STARS: Stage-Adaptive Responsibility Separation for Shadow Removal in Remote Sensing Images
Organizations: Harbin Institute of Technology · Northeast Forestry University
Abstract
Shadows are a ubiquitous radiative attenuation phenomenon in remote sensing images and can severely affect high-resolution remote sensing interpretation tasks. In shadow removal, although introducing a mask prior provides explicit positional information for the restoration network, it also causes non-shadow regions to be modulated by the shadow prior, producing a systematic brightness shift; notably, this phenomenon is independent of whether the mask is accurate. In remote sensing images, non-shadow regions usually occupy a larger proportion, so a small shift over these regions incurs a much higher cost on whole-image metrics than an equal gain in shadow regions. Adopting a task-decoupling perspective, we therefore decompose shadow removal into two tasks, shadow-region restoration and non-shadow-region identity, within a cascaded architecture, and handle them sequentially. Based on this insight, we propose STARS, a cascaded shadow removal framework that separates regional responsibilities across stages. The method retains the advantages of the shadow mask while resolving the degradation of non-shadow regions, without requiring a complex parallel dual-branch structure. Meanwhile, to introduce large-model prior knowledge into remote sensing tasks, we further design a Feature Selection and Fusion Adapter (FSFA). We conduct extensive experiments on three benchmark datasets, UAV-SC+, ISTD+ and WSRD+ 2026, demonstrating the superiority of STARS. The framework shows that decoupling regional responsibilities can resolve the intrinsic contradiction of mask priors, providing a new direction for image restoration with region heterogeneity.
Figures & tables
| UAV-SC+ dataset Method PSNR (dB) SSIM LPIPS S. N.S. All Mask-ShadowGAN [ 31 ] 27.57 20.45 19.13 0.737 0.5754 G2R-ShadowNet [ 32 ] 24.69 17.98 16.83 0.718 0.3376 DMTN [ 18 ] 31.60 24.15 23.02 0.819 0.3059 DHAN [ 33 ] 33.06 25.17 24.13 0.862 0.3005 ShadowFormer [ 8 ] 32.71 24.37 23.44 0.835 0.3053 TBRNet [ 34 ] 31.22 23.66 22.57 0.772 0.3103 NeFour [ 35 ] 33.11 25.32 24.25 0.872 0.2989 RASM [ 19 ] 33.34 25.73 24.63 0.874 0.2994 RSMamba [ 17 ] 33.20 25.48 24.41 0.874 0.2774 Ours 34.57 28.08 26.65 0.897 0.0772 | ISTD+ dataset Method PSNR (dB) SSIM LPIPS S. N.S. All SP+M-Net [ 26 ] 36.16 35.84 32.33 0.9479 0.0716 Param+M+D-Net [ 36 ] 33.59 34.33 30.12 0.9420 0.0759 Fu et al. [ 37 ] 36.41 31.01 29.43 0.8404 0.1673 LG-ShadowNet [ 38 ] 32.65 33.36 29.20 0.9335 0.0920 DC-ShadowNet [ 39 ] 32.20 33.21 28.76 0.9219 0.1112 G2R-ShadowNet [ 32 ] 34.01 34.62 30.49 0.9330 0.0868 BMNet [ 40 ] 37.30 35.06 32.30 0.9551 0.0567 SG-ShadowNet [ 41 ] 37.41 34.95 32.41 0.9524 0.0594 Ours 38.72 37.47 34.22 0.9612 0.0206 |
Shadow Removal
Shadow removal restores plausible, shadow-free images while preserving scene content, color, texture, and boundary detail.
See all 16 papers