Research And Application Of License Plate Recognition Technology Based On Deep Learning

With the development of economy and the advancement of technology,there are more and more car ownership in China.There are huge demand for automatic license plate recognition in China's road traffic intelligent control and parking lot management.The traditional license plate recognition divides the license plate recognition into three modules: license plate location,character segmentation and character recognition.The decentralized module design method improves the complexity of the algorithm,and the dependence between the three modules will cause the error accumulation problem,and finally the accuracy of the license plate recognition will be reduced.This paper proposes a license plate recognition system(LPR-YOLO)based on YOLOv3.The system is divided into two parts: license plate location and character recognition.The system has the advantages of improving the accuracy of license plate location and classification,reducing the recognition error caused by character segmentation,and improving the accuracy of character recognition.The experimental results show that the proposed license plate recognition system(LPRYOLO)is highly adaptive,and after 200,000 rounds of training,a faster convergence speed is obtained.The average ratio of license plate location(IoU)reached 85.95%,the total loss function converges to 0.07.The main work and innovations of this paper are as follows:(1)In order to adapt to the specifications and characteristics of the license plate,a license plate location system was constructed.Based on the YOLOV3 target recognition model,this paper proposes an end-to-end network model(LPR-YOLO)for license plate recognition.In order to enhance the sensitivity of the model to the license plate size and character size,the license plate recognition system proposed in this paper redesigned the feature extraction method in the license plate prediction stage in addition to the redesigned feature extraction network.(2)According to the shape characteristics of the license plate,construct a new prediction frame.In order to speed up the multi-scale prediction speed and improve the classification prediction accuracy,a new prediction frame anchor point pair is measured by the clustering method.And in order to accelerate the convergence speed of the model,the composition of the loss function is optimized.According to the shape characteristics of the license plate,the Anchor value of LPRYOLO is re-evaluated.In addition,because the license plate classification is exclusive,this paper uses the Softmax classifier to classify and classify the license plate in the LPR-YOLO model.Finally,the Softmax classifier is used to classify the license plate.Based on YOLOv3,this paper redesigned the classification loss function.(3)Improve the recognition accuracy of confusing characters.In the process of recognizing characters,since the license plate may have problems such as tilting and blurring,this may cause the license plate characters to be deformed,causing confusing character recognition errors.Therefore,after the classification and classification of characters,this paper continues to study the secondary recognition of confusing characters.The secondary recognition of characters mainly relies on the shape of characters,and performs partial projection or global projection for verification and recognition.Experiments show that the improvement effectively improves the recognition accuracy of confusing characters.