On-Line tamil characater recognition using neural networks

Abstract

In this thesis, we design a neural network-based online Tamil character recognition system. The thesis focuses on the following areas: Character representation, Neural network architecture, and Training algorithms for neural networks. These issues are important since they affect the recognition accuracy of the system. Experiments with different character representations, various neural network architectures, and training algorithms were carried out. To begin with, we used a character representation that was previously used for recognizing digits and uppercase hand-printed English characters. This representation required a large network with several hidden layers and many connection weights for recognizing the Tamil characters, and learning was also slow. It was found that a single hidden layer network is sufficient to solve the Tamil character recognition problem if wavelet features are used for representing the characters. We also observed that if we use wavelet features for representing the characters, simple algorithms for avoiding overfitting also work well. The outcome of the thesis is a highly accurate online recognition system for the entire set of Tamil characters.