Dong Wei ; Zhou Han ; Tian Yuqiong ; Sun Jingke ; Liu Xiaohong ; Zhai Guangtao ; Chen Jun

Abstract

Shadow-affected images often exhibit pronounced spatial discrepancies incolor and illumination, consequently degrading various vision applicationsincluding object detection and segmentation systems. To effectively eliminateshadows in real-world images while preserving intricate details and producingvisually compelling outcomes, we introduce a mask-free Shadow Removal andRefinement network (ShadowRefiner) via Fast Fourier Transformer. Specifically,the Shadow Removal module in our method aims to establish effective mappingsbetween shadow-affected and shadow-free images via spatial and frequencyrepresentation learning. To mitigate the pixel misalignment and further improvethe image quality, we propose a novel Fast-Fourier Attention based Transformer(FFAT) architecture, where an innovative attention mechanism is designed formeticulous refinement. Our method wins the championship in the Perceptual Trackand achieves the second best performance in the Fidelity Track of NTIRE 2024Image Shadow Removal Challenge. Besides, comprehensive experiment result alsodemonstrate the compelling effectiveness of our proposed method. The code ispublicly available: https://github.com/movingforward100/Shadow_R.