Research On License Plate Recognition Technology Based On Deep Learning

License plate recognition is an extremely important part of intelligent transportation system,which is also indispensable in the concept of smart city and has a high value of research and application.After a long period of research,the outstanding research results of license plate recognition technology can solve the problem of general scene of license plate recognition,such as traffic fixed bayonet,garage entrances and residential access.However,the license plate could not accurately located and identified in some cases such as the license plate tilt,the distortion of the license plate,the poor lighting conditions,low pixel resolution and so on in natural scenes.In order to solve these problems,this paper uses the method of target detection based on deep learning to study the following aspects.Firstly,this paper proposed a method of target detection based on deep learning to locate the license plate,and improve the convolution neural network structure.During the training of the model,30521 pictures collected on-site were manually marked with the image annotation software.In order to increase the number of training samples,the image was mirrored and scaled at random.The datum collected from different traffic bayonets were tested by using a trained model.By comparing the test results and the method of license plate location using gray image,the robustness of the target detection method based on the deep learning in terms of the license plate location was verified.Secondly,the same method of the deep learning based on the target detection was used to detect the license plate characters,and the final license plate recognition results were obtained by the corresponding sorting processing of the mainland license plate and Hong Kong,Macao and Taiwan license plate.In the training process of the model,21269 pictures were manually marked with the image annotation software.For the character in the license plate was asymmetric,the image was not enhanced.The data network was trained and tested by using different network structures,and the optimal network model was selected by comparative verification image data collected from the field.Finally,the combination of license plate location and license plate recognition was used to test the recognition speed of the whole system.The test result shows that the recognition speed is 0.12 seconds per picture,which is superior to the traditional license plate recognition system.In addition,the system can be combined with vehicle detection and vehicle identification.The image to be detected first carries on the vehicle detection and the vehicle identification,then carries on the license plate detection and the recognition.These results can be used to build large data platform,real-time upload detection of vehicle models,license plates and location information,which is of great significance for building intelligent city and realizing the Internet of Everything.