Generative Difference Image for Blind Image Quality Assessment

Abstract

Image quality usually refers to the degree of error of the distorted image relative to the reference image in the human visual perception system. Image quality assessment is to score the image quality objectively. No-reference image quality assessment is limited to distorted image information, which is more challenging in the field of computer vision. In this paper, we proposed an approach based on difference image generation to address this problem. First, by removing the up-sampling layer and batch normalization layer in the Super-Resolution Generative Adversarial Network (SRGAN) to build a difference image generation model, and applying the content loss function to optimize the model. Then, the regression network is constructed based on the convolutional neural network (CNN). The regression network contains 4 convolutional layers and 2 fully connected layers and learns the correlation between the generated difference image and the image quality score to predict the distorted image quality. Finally, comparative experiments were evaluated on three public datasets. Compared with the previous state-of-the-art methods, our method obtains similar results on the LIVE dataset and achieves significant improvement on the TID2013 and CSIQ datasets. The results demonstrate that our proposed approach achieves state-of-the-art image quality prediction.