Automatic prototype stroke generation based on stroke clustering for on-line handwritten Japanese character recognition

An automatic method for generating prototype strokes is proposed for on-line handwritten character recognition. The method consists of two steps: an intra-category step and an all-category step. In the first step, which has already been proposed by the author, handwritten stroke examples are clustered to obtain prototype strokes for each category. In the second step, these prototypes are merged to obtain common prototype strokes for all categories. This two step approach alleviates the difficulty in determining initial clusters for a large number of examples. To assess the quality of them, recognition experiments are conducted to see the relationship between the number of prototypes and the accuracy. It is found that the two relationships in kanji categories and in non-kanji ones differ from each other. This observation indicates that a set of prototype strokes for Japanese characters can be made of common prototypes for all kanji categories and prototypes for each non-kanji categories