Variable Rate ROI Image Compression Optimized for Visual Quality

Yi Ma, Yongqi Zhai, Chunhui Yang, Jiayu Yang, Ruofan Wang, Jing Zhou, Kai Li, Ying Chen, Ronggang Wang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 1936-1940

With the development of compression technology, objective metrics (e.g. PSNR, MS_SSIM) cannot satisfy our need, especially in extreme low bit-rate compression, thus more attention is being paid on perceptual quality. Since people have different standards for objective evaluation. For this reason, we simplify the topic with the consideration that people will strict more on interested region, so a ROI(region of interest) based image compression model is proposed with team name 'Sub201'. For the ROI, we expect its reconstructed part to be more accurate, while the background, server distortion is tolerable, and fake texture can be generated. Firstly, a weighted mask from saliency map is used. Secondly, to balance the difference of ROI and background area, different losses are applied separately. What's more, GAN and LPIPS are utilized to generate more texture in background. At last, variable rate method is adopted to realize rate control, and it performs well with perceptual metric. Experiment shows that our method can achieve better performance both in visual and objective quality.