Hand-written Character Recognition System Using Uniform Division of Training Samples on subspace Method

Abstract

In spite of the fact that subspace method can approximate the distribution of categories precisely, only a few attempts have so far been made at applying it in hand-written character recognition. The subspace method proposed by Watanabe (1969) offers the basic concept of subspace construction, but the issue of how to use the limited samples to construct effective subspace to avoid the problem of mis-recognition caused by the fact that the subspace of a category is almost parallel to the mean vector of the category remains unresolved. This problem leads to the mis-recognition of the samples far from the mean vector. To cope with this problem, Abe, Nemoto and Sun (1995) have proposed the Combination method (CM), which constructs the subspace from several groups including different number of samples divided from the whole training samples. CM obtained a high recognition rate of 97.76% with respect to ETL9B, the largest database of hand-written characters in Japan. ETL9B was published by Electro-technical Laboratory of Japan in 1985. It includes 2,965 categories of Chinese characters and 71 categories of Japanese Kana, there are 200 data sets in the database where each set has one sample per category. Total number of samples is 607,200. The issues that need to be dealt with next are how to improve the recognition accuracy and how to accelerate the recognition speed. In this paper, we propose a new method called Uniform Division Method (UDM), which uses the uniformly divided training samples to construct subspace. Compared to CM given earlier, UDM is very simple and effective enough to improve the accuracy of recognition; as a result, we obtained a recognition rate of 98.64% for ETL9B compared to the 97.76% for CM. This is the first time that such a high recognition rate has been obtained by making good use of subspace method. Furthermore, the computation required for UDM is less than a half of that of CM. The UDM algorithm and the experiments with ETL9B will be described in this paper.