Markov random field models for handwritten word recognition

We propose the use of a Markov random field model for handwritten word recognition. The main advantage of Markov random field models is that they provide flexible and natural models for the interaction between spatially related random variables in their neighborhood systems via clique functions. In our scheme, Gabor filters are adopted for feature extraction. A fuzzy neighborhood system is proposed and fuzzy matching measurements are developed to cope with the variability of handwritten word shapes. A relaxation labeling algorithm is used to maximize the global compatibilities of Markov random fields. The influence of neighborhood sizes and the iteration number on recognition rates of the system is investigated. Our initial experiments have shown encouraging results.