Recognition of Online Handwritten Gurmukhi Strokes Using Support Vector Machine

Abstract

Pen-based interfaces are becoming more and more popular and play an important role in human-computer interaction. This popularity of such interfaces has created interest of lot of researchers in online handwriting recognition. Online handwriting recognition contains both temporal stroke information and spatial shape information. Online handwriting recognition systems are expected to exhibit better performance than offline handwriting recognition systems. Our research work presented in this thesis is to recognize strokes written in Gurmukhi script using Support Vector Machine (SVM). The system developed here is a writer independent system.

First chapter of this thesis report consist of a brief introduction to handwriting recognition system and some basic differences between offline and online handwriting systems. It also includes various issues that one can face during development during online handwriting recognition systems. A brief introduction about Gurmukhi script has also been given in this chapter In the last section detailed literature survey starting from the 1979 has also been given.

Second chapter gives detailed information about stroke capturing, preprocessing of stroke and feature extraction. These phases are considered to be backbone of any online handwriting recognition system.

Recognition techniques that have been used in this study are discussed in chapter three. In this chapter Support Vector Machine is discussed and two cross validation techniques namely holdout and k-fold have been discussed.100 words that are used in this study are also given in this chapter.

Chapter 4 contains the results that have been calculated after applying Support Vector Machine in MATLAB for three different writers. These results have been calculated for each partitioning techniques. For holdout 70%, 60% and 50% training sets have been used for stroke recognition. In k-fold technique 10-fold has been used. Both of these partitioning techniques have been used for both non preprocessed and preprocessed strokes. Also results for recognition of 1000, 2000 and 3000 preprocessed strokes have been shown in this chapter. Last chapter of this work concludes all the work that has been done and gives some directions in which more work can be done to improve the recognition system.