Research On Unconstrained Scene License Plate Recognition Method Based On Deep Learning

Intelligent transportation is an important part of smart cities.With the continuous advancement of intelligent transportation technology,automatic license plate recognition system has attracted more and more attention.However,the existing automatic license plate recognition system cameras are often deployed at fixed positions and angles,with a single scene.In actual unconstrained scenarios such as vehicle driving and road safety,the license plate pictures captured by the camera often have complex backgrounds,different shooting angles,and bad weather.In addition,vehicles of different colors,types,and character layouts,such as new energy vehicles and coach vehicles,may appear in the actual traffic scene.To solve the above problems,this paper studies the automatic license plate recognition method in unconstrained scenarios,aiming to obtain an automatic license plate recognition algorithm model with high recognition accuracy,strong real-time performance and good robustness.The main work is as follows:(1)An end-to-end single license plate automatic recognition model with a lightweight architecture is proposed.The model is trained in an end-to-end manner,and a spatial transformation network is introduced to correct the tilted license plate,and a data enhancement strategy based on prior information is proposed to improve the recognition accuracy and robustness of the model in various scenarios.Results show that the introduced spatial transformation network improves the recognition accuracy by an average of 0.6 percentage points on the CCPD dataset.At the same time,the data augmentation strategy based on prior information further improves the recognition accuracy by 0.2 percentage points.(2)To solve the problem of recognizing multiple license plates in a single image in unconstrained scenarios,this paper proposes a cascaded model that can identify different types of individual license plates as well as multiple license plates.The model detects and screens out effective vehicle targets through YOLOv5 l and post-processing methods.To improve the real-time performance,the license plate detection and correction modules are designed as parallel branches,Res Net18 and BLSTM networks are used to further improve the accuracy of license plate character recognition.The model can show higher recognition speed and recognition accuracy on multiple sub-data sets of CCPD and AOLP and multi-license plate data sets collected by the network,and has stronger robustness.To intuitively demonstrate the excellent performance of the model,this paper also provides a client display page,where inputting the test image can retrieve information on vehicles,corrected license plates,and license plate characters.