Research On The Detection Algorithm Of Banknote Security Line Based On Convolutional Neural Network

Most of the circulating banknotes in China have anti-counterfeiting designs with metal security lines.In the distribution of banknotes,the distance between the security line and the printed image drifts back and forth within a certain range.The existing banknote image detection system in the printing industry has made targeted attempts on the basis of traditional pattern recognition algorithms in order to overcome the interference of security line drift on banknote printing quality control.According to the priori characteristics of the window opening security line,the threshold segmentation algorithm is used to segment the window block within the security line detection area.When learning high and low threshold samples,the security line window opening part is completely shielded,and the fluctuation of printed image within the tolerance range is counted to form a complete face high and low threshold template.This method can detect the defects between the window blocks(such as the white dots of the offset printing shading,the ink dots and the dirty of the intaglio printing),and the four directions of the window opening security line.As a national business card,RMB has strict quality requirements.Due to the production process of the banknote paper,the non-opening area of the anti-counterfeiting window opening security line has a lower gray level than other areas.In order to control the defect of the exposed security line in the non-windowed area,the banknote image detection system using the traditional algorithm usually uses the method of reducing the "fishnet aperture",that is tightening the detection parameters of the area.As a result,the false alarm rate of defects in this area in the detection system has increased significantly,and the cost of production obsolescence remains high.In the past ten years,with the vigorous development of deep learning technology,convolutional neural networks have a wide range of applications,especially their advantages in digital image semantic segmentation.The author tries to use deep learning technology for the quality inspection of the security line of banknotes.Different from the conventional line segment target detection,the designed network focuses on the abnormal morphological features of the security line and screens it out from the wrongly discarded images.The main research task of this paper is to aim at the problem of high false waste in the security line area of the small banknotes image detection system currently used by the banknote printing enterprises,drawing lessons from the classic convolutional neural network structure,design a network model suitable for target detection in the security line area of the banknote window,it is used to assist the traditional pattern recognition detection system to reduce the detection error rate and try to ease the workload of subsequent manual image confirmation.Using the Windows version of the Pytorch framework(ver1.1.0),the Python programming language(ver3.7)is used to build a convolutional neural network on the Anaconda platform.The network structure draws on the relevant advantages of classic semantic segmentation models such as FCN,Unet,and Segnet,and designs and builds a convolutional neural network model that assists the detection of security lines of banknotes.In the main structure of the network,skip connections are used for many times,the feature location information of low level and feature semantic information of high level are integrated gradually.The network model performs semantic segmentation on the input image,and finally realizes pixel-level semantic classification,and reduces the final error rate of the detection system through the secondary classification of all defective images.In the case of a small training sample scale,additional data enhancement operations such as image flipping,perspective transformation,filtering processing,and adding noise are performed in each training iteration to expand the data characteristics and sampling range.In the process of model training,the existing problems are analyzed according to the training results of multiple and different deployment,and the network settings are adjusted and optimized.The Focal loss statistical function is introduced as the calculation function of the loss layer to solve the problem that the number of pixels related to the security line is relatively small.The test results show that the convolutional neural network model designed and trained in this paper has a good auxiliary effect on the banknote image detection system,and can significantly reduce the false waste of the banknote window opening security line area under the premise of ensuring that the product does not leak.