A New Information Combination Approach for Character Recognition with a Limited Lexicon

This paper presents a new information combination approach for character recognition by combining the sample-based similarity measure and the posterior probabilities of DHMMs (Discrete Hidden Markov Models). In the new method, a prototype is obtained for each class at the training stage besides an HMM. At the recognition stage, the sample similarity between an unknown sample and the prototype for a special class is calculated and normalized after feature extraction module. Then the normalized similarity measure is combined with the traditional DHMMs for classification. Experiments on off-line handwritten Chinese amount in words recognition show that the new method can effectively improve the recognition accuracy of the DHMMs-based single classifier, but the recognition speed declines little.