Research And Implementation On RMB Banknote Classification And Stain Identification Algorithm Based On Neural Network

With the development of modern economy in our country,the amount of RMB banknote is increasing rapidly.Stain is inevitable while the RMB banknotes are using and that will cause a big affection.Thus the bank needs an automatic program to classify the RMB banknotes by different denominations and to identify if there are stains on RMB banknotes or not.Under this background,programs to classify RMB banknote images based on neural network and to identify stains exist or not based on computer vision are studied.Different neural networks need input images with different sizes.So use computer vision technology to cut and compress the original input RMB banknote images using bilinear interpolation skill in order to supply input images with proper sizes for different neural networks.The algorithm will be too complex using feature points and contour detection not only because the original input RMB banknote images have too many lines also the light intensity and the drift of image will cause affection easily.Build and train neural networks to classify the input RMB banknote images are used in this paper.An FCN?LeNet-5 and AlexNet are adjusted and trained in this paper.To improve classification accuracy,this paper designed a program to output classification results of the RMB banknote images using the output classification results of neural networks above.Program tested in Matlab shows 100% classification accuracy.The pixel difference in stained area of stained RMB banknotes is larger than the pixel difference in the same area of unstained RMB banknotes because stain shows a lower pixel on the input RMB banknote images.So setting standard pixel difference to identify RMB banknote images have stains or not is used in this paper.This paper uses normalization to reduce the affection of unstable light intensity.Use statistics when setting the standard pixel difference to reduce the affection of the drift of images.Set thresholds in identification program to improve identification accuracy.Program tested in Matlab shows 99.5% identification accuracy.Use TX2 as platform.Transplant Matlab program in Windows to C++ program in ubuntu.Test shows 99.5% identification accuracy and 14.5ms time cost each.