Markov random fields for handwritten Chinese character recognition

In this paper, we propose a statistical-structural scheme for Chinese character modeling based on Markov random fields (MRFs). We use 2-D Gabor filters to extract directional stroke segments from images of Chinese characters, where each stroke segment is associated with a state in Markov random field models. The structural information is described by neighborhood system and pair-state clique potentials; meanwhile the statistical information is represented by single-state probability density functions (pdfs). Extensive experiments on similar characters have been carried out on the database ETL9B. The experimental results confirm that Markov random field models are effective in modeling both statistical and structural information of Chinese characters, and works well for handwritten Chinese character recognition.