Research And Application Of Deep Learning In License Plate Recognition

Taking intelligent transportation as an entry point,achieving a smart city is one of the directions for future urban development.As an important part of intelligent traffic management,license plate recognition plays an active role in alleviating traffic congestion,saving manpower and time costs,improving work efficiency,and improving management models.The traditional license plate recognition algorithm has a great influence on the image quality.It needs a fixed camera shooting angle and a clear shooting image.Therefore,it is mostly used in some simple scenes such as parking lots,toll stations,and so on.As computer hardware and software infrastructures improve,it becomes possible to use more sophisticated algorithms to achieve better recognition performance.The license plate recognition algorithm in this paper is divided into two parts: license plate target detection and license plate character recognition.The main tasks are as follows:(1)The research status of the license plate recognition technology at home and abroad and the target detection technology based on deep learning are reviewed,and the future development trend of license plate recognition technology using deep learning algorithm is analyzed.The basic neural network algorithm is studied and analyzed,which provides theoretical support for the subsequent neural network design.(2)According to the license plate target detection task,the specifications and characteristics of the domestic vehicle number plate were studied,and the advantages and disadvantages of the traditional target algorithm were analyzed.The YOLOv2 deep learning model was modified as the technical measure of the vehicle license plate target detection.Compared with the traditional algorithm,this algorithm does not need to do tedious preprocessing on the input image,supports multi-size images,and achieves multi-scale detection and small-object detection.The detection accuracy is high,and for a 1080 P video stream,the processing speed reaches 50 fps at 1080 Ti.(3)For the license plate character recognition task,the advantages and disadvantages of the traditional algorithm are analyzed.Based on the analysis of current technology requirements,an end-to-end license plate character recognition algorithm based on deep learning is designed.Compared with traditional algorithms,this algorithm does not require tilt correction and character segmentation for license plate images,and it also has a good performance for severely worn characters.The algorithm of this paper realizes the accurate recognition of license plate characters,and satisfies the requirement of intelligent traffic on the accuracy of license plate recognition system.