Title: PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing

Abstract:

Real-world image dehazing (RID) focuses on eliminating haze-induced degradations from actual scenes. This objective is notoriously difficult to achieve because of uneven haze distribution, spatially dependent color distortions, and a notable lack of paired data containing both hazy and clean images. To address these issues, we introduce Proximal Scattering Atmosphere Reconstruction (PSAR) within the PRISM framework. This physically grounded approach simultaneously reconstructs the clear scene and the associated scattering variables according to the atmospheric scattering model, thereby enhancing interpretability in complex, real-world environments.

To narrow the gap between synthetic and real data, we have developed an online non-uniform haze synthesis pipeline alongside a Selective Self-Distillation Adaptation (SSDA) strategy tailored for unpaired real-world scenarios. This mechanism allows the model to selectively absorb insights from high-quality perceptual targets while utilizing its inherent understanding of atmospheric scattering to detect residual haze and drive self-refinement. Our experiments on established real-world benchmarks confirm that PRISM delivers competitive results for RID tasks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC