Abstract

Paper checks play an important role in our modern commercial society. However, Reading paper checks and typing into computers manually is a time-consuming and labor-intensive task. Therefore, accurate and efficient Check Reading Systems (CRS) are in high demand by banks. In this thesis, a CENPARMI CRS is introduced. Its courtesy amount recognition module is then presented in depth. The module makes use of a new feedback-based segmentation algorithm. The original segmentation algorithm can segment the inputted numeral string only once. The new algorithm can adjust the parameters of the segmentation algorithm according to the feedback information from the digit recognizer and can re-segment the inputted numeral string multiple times if necessary. Two rejection strategies in the new segmentation algorithm are also presented. The first strategy is used for rejecting unreasonable segmentations, which can be accepted by the individual digit recognizer. The second strategy is used to avoid accepting a digit string as one digit. Moreover, three algorithms in pre-processing and post-processing modules are proposed: a new border noise removal algorithm, an extra punctuations removal algorithm, and an implicit decimal points detection algorithm. Finally, a Convolutional Neural Network (CNN) recognizer is integrated into the CRS to recognize "00" and "000" directly in order to avoid segmenting "00" and "000". The experiment, based on the Bell Quebec check database, has shown that the courtesy amount recognition rate of the CRS has improved from 41.2% to 74.3%. The CRS can produce preliminary courtesy amount recognition results, which can be used for validation with the legal amounts to suppress the substitution rate.