Handwritten numeral recognition using a small number of fuzzy rules with optimized defuzzification parameters

Handwritten numeral recognition using a small number of fuzzy rules with optimized defuzzication parameters is presented in this paper. The technique deals with the problem of performance degradation resulting from the minimization of a fuzzy rule set. In the learning phase, the Kohonen self-organizing map (SOM) algorithm is applied to organize the feature map. The resulting weight vectors are considered as prototypes that, together with corresponding variances, are used to determine fuzzy membership functions. Fuzzy rules are then generated by learning from the same SOM prototypes and the defuzzication parameters are determined by training a three-layer feedforward network. Experiments on NIST Special Database 3 of more than 20,000 handwritten numerals show that with the optimized defuzzification parameters, the classier based on a small number of fuzzy rules can achieve a classification rate of 96.3%.