Handwritten Kanji recognition with determinant normalized quadratic discriminant function

This paper describes two approaches to increasing the accuracy of character recognition. One possible approach is to improve quadratic discriminant analysis. If a sample covariance matrix is obtained by using a relatively small set of training data, estimation errors of its determinant differs from class to class. This is thought to lead a deterioration in performance. To cope with this problem, this paper outlines an approach based on normalization of the determinant of each class covariance matrix. Another approach is to use features which reflect differences in the shape of characters. The results of a benchmark test in which various feature transformation methods and discriminant functions were compared are reported. The tests confirmed that combination of normalizing quadratic discriminant function for the determinant and the use of the common difference principal components proposed by the author gives the best accuracy.