A Physics-Guided Diffusion Framework for Mars Rover Image Restoration Under Atmospheric Dust

Atmospheric dust reduces the contrast and visibility of Mars rover images, while strictly aligned clean–dusty image pairs remain scarce. To address this problem, we propose a two-stage physics-guided diffusion framework. In the first stage, we propose GenDust, which combines synthetic pairs based on an atmospheric scattering model with observed unpaired dusty images to construct training data that better match real dust degradation. In the second stage, we propose Physics-Guided Diffusion-based Mars Dust Skip-step Restoration (DiffMDSR) and design PhysNet to estimate spatial transmission as a physical prior for latent diffusion restoration. We further design a codebook-based latent constraint and timestep remapping strategy to regularize skip-step sampling. The framework was evaluated on 200 real dusty Curiosity Mastcam images, 30 controlled Perseverance Mastcam-Z samples with known clean references, and 30 independent dusty Mastcam-Z images. In the controlled experiment, DiffMDSR achieved a peak signal-to-noise ratio of 31.44 dB, a structural similarity index measure of 0.870, and a learned perceptual image patch similarity of 0.182, showing improved consistency with the corresponding clean references. The independent Mastcam-Z experiment further assesses transfer to previously unseen mission and camera conditions. The restored images are intended mainly for scene inspection and cross-observation comparison, while the equivalent optical depth derived from PhysNet-estimated transmission characterizes the relative image-space attenuation.