Research On Image Quality Evaluation Based On Non-referenced Image

In the era of visual information,digital image has become the most extensive and efficient information carrier in daily life.However,in the process of image processing,transmission and compression,there are always a variety of distortions that affect the image quality.Therefore,it is particularly important to study algorithms that can accurately evaluate image quality.Due to the lack of reference undistorted image in practice,it is of great theoretical and practical significance to study the relationship between image distortion and image quality,and to build a non-reference image quality evaluation model consistent with human visual system.In this paper,three kinds of non-reference image quality assessment methods are proposed: multi-scale non-reference image quality assessment,image quality assessment based on adaptive weight and image quality assessment based on multi-task generative adversarial Network.The detailed research contents are as follows:(1)Most of the non-reference unreferenced image quality evaluation only pays attention to the overall information and ignores the important local information.And the difference between the distorted image and the undistorted image in detail is exactly where the distortion occurs.Through multi-scale,the distorted image can be analyzed and compared under different resolutions,which makes the detail difference of the image with different degree of distortion more prominent.Therefore,the algorithm of multi-scale non-reference image quality evaluation is proposed.Firstly,the Gaussian pyramid is used to construct the scale image to reflect the local structure changes of the distorted image.Then,two network structures(MS-F and MS-C)are formed according to the different objectives of the fusion,in which MS-F fusion is the scale feature and MS-C fusion is the prediction result.Finally,the network training is divided into two stages to better learn the whole image rather than the quality score of the image patch.Experiment results on different datasets to show the accuracy and generalization ability.(2)Because the final observer of the image is human,the image quality assessment requires that the predicted quality score is highly consistent with the human subjective score,so it is necessary to meet the human visual system's perception of image when building the model.However,for the dataset used in image quality assessment,researchers usually process all the image patches equally after non-overlapping segmentation,which is not in line with the visual characteristics of different content image patches should be paid different attention.Therefore,this paper proposes the image quality assessment based on adaptive weight.The single task network is transformed into a multiple task network,and the quality score of image patch and its corresponding weight are learned at the same time.Local quality score is transformed into global quality score by adaptive weight method.Experimental results show that this method can effectively improve the accuracy and achieve a high consistency with the subjective quality score.(3)Because of the undistorted image,the accuracy and generalization ability of the full-reference image quality assessment is often higher than that of the non-reference image.In order to narrow the gap between the two,this paper proposes a non-reference image quality assessment based on multi-task generative adversarial network.Firstly,the quality map generated by the auxiliary circuit is used to improve the reliability of the generated hallucination image,and then the hallucinated image is used as the missing reference image to convert the non-reference image quality evaluation into full-reference.By analyzing the difference between distorted image and hallucinated image to learn the characteristics related to subjective quality score,the problem that non-reference image quality assessment can not simulate human visual process is fundamentally solved.The experimental results show that the generated hallucinated image can restore the undistorted image well,and the method can accurately evaluate the image quality.