A contextual postprocessing for optical Chinese character recognition

A method of improving the recognition rate for optical Chinese characters by incorporating statistical language knowledge for post processing is described. Character occurrence probability p(c_i), and a set of, first-order conditional probabilities, p_k(c_i/c_j), where c_i follows and is apart from c_j by k-1 intermediate characters (1⩽k⩽N), are compiled and regarded as contextual knowledge, and a delayed decision greedy algorithm is suggested to search for a suboptimal solution along a trellis-like structure for a real-time implementation. Both the computing and the storage demands of this approach are much less than that of the Viterbi algorithm with N-th order conditional probabilities, and only a minor performance degradation is experienced