Markov random field-based hierarchical handwritten Chinese character modeling

Abstract

This dissertation proposes a statistical-structural character modeling method based on the Markov Random Fields(MRFs) in the Handwritten Chinese Character Recognition(HCCR) problem. In the MRF framework, we view the character recognition as a labeling problem, as how well a given observation matches a character model. The MRF framework can represent both statistical and structural information of the Chinese character by the neighborhood systems and clique potentials. The neighborhood system denotes the most important stroke relationships. The clique potential is composed by prior clique potential based on our prior knowledge and likelihood clique potential represents both statistical and structural information, which is derived from Gaussian Mixture Models(GMMs). We add the radical information into our prior knowledge, thus form a hierarchical character structure in which radicals constitute characters, and strokes constitute radicals. With the help of radical structure, we can easily grasp the most important stroke relationships. We implemented a real-world application of character recognition. In the proposed HCCR system, we extract candidate strokes from character image by minimizing the single-site likelihood clique potentials, and find the best structural match between candidate strokes and stroke models by the relaxation labeling algorithm. The experiments done on the Korea Advanced Institute of Science and Technology (KAIST) character database demonstrate the practicability of proposed approaches.

Date of Award	4 Oct 2010
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Zhi-Qiang LIU (Supervisor)