Image Quality Assessment Based On Feature Learning

Information is an essential element of human life,and human visual system is the primary approach to obtain information among all the ways.Image serves as an important information source of human perception and machine recognition,while on one hand has brought great economic and social benefits,but it also faced the problem of image quality during the process of collection,compression,processing and transmission on the other hand.The study of image quality evaluation can be helpful to adjust the system parameters in real-time and achieve better results,so the assessment of image quality is very importantand meaningful.The basic meaning of image quality refers to subjective evaluation of visual perception about an image.However,it is not accepted by researchers due to inconvenient,time-consuming and expensive in practical applications,Therefore,more attentions have been paid to objective image quality assessment methods and applications.The main purpose of objective image quality evaluation methods is to construct a system which comparable to human subjective evaluation,and gives results consistent with human beings.According to the dependence of the reference image,the objective image evaluation methods can be divided into three categories,which are the quality evaluation of the full reference image,the quality evaluation of the reduced reference image and the quality evaluation of the no reference image or blind image quality assessment.Blind image quality assessment(BIQA)is the most challenge and difficult problem in the filed of image quality assessment,Different methods have been put forward in terms of this issue.In this paper,we design a blind image quality evaluation system from the perspective of feature learning.More specifically two methods which named the discriminative sparse representation and the deep convolution neural network,are presented to illustrate how to apply the trained features from a large number of samples to the image quality evaluation.The main work is summarized as follows:(1)The research background and significance of image quality evaluation are introduced,and the research status of image evaluation method is reviewed.Moreover,this paper describes the theoretical basis of feature learning method.The sparse representation problem and methods of generating dictionary based on the dictionary learning are introduced mainly.On the other hand,some aspects of deep learning such as development history,working principle,main network layer and optimizationmethods are described.Then the similarities and differences between the two are compared in detail.(2)The dictionary learning based on sparse representation can be a good learning characteristics of the image,so a new method is proposed to evaluate the quality of discriminative dictionary,the evaluation results are obtained by the positive and negative quantified coefficients.Then,the public standard data sets of different degrees of Gaussian white noise and different doses of CT images are tested.The experimental results show that the proposed method is more reasonable and consistent with subjective evaluation.(3)Furthermore,the quality evaluation method based on discriminative sparse representation is extended to the problem of image blur degradation.A corresponding discriminative dictionary is trained from blur images,and then quantitative validations are processed with natural images and DR images data sets,the test results show that this algorithm is effective for the evaluation of the blured images.At the same time,a single blured dictionary is trained from the view of signal decomposition,and it is applied to the problem of image blur region detection.The final results show that the method is effective.(4)A new method of image quality assessment based on deep learning is proposed.In terms of the noise/artifacts,an image to be evaluated is mapped into the degenerate residual part by the convolution network model,and then the image is evaluated by analyzing the residuals.The results of a variety of natural images as well as CT image data show that the proposed method has the advantages of high prediction accuracy and computational speed.