Fast and precise discriminant function considering correlations of elements of feature vectors and its application to character recognition

During the late few years, research in recognition of handwritten Chinese and Japanese characters has matured significantly. However, in order to obtain high recognition rate, most character recognition systems have paid extremely expensive computational cost. For high performance character recognition systems, how to reduce the expensive computational cost is a very important problem now. Discriminant function is a very important factor for precise pattern recognition. The Mahalanobis distance is considered as an effective function. However, to calculate the Mahalanobis distance precisely, extremely large number of training samples are needed. In this paper, by investigating the relationship of elements of feature vector, a new discriminant function, called vector-divided Mahalanobis distance, is proposed. With the proposed method, high recognition performance can be obtained with less computational cost. Because the proposed method partitions high dimensional feature vector into several small number dimensional vectors, the ratio of the number of training samples to the number of dimensions becomes larger. This method is especially effect in the case of lack of training samples. The effectiveness of the proposed method is shown by the experimental results with the database ETL9B.