Segmentation of Handwritten Kanji Numerals Integrating Peripheral Information by Bayesian Rule

We have developed a new segmentation method for Kanji numerals. Segmentation of Kanji numerals written vertically i s dZcu l t because part of one Kanji numeral pattern can be read as another Kanji numeral character. In our method, Kanji numerals are segmented correctly by using peripheral in for ma tion and the similarity given by the character classifier. However, the usefulness of this information and the similarity differs depending on the character categories. Thus we represented the degree of usefulness by a likelihood ratio that i s selected according to the character categories. The likelihood ratios are integrated ideally by using the Bayesian rule. When implementing this method, we achieved character segmentation accuracy of 77%, and address recognition accuracy of 63%. In both cases our obtained accuracy was much higher than with a con ven tional segmentation method.