The Research Of License Plate Recognition Method Based On Deep Learning In Natural Scenarios

As an important part of intelligent transportation system,automatic license plate recognition technology has always been a research hotspot in the field of computer vision.With the development of deep learning,new requirements proposed in safe cities and unmanned driving for license plate recognition technology,as well as the upgrading of challenges brought by complex license plate images taken by smart phones and mobile terminal devices,license plate recognition technology is facing new challenges and brings new research opportunities.In recent years,there have been numerous researches for license plate recognition,but most of those methods fail to analyze the characteristics of license plate images under natural scenerios in a comprehensive and detailed way.The recognition speed of traditional methods is fast,but they are mostly used for recognizing license plate images under controlled conditions,and not robust for varing scenerio factors;Those methods based on deep learning can perform better in the changeable natural scenerio by training and learning a large number of samples.However,the irregular distortion of license plate in natural scenerios is seldom taken into account,so there is a great room for improvement.In the natural scenerio,the interference of complex environment,uneven license plate image quality,irregular distortion of license plate shape,design rules of different countries and other influencing factors make license plate recognition very challenging.In order to improve the overall recognition result of license plate recognition in natural scenerios,we fully consider the shape characteristics of license plate,and propose a distorted license plate detection model in natural scenarios,named DLPD-Net(Distorted License Plate Detection Network).On this basis,a complete license plate detection and recognition system is proposed.The main work of this thesis is as follows:(1)License plate recognition mothods are summarized and analyzed.This thesis describes and analyzes license plate recognition methods in recent years in detail,and points out the shortcomings of existing methods by summarizing the characteristics of license plate image,which lays a foundation for the further study of license plate recognition.(2)A distorted license plate detection model in natural scenarios is proposed.The model uses encoder-decoder to extract feature maps,predicts the license plate center based on heat map and offset map,and then finds the license plate's corners based on affine transformation,finally corrects the distorted license plate to a plane rectangle close to the front view,so as to realize the distorted license plate detection in various natural scenarios.The validity and robustness of our model are verified by comparing the experimental results on AOLP and CD-HARD datasets.(3)A license plate detection and recognition system in natural scenarios is designed.Based on DLPD-Net model,a license plate detection and recognition system is proposed.Through the combination of several modules,license plate recognition can be realized in various distorted conditions.It has strong robustness and a very good recognition result in complex natural scenarios such as shielding,dirt and image blurring.We use this technology in myanmar license plate recognition system,which has good recognition rate and application performance.