SwinMDR: Swin-Transformer-based Multi-Degradation Restoration for Planetary Images

Planetary remote sensing images are essential for analyzing the surface structures and geological features of celestial bodies, but their quality is often compromised by multi-degradations such as down-sampling, noise, and blurring due to camera limitations and transmission challenges. Existing deep learning methods, while effective for single-task restoration, struggle to address the multi-degradation in planetary data. This paper introduces SwinMDR (Swin-Transformer-based Multi-Degradation Restore Network), a unified restoration model employing a multi-head, multi-tail architecture with task-specific modules and a Swin-Transformer-based backbone. The MHTR-Net component performs targeted restoration for super-resolution, denoising, and deblurring, while the novel attention-driven CTCF-Module fuses outputs, thereby enabling end-to-end image restoration. Experiments on Ceres-1K demonstrate that SwinMDR achieves competitive performance in both objective assessment and perceptual quality.