Stroke Segmentation of Chinese Characters Using Markov Random Fields

Abstract

This paper presents Markov random fields (MRFs) to segment strokes of Chinese characters. The distortions caused by the thinning process make the thinning-based stroke segmentation difficult to extract continuous strokes and handle the ambiguous intersection regions. The MRFs reflect the local statistical dependencies at neighboring sites of the stroke skeleton, where the likelihood clique potential describes the statistical variations of directional observations at each site, and the smoothness prior clique potential describes the interactions among observations at neighboring sites. Based on the cyclic directional observations by Gabor filters, we formulate the stroke segmentation as an optimal labeling problem by the maximum a posteriori (MAP) criterion. The results of stroke segmentation on the ETL-9B character database are encouraging.