Bank Cheque In Handwritten Application Domain String Recognition

Handwritten Chinese string recognition is one of the most important parts of the bank note automatic processing system. In this paper, the recognition of handwritten Chinese string in the usage area on bank check is studied. A concrete and effective solution is presented. The following is the main content of the paper.The preprocessing for the string image which includes removing seals by gradation, removing frame lines and locating the string accurately.The segmentation of the string image. In this stage, a contour tracing method is firstly applied to get the connected domains. After combining the connected domains up and down, left and right, we get the initial char parts. Those char parts which has touching characters are detected before being cut into smaller ones. At last, several cutting plans are generated through combining neighboring char parts.The normalization and feature extraction of single Chinese char. A nonlinear normalization method based on stroke density equalization is improved in order to reduce differences between gray char images within the same class. Then we extract eight direction gradient features on the gray char image separated by fuzzy grid.The classification of single Chinese char. A three-step classifier is designed for the concrete situation. First, the initial candidates are selected by the position of the char in the string. Then the candidates are reduced by a rough classification method. At last, we use a pricier classifier to get the final eight candidates for each single char.The post processing of the Chinese string. We first applied a SLM (Statistical Language Model) based method. The string matching method is then proposed which can highly improve the string recognition rate.We use the complete solution described in this paper to recognize1000testing samples while the training set containing3000string images. The correct rate is75.6%.