Compression of Printed English Characters Using Back Propagation Neural Network

Abstract

In this paper, image data compression algorithm is presented using back propagation neural networks. Back propagation is used to compress printed English characters by training a net to function as an auto associative net (the training input vector and the target output vector are the same) with fewer hidden units than there are in the input or output units. The input and the output data files are formed in +1 and −1 form. The network parameters are adjusted using different learning rates and momentum factors. Mainly, the input pixels are used as target values so that assigned mean square error (MSE) can be obtained, and then the hidden layer output will be the compressed image. The proposed algorithm has been implemented in MATLAB to simulate the algorithm for the English characters A, B, C, D, E. The results obtained, such as compression ratio, mean square error, number of epoch for different learning rates and momentum factors are presented in this paper. Hebbian learning rule and Delta learning rules are used to train the network. Sigmoidal function, binary sigmoidal function and bipolar sigmoidal function are used in feed forward net and back propagation net respectively.