Research On License Plate And Vehicle Type Recognition Based On Convolutional Neural Network

Due to the leap growth of vehicles on the road,lots of problems such as traffic violations,traffic congestion,traffic accidents,and parking difficulties have become increasingly prominent.Therefore,it is urgent to accelerate the construction of intelligent transportation systems.As an important part of vehicle information recognition,license plate and vehicle type recognition play a important role in intelligent traffic systems.The license plate and vehicle information can effectively assist vehicle retrieval and comparison,help identify traffic violations,and rationally divert vehicles.In order to further achieve accurate and efficient license plate recognition and vehicle type recognition,the deep learning method is used to build a convolutional neural network for license plate and vehicle type recognition in this paper,and simulation experiments are carried out on the corresponding data set.This paper mainly conducts research from the following two aspects:In the aspect of license plate recognition:this paper first designs and implements an improved PSPNet to locate the license plate area.Based on the original PSPNet,the skip connection structure of the UNet is introduced,so that the PSPNet combines the high-level semantic information and low-level location information in the backbone feature extraction network,thereby obtaining a more accurate distinction between license plate areas and non-license plate areas result.The transfer learning method is used to transfer the trained parameters of ResNet50 on the ImageNet data set to the improved PSPNet model,and experiment on the labeled CCPD data set.Experimental results show that the MIoU of the improved network reaches 97.07%,and the F1 score reaches 98.6%;Secondly,the weight method is adopted to determine the four vertices of the license plate,and the perspective transformation method of Opencv is used to correct and segment the license plate;Finally,the vertical projection method is adopted to segment the characters of the obtained license plate image,and the improved LeNet-5 is designed to recognize the segmented characters.Specifically,based on the original LeNet-5,this article increases the depth of the network,enhances the feature extraction capabilities of the network,and changes the number of convolution kernels in the convolutional layers and the number of neurons in the fully connected layer to better match the license plate recognition task,and apply the Dropout method to optimize the network to reduce the risk of network overfitting.According to the improved LeNet-5 network simulation results on the segmented license plate character data set,the accuracy of single character recognition can reach 99.24%.In the aspect of vehicle type recognition:The improved YOLOv4 target detection algorithm is designed in this paper to identify vehicle type.First of all,this article uses the lightweight convolutional neural network MobileNetV3 as the Backbone of the original YOLOv4 to reduce the amount of model parameters.Secondly,all the 3*3 standard convolutions in the Neck and Head modules of the original YOLOv4 are replaced with deep separable convolutions to further reduce the amount of model parameters.Then,in order to improve the recognition accuracy of the model,the K-means algorithm is used to reset the prior frame.Finally,the transfer learning method is used to transfer the trained parameters of MobileNetV3 on the ImageNet data set to the improved YOLOv4 model.The simulation results on the BIT-Vehicle dataset show that the improved vehicle identification method mAP is 96.17%,the parameter amount is about 53.77MB,and the detection speed is increased by 26%compared with the original YOLOv4.