Construction of Multi-Template Dictionary for Pattern Recognition by Decreasing Region Overlaps

In statistical pattern recognition problems, it is important to consider distribution of patterns. In this paper, we propose an algorithm to construct a multi-template dictionary for precise and efficient pattern recognition by considering the distributions. First, by considering relation between distributions of different categories, the possibility that a pattern may be misclassified is investigated. Then an algorithm of dividing a category into subclasses to reduce the possibility of misclassification by decreasing region overlaps is proposed. The proposed algorithm is applied to Japanese character recognition problem which requires high dimensional feature vectors. Experimental results show the effectiveness of the proposed algorithm.