Hu Keun Kwak, Yeong Woo Choi, Kyu Sik Chung

The KIPS Transactions:PartB , Vol. 8, No. 1, pp. 10-19, Feb. 2001
10.3745/KIPSTB.2001.8.1.10,   PDF Download:

Abstract

Phoneme segmentation methods for handwritten Hangul developed so far have following limitations : 1) segmentation results depend heavily on the result of thinning, since most of the methods have applied to the thinned image; 2) they cannot handle the cases of touched phonemes where no or only bad feature points are found, since most of the approaches have developed by assuming that good feature points are found around the touching area of the phonemes. In order to overcome such limitations, we propose a new method of phoneme segmentation by applying the segmentation method to the input image directly without thinning and by developing methods to handle even the cases where no or bad feature points are found. Our method divides the cases of touched phonemes into several types and develops the segmentation method in each case based on structural information of vowels and their relative positions, touching shape and its heuristic rules. Our method proceeds as follows 1) vowels are traced in each character image 2) touching cases are classified according to vowel structure 3) for each case, touching type is decided; and 4) the corresponding segmentation method is applied. PE92 handwritten Hangul database is used for testing the segmentation method, and segmentation rate of 89.5% was obtained. This result shows effectiveness of the proposed method.

Statistics
Cite this article