Perceptual Friendly Variable Rate Image Compression

Yixin Gao, Yaojun Wu, Zongyu Guo, Zhizheng Zhang, Zhibo Chen; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 1916-1920

In this paper, we study high fidelity variable rate compression framework. Both conventional and learned codecs in prior works are optimized for objective quality commonly measured by PSNR or SSIM, leaving perceptual quality optimization underexplored. Besides, to circumvent the need of training separate models under different rate conditions, we design a novel coding framework to support variable rate compression. Aside from the variable rate functionality, we propose an adaptive bit allocation unit to strengthen rate-distortion optimization across different rates. Extensive experimental results demonstrate that our proposed approach achieves better subjective quality than methods optimized by the objective metrics such as MSE, and MS-SSIM on CLIC 2021 validation dataset.