HANDWRITTEN CHARACTER RECOGNITION USING RADIAL BASIS FUNCTIONS

Abstract

This thesis presents a character recognition system implemented using radial basis function networks. A recognition system was first implemented to deal with handwritten digits before it was extended to cover the entire alphanumeric character set. In the handwritten digit recognition system, an input image was first passed through an elaborate pre-processing routine before its features were extracted. Using a radial basis function network as classifier, recognition rates of close to 99% was achieved. In the handwritten alphanumeric character recognition system, a dual classifier approach was adopted. In this system, two separate radial basis function classifiers, each extracting a different set of features from the input image, were used. Each classifier then makes its own independent classification decision on the input image before these decisions were combined to give an overall system classification output. Recognition rates of up to 97.42% were achieved using this approach. The radial basis function classifiers mentioned earlier were all trained using a class-based clustering algorithm. While this method has proved effective in classification problems, other alternative methods of training such networks, such as the use of genetic algorithms, are possible. In this thesis, a genetic algorithm was used to search for an optimal set of centers for the radial basis function classifier. Its performance on the handwritten digit recognition task was evaluated and compared with the results achieved using the class-based clustering algorithm. It was shown that with this method, recognition rates of better than 97% can be achieved.