A Diffusion Model With Physically Plausible Gradient for Remote Sensing Shadow Removal

Remote sensing imagery is important for geographical object exploration, but shadow contamination consistently challenges the image formation quality and subsequent applications. Although the diffusion model significantly advances the shadow removal field, current paradigms ignore the physical property of shadow images and thus lose the desired interpretability. To bridge this gap, we propose SR-Diffusion, a shadow removal diffusion model that collaborates with infrared thermal distribution, chromaticity, and illumination intensity regulations. The core insight of SR-Diffusion is to inject nearly all available physical priors into the noise during the reverse process, thus providing desirable generative paths in noisy environments. Specifically, we leverage a modal translation (visible $\mapsto $ infrared) scheme to explore the cross-domain mapping, thus providing the thermal spectrum. Simultaneously, we introduce a novel horizontal/vertical-intensity (HVI) space to decouple the visible modality into chromaticity and illumination. Coupled with a gradient guidance, the above physical constraints are embedded into the noise, which contributes to generating stable shadow-free images. Comprehensive experiments demonstrate that SR-Diffusion outperforms state-of-the-art shadow removal methods.