Research On Vehicle Location And License Plate Character Recognition Algorithm Based On Deep Learning

With the improvement of scientific and technological productivity and economic development,the pace of people’s life is getting faster and faster,and the demand for vehicles and other means of transportation is also increasing.The increasing number of vehicles has undoubtedly put tremendous pressure on urban traffic.Today,when traffic intersections are frequently congested and traffic accidents emerge one after another,the construction and management of intelligent transportation systems are particularly important.The monitoring of traffic intersections is usually achieved through surveillance cameras,which contain a large amount of video and image data.Therefore,intelligent image detection technology is of great significance in the intelligent management of urban traffic.Vehicles and license plates,which are important objects of concern in intelligent transportation,have been researched on target detection and character recognition algorithms based on deep learning.In order to enable the research results to meet the actual needs of my country’s traffic management,the research mainly uses surveillance cameras at the Forehead Bay Interchange in Wuhan enters the third ring road from west to east as the research material.5,000 key image frames are extracted from traffic videos to make the data set,which contains 9 target types: small cars,taxis,vans,tricycles,two-wheeler,trucks,buses,and blue and yellow license plates.Research on the target detection algorithm based on deep learning is carried out on this data set.The research algorithm is mainly based on the YOLOv4 network,and compared with the classic target detection algorithm Faster R-CNN and SSDNet algorithm.In order to recognize the license plate characters corresponding to the vehicles in traffic monitoring,the license plate in the surveillance video frame was intercepted,and a total of 2,717 license plate images were collected and made into a license plate data set.For the license plate data set,three convolutional neural network models are designed and the ZNet network model with the best training effect is selected through comparative experiments.In addition,experiments on other license plate character detection algorithms have been conducted to compare and verify the practicability and superiority of the license plate character recognition algorithm based on convolutional neural network in actual traffic photography.The YOLOv4 network algorithm is used to train and verify the vehicle data set,and the accuracy of the verification set can reach up to 95.21%;In the data processing of the YOLOv4 algorithm,certain restrictions are placed on the edge target features of the image,and a data enhancement method called image stitching method suitable for this data set is proposed.The experimental results show that compared with the original algorithm,the detection accuracy of the verification set can be improved by about 1%,and the accuracy of the algorithm under the 50% threshold is as high as 96.39%,and the number of frames processed per second is about 29.2.Compared with other classic target detection algorithms,the improved YOLOv4 network has certain advantages in terms of comprehensive detection efficiency and accuracy.The license plate recognition experiment based on convolutional neural network was carried out on the license plate data set,and the best-performing ZNet network model has the highest accuracy rate of95.28% on the verification set.In the network training,the training strategy of alternating data augmentation and non-data augmentation is discovered.This training strategy enhances the generalization ability of the network to a certain extent,successfully increasing the verification accuracy of the data set to about 96.65%.Finally,the improved YOLOv4 target detection algorithm is combined with the license plate character recognition algorithm based on ZNet to realize the unified detection of vehicles and license plates in traffic images.