Research On The Algorithm Of License Plate Recognition Based On Complex Scene

The rapid economic development has optimized the convenience of transportation,and while improving the efficiency of travel,the problem of traffic control has become increasingly prominent.As far as traffic control is concerned,the problems of vehicle charging,information collection and illegal tracking are also becoming more and more prominent.Now,The most way to solve the traffic problem is intelligent license plate recognition system,as well as some application scenarios that arise can be served and secured by some number plate recognition systems.they are all about getting information to identify vehicles and ultimately solve problems effectively.The presence of complex scenes such as weather,occlusion,Stain and long distance can make identification more difficult.This paper focuses on the analysis of the complex scene license plate recognition algorithm:(1)Experimental data preprocessing.So as to improve the balance of license plate data and avoid the influence of noise on license plate data,this paper uses(Multi-Scale Retinex,MSR)algorithm to enhance the license plate images so that they can be well integrated into the network training.Meanwhile,in order to improve the network performance of license plate recognition algorithm,this paper generates some license plate data for experiment using DCGAN network.(2)Study and improve the license plate positioning algorithm of YOLOv3.This paper adopts a network model based on YOLOv3.In this paper,a YOLOv3-based network model is used.First,a new streamlined feature extraction network,Dark Net-41,is proposed to improve the detection efficiency of the model and reduce the computational consumption.Then,the multi-scale feature fusion of YOLOv3 is changed to improve the accuracy of license plate localization,and the attention mechanism convolution module Inception-SE is added to improve the localization ability,and finally,a new effective detection model is obtained by using the Kmeans++ algorithm to optimize the anchor point size.(3)Research and improve the license plate recognition algorithm of CRNN network.In this paper,an end-to-end SCRNN network for license plate character sequence recognition is proposed.Among them,the added STN network avoids the influence of uncertainty factors brought by the skewed license plate sequence,while the replaced BLSTM model can improve the license plate character recognition in complex scenes.In this paper,an experimental study is conducted by proposing and studying the above three modules.The experimental analysis shows that the algorithm m AP(average accuracy)reaches 98.84% and the detection speed reaches 36.4 frames/second.Meanwhile,the improved CRNN algorithm model is used for license plate character sequence recognition,and the recognition rate reaches 99.2%.It can better complete the license plate recognition under the complex background,at the same time,it verifies the shortcomings of the intelligent license plate recognition system,and has better applicability and real-time.