Research On Key Technologies Of Licence Plate Image Quality Assessment In Traffic Surveillance Video

The license plate recognition algorithm is an important research topic in ITS and has played an increasingly important role in many applications over the past few decades.However,at present,the latest license plate recognition algorithms still can not perform well in different scenes,including sunny days,nighttime,and complex backgrounds with different colors.Therefore,the sharpness evaluation and pre-classification of the license plate image is an effective way to enhance the adaptability of the license plate recognition algorithm in different scenarios.The main contents of this article include the following four aspects:(1)From the perspective of image quality assessment,the features commonly used to describe the quality of non-referenced images are extracted from the license plate images,and then the quality of license plate images is estimated and graded.Finally,the analysis of common non-reference image quality assessment algorithm can not be well applied to the license plate image reason.(2)From the perspective of image restoration,blur kernel estimation based on regularized image blind deblur algorithm is used to classify the obtained blur kernel parameters.Finally,the analysis of the reasons why the image blind deblur algorithm can not be applied to the license plate image well.(3)On the basis of sparse representation and reconstruction error,this paper proposes a license plate image quality classification algorithm,which can effectively divide license plate images into two categories: common quality and low quality.By making an overcomplete dictionary of two types of license plate images and extracting the error of the test license plate image reconstructed from the two dictionaries as a feature vector and classifying them by a support vector machine,the definition category of the test license plate image can be accurately obtained.(4)Using convolutional neural network method to achieve a license plate image quality classification model.Firstly,a convolutional neural network model with three convolution layers and two full connection layers is constructed to classify the license plate images quality,and a high classification accuracy is achieved.Then by changing the number of convolution layers and changing the learning rate,the network model is adjusted and optimized so that the convolution neural network can optimize the accuracy of the license plate image classification.