Application of bidirectional probabilistic character language model in handwritten words recognition

This paper presents a concept of bidirectional probabilistic character language model and its application to handwriting recognition. Character language model describes probability distribution of adjacent character combinations in words. Bidirectional model applies word analysis from left to right and in reversed order, i.e. it uses conditional probabilities of character succession and character precedence. Character model is used for HMM creation, which is applied as a soft word classifier. Two HMMs are created for left-to-right and right-to-left analysis. Final word classification is obtained as a combination of unidirectional recognitions. Experiments carried out with medical texts recognition revealed the superiority of combined classifier over its components.