A Training Method for Image Compression Networks to Improve Perceptual Quality of Reconstructions

Jooyoung Lee, Donghyun Kim, Younhee Kim, Hyoungjin Kwon, Jongho Kim, Taejin Lee; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 144-145

Recently, neural-network based lossy image compression methods have been actively studied and they have achieved remarkable performance. However, the classical evaluation metrics, such as PSNR and MS-SSIM, that the recent approaches have been using in their objective function yield sub-optimal coding efficiency in terms of human perception, although they are very dominant metrics in research and standardization fields. Taking into account that improving the perceptual quality is one of major goals in lossy image compression, we propose a new training method that allows the existing image compression networks to reconstruct perceptually enhanced images. By experiments, we show the effectiveness of our method, both quantitatively and qualitatively.