Research And Implementation For Ocr-based Check Recognition System

With the development of social economics and information technologies, commercial banks need to deal with a great lot of checks every day. However, the checks are being processed manually, which can not meet the needs of social development. The check recognition technique uses OCR algorithms to recognize the data and characters on the check. The main contributions of this paper are summarized as follows:(1) Present an OCR-based check recognition system called CheReS. The check recognition algorithms includes preprocessing, check layout analysis, character string extraction and character segmentation, feature extraction and character recognition, results cross-validation. The CheReS automatically achieves recognition procedure and obtain the results. The experimental results demonstrate that our system can process checks efficiently and recognize characters in different check fields exactly.(2) Realize a check layout analysis algorithm based on rectangle line detection. We first create the check template with marked recognition regions, and then achieve rectangle line detection by a line-growth method based on window scanning. So we locate the template in the image and obtain the position of recognition regions according to the detected rectangle lines. This paper uses algorithms including projection histogram analysis and connected region analysis to robustly achieve character segmentation for three abnormal situations. (3) Present a character segmentation algorithm based on connected region analysis and character recognition. To make the character segmentation results more accurate, the method provides evaluation and feedback results during character segmentation. For the numeric characters are restricted by black grids, we realize a character segmentation method based on grid membership analysis.(4) Realize a feature extraction method for Chinese characters based on membership analysis on four directions and a multiple classifier fusion algorithm based on Minimum Mean Square Error for Chinese character recognition. The artificial neural network algorithm is used to recognize numeric characters in this paper. To improve the accurateness of character recognition, we present a cross-validation method for digit characters with useful information from multiple recognition regions.