Bao Jingwei ; Hao Jinhua ; Xu Pengcheng ; Sun Ming ; Zhou Chao ; Zhu Shuyuan

Abstract

High-resolution (HR) images are commonly downscaled to low-resolution (LR) toreduce bandwidth, followed by upscaling to restore their original details.Recent advancements in image rescaling algorithms have employed invertibleneural networks (INNs) to create a unified framework for downscaling andupscaling, ensuring a one-to-one mapping between LR and HR images. Traditionalmethods, utilizing dual-branch based vanilla invertible blocks, processhigh-frequency and low-frequency information separately, often relying onspecific distributions to model high-frequency components. However, processingthe low-frequency component directly in the RGB domain introduces channelredundancy, limiting the efficiency of image reconstruction. To address thesechallenges, we propose a plug-and-play tri-branch invertible block(T-InvBlocks) that decomposes the low-frequency branch into luminance (Y) andchrominance (CbCr) components, reducing redundancy and enhancing featureprocessing. Additionally, we adopt an all-zero mapping strategy forhigh-frequency components during upscaling, focusing essential rescalinginformation within the LR image. Our T-InvBlocks can be seamlessly integratedinto existing rescaling models, improving performance in both general rescalingtasks and scenarios involving lossy compression. Extensive experiments confirmthat our method advances the state of the art in HR image reconstruction.