Bi-character model for On-line Cursive Handwriting Recognition

Abstract

This paper deals with on-line cursive handwriting recognition. Analytic approach has got more attraction during the last ten years. It relies on a preliminary segmentation stage, which remains one of the challenges and might have a strong effect to the correct recognition rate. The segmentation  aims  to  cut  the  ink  strokes into a set of small pieces,  called graphemes. The recognition process tries to combine them to build  different segments of cursive pattern, which correspond to individual characters in the strokes.This is not a trivial process because there is no effective algorithm to decide which grapheme belongs to which character.  Traditionally, the recognition process makes different assumptions about word segments which corresponding to the characters presenting in the cursive  handwriting  pattern. Then, the recognition process chooses the best possibility based on the probabilities of the recognition results. However, there is very little information to validate or re-evaluate that “the best possibility” is appropriate in the real world. In order to overcome this problem, thispaper introduces a bi-character model, where each character is recognized jointly with its neighbor. It offers a possibility to validate a segment of word (with its neighbor) to see if it is a correct segmentation (respecting to a character). The experimental investigation  on a standard dataset illustrates that the proposed model has a significant  contribution to improve the recognition rate. In fact, the recognition rate is move from 65% to 83% by using the bi-character model.