Research And Implementation Of License Plate Recognition Algorithm In Complex Traffic Scene Based On Deep Learning

With the increasingly prominent urban traffic problems,intelligent traffic management is welcomed by urban managers.As one of the core elements of intelligent traffic management,license plate recognition is widely used in various traffic scenes,which greatly improves the level of technology and information of traffic management.However,in complex traffic scenes,license plate recognition is prone to noise interference,especially when the license plate has a certain degree of tilt,resulting in the degradation of detection and recognition effect.Aiming at the above problems,this paper studies the license plate recognition technology based on deep learning.The main work and innovation of this paper are as follows:1:Establish datasets of license plate detection and license plate character recognition.In terms of the dataset of license plate detection,8200 license plate images under various traffic scenes of CCPD are selected in this paper.In order to improve the defect of the concentration of Anhui license plate,1800 license plate images from 30 provincial administrative regions(excluding Anhui,Hong Kong,Macao and Taiwan)are collected in various ways.In terms of the dataset of license plate character recognition,on the basis of 10000 license plate region images which detected and saved,18000 license plate region images are generated by GAN network.2:Construct the license plate detection and correction algorithm based on YOLO-stn.This method is based on the YOLOv2 model,and the STN module is embedded in YOLOv2 to realize network reconstruction,make the correction of tilt license plate,and improve the performance of license plate detection finally.After testing,the detection performance of YOLO-stn in complex traffic scenes has both high accuracy and high real-time,especially can improve the detection effect of tilt license plate.3:Construct the algorithm of license plate character recognition based on CRNN.Thismethod avoids the intermediate process of license plate character segmentation and realizes the end-to-end license plate character recognition.After testing,the accuracy and recognition speed of this method are better than that of open source LPRNet.4:Realize the license plate recognition system based on Jeston Nano.In this paper,a self-defined license plate recognition algorithm is implemented on Jeston Nano hardware platform.Test results show that the performance of the algorithm in Jeston Nano can meet the accuracy and real-time requirements of the landing of license plate recognition algorithm.In summary,this paper has carried out research on license plate detection and license plate character recognition.The tests show that this method can effectively improve the accuracy and speed of both license plate detection and license plate character recognition,and meet the accuracy and real-time requirements of algorithm landing.