Research And Implementation Of License Plate Recognition In Low-quality Surveillance Video

Intelligent transport systems effectively improve the efficiency of traffic and transportation,and reduce traffic accidents by real-time monitoring and intelligent management of road conditions.As an important part of the intelligent transport systems,the license plate recognition makes the intelligent transport systems more convenient.However,plenty of factors cause the low quality of surveillance video,such as the motion blur of vehicles,the defocus of equipment,the low pixels of monitoring devices,the influence of rain and snow weather,the darkness of the shooting environment or the illumination of strong light.These factors will increase the difficulty of license plate recognition.In this paper,we analyze the difficulties of license plate recognition in surveillance video of low quality,and make research and implementation of key technologies such as license plate image preprocessing,license plate character segmentation,license plate character recognition and End to End license plate recognition.The main results of the research are as follows:Firstly,this paper introduces the theory of license plate recognition,and then two frameworks——the license plate recognition framework based on character segmentation and the End to End license plate recognition framework are proposed and analyzed.Secondly,in view of the license plate recognition framework based on character segmentation,we propose a character segmentation algorithm based on sliding window and a character recognition algorithm using Dictionary Pair Learning.Segmentation algorithm based on sliding window combines sliding window with recognition errors.Once sliding recognition is done,the window with minimum recognition error is segmented,which can achieve better segmentation effect for license plate in low quality surveillance video.In the recognition link,the dictionary pairs learning model is used to improve the traditional dictionary learning,and the computation speed is greatly reduced in the case of ensuring the recognition rate.Thirdly,we improve the above character segmentation algorithm and character recognition algorithm respectively.Then we propose the character segmentation algorithm based on connected domain with regression model,and the character recognition algorithm based on convolutional neural network.The character segmentation algorithm based on connected domain analysis with regression model can extract character width and character location adaptively,and can get accurate segmentation results for many kinds of low quality vehicle license plates.In recognition link,we use convolution neural network to build 520 thousand license plate character training library and 10 layer network structure in order to optimize the parameters of common network according to the character of license plate,and achieve high recognition rate for license plate characters in low quality surveillance video.Fourthly,we use deep learning to realize End to End license plate recognition."Convolutionactivation-pooling" in convolutional neural networks can extract feature.To realize multi-label learning,we build a deep convolution network composed of three "convolution activation-pooling" modules and a fully connected network.Better recognition results can be obtained from the license plate recognition based on character segmentation.