基于生成感知差异的无参考图像质量评价模型

人眼视觉系统中的视觉感知差异是图像质量评价过程中的重要组成部分，通过感知失真图像与原始图像之间的视觉差异，可对图像的失真程度进行判断，然而在无参考图像质量评价中无法获取原始未失真的图像，且缺乏对失真图像的视觉感知差异。通过对深度学习中的生成对抗网络进行分析，提出一种基于生成视觉感知差异的无参考图像质量评价模型。利用生成对抗网络产生与失真图像相对应的视觉感知差异图像，并将其与失真图像输入质量评价网络以进一步学习图像的失真信息，从而达到评估图像质量的目的。在TID2008和TID2013数据库上的实验结果表明，与CNN、SOM、CORNIA等模型相比，该模型能够使失真图像质量预测准确度提升1个百分点以上，且对不同种类失真也表现出良好的预测性能。

No-Reference Image Quality Assessment Model Based on Generated Perceptual Difference

The visual perception difference in the Human Visual System(HVS)is key to image quality evaluation.By sensing the visual difference between the distorted image and the reference image,the distortion degree of the image can be judged.However,this method is inadequate for No-Reference Image Quality Assessment(NR-IQA)where the original undistorted image cannot be obtained for the judgement of visual perception difference.By analyzing the Generative Adversarial Network(GAN)in deep learning,this paper describes a NR-IQA model based on generated perceptual differences.The model employs GAN to generate the visual perception difference image corresponding to the distorted image,and then inputs both kinds of images into the Image Quality Assessment(IQA)network to further learn the distortion information of the image,completing the evaluation of the image quality.The experimental results on TID2008 and TID2013 databases show that compared with CNN,SOM,CORNIA and other models,the proposed model can improve the prediction accuracy of distorted image quality by more than 1 percentage point,and shows excellent performance for different types of distorted images.