IJSTE - International Journal of Science Technology & Engineering | Volume 1 | Issue 10 | April 2015

ISSN (online): 2349-784X

Weighted Zone based Kannada Character

Recognition Using Crack Feature Extraction
Technique And Support Vector Machine
Classification
Roopananda M K
M.Tech Student
Department of Computer Science & Engineering
UBDT College Of Engineering Davanagere Karnataka, India

Abstract
Many methods and technologies evolved in recognizing optical character but very less research made on recognizing handwritten
character because of the complexity of variations of individual way of writing style. Many systems are available for recognizing
optical character. In this paper, I have worked recognizing the character using crack feature extraction and well-known SVM
classifier but worked on new concept of WEIGHTED zone based. More weight need to be given on the zone which has
maximum number of character overlapping on that zone thereby increasing the efficiency and maximum probability of finding a
part of character.
Keywords: Crack Code, Support Vector Machine, Optical Character Recognition
________________________________________________________________________________________________________

I. CHARACTER RECOGNITION
Efficient methods involved in recognizing character that might be either optical character or handwritten character and also
differentiated into online and offline character recognition. Online recognizing is the one where as and when a character is
written it is recognized. Offline is the one where a digitalized image is fed as an input and thereafter preprocessing step is carried
out and then fed its output to feature extraction technique and SVM to classify the image and in recognizing a character.

II. A SURVEY
Handwritten Character Recognition is one of the emerging technologies in the field of archeology and crime detection of
recognizing the handwritten of either victim and suspects. Few are said below. Handwritten character recognition of Devangiri
Script using gradient feature technique [1]. Sobel and Robert operators are used for extracting gradient features and fed as input
to SVM to find and study accuracy. Modified quadratic discriminant function used for recognizing Offline Bangla Handwritten
Character [2]. Extracting Feature is purely on directional basis. For extraction technique, crack code concept is used where
number of vectors is used for extracting features. Fuzzy model based can be used for recognizing the Bangla Character[3].
Very Less attention has made in work and research of Kannada Handwritten Character Recognition. Pal et. al [4] has proposed
a technique of quadratic classifier based for recognizing offline handwritten character recognition of popular three south Indian
languages Kannada, Tamil and Telgu. Kannada Character Recognition using FLD based technique is yet another way of popular
way of recognizing a character [5].

III. SYSTEM ARCHITECTURE

Zone based handwritten character recognition has four steps.
1) Database Creation.
2) Pre processing
3) Feature Extraction
4) Classification
All the above four steps are very important and have equal importance that will contribute in recognizing character. All the four
steps involved above will be discussed in detail one by one.

All rights reserved by www.ijste.org

179

Weighted Zone based Kannada Character Recognition Using Crack Feature Extraction Technique And Support Vector Machine Classification
(IJSTE/ Volume 1 / Issue 10 / 034)

A. Database Creation:
Before carrying out preprocessing step first we need to create a database for making preprocessing step easier. Creating database
for handwritten character recognition is very difficult and it is no inbuilt way or tool to create database. One need to collect
different samples from people belonging to different department and should make database. Collecting database is difficult job as
their may be chances of unconstrained handwriting , broken strokes, poor binarization. Since there is no standard database for
handwritten kannada characters, we need to create our own database and store in a database

Fig. 1: Handwritten Kannada Characters

B. Pre-processing:
In preprocessing step, we need to read a digitalized image i.e. raw image. Usually raw image contain noise and inaccuracy. Noise
can be removed from median filtering thrice. After removing noise we need to do segmentation where we go for doing histogram
equalization, plotting histogram and image binarization is carried out using well-known Otsus method.
C. Feature Extraction:
Feature extraction is one of the difficult step in overall recognizing a character. As there are different techniques for extracting a
feature but the one technique I used is very familier one i.e... Crack Code
1) Crack Code:
It is one of the way to encode boundry by connected straight line of length and direction. Boundry line between foreground and
background that form or give raise to crack and encoding this leads to crack code concept.

Fig. 2:

Fig. 3

2) Steps Involved In Feature Extraction:

1) Step 1: Divide input image into 8x8 zones (thereby we will get 64zones)
2) Step 2: For each zone, we need to calculate or compute crack code by traversing the character portion in zone in
anticlockwise direction.
3) Step 3: For each zone, we need to calculate feature using eq(1) and eq(2)
4) Step 4: Repeat steps 1 and 2 till we get feature vector of size 512

All rights reserved by www.ijste.org

180

Weighted Zone based Kannada Character Recognition Using Crack Feature Extraction Technique And Support Vector Machine Classification
(IJSTE/ Volume 1 / Issue 10 / 034)

Fig. 4:

D. Support Vector Machine Classification:

Support Vector Machine is the best classifier that can be used for classification and for recognizing the character. SVM is used
to find out the optimal hyper-plane which maximizes the distance, margin and between the nearest of both the classes.
Given the training database M data {Xm | m=1,2,3,.M} the linear svm classifier is given as:

1) Weight Concept:
I have worked on weighted zone based character recognition, where the concept go like this. If we place all the images one above
the other as in stack format then we need to find on which zone more number of characters overlap thereby more number of
characters can be found in a same classes there we should give more weight, whereas less weight need to be given in other case
thereby we move towards achieving more accuracy in finding a character.

IV. CONCLUSION
Crack code feature extraction technique and SVM classifier combination gives best result in recognizing an character where we
get accuracy upto 80% and with a new added concept of weighted zone base gives 85% of accuracy compared to normal zone
based character recognition.

REFERENCES
[1]
[2]
[3]
[4]
[5]

Holambe, A.N, Thool, R.C and Jagade, S.M, "Printed and Handwritten Character & number Recognition of Devanagari Script using Gradient Features".
International Journal of Computer Applications vol. 2(9), 975-8887,2010
Pal. U, Wakabayashi and T. Kimura. F "Handwritten Bangia Compound Character Recognition Using Gradient Feature", In: 10 th International Conference
on Infonnation Technology , 17-20, pp. 208213, December-2007
Hanmandlu. M, Ramana Murthy and O.V Madasu V.K. "Fuzzy Model based recognition of handwritten Hindi characters",9th Biennial Conference of the
Australian Pattern Recognition Society on Digital Image Computing Techniques and Applications, IEEE pp. 454-461. December 3-5, 2007
Pal U, Shanna N, Wakabayashi T and Kimura. F, "Handwritten Character Recognition of popular South Indian Scripts", In: Doennann, D., Jaeger, S. (eds.)
SACH 2006. LNCS, vol. 4768, pp. 251-264. Springer, Heidelberg, 2008.
Niranjan. S.K, Hemantha Kumar and G. Aradhya, M,"FLD based Unconstrained Handwritten Kannada Character Recognition", International Journal of
Database Theory and Application 2(3) September 2009

All rights reserved by www.ijste.org

181