Study Of The RMB Series Number Recognition Technology Using Single CIS

Now there are a lot of counterfeit banknotes on the market, which not only cause heavy losses of property to the people, but also disrupt the finance markets. As the RMB serial number is unique, this feature could be used to distinguish counterfeit banknotes. Now the financial devices, which have the function of recognize series number, are usually equipped with double CIS(Contact Image Sensor) or CCD(Charged Coupled Device). The CCD is so expensive that it can't widely available. In some financial facilities, there are other parts on both sides of the paper. In some financial devices, there are other parts on both sides of the banknotes, which can't meet the installation requirements of the double CIS. This drawback has been solved in this dissertation which adopts the method of using single CIS, and the recognition rate is higher than 99.7% which can meet the requirements of the national, and the rate can over 500 every min which reach the requirement of ATM.The captured image which collected by single CIS having both sides of RMB which results in both the background of the image is complex and the noise of the image increasing, thus the image preprocessing is particularly important. Preprocessing in this dissertation include edge detection, tilt correction, noise removal, image binaryzation, series number region extraction, character segmentation and normalization. And the noise removal and binaryzation are the most important operation. The noise removal is to remove salt-and-pepper noise and noise isolation. Since the image background is complex, we use a local dynamic binaryzation method in this dissertation with both rate and binarized results are very good. Ten character images can be gotten after preprocessing.According to the character of series number, three kinds of character recognition algorithms, including commonly used template matching method, the method basing on character structural features and self-learning artificial neural network method, are designed and realized in the dissertation. Because of the gray value of the character image is only two, this characteristic is used to simplify matching rate calculation process, which greatly reduces the computation of the template matching method. According to the characters of the characters, a stable characteristic tree is created. According to the characteristics of the digital set, the artificial neural network with fast training speed and high recognition rate is established, and the recognition rate can reach 99.9%.Based on the advantages of these three kinds of character recognition algorithms, the recognition system of series number is established. The first four characters using the template matching method combined with method of characteristics to identify, which make up the shortcoming of the similar font character recognition rate is low when using template matching method. After 6 characters, the artificial neural network algorithm is used to identify, which full play the advantages of neural network method for digital character recognition rate.Two improvements of recognition algorithm are analyzed. According the point that there are only one letter between the second and the forth character to improve the union algorithm, making full use of the result of template matching method to make sure the the number of letter is only one, which can avoid recognition error efficiently. And the other one is that some improvements of standard deviation back propagation algorithm are proposed, including adding momentum and the adaptive learning rate, which can improve the learning rate fifty percent faster.According to the result of experimental, the system recognition rate is 99.73%, which reaches the national standard. System identification rate is 650 every min, which reach the requirement. Currently, the recognition system designed in this dissertation has been used in ATM on the market.