment And Density Based Hadwritten Marathi Numeral Recognition; Indian Journal of Computer Science and Engineering (IJCSE); ISSN : 0976-5166 Vol. 3 No.5 Oct-Nov 2012.

128. S.P.Kosbatwar , S.K.Pathan; 2012; Pattern Association for character recognition by Back-Propagation algorithm using Neural Network approach; International Journal of Computer Science & Engineering Survey (IJCSES) Vol.3, No.1, February 2012.

129. S.Ramasundaram, S.P.Victor ; 2010; Text Categorization by Backpropagation Network; International Journal of Computer Applications (0975 – 8887) Volume 8– No.6, October 2010.

130. S.V. Rajashekararadhya, P. Vanaja Ranjan; 2008; Efficient Zone Based Feature Extration Algorithm For Handwritten Numeral Recognition Of Four Popular South Indian Scripts; Journal of Theoretical and Applied Information Technology.

131. S.V. Rajashekararadhya, P. Vanaja Ranjan; 2009; A Novel Zone Based Feature Extraction Algorithm for Handwritten Numeral Recognition of Four Indian Scripts; Digital Technology Journal 2009, Vol. 2, PP. 41{51, ;ISSN 1802-5811 (print), ISSN 1802-582X (online).

147 Bibliography

132. Sandhya Arora, Debotosh Bhattacharjee, Mita Nasipuri, Dipak Kumar Basu, Mahantapas Kundu; 2008; Combining Multiple Feature Extraction Techniques for Handwritten Devnagari Character Recognition; IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.

133. Sandip N.Kamble, Megha Kamble; 2011; Morphological Approach for Segmentation of Scanned Handwritten Devnagari Text; International Journal of Computer Science & Technology Vol. 2, Issue 4; ISSN : 0976-8491 (Online) | ISSN : 2229-4333(Print).

134. Sang Sung Park, Won Gyo Jung, Young Geun Shin, Dong-Sik Jang;2008;Optical Character Recognition System Using BP Algorithm;IJCSNS International Journal of Computer Sci 118 ence and Network Security, VOL.8 No.12, ;

135. Sangame S.K., Ramteke R.J., Rajkumar Benne; 2009; Recognition of Isolated Handwritten Kannada Vowels; Advances in Computational Research, ISSN: 0975–3273, Volume 1, Issue 2, 2009, PP-52-55.

136. Satish Kumar; 2009; Performance Comparison of Features on Devanagari Hand-printed Dataset; International Journal of Recent Trends in Engineering, Vol. 1, No. 2.

137. Satish Kumar; 2010; An Analysis of Irregularities in Devanagari Script Writing – A Machine Recognition Perspective; (IJCSE) International Journal on Computer Science and Engineering Vol. 2, No. 2, 2010, 274-279;ISSN:0975-3397.

138. Shailedra Kumar, Shrivastava Sanjay, S. Gharde; 2010; Support Vector Machine for Handwritten Devanagari Numeral Recognition; International Journal of Computer Applications (0975 – 8887) Volume 7– No.11, October 2010.

139. Shailendra Kumar, Shrivastava Pratibha, Chaurasia; 2012; Handwritten Devanagari Lipi using Support Vector Machine; International Journal of Computer Applications (0975 – 8887) Volume 43– No.20, April 2012;

140. Shazia Akram Dr. Mehraj-Ud-Din Dar Aasia Quyoum; 2010; Document Image Processing - A Review; International Journal of Computer Applications (0975 – 8887) Volume 10– No.5, November 2010.

141. Shraddha V. Shelke, D. M. Chandwadkar; Automatic System for Recognition of Handwritten Character Using Multiscale Neural Network; International Journal of Computer Applications (0975 – 8887).

142. Snehal S.Patwardhan, R.R. Deshmukh; 2015; A Review on Offline Handwritten Recognition of Devnagari Script; International Journal of Computer Applications (0975 – 8887) Volume 117 – No. 19, May 2015.

143. Soumen Bag, Gaurav Harit; 2013; A survey on optical character recognition for Bangla and Devanagari scripts; Sadhana Vol. 38, Part 1, PP. 133–168. Indian Academy of Sciences.

148 Bibliography

144. Stuti Asthana, Farha Haneef, Rakesh K Bhujade; 2011; Handwritten Multiscript Numeral Recognition using Artificial Neural Networks; International Journal of Soft Computing and Engineering (IJSCE) ISSN: 2231-2307, Volume-1, Issue-1, March 2011.

145. Sukhpreet Singh, Renu Dhir; 2012; Recognition of Handwritten Gurmukhi Numeral using Gabor Filters;International Journal of Computer Applications (0975 – 8887) Volume 47– No.1, June 2012.

146. Suryaprakash Kompalli, Srirangaraj Setlur, Venu Govindaraju; 2009; Devanagari OCR using a recognition driven segmentation framework and stochastic language models; Springer IJDAR (2009) 12:123–138 DOI 10.1007/s10032-009-0086-8.

