Research And Application Of License Plate Detection Method Base On Deep Learning

With the improvement of per capita car ownership in China,vehicle management is becoming more and more important.One of the most important steps of vehicle management is license plate detection.At present,although there are many in-depth learning based license plate detection methods at home and abroad,their research is aimed at foreign datasets or specific scenes in China.There are many kinds of Chinese license plates in the natural scene,and the shooting distance is far and the light changes greatly.However,there are few researches on the detection effect of Chinese license plates in the natural scene.In this paper,we mainly study the effect of the existing license plate detection method based on deep learning on Chinese license plate detection in the natural scene,and apply it in the actual scene.This paper summarizes the existing traditional license plate detection methods and license plate detection methods based on deep learning,and makes in-depth analysis and Research on the principle or network structure of each method.Most of the vehicle data sets currently published are foreign vehicles,and the data sets containing Chinese vehicles are few,with a single category and scenario,which is not a natural scenario.Therefore,this paper marks a data set NSI(natural scene images)composed of Chinese license plates captured by surveillance cameras in natural scenes.In the NSI data set,the existing mainstream license plate detection methods are compared in detail,and the experimental results are analyzed to summarize the advantages and disadvantages of each method.According to the experimental results,the key point detection method based on MTCNN network detection method and the network detection method based on YOLOv3 are combined to provide the detection of four corner points of license plate on the basis of high detection rate.In addition,this paper also proposes a method based on deep learning two classification combined with traditional feature verification to distinguish license plate and non license plate,and verifies its effectiveness through experiments.Finally,this paper designs and implements the license plate desensitization and labeling system based on the YOLOv3 license plate detection method and MTCNN key point detection combined with the method of distinguishing license plate and non license plate,develops the video desensitization,image desensitization,license plate labeling and personalized training functions to improve the efficiency of license plate desensitization and labeling,and provides HTTP interface for license plate detection to meet the user’s second driving Demand.