| With the improvement of the national economic level,people’s travel mode has gradually evolved to be dominated by cars.However,due to unreasonable road planning and traffic violations,road safety problems are becoming more and more serious,the phenomenon of illegal parking and illegal driving are more severe,which can lead to traffic congestion,traffic paralysis,or even worse,people would die in traffic accidents.As the “identity card”of a vehicle,license plate number can uniquely identify the vehicle,it plays an important role in the identification of the vehicle in traffic violations.Therefore,it is very important to identify a car accurately and rapidly in today’s traffic scenarios.In recent years,the way to solve traffic violations is using the law enforcement instrument to record some information in the scene,such as taking relevant photos and videos,in which the backgrounds are complex,and most of the photos are oblique,leads to low accuracy tested by traditional license plate recognition system under this law enforcement instrument scene,and of course the enforcement efficiency is low.The traditional ways can’t satisfy the law enforcement scenario.Thus,this paper did some research on license plate recognition problem under the law enforcement scenario,try to use deep learning-based method to solve it,and improve the enforcement efficiency.The main work of this paper includes the following four aspects:Firstly,this paper proposes a new license plate recognition process which includes three parts:license plate area detection,license plate key point detection and end to end license plate recognition.The previous recognition process is basically vehicle detection,license plate detection,license plate character detection or segmentation,and character classification,in which there is a chain effect,that each part of the process will have a direct impact on the subsequent process.Therefore,to simplify the whole process and improve the overall license plate recognition accuracy,this paper proposes a new license recognition process.Secondly,this paper made a high-quality dataset for plate recognition.In order to balance the car number of each province in license plate recognition module,we made a total of11550 license plate pictures includes one-layer yellow plate,two-layer yellow plate,blue plate,police car plate,new energy plate,as well as Hong Kong and Macao car plate,we add some light and shade in the plate and make it oblique with arbitrary angle,this batch of data varies in all kinds and in huge amount,its effectiveness has been proved in the license plate recognition experiments.In addition,this paper optimized a popular object detection algorithm and applied it into the license plate recognition under law enforcement scenario.This paper takes YOLOv4-Tiny as the base network,and made some optimization upon it.The experimental results show,that the model trained by the optimized network performed better in plate area detection and plate recognition,the accuracy is higher,in the meantime,model size and time-consuming didn’t change too much.Finally,this paper adopted Flask as the system’s framework,which is code-concise and easy to expand.Combined with the MVC pattern,this paper developed a license plate recognition system.Upload an image file or a bunch of images,the system will do some calculations and display the results,and store relevant information in the My SQL database.In addition,these data can also be used to analyze the deficiency of the system,contribute to do some optimization later.Focus on the license plate recognition problems under law enforcement scenario at present,this paper put forward a new plate recognition process and did some optimization on the object detection algorithm to solve the tilted license plate recognition problem,and did some experiments to approve the feasibility and effectiveness of the proposed process and the optimized network,meet the requirement of enforcement efficiency and achieved the aim of the reservation. |
PDF Full Text Request