Video License Plate Detection And Recognition In The Open Environment

License plate detection and recognition has been researched more than 30 years,and has been widely used in some simple scenes,but in the open environment where cameras are not fixed,poor lighting,etc.,the effect of existing license plates detection and recognition algorithm will drop dramatically.In recent years,deep learning has been widely used in the field of computer vision.It is far superior to traditional algorithms in terms of robustness and accuracy.Therefore,this paper attempts to apply deep learning to license plate detection and recognition,and achieve real-time and accurate results in the open environment.This paper aims at the detection and recognition of video license plate in the open environment,considering the real-time and accuracy,combined with the actual application scenarios,the license plate detection algorithn LPDNet(License Plate Detection Network)based on YOLOv3(You Only Look Once v3)and the license plate recognition algorithm LPSRNet(License Plate Sequence Recognition Network)based on CRNN(Convolutional Recurrent Neural Network)are proposed.The main work and innovations of this paper are as follows:(1)Establish license plate detection and recognition benchmark datasetAt present,the only large dataset of Chinese license plates is CCPD published by the University of Science and Technology of China,but the shooting scene is fixed,the license plate has high similarity and each image contains only one license plate,which is not suitable for the benchmark dataset.In the aspect of license plate detection,this paper establishes a detection data training set which contains 5854 images by collecting and labeling images by myself,and establishes a detection benchmark testing set which consists of 1050 images of easy,hard and extreme;In the aspect of license plate recognition,this paper proposes a method based on data synthesis to construct a license plate recognition training set,and constructs a license plate recognition testing set composed of 500 license plate images.(2)License plate detection algorithm based on YOLOv3Based on the improvement of YOLOv3,this paper proposes LPDNet for license plate detection,and experiments proves that the algorithm can achieve good detection results in the open environment such as poor illumination,license plate tilt,motion blur,and occlusion,and it has real-time performance.(3)License plate recognition algorithm based on CRNNBased on the improvement of CRNN,this paper proposes LPSRNet for license plate character sequence recognition.Experiments show that the algorithm can achieve good recognition effect on the license plate under the conditions of blur,poor illumination and tilt,and it has real-time performance.Experimental has shown that the proposed LPDNet license plate detection algorithm can achieve the precision of 100%,the recall of92.65%and the detection speed of 16FPS.The LPSRNet license plate recognition algorithm can achieve the accuracy of 94%and the speed of 3ms for one license plate image.The final model of license plate detection and recognition can achieve the accuracy of 86.65%and the speed of 10FPS,which is better than other license plate detection and recognition algorithms.