Handwritten numeral recognition with the improved LDA method
Takahiko Kawatani, H. Shimizu, Matthew McEachern · 1996
We describe the improved LDA (learning by discriminant analysis) method proposed originally by one of the authors and its application to handwritten numeral recognition. In the LDA method, the discriminant function obtained by applying the Fisher linear discriminant analysis is superposed onto the original distance function such as the weighted Euclidean distance. Two improvements were made. One is the introduction of iteration to reduce the number of patterns which turn out to be misrecognized after learning. The other is to reduce the influence of asymmetry which takes place because of using quadratic terms in the discriminant function as linear terms. Experiments using the NIST database proved that both are effective to improve recognition accuracy. The misrecognition rate for the test data reduced to about 20% of that before learning. We obtained 99.02% recognition rate for the test data of which the recognition rate by humans is about 99.3%.