Chinese License Plate Recognition System Based On YOLO V4 And CRNN+CTC Algorithm

With the economic and social development of 40 years of reform and opening up,the current domestic motor vehicle ownership has exceeded 400 million.How to effectively and quickly manage the huge motor vehicles is a key problem that needs to be solved urgently in the development of social transportation.At present,the traditional license plate recognition system has been seen everywhere in social development.However,it is subject to the limitations of traditional algorithm models or the bottleneck of equipment performance.The detection results of the algorithm are not ideal,and there are characteristics such as false detection and slow recognition speed.Based on this,this thesis proposes an end-to-end license plate recognition scheme,which replaces the original license plate segmentation and recognition scheme with the license plate text sequence recognition problem.While reducing the model training parameters,it also speeds up the detection speed and improves the detection and recognition accuracy.The following are the main contributions of this thesis.For the license plate location algorithm: The improved YOLO v4 algorithm model proposed in this paper extracts license plate information.On the original Darknet53 network structure,a lightweight CSP structure is introduced in the feature fusion layer of the network.The results of the two convolutional layers in the lightweight CSP structure are directly fused,which effectively reduces the amount of computation and avoids the use of batch normalization operations alone,further reducing the computational pressure of the network model.By introducing the Dense Net network in the extraction layer,each network layer is directly connected to each other to ensure that the feature information obtained by each layer can be transmitted to the next layer most effectively and to the greatest extent,avoiding the problem of gradient disappearance due to the deepening of the number of network layers.The K-means++ clustering algorithm is introduced to solve the problem that the original clustering algorithm of YOLO v4 is unstable.Finally,the separable convolution operation is adopted,which effectively reduces the number of parameters of the network model,and also reduces the cost and time of the operation.Through experiments on the CCPD dataset and a mixture of autonomously collected datasets,the final results show that the localization accuracy is as high as98%.License plate character recognition: In this thesis,a segmentation-free license plate recognition scheme is adopted.Different from the traditional segmentation scheme of license plate recognition,this thesis regards the image information of the entire license plate as a unified whole,regards it as a speech sequence problem,and trains and recognizes the model.Based on the improved CRNN network model,the network model structure of bidirectional LSTM is used to analyze and extract semantics.Compared with traditional detection models such as template matching and classification,the performance has been improved compared with the past,and the recognition accuracy has reached97% on the CCPD dataset and a mixture of autonomously collected datasets.