| As the society developing rapidly,the number of motor vehicles steadily surge so that it brings great challenges to traffic management.The license plate is the most important standard for identifying vehicles.Therefore,as an application of image processing,computer vision and deep learning technology in the field of intelligent transportation,the license plate recognition system is an indispensable part of the intelligent transportation management system.A growing number of illegal parking behaviors exist in the city.The owners do not obey the traffic regulations,for example,parking ineligibly.Such a behavior further exacerbates the inherent contradiction in the urban traffic with more cars and fewer roads,seriously affecting the urban road Traffic order,and even threatening the safety of citizens traveling.Therefore,how to accurately identify illegal vehicle license plate numbers,generate illegal notices,and assist traffic police to manage illegal parking behaviors through mobile terminals and artificial intelligence has become a technology staple.Most of the existing license plate recognition algorithms adopt a centralized training method,and the training data need to be concentrated in a definite machine or a single data center.The condition that the illegal parking license plate training data collected by the traffic police involves personal information can easily cause privacy leakage during the centralized training.In addition,most mainstream license plate detection and recognition algorithms are based on deep convolutional neural networks to complete the design of the model.The privacy leakage of existing algorithms and the large number of model parameters make it difficult for these algorithms to be effectively applied to mobile terminals.Solving these problems has high practical application value.This paper starts with the application scenarios of traffic police punishing roadside violations,and studies the license plate detection and recognition algorithm under the condition of inclined license plates.This paper mainly consists of the following innovative work.(1)The algorithm model of license plate detection is studied.This paper gives a brief overview of the classic two-stage target detection algorithm Mask-RCNN and the single-stage detection algorithm YOLO.Given the advantages of the two detection algorithms and the application scenarios of mobile devices,a license plate detection algorithm Yo-Mask based on the Mobilenet network is proposed.The algorithm has a favorable detection speed while maintaining a high recall rate.Simultaneously,the model containing a handful of parameters can be effectively deployed on mobile devices.(2)The license plate recognition algorithm based on federated learning is studied.This paper briefly introduces the basic principles of federated learning,and on the basis of this principle,a license plate recognition algorithm framework based on federated learning that completes model training on the user’s mobile phone is designed.During the whole algorithm training process,the server will not involve the user data.Only the updated parameters will interact without any leakage of user data. |
