On-line Handwritten Japanese Text Recognition by Improving Segmentation Quality

This paper describes a method of on-line handwritten Japanese text recognition by improving segmentation quality. The method produces hypothetical segmentation points according to features such as distance and overlap between adjacent strokes. Moreover, it extracts multidimensional features from these hypothetical segmentation points and applies an SVM to the extracted features to produces segmentation point probabilities. It constructs a candidate lattice traversing the hypothetical segmentation points, and evaluates the likelihood of the text candidate paths in the candidate lattice composed of character pattern size, character pattern inner gap, character recognition, single-character pattern position, paircharacter patterns position, character segmentation point probability and linguistic context. The likelihood of the text candidate paths are weighted using the number of hypothetical segmentation units with the weighting parameters trained by a genetic algorithm. An experiment on the database HANDS-Kondate_t_bf-2001-11 shows that this method improves segmentation rate and character recognition rate remarkably.