Vehicle License Plate Recognition Based On Deep Learning

Vehicle license plate recognition is an important technology in Smart City,and it is widely used in many areas of intelligent traffic management.At present,license plate recognition has a good performance in some specific scenes.However,the license plate location and recognition effects in natural environments caused by inconsistent camera angles,inconsistent vehicle positions and insufficient image lighting,etc.are still unsatisfactory.Deep learning has had a transformative impact on the development of computer vision,thus all the best-effect target detection and optical character recognition algorithms use deep learning currently.Therefore,this thesis attempts to build a license plate recognition algorithm which is faster,more accurate,and more robust in complex natural environments based on deep learning technology.This thesis uses the CCPD data set published by the University of Science and Technology of China,vehicle pictures taken by mobile phones and road cameras to construct a vehicle image data set covering various complex scenes.The author divides the license plate recognition algorithm into two sub-tasks: license plate location and license plate character recognition,focusing on the algorithm design of the two tasks.In order to take into account the recognition speed and accuracy at the same time,the author proposes an LPL-YOLO license plate location algorithm and a GCRNet license plate recognition algorithm.The specific work and innovations are introduced below.For the license plate location part,based on the YOLOv3 algorithm,this thesis adds the SPP module,optimizes the algorithm according to the characteristics of the license plate location,and designs the LPL-YOLO network for small objects,improving the prediction accuracy of the target frame position,width and height by transforming the loss function.Also,a method for extracting anchor box size based on Canopy +K-Means is proposed.In addition,a model compression method based on sparse training is adopted to cope with large model parameters' data,and the model parameter quantity is reduced by more than 98%,almost without changing the accuracy.For the license plate character recognition part,this thesis designs a method for enhancing the license plate data.The recognition algorithm is in the structure of CNN + Bi GRU +CTC,which can recognize license plate in indefinite length without segmenting characters.Besides,a GCR module that effectively fuses global CNN features and local GRU features is designed,and a GCRNet license plate recognition algorithm is proposed.Experiments show that the LPL-YOLO license plate location algorithm and GCRNet license plate recognition algorithm have real-time detection speed,higher prediction accuracy and more robustness.They also have a good recognition effect on vehicle images such as severe license plate tilt,too bright or insufficient lighting and low graphics resolution.Among them,the LPL-YOLO license plate location algorithm can achieve 100% accuracy and recall in common scenarios,the GCRNet license plate recognition algorithm can reach 99.38% license plate recognition accuracy,and the final comprehensive algorithm of license plate location and character recognition can reach the recognition accuracy of 98.2% with the recognition speed of 0.06s per picture.