Research On Real-Time License Plate Recognition Algorithm Based On End-to-End

As one of the key technologies of the Intelligent Transportation System,the license plate recognition system has developed rapidly in recent years with a lot of achievements.In particular,the incessant emerging of new deep learning methods promotes the update of license plate recognition algorithms.Deep learning has a strong generalization and adaptability.The license plate recognition algorithm based on deep learning can complete the recognition under complex conditions such as blurred image,uneven illumination and tilted license plate.However,its huge computational load and slow recognition efficiency make it difficult to deploy the algorithm to the front of the device.Reducing the network structure of the deep learning model reduces its computational load,but also reduces the recognition accuracy of license plate features.A license plate recognition algorithm based on deep learning can hardly achieve the best of both accuracy and speed.The universal target detection method pays attention to a wide range of natural target,but the license plate detection only needs to distinguish the foreground and background of the easily recognized license plate;the shape of the license plate is long and the receptive field of deep feature map pixels is square;the region proposals for license plate character recognition are often horizontal rectangular boxes,while the horizontal rectangular boxes can easily be used to intercept the image license plate with redundant background information.These redundant calculations give the license plate recognition algorithm possibility to speed up.For this reason,thesis does in-depth research on the implementation speed of license plate recognition algorithm,the main work is as follows:(1)The real-time performance of the license plate detection algorithm is studied.For the license plate detection module,the model complexity analysis is carried out.On the one hand,the amount of computation and parameters of deep learning model are calculated,and the time spend law of the convolution neural network structure layer is discovered.On the other hand,the shape characteristics are studied for a specific license plate target.The improved five-dimensional vector is used to express the shape of the license plate,and the speed optimization idea is provided with the receptive field as the starting point.(2)Design and implement an end-to-end real-time license plate recognition algorithm.Taking the license plate image as a specific target,using the small region features of the separated license plate,the license plate detection is completed with fewer hierarchical backbone networks and named as Separated Plate Detection Network(SPDNet).The matching method of SPDNet is designed,and the functions of separating feature detection,recombination,shaping and filtering are completed.A set of end-to-end real-time license plate recognition algorithm with lightweight network structure,fast detection and recognition speed,and flexible detection performance is put forward.After removing some redundant computations,the proposed end-to-end learning license plate recognition algorithm meets the real-time requirements,and effectively improves the speed of license plate recognition algorithm implementation without significant difference from popular algorithms in recognition accuracy.