Self-Asymmetric Invertible Network for Compression-Aware Image Rescaling

High-resolution (HR) images are usually downscaled to low-resolution (LR)ones for better display and afterward upscaled back to the original size torecover details. Recent work in image rescaling formulates downscaling andupscaling as a unified task and learns a bijective mapping between HR and LRvia invertible networks. However, in real-world applications (e.g., socialmedia), most images are compressed for transmission. Lossy compression willlead to irreversible information loss on LR images, hence damaging the inverseupscaling procedure and degrading the reconstruction accuracy. In this paper,we propose the Self-Asymmetric Invertible Network (SAIN) for compression-awareimage rescaling. To tackle the distribution shift, we first develop anend-to-end asymmetric framework with two separate bijective mappings forhigh-quality and compressed LR images, respectively. Then, based on empiricalanalysis of this framework, we model the distribution of the lost information(including downscaling and compression) using isotropic Gaussian mixtures andpropose the Enhanced Invertible Block to derive high-quality/compressed LRimages in one forward pass. Besides, we design a set of losses to regularizethe learned LR images and enhance the invertibility. Extensive experimentsdemonstrate the consistent improvements of SAIN across various image rescalingdatasets in terms of both quantitative and qualitative evaluation understandard image compression formats (i.e., JPEG and WebP).