Research On License Plate Location Algorithm In Complex Scene

With the increasing number of cars in the city and the increasing traffic flow,it brings some challenges to the intelligent transportation system of our country.Therefore,the research of license plate location and recognition technology can not stop,especially the license plate location technology is the most critical part in the whole research field.At the same time,with the change of the surrounding environment and the interference of uncertain factors,traditional license plate location methods in fixed scenes are gradually difficult to meet the needs of more and more complex actual scenes.This paper provides two different solutions to the problem of license plate location in complex scenes based on the classic feature design engineering and deep learning has obtained the premise of great success in the field of target detection.The specific work contents are as follows:(1)Aiming at the situation that the location accuracy of single feature is poor in complex scenes,this paper designs a license plate location algorithm based on multi-features fusion with minimum error rate criterion.The whole localization algorithm is divided into two stages.The first stage is the Hypothesis Generation stage: eight feature descriptors are designed based on the color,character,circle,corner and texture of the license plate respectively,and the corresponding candidate license plate are generated;The second stage is the Hypothesis Selection stage: eight features are detected for each candidate frame,and multiple features are fused according to the minimum error rate.Finally,the confidence of each candidate license plate is calculated,and the candidate license plate with the maximum confidence is selected as the final algorithm frame.The experimental results show that the license plate location algorithm based on multi-features fusion has higher accuracy than that based on single feature in different scenes.(2)We propose anther license plate location detection algorithm based on improved Faster-RCNN.Firstly,the K-means clustering algorithm is used to cluster the real license plate,and the size and proportion of Anchor are redesigned,so as to adapt to the license plate location tasks of different sizes and shorten the Faster-RCNN training time.Then,the backbone network VGG16 is replaced by Res Ne Xt50 which has deeper network and less parameters.Finally,channel attention model(SENet)is introduced into the backbone network to form a new backbone network SE-Res Ne Xt50;Through experimental comparison and analysis,the improved algorithm is better than the original Faster-RCNN that can achieve higher accuracy in license plate data set of complex scenes.