Research On Serial Number Of Banknotes Identification Algorithm

There are a lot of noise in the grayscale images scanned by equipment because of stains andcreases caused when the paper money is used. In order to better achieve the identification of the serialnumber, the banknote image should do denoising processing. There are two kinds of significant noisein the grayscale image of the banknote, one is the speckle in irregular shape and position, the other isthe straight line creases in uncertain position. In this paper, some ideas on image processing areannounced to remove these noise. An algorithm by regional average denoising is proposed on thespeckle in irregular shape and position. And taking advantage of crease’s penetrability in banknotes isto remove algorithm crease.The recognition algorithm is divided into two parts, head word recognition and numberrecognition, by analyzing the characteristics of the serial number of RMB. First three words of serialnumber is head word composed of letters and numbers. Algorithm based on BP neural network is toachieve the classification of the letters and numbers whose structure is various and complex. For theconfusing characters, a method of secondary identification based on BP neural network is proposed.In order to avoid that the training of BP neural network waste quite a long time and fall into localminimum, the BP neural network is optimized by immune algorithm. Last seven words of serialnumber is only composed of numbers. In order to achieve the classification of these numbers, acomplex identification approach with multi-feature is designed.Recognition experiments by MATLAB for1000serial number images containing differentprocedures defaced show that, the recognition rate is95.8%and the recognition speed is about0.8sceond per image on the PC with2.2GHz Pentium dual-core processor.The results of experiments show that the ideas on the aspects of the image processing proposedin this paper can drastically improve the picture quality and be great helpful to the identification.Compared with general neural network, the training of neural network in the recognition algorithmproposed in this paper has faster convergence speed, higher recognition rate and recognition speed.Compared with traditional matching recognition algorithm, the method proposed in in this paper hasbetter efficiency.