A segmentation system for touching handwritten Japanese characters

Abstract

The present character recognition system needs the segmentation process as preprocessing for an input image of the touching character string. The process divides it into some isolated characters. To improve the total performance of the character recognition system, it is required to lessen the number of retrieval path in the candidate character lattice. We proposed the segmentation method based on connecting condition of the neighbor lines to resolve these problems in the previous paper. In this paper, we modify, our segmentation algorithm, and we evaluate the performance of these segmentation methods by a direct evaluation system with permissible degree and an indirect evaluation using the character recognition. We confirm the usefulness of our new method for 456 touching handwritten Japanese "Kanji" image with 948 characters by the comparison of three segmentation methods. The correct segmentation rates are 61.4% (conventional method), 75.1% (previous method), and 83.0% (proposed method).