Key Technology For Bill Identification Based On Deep Learning And Object Detection Of Machine Vision

As an important means of financial circulation,bills and its visual identification technology and anti-counterfeiting identification technology are important guarantees for financial security.This article is titled "Key Technology of Bill Identification Based on Deep Learning Machine Vision Target Detection",researching the use of object detection methods and applying to bill anti-counterfeiting identification scenarios,focusing on the rapid identification of bill images based on regression methods,anti-counterfeit feature recognition based on regional proposals,and the extraction of bill image text information based on small target detection method,then developed a bill detection system based on deep learning and object detection,which promotes the automation and intelligence of visual inspection and the intelligent inspection of manufacturing engineering.The research work was supported by the Guangzhou City Science and Technology Plan Project(No.201802030006).The paper studies the key technology of bill identification based on deep learning.It discusses the domestic and foreign research progress of related research content from three aspects: object detection based on classic method,object detection based on deep learning,and bill identification technology.The main point of this paper are as follows:? Design key technical solutions for bill identification based on deep learning.Analyze the different speeds and accuracy requirements of different stages of bill identification based on object detection,design the bill identification technology scheme and identification process,analyze the key technology of bill recognition technology,multiple anti-counterfeit feature detection,image text information extraction.? Study the bill recognition and rapid positioning technology based on regression method.Analyze the mechanism of bill recognition and rapid location detection technology,point out that the backbone network computation is the bottleneck of the detection speed,and study the backbone network optimization method from three aspects: backbone network depth,convolution kernel scale,and network design method.By using a smaller-scale backbone network(10-20 layers deep)in bill recognition and location detection,using 1 × 1 and other small-scale convolution kernels,and using global average pooling instead of fully-connection layer,the backbone network computation is effectively reduced and the speed of bill recognition and positioning detection is improved.? Study the bill anti-counterfeit feature detection technology based on regional proposal method.Analyze the technical mechanism of bill anti-counterfeit feature detection based on regional proposal method,point out that the problem of sample imbalance during algorithm training affect the accuracy of bill anti-counterfeit feature detection,and do research from difficult training sample mining methods,loss function settings,and prior anchor box settings to improve the detection accuracy.Retrain the model by filtering and using difficult training samples,introducing positive and negative sample weight coefficients and difficult training sample weight coefficients in the loss function to adjust the proportion of sample loss terms,and introducing anchor box position and shape prediction sub-network can effectively alleviate the problem of sample imbalance and improve the detection accuracy of bill anti-counterfeiting features.? Study the bill text information detection technology based on small target detection.First,analyze the mechanism of bill text information detection,point out the differences and technical difficulties of text information detection and general target detection tasks,and two aspects of the algorithm detection effect improvement method are studied to effectively improve the text detection effect of the bill,include multi-convolution fusion method and two-way LSTM text feature sequence recognition method.? Application of bill detection technology based on deep learning and object detection.Build a bill detection system and detection software platform,select the hardware components of the system,briefly describe the workflow of the bill detection system in practical applications,and apply it to bank bills,checks and other bills to comprehensively evaluate the application of the bill detection system.