Support Vector Machines for Odiya Handwritten Numeral Recognition

Abstract

Handwritten Numeral recognition has its own importance over various fields like automatic pin code reading, bank check processing, form processing, etc. Handwriting numeral recognition is a subset of character recognition. But handwriting recognition has been always a complex task due to variations of handwriting shape, size, and stroke among different writers. This paper focuses on automatic offline recognition of isolated odiya handwritten numerals recognition using two features extraction techniques - Fourier Descriptors and Normalized Chain Code. Once the features are extracted for the numerals, the features are fed into two different machine learning technique i.e., Back Propagation Neural Network (BPN) and Support Vector Machines (SVM) for classification and recognition. A total set of approximately 20000 handwritten numerals data from 200 different classes of people are collected and are considered for the experimentation of the proposed methods. The classification results from the two machine learning techniques are then compared. We have obtained a recognition accuracy of 83.33% & 93.4% with fourier descriptor feature extraction techniques for BPN and SVM respectively. In case of normalized chain code feature extraction technique we obtained 83.63% and 93.57% for BPN and SVM respectively. The experimented result shows SVM outperform the BPN with both the feature extraction techniques for recognizing and classifying odiya numeral.

Keywords: Handwritten Numericals, Feature Extraction, Fourier Descriptors, Normalized Chain Code, Back Propagation Neural Network, Support Vector Machines.