Research On The Key Technologies For Multi Currency Recognition Based On Image Spectral Features

With the development of the commodity economy,currency appearance makes human financial transactions more convenient,and the speed of commodity circulation is getting faster and faster.Paper currency plays an important role in people's daily business activities.It is the basis of commodity exchange,and it is widely used in all fields of commodity exchange.With the rapid economic development of various countries in the world,commercial activities are very frequent,which makes the circulation of banknotes increase year by year.There are a lot of cash transactions in the world every day,which brings heavy banknotes processing work to the banking sector.Due to the bad working environment,slow counting speed and error counting,the traditional manual counting can not meet the demand of social development for banknotes.The emergence of banknote sorting equipment is gradually solving the problems faced by the central bank and commercial banks.The application of sorting technology not only greatly reduces the labor intensity of bank staff,but also greatly improves the work efficiency of staff,and significantly improves the counting quality of banknotes.Although banknote sorting technology has been greatly developed in recent years,the detection technology of fiting and unfitting banknotes,stains and handwriting can not meet the banking needs.The main reasons are described as follows.Firstly,the fiting and unfitting banknotes are a vague concept,which different people understand differently.Secondly,the complex patterns and textures of banknotes bring great difficulty to banknotes detection.Thirdly,the location of stains is not fixed,the depth is different,and the shape is different,which also increases the difficulty of detection.Finally,the different thickness,color,shape and size of handwriting make it even more difficult to detect banknotes.In view of the above problems,a recognition method of fiting and unfitting banknotes is proposed based on the spectral characteristics of banknotes images.This method takes the multispectral banknote images collected by CIS image sensor as the research object.According to the characteristics of fiting and unfitting banknotes,stains and handwriting,the detection of fiting and unfitting banknotes is divided into the whole fiting and unfitting banknotes and local dirty banknotes.A texture feature extraction method based on gray level co-occurrence matrix(GLCM)of banknote images is proposed to analyze the influence of different light sources on the detection of fiting and unfitting banknotes,and a new WOA-MLSVMs recognition method of fiting and unfitting banknotes is proposed baed on the image texture features.In addition,a convolution neural network(CNN)based multi-spectral recognition method for fiting and unfitting banknotes and a Gaussian confidence based multi-spectral detection method for banknotes contamination are proposed.The main research work and contributions of this paper are described as follows:Firstly,an embedded digital image processing platform architecture based on FPGA and CPU is proposed,and an embedded digital image processing unit based on XC7Z020 core FPGA chip is designed and implemented.Firstly,the platform collects the banknote image S through CIS image sensor,and carries out image pixel correction and drift correction on the hardware.Secondly,the line scan detection algorithm based on search is adopted to detect the edge of the collected banknote images,and the affine transformation algorithm is used to normalize the banknote images.Thirdly,a face recognition algorithm based on multispectral template matching is used to recognize the basic information of banknotes.Finally,the reliability of the platform is verified by experiments,which provides a favorable support for the follow-up research.Secondly,a texture feature extraction method based on gray level co-occurrence matrix(GLCM)of banknote images is proposed,which effectively reduces the influence of different light sources on the detection of fiting and unfitting banknotes.This method uses the double-sided reflection images of banknotes under red,green,blue,infrared and ultraviolet light,as well as the green and infrared transmission images as the research objects.Firstly,GLCM is used to extract the texture characteristic parameters,such as energy,entropy and inertia,to describe the visual characteristics of the fiting and unfitting banknotes.The texture characteristics of the double-sided images reflected by red light and ultraviolet light are analyzed and determined,which have no obvious effect on the recognition of the fiting and unfitting banknotes.Then,the texture features extracted from other 8 images are used to establish the recognition model of fiting and unfitting banknotes,and five essential dimension estimation methods and 17 data dimensionality reduction methods are combined to determine the essential dimension and the optimal dimensionality reduction method.Finally,MLSVMs based on whale optimization algorithm(WOA)is used to realize the recognition of old and new banknotes under full spectrum,and the simulation results show the effectiveness of the proposed method.Thirdly,a recognition method of fiting and unfitting banknotes under multi-spectral based on convolutional neural network(CNN)is proposed.This method adopts the double-sided reflection images under green and blue light,as well as the transmission images under green and infrared light as the research objects.Firstly,a convolutional neural network is constructed,which is suitable for the recognition of fiting and unfitting banknotes.The network has 9 layers of depth,including three convolution layers,three pool layers and three full connection layers.Secondly,the loss function of the convolutional neural network and the training method of forward propagation training and reverse propagation training are described.Finally,the network parameters of 100 yuan 50 yuan and 20 yuan in 2005 RMB are trained through experiments.The effectiveness and accuracy of the method are verified by comparison experiments.Fourthly,a multispectral banknote contamination detection method based on Gaussian confidence is proposed.This method uses the double-sided reflection images under red light,green light,blue light and infrared light as the research object.Firstly,Gaussian confidence feature extraction algorithm is used to model and extract features from the multi-spectral banknotes images.Secondly,the edge tracking image topology analysis algorithm is used to detect the dirty features of banknotes,and the quantitative calculation after feature extraction is realized,and the quantitative index of dirty banknotes is given.Finally,the banknotes with stains and handwriting is taken as an example to show the processing results of the images collected under different spectra of the dirty banknotes in each stage,and the effectiveness of the method is verified.