LBG Method and Vector Quantization for Recognition of Punjabi Handwritten Strokes in Hidden Markov Models

Abstract

Handwritten stroke recognition problem is being solved in the work specifically for Punjabi language. Fifty three different types of strokes, which constitute most of the Punjabi characters, are considered here. 242 different instances of each such stroke drawn by various people are used for the training and testing experiments. The recognition system proposed in this report performs quite well yielding high levels of recognition accuracy. This system will help in solving handwritten character recognition which would be a great tool for conversion of handwritten documents into computerized textual documents and recognizing handwritten phrases. Out of many machine learning models we have used hidden markov models (HMMs) for solving this problem which is a doubly embedded stochastic model. HMM performs fairly well in recognizing Punjabi strokes. We have created model for each stroke in our system