Detection and correction of recognition errors in check reading

Abstract

An important subtask in the automated reading of bank checks by an OCR program is the combination of the recognition results of the courtesy and the legal amount. In this paper we present some new techniques to solve this problem. First, a systematic approach for the translation of the legal amount into a digit string is described. The technique is based on syntax-directed translation. Once the result of the legal amount recognizer has been translated into the corresponding digit string, it can be easily compared to the result of the courtesy amount recognizer. Thus, inconsistencies between the legal and the courtesy amount can be detected and precisely located at the level of individual subwords in the legal amount and digits in the courtesy amount. Moreover, recognition errors can be potentially corrected. In experiments with real data from Swiss postal checks a significant improvement in the acceptance rate could be achieved without increasing the error rate of the overall system.