Research On Vehicle License Plate Recognition System Based On Deep Learning

In the increasingly frequent use of vehicles in the society,smart vehicle management is also becoming more and more important in our traffic management.The license plate is the only accurate vehicle identification,and its identification is an indispensable step for management,and license plate recognition.Timeliness and accuracy are very important for vehicle management.A good license plate recognition system can greatly reduce the workload of vehicle management and improve the efficiency and quality of traffic management.Therefore,it is necessary for the study of license plate recognition.Many scholars have also proposed many different algorithms and implementations for license plate recognition.This paper analyzes and researches various algorithms proposed for collecting key technologies at home and abroad for license plate recognition.Combined with the recent deep learning in the image field,this paper proposes a vehicle license plate recognition method based on deep learning.And innovations are as follows:1.For license plate location,coarse positioning based on color information method is proposed,and a method of combining regional characters based on character features is proposed to screen the candidate areas that have been initially selected,and in addition to reducing the noise of the pictures,the Morphological methods are also used.2.For tilt correction,we propose a method based on deep learning to classify license plates according to different degrees of tilt and angles,then construct a deep learning network with this data set,identify them,and identify the type of tilt.It is corrected by a preset affine transformation.This method has a speed accuracy rate of 96.4%,3.For character cutting,we use a method based on projection information and cut characters from both sides according to character transition characteristics.4.For character recognition,we also use a deep learning method based on the characteristics of license plate characters to construct a convolutional network with an accuracy of 99.8%.5.Finally,we conducted forward acceleration research on deep learning,completed the implementation of C code in the network,and conducted an HLS-based acceleration study on the core convolutional code part.