The Dual Detecting System Of Banknote Image Based On Convolution Neural Network

The quality requirement of RMB printing is much higher than other presswork.In the printing enterprises,the existing RMB detection system is based on the template matching technology to detect the printing defects.In the actual production environment,the existing system has certain technical limitations,such as poor detection ability for specific defects(thin track,Lack of printing on metal lines,ink stains and so on),and high error detection rat(the ratio of error detection number to actual waste number reaches 10:1).The Dual Detection System of Banknote Image based on Convolutional Neural Network is designed in this paper.Firstly,when the image of training set of a specific defect type is insufficient,the printing defect is simulated to realize data augment.In the experiment,existing detection results were divided into infrared light source image and white light source image in the HSV color space,and then network model design and training were conducted for sample images generated under different light sources.The experimental results show that the weak points in the current production environment are solved by the system designed in this paper.Main contribution of this paper: the use of convolutional neural network for the detection of RMB image was first proposed in the printing industry,and there is no similar detection method in the industry.There are three innovations in this paper:1.According to the shortage of training set for specific types of RMB defects,data augment is carried out by using opencv,the long and short axes of the ellipse were set by random variables to simulate the printing defects of blank and ink.By setting the length and width of the vertical line in random variables,the printing defects of the ‘thin track' were simulated to realize the image enhancement method suitable for RMB detection.2.According to the print area of infrared light source image in the process of production with white light source image detection requirements,consult RGBD image processing method,first the image is divided into infrared light source images and white light source image(image in the HSV color space,hue and saturation are 0)in HSV color space,the through different network structure,the final test results under different network structure to integration.3.Detection of images of white light source and infrared light source in the area of banknote printing.For the image of white light source.the network structure of 17 convolution layers,7activation layers and 2 polling layers is designed by referring to the core idea of ‘ResNet'.By adjusting the image size of input layer,pooling layer function,activating layer function and reducing convolution layer,the network structure and parameters with the highest accuracy of RMB detection were obtained.Finally,the structure is compared with top-1 accuracy results of classical network models such as ResNet,GoogleNet and VggNet,to demonstrate that the network structure and parameter configuration designed in this paper is more suitable for image detection under the white light source of RMB.For the infrared light source,two sets of inception structure are adopted,with a total of 15 convolution layers and 4 pooling layers.