Neural network model for invariant pattern recognition

Abstract

A pattern recognition system using Artificial Neural Network (ANN) classifier is proposed. The system is intended to recognize translated, rotated aod scaled versions of the exemplar patterns. The model proposed for this system consists of two parts. The first is a preprOCessor and the second is an ANN classifier. Preprocessing is done in two stages. In the first stage, projection from each active bit of the pattern is taken in such a way that for any rotated or scaled version of the exemplar pattern, the projected values become cyclically shifted of those of the exemplar pattern. For translated version of the exemplar pattern, the projected values remain same. The second stage performs Rapid Transform (RT) on the projected values. Thus, for an ideal case when rotated or scaled pattern is assumed to be noise free and distortionless, the outputs of the preprocessor are invariant to rotation, scal~g and translation. However, in practical case, rotation and scaling always insert some amount of noise in the input pattern. Therefore, the preprocessed outputs in response to a rotated or scaled pattern appear to be somewhat different from those of exemplar pattern. The second part of the system is an ANN classifier trained by backpropagation algorithm chosen especially for its good ability to deal with variation of inputs within the same category. The outputs of the preprocessor are then fed into the ANN classifier. In spite of the variation present in the preprocessed outputs due to rotation and/or scaling, the ANN classifier is expected to classify the input pattern correctly. The proposed model is tested extensively with teo numeric digits (0 - 9). With these patterns, the proposed model achieves considerably good degree of invariance to rotation and scaling. The performance of the system also depends on the number of input and hidden units of ANN classifier. The effect of classifier's size on the performance of the system is also studied.