Research And Realization Of Renminbi Clearing Method Based On Image Analysis

The paper currency sorter is an important tool for banks and other financial institutions in paper currency sorting, counting, and identifying counterfeit money. It can do the banknote sorting work fast and efficiently instead of manual work. The critical technology of bills clearing system is based on the image analysis and processing of paper currency.The paper use reflection, transmission and color images of RMB and theories that include digital image processing, machine learning and neural network to achieve these algorithms include preprocessing of RMB images, detecting note’s direction and denomination, identifying new and old bills and recognition of RMB crown word number. The main contents are as follows:(1) In the image preprocessing module of this paper, firstly, according to the gray scale difference between target and background, we put forward a fast method to find these dots that in the RMB image edges. Then the least square method is applied for fitting a straight line. Finally, we obtain four boundary lines of the banknote image. The paper correct banknote image by translating pixels instead of a normal algorithm that rotate pixels for tilt correction for banknotes. The computation loads of the method is low, so it’s speed is faster than rotating pixels.(2) In the module of detecting note’s direction and denomination, the grid feature is applied. The features are obtained from banknote gray and color image. When we extract features from color image, the RGB space and H in HSV space is used. For classifier section, we select LVQ for classification. In this paper, a simple and reliable bill identification system is designed that the value of the bill were successively oriented identification(3) In the part of detecting new or old bills, the RMB transmission image is analyzed. The paper get feature from the histogram of the transmission image. We used the classification include SVM and KNN.(4) In RMB character recognition module, we used eight directions gradient feature and Gabor feature as character feature, and applying linear discriminant analysis (LDA) on the extracted character feature further dimensionality reduction. The character recognition classifier, using convolution neural network (CNN), support vector machine (SVM) and modified quadratic discriminant function (MQDF), designed a three-floor cascade classifier, and refused to recognize the credibility of not high of samples to ensure the accuracy and reliability of the character recognition system. On NUST-RMB2013 datasets, when the case was 2.61% of refusing recognition, character recognition accuracy rate can reach 100%.