TradMS‐ResGAN: A Sequential Multi‐Stage Traditional Framework With Residual Multiscale GAN for Fast and Robust Sand‐Dust Image Restoration

Sand‐dust degradation severely impairs visibility, colour fidelity, contrast, and structural detail, significantly affecting outdoor surveillance, autonomous navigation, and remote sensing applications. This paper presents TradMS‐ResGAN, a physics‐guided sequential restoration framework that integrates a Dust‐Aware Degradation Prior (DADP)‐driven structured enhancement pipeline with a lightweight multi‐scale residual GAN. Unlike conventional end‐to‐end approaches, the proposed method models sand‐dust degradation as a structured composition of wavelength‐dependent chromatic distortion, contrast suppression, dynamic range compression, and visibility attenuation, which are progressively mitigated through six physically interpretable stages based on global channel‐wise statistical correction and adaptive enhancement principles. These stages include chromatic compensation, contrast restoration, tone correction, degraded region refinement, edge‐preserving enhancement, and multi‐exposure fusion, forming a deterministic inversion pathway prior to data‐driven refinement. A lightweight residual GAN is then employed to remove residual artefacts and reconstruct fine textures while maintaining computational efficiency suitable for embedded deployment. Experimental evaluations on both paired synthetic and unpaired real‐world sand‐dust datasets under diverse atmospheric conditions demonstrate consistent improvements in visibility restoration, structural preservation, and perceptual quality. Further analysis, including runtime, parameter efficiency, and embedded deployment on Raspberry Pi 4 and edge‐class devices, confirms the practicality of the proposed design. By explicitly decomposing degradation physics into a structured sequential inversion process and coupling it with efficient residual learning, TradMS‐ResGAN provides a robust and computationally efficient solution for real‐world sand‐dust image restoration.