Research On Vehicle License Plate Localization In Complex Scene And Its Application

License plate localization, the core component of license plate recognition system, is very valuable in both academic development and potential applications. Though much progress in recent years, it still a challenging problem in complex scene, such as low luminance, low resolution and inclination scene of vehicle. To solve these problems, the study of article is divided into three parts:firstly, the traditional license plate localization method, which based the gray binarization image, was modified and implemented in the article, the algorithm via segmenting the gray scale image into lots of grid block, then calculating the threshold and binarization of each grid block, the binary image about the license plate area is relatively obvious. Then the plate candidate regions come out via strokes jump method and connected region detection. Filtering the no plate candidate regions based neural network classify. But the false positive rate of method is high, and cannot solve the license plate location problem under complex scene. It cannot be worked out to just based the method of expert system.Secondly, In order to solve the problem, a method which based on convolutional neural network to solve the license plate localization in the complex scene was proposed. The paper designed and programmed a new fully convolutional neural network, which got the license plate location accurately via learning corner regression-based. The major process is to make the two corners of license plate area as the output target of a convolution regression neural network model. As the model does not contain the full connection layer, that the model can handle any scale of the image. To guarantee the effective training in our model,4,5000 sample images are annotated by one person. Meanwhile, the annotated sample images are processed by four operators, including translating, scaling, rotating and noising, to increase the number and diversity of the training samples. Extensive experiments on the newly created datasets by us, traffic monitoring dataset and the complex scene dataset, demonstrate the effectiveness of the proposed method against other two license plate localization methods.The method based on license plate locating model showed more fall-out ratio, as the article also presented the methods to assist the plate locating based on vehicle positioning and auxiliary license plate location and license plate space information, vehicle positioning method is also based on the convolution neural networks, whose output target object is the binary image making the vehicle as the fore ground. According to the vehicle location to filter the plate candidate regions. At the same time, we used the one car license plate spatial information and made use of the license plate locating model of angular point confidence values to filter the license plate candidate regions. Experiments show that the secondary positioning method of license plate localization algorithm accuracy is higher, lower fall-out ratio, which could be a license plate locating auxiliary strategy.All in all, after the article got the conclusion that the traditional plate localization method could not solve the problem of license plate localization in complex scene, proposed a model which learning corner regression-based fully convolutional neural network for license plate localization. In additional, we made the vehicle license plate location and license plate spatial information as an auxiliary strategy to accurate positioning. Experiments proved method proposed in the paper effected well in complex scene, high robustness, and achieved the real-time effects on the time complexity.