147. Sushama Shelke, Shaila Apte; 2011; Multistage Handwritten Marathi Compound Character Recognition Using Neural Networks; Journal of Pattern Recognition Research 2 (2011) 253-268.

148. Sushama Shelke, Shaila Apte; 2010; A Novel Multi-feature Multi-Classifier Scheme for Unconstrained Handwritten Devanagari Character Recognition; 12th International Conference on Frontiers in Handwriting Recognition 978-0-7695-4221-8/10 2010 IEEE.

149. Sushama Shelke, Shaila Apte; 2011; A Multistage Handwritten Marathi Compound Character Recognition Scheme using Neural Networks and Wavelet Features;International Journal of Signal Processing, Image Processing and Pattern Recognition Vol. 4, No. 1, March 2011.

150. Swapnil A. Vaidya, Balaji R. Bombade; 2013; A Novel Approach of Handwritten Character Recognition using Positional Feature Extraction; International Journal of Computer Science and Mobile Computing IJCSMC, Vol. 2, Issue. 6, June 2013, pg.179 – 186;ISSN 2320–088X.

151. T. Mondal, U. Bhattacharya, S. K. Parui, K. Das; 2010; On-line handwriting recognition of Indian scripts – the first benchmark; 2010 12th International Conference on Frontiers in Handwriting Recognition; 978-0-7695-4221-8/10 2010 IEEE.

152. U. Bhattacharya, B. B. Chaudhuri; 2005; Databases for Research on Recognition of Handwritten Characters of Indian Scripts; Proceedings of the 2005 Eight International Conference on Document Analysis and Recognition (ICDAR’05) 1520-5263/05 2005 IEEE.

153. U. Pal. B.B. Chaudhari; 2001; Machine Printed and handwritten text lines Identification;Elsevier Science Pattern Recognition Letters 22(2001) ;0167-8655/01.

154. U. Pal., B.B. Chaudhari; 2004; Indian Script Character Recognition : a Survey;Pattern Recognition (37) 2004.

155. V.N. Manjunath, Aradhya S.K., Niranjan G., Hemantha Kumar; 2010; Probabilistic Neural Network based Approach for Handwritten Character Recognition; Special Issue of IJCCT Vol.1 Issue 2, 3, 4; 2010 for International Conference [ACCTA-2010], 3-5 August 2010.

149 Bibliography

156. Vandana Korde, C Namrata Mahender; 2012; Text Classification And Classifiers: A Survey; International Journal of Artificial Intelligence & Applications (IJAIA), Vol.3, No.2, March 2012.

157. Vandana M. Ladwani, Latesh Malik; 2010; Survey Of Various Approches Towards Handwritten Devanagari Word Recognition; International Journal on Computer Engineering and Information Technology; ISSN 0974-2034

158. Ved Prakash Agnihotri; 2012; Offline Handwritten Devanagari Script Recognition; I.J. Information Technology and Computer Science, 2012, 8, 37-42.

159. Veena Bansal, R. M. K. Sinha; 2000; Integrating Knowledge Sources in Devanagari Text Recognition System; IEEE Transactions On Systems, Man, And Cybernetics—Part A: Systems And Humans, Vol. 30, No. 4, ;1083–4427.

160. Vijay Kumar, Pankaj K. Sengar; 2010; Segmentation of Printed Text in Devanagari Script and Gurmukhi Script; International Journal of Computer Applications (0975 – 8887) Volume 3 – No.8, June 2010.

161. Vijaya Rahul Pawar, Arun Gaikwad; 2014; Multistage Recognition Approach for Offline Handwritten Marathi Script Recognition; International Journal of Signal Processing, Image Processing and Pattern Recognition Vol.7, No.1 (2014), PP.365-378 http://dx.doi.org/10.14257/ijsip.2014.7.1.34.

162. Vikas J Dongre, Vijay H Mankar; 2010; A Review of Research on Devnagari Character Recognition; International Journal of Computer Applications (0975–8887) Volume 12– No.2, November 2010.

163. Vneeta Rani, Pankaj Kumar; 2013; Problems of Character Segmentation In Handwritten Text Documents written in Devnagari Script; International Journal of Advanced Research in Computer Engineering & Technology (IJARCET) Volume 2, Issue 3, March 2013;ISSN: 2278 – 1323.

164. Xian Wang, Venu Govindaraju, Sargur Srihari; 2000; Holistic recognition of handwritten character pairs; Elsevier Science Ltd Pattern Recognition 33 (2000) 1967}1973 0031-3203/00.

165 http://marathi.tripod.com/marathi.html 166 http://learnmarathithroughai.weebly.com/combining-consonants-in-marathi 167 http://mindurmarathi.com/marathi-vocabulary 168 http://marathi.tripod.com/marathi.html