Contextual word recognition using probabilistic relaxation labeling

摘要

A postprocessing method based on probabilistic relaxation labeling is described. The initial label probabilities of this process are determined from the similarities between characters in the input and model characters in the database. The constraint relations for this relaxation process are determined from the transition probabilities between input characters. The transition probabilities themselves are determined from frequencies of letter pairs in the underlying language. Preliminary results are presented that show the performance of this postprocessing method on printed text.