Research On License Plate Detection Method Based On Deep Learning In Unrestricted View

License plate detection and recognition plays an important role in the intelligent transportation system,and it is the necessary prerequisite for vehicle fine recognition,license plate recognition and other processing.Therefore,it is of great significance to study the license plate detection and location in complex and unrestricted scenes.In recent years,license plate detection has attracted the attention of all walks of life,and has achieved good results in practical application.However,due to various adverse conditions,it is still difficult to meet the requirements of practical application in the actual scene.Therefore,how to effectively improve the accuracy and speed of license plate detection algorithm is of great significance for the development of intelligent transportation system.Thesis focuses on two aspects of improving the accuracy and speed of license plate detection algorithm.The main work is as follows:(1)At present,the license plate detection models are better for the light conditions,and the scene with a single perspective is better.There are still many problems for the license plate detection under the unrestricted perspective.In order to solve this problem,thesis proposes an extended convolutional attention module.In this module,firstly,three convolutions with different expansion rates are performed on the feature map output from the basic feature extraction network,and then the generated feature map is fused and output;secondly,dual attention is paid to the feature map output from the basic feature extraction network in terms of channel and space,and finally the generated feature map is fused.At the same time,most of the open data sets are single in type,too few in number and fixed in perspective,so they are not universal.Therefore,thesis proposes a method of synthesizing license plate data based on 3D scene rendering.The algorithm uses 3D scene modeling to simulate the change of illumination intensity,angle and distance of license plate in the real environment to synthesize the multi scene comprehensive data set.In thesis,the proposed method is tested on the CCPD data set and the synthetic virtual data set,and the results show that the method improves the accuracy of license plate detection significantly.(2)Considering the real-time performance of the algorithm.Although the accuracy of the algorithm based on the extended convolution attention is very good,it is not very good in speed.The main reason is that the two-stage Faster R-CNN algorithm is improved,which consumes a lot of time in CNN feature extraction and RPN region recommendation.In order to improve the speed of license plate detection.In thesis,a vehicle license plate detection method based on depth separable convolution mobilenetv2 is proposed.Mobilenetv2 is a deep separable convolutional network,which uses the deep separable convolution instead of the traditional convolution,reduces the number of network parameters,and improves the speed of the network.Therefore,we replace vgg-16,the basic network of SSD,with mobilenetv2.At the same time,we also do the dilation convolution operation after each convolution layer of SSD multi-scale feature fusion,which can better combine the semantic information of high-level and low-level features,and help the network predict the location of the target.The experimental results also show that compared with the algorithm based on the expanded convolutional attention module,the speed of this combination is increased by about 4 times under the condition of little precision reduction.