| Various intelligent management such as highway toll collection system,speeding violation supervision,highway inspection all rely on vehicle license plates.Real-time and accurate license plate recognition is the basis for realizing various managements.Traditional license plate recognition technologies rely on intuitive features,which have high limitations on image clarity and complexity.The license plate recognition algorithms based on deep learning can automatically extract detailed features and improve prediction accuracy.With the development of electronic technology and GPU,it is possible to use deep learning algorithms to detect and recognize license plate images in real time.However,this method generally relies on a fixed license plate style,which limits its application value.The license plate recognition technology combines artificial intelligence,image processing,computer vision,pattern recognition,and deep learning to explore the feasibility and practicability of deep learning-based license plate recognition tasks in complex scenarios.The following research work has been carried out in this paper.(1)A method for reasonable expansion of the data set is proposed.In order to solve the problem of poor generalization performance after training due to the uneven distribution of the data set used for training,data enhancement algorithms such as random rotation and blur of the labeled license plate area are used.It also collected a part of the license plate pictures from the network and marked them for training and testing.In order to solve the problem of insufficient character recognition network training data volume and single picture type,it is proposed to use Open CV-based license plate synthesis technology to generate different types of license plates,increasing the amount of network training data.(2)A pyramid model generation method for multi-task convolutional neural network MTCNN in license plate detection task is proposed.Aiming at the problem that the license plate detection scheme based on MTCNN in complex scenarios takes too much time in the entire detection phase in the first stage,by comparing the accuracy and regression rate changes of the model at a single scale,the size distribution model is determined to improve the multi-scale generation Pyramid model of pictures to reduce model processing time and improve detection accuracy.According to the aspect ratio characteristics of the license plate,the input size and convolution kernel size of the three MTCNN sub-networks are improved.Finally,adjust the network output and improve the loss function according to the tasks of each MTCNN network.(3)Research and implement the character recognition network LPRNet.Aiming at the problem that characters are difficult to segment,LPRNet is used to extract features,and then the decoding results are obtained through CTC and cluster search respectively,and finally the final license plate number is spliced.In order to prevent skewed license plates from affecting the recognition results,a spatial transformation network is used to correct the license plates detected by MTCNN.After the above research work,this article mainly uses three public data sets and one network collected data set as training and test cases,namely Caltech Cars,English LP,CCPD and Inter LP,using the improved MTCNN cascade to detect license plates in this paper,and using LPRNet recognize characters.Experimental results show that in multiple scenarios,the success rate of license plate image positioning can reach 94.36%,the recognition accuracy of all characters can reach 81.9%,and the recognition rate of more than 5 characters can reach 97.3%.And the average end-to-end processing time of each license plate is about 213 ms,which can meet the application requirements of the actual scene.In future work,we will further study the expansion of the license plate recognition algorithm in performance,evaluation and application,and further combine the social security issues to further identify the vehicle and owner information. |
PDF Full Text Request