Research Of No-reference Image Quality Assessment Algorithm Based On Deep Learning

Image quality assessment can provide an important index and basis for evaluating the performance of image algorithm and optimizing image processing system.No-reference image quality assessment aims at accurately predicting the visual quality of any image without a reference image.Concerning the issue and its application in image denoising algorithm,we conduct exhaustive study.The main content of this paper can be summarized as the following four points:(1)Considering the distortion information in the color components of the image,we study the expression of the image in the HSV color space,and design a CNN model based on the hue component and gray information to extract the distortion features.(2)Through the study of the human visual system,we find that human is more sensitive to the distortion of salient regions.Therefore,we propose a novel no-reference image quality assessment method based on visual saliency weighting.The experiment results on the standard image database show that the objective quality predicted by this method is highly consistent with the human subjective perception quality.Considering the problem of multiply distorted image quality assessment,we fine-tune the CNN model trained on single distorted image database.The experiment results show that it can accurately predict the multiply distorted image quality.(3)Considering the application of no-reference image quality assessment in image processing system,we apply the image quality assessment model to the image denoising algorithm,and provide the basis for parameter selection,so as to achieve the parameter optimization.The experiment results show that the no-reference image quality evaluation method based on CNN can accurately evaluate the quality of denoised image.Therefore,we propose a parameter selection framework based on no-reference image quality assessment,which optimizes the parameter of denoising algorithm based on image quality and improves the algorithm performance.