Character recognition with fuzzy features and fuzzy regions
Eric R. Mertooetomo, Jianhua Chen · 2002
The authors propose a method for character recognition using fuzzy features and fuzzy regions in a neural network. The method is robust to noise and distorting, scaling, and shifting of the patterns within their pixel frames, yet it is mathematically simple. The fuzzy neural network presented in the paper consists of three layers: a layer for feature extraction, for regional emphasis of features, and for classification. They extract features from regions of the characters in which they are most likely to occur. To make the system robust these regions are fuzzified, giving higher weight to areas where the features are most likely to occur and lower to areas where the features are rare. Sample patterns from the literature have been used for training of the network to obtain the minimal set of distinguishing features with their associated measures and to determine the optimal slopes of these linear regions. The network has been tested using patterns from the literature. Its performance is comparable for distorted and noisy patterns and superior for shifted, partial, and down-scaled samples.