Handprinted numeral recognition with the learning quadratic discriminant function

To enhance the recognition performance for character recognition, it is necessary not only to describe the distribution in each class as exactly as possible, but also to emphasize the distribution differences between the classes. The author realizes these approaches by applying the LDA (learning by discriminant analysis) method to the quadratic discriminant function (QDF) and the modified quadratic discriminant function (MQDF2). In the LDA method, the discriminant function obtained by applying linear discriminant analysis is superposed onto the original discriminant function. Both QDF and MQDF2 have a problem in that their performance is degraded if the distribution deviates from the normal distribution. The LDA method overcomes this problem. Experiments are performed for handprinted numeral recognition and MQDF2 yields the best result; the maximum recognition rate is 99.67%. The LDA method reduces the misread rate by 30%