Character Recognition with Mahalanobis Distance Based on Between-cluster Information

In the case of using the Mahalanobis distance as discriminant function, usually the covariance matrix calculated from training samples is used. However, it is extremely difficult to prepare enough training samples if the dimension of feature vector is large. Therefore, estimated eigenvalues and eigenvectors of covariance matrix will include errors that cause misclassification. In this paper, a new method to construct an effective discriminant function is proposed by considering between-cluster information. In the proposed method, if the number of calculable eigenvalues and eigenvectors is not enough because of less training samples, some new axes are constructed and then the pseudo-variances are computed based on the between-cluster information. The effectiveness of this method is shown by the experiments with handwritten characters. key words character recognition, mahalanobis distance, between-cluster information, estimate error, ETL9B