Research On Vehicle License Plate Location And Recognition Algorithm Based On Deep Learning

Motor vehicles to a large extent to solve People’s Daily travel needs,but at the same time,there will inevitably appear many traffic accidents and vehicle control problems.The increase of motor vehicles brings challenges to intelligent traffic management to a certain extent.The main information of the motor vehicle is reflected in the license plate,which is the most intuitive identification of the vehicle.Therefore,the license plate location and recognition algorithm plays an important role in intelligent traffic management and dealing with traffic accidents.Chinese vehicle license plates consist of Chinese characters,letters and numbers arranged in a fixed-size box according to the rules of China’s DMV.License plate location and recognition algorithm can be divided into traditional image processing method and deep learning method with learning ability.The multi-task method of character segmentation and recognition based on image processing is not effective for the recognition of character adhesion,uneven lighting,blurred license plate,damaged license plate and slanted license plate.Deep learning can learn license plate features from a large amount of data,so the method of recognizing license plate characters using deep learning can solve the shortcomings of traditional methods to a large extent.The license plate data set adopted in this paper is the open source CCPD data set and some license plate images collected by ourselves,which constitute an abundant license plate data set distributed in various scenes.In this paper,the license plate recognition model is composed of license plate location and license plate character recognition,and the algorithm of these two parts is designed to have better robustness,higher accuracy and faster speed.The main work of this paper includes:(1)In the license plate location model,this paper proposes to introduce Dilated Convolution on the basis of U-NET network to form a pixel-level license plate location model based on DU-NET.First,low level feature information is extracted through subsampling,then the features of corresponding channels are transposed convolution and jump stitching,and finally restored to the size of the original image,so as to achieve pixel-level classification and correction,so as to obtain the license plate area.This model can accurately locate in various complex scenes.The experimental results show that the model can not only locate the license plate images in various scenes,but also correct the distorted and deformed license plate images.(2)In the license plate character recognition model,based on the idea that the license plate character is regarded as a text sequence,this paper proposes a license plate character recognition algorithm based on PGC-Net,and constructs a license plate character recognition model that is coordinated by multiple network modules and efficiently solves a task.For the license plate character recognition algorithm based on PGC-Net,this paper can be divided into three parts for design.Firstly,a P-CNN network based on the combination of Max pooling and Mean pooling technologies is designed to extract the image features of the license plate area located by the license plate location model.Then,in order to extract the sequence features of license plate characters,an improved two-way GRU network based on RNN network was designed to extract license plate character sequence features from the license plate feature images extracted from P-CNN network.Finally,when the sequence information of license plate characters is obtained,it is necessary to recognize these sequence features.In this paper,CTC character decoding module is introduced to conduct multi-task learning and classification of the timing features of license plate characters.Experimental results show that the proposed model improves the accuracy,speed and robustness.