Research On Vehicle Information Detection And Recognition Based On Deep Learning

In order to better assist traffic management departments to manage traffic problems,researchers began to use computer vision method to detect and identify illegal vehicles in traffic road.The detection and identification of vehicle information including plate number,vehicle logo and vehicle type is a key step of vehicle identification.In computer vision,Deep Learning has great advantages over the traditional Machine Vision Methods in terms of speed and accuracy.Therefore,this paper uses the deep learning method to detect and identify vehicle information in images,and the specific research contents are as follows:First,the information of the vehicle including license plate,vehicle logo and vehicle face in the image is detected.This paper selects YOLOV3 algorithm as the main detection algorithm.This paper come up with a vehicle information detection algorithm based on YOLOV3-fass algorithm aim at problems such as the poor stability of model,slow detection speed and small object missing detection of YOLOV3 algorithm.The new algorithm can address some problems existing vehicle information detection technology such as detection speed,accuracy and stability.In this algorithm,we deleted some residual structures,reduced a number of channels of convolutional layer,added a down-sampling branch,and added three scale-hopping connection structures based on Darknet-53 network structure,and added one detection scale and calculated twelve groups of anchor frame values through the means of K-means clustering algorithm combined with manual setting.Finally,we fine-tuned YOLOV3-fass algorithm through the migration learning mechanism.The experimental verified on the self-built vehicle dataset show that YOLOV3-fass algorithm can detect vehicle information more accurately,efficiently and stably.Then,this paper determine the vehicle brand according to the vehicle logo information.This paper designs RDCNet dual anti-fuzzy identification model to solve the problem of fuzzy vehicle logo recognition in the self-built data set based on the Res Net18 model,proposes the turn-relu activation function to solve the problem of slow convergence speed,add a random data enhancement processing device to input of algorithm model to solve the problem of small scale of self-built vehicle mark dataset and great influence of light on vehicle logo identification.Experimental verified on the self-built vehicle logo dataset shows that the RDCNet model has anti-fuzzy performance,the vehicle logo recognition rate reaches 99.62%,and the turn-relu activation function significantly improves the convergence speed of the model.Finally,this paper completes the vehicle type identification based on the vehicle face after the vehicle brand was determined.this paper proposes a vehicle type recognition algorithm based on vehicle face image region divide-and-conquer strategy of RPNet model in allusion to the problem of high similarity of vehicle faces of two adjacent years within the same brand.RPNet is a model structure composed of two parallel convolution modules based on the classical model,it uses an improved fusion down-sampling module to down-sampling processing of feature map.The region divide-and-conquer strategy of vehicle face image is to divide the vehicle face into five area images according to prior experience,and extract the fine features of each area image using RPNet model.Experimental verified on the self-built vehicle face dataset shows that the vehicle type identification accuracy of the algorithm in this paper is much higher than classical model,and the validity of the turn-relu activation function is verified again.