Perceptual JPEG artifact removal using weighted sum of IQAs as loss function

Abstract

The challenge with existing JPEG artifact removal methods is that the restored images are excessively smoothed, resulting in low perceptual quality. This is due to the use of the L1 or L2 norm as the loss function, which has a low correlation with perceptual quality. A common approach to solving this problem, the adversarial generative network (GAN), makes the learning process unstable. In this paper, we propose a new perceptual JPEG artifact removal method by using the weighted sum of multiple IQAs as a loss function. Furthermore, we modify the upscaling architecture of the existing method to prevent periodic artifacts caused by changes in the loss function. Experimental results show that the proposed method significantly improves the perceptual quality of artifact removed images quantitatively and qualitatively compared to the existing baseline.