Neural Network Approach for Recognition of Printed Telugu Characters

摘要

Automatic recognition of Indian language characters has a lot of applications in India. As an experiment for character recognition, a sub set of printed Telugu characters with 16 vowels is considered. The existing methods like statistical or syntactical approaches have disadvantages to classify characters which are noised and rotated. Neural Network techniques are used to over come these problems. After preprocessing i.e.binarization, thinning and size normalization, the entire image is given as input to the neural network. Various phases in character recognition are creation of Telugu alphabet database, training the neural net and testing the performance of the network. For training the neural network, Backpropagation with momentum and variable learning rate is used. The performance of the neural network with and without hidden layers is also discussed. The performance of the neural network with various input sizes 64 × 64, 32 × 32, 16 × 16 is also discussed. Original characters and noised characters are recognized accurately but few rotated characters are not recognized. We claim more than 90% recognition using neural network techniques.