Research On Vehicle Key Information Detection And Recognition Method Based On Deep Learning

With the continuous improvement of the people’s material living standards and the increasing number of cars,traffic control problems such as refitting vehicles without permission and using fake license plates are becoming more and more serious.It is insufficient to deal with the actual problems only by using the characters of license plates as vehicle identification.At this time,it is particularly important to combine other attributes of the vehicle to determine the identity information of the vehicle.This paper combines deep learning technology to realize vehicle target detection,license plate recognition and vehicle color recognition.It has achieved good test results when using actual traffic image data sets for testing.Firstly,in the aspect of vehicle target detection that is to obtain vehicle location and vehicle type information,in view of the traditional target detection algorithm has poor environmental adaptability and complex structure,this paper proposes an improved Center Net vehicle target detection model,which is uses Center Net as the basic network.On the one hand,the BFP network is used to enhance the model’s ability to extract image features,and on the other hand,the Mobile Net network is used to reduce the number of model parameters.In order to enrich the data set,the data set is expanded by means of random flipping,color jittering and other data enhancement methods.The adaptive adjustment of learning rate is carried out for different parameters in the back propagation by using Adam optimization algorithm.The test m AP value on the actual traffic data set can reach 94.5%.Secondly,in the aspect of license plate recognition,for the Canny edge detection algorithm needs to artificially set high and low thresholds,prone to false edges and other issues,this paper uses the Otsu method to automatically calculate the high and low thresholds in order to improve the accuracy of license plate positioning.In order to solve the problem of low accuracy of character recognition by simple convolution neural network,this paper constructs LPRNet character recognition network by adding residual connection in the structure of Inception network.When training LPRNet,in order to accelerate the convergence speed of the network,the training process is optimized by momentum.On the basis of the original model,the images with errors are merged with the previous training set for secondarytraining for continuously improve the accuracy of the model.The average recognition accuracy on the actual traffic data set can reach 97.3 %.Finally,in the aspect of vehicle color recognition,for the problem of low accuracy using color quantization templates in the HSV color space,this paper constructs convolutional neural networks on RGB,HSV,and Lab color spaces respectively.The VGGNet network is used as the basic network for color feature extraction,and VGGNet is improved by using dense connections.Xavier is used to uniformly initialize the parameters to achieve the goal of making the variance of each layer output as same as possible.The average recognition accuracy on the actual traffic data set can be up to 97.04%.