Improving the Recognition Accuracy of Text Recognition Systems Using Typographical Constraints

Spelling correction techniques can be used to improve the recognition accuracy of text recognition systems. In this paper a new spelling-error model is proposed that is especially suited to the correction of recognition errors occurring during the recognition of printed documents. An implementation of this model is described that exploits typographical constraints derived from character shapes. In particular, the fact is used that vertical strokes in character images are seldom misrecognised. Experimental results show: 1) that the sizes of candidate word sets are substantially reduced; and 2) that the probability that the wrong candidate word is chosen is reduced by an average factor of approximately 2 when compared to spelling correction techniques without the use of typographical constraints.