MARSWIN: A Cross-Shot Swin Transformer Network for Super-Resolution on Mars MoRIC Image

Dust storms and fluctuating lighting conditions on Mars often obscure Martian surface information and degrade the quality of remote sensing images, limiting their applications for scientific exploration. Super-resolution (SR) has emerged as a promising technique due to its ability to enhance image quality efficiently. However, applying SR to Martian images faces significant challenges, including the limited availability of large-scale datasets and the difficulty of designing accurate degradation models. To address these issues, we consider the unique properties of overlapping Mars remote sensing images and design a novel cross-shot SR network, termed as MARSWIN. Specifically, a physically guided Mars moderate degradation model that considers atmospheric scattering is first proposed. In addition, to ensure the alignment of overlapping regions, we design an adaptive cross-shot attention module to utilize the interimage correlations and reconstruct the image region obscured by dust. An alignment loss is employed to further enhance the consistency constraints during image reconstruction. Extensive experiments on the Mars Moderate Imaging Camera dataset demonstrate the effectiveness of our framework. Compared to state-of-the-art SR methods, MARSWIN achieves an improvement of 2.20 dB in peak signal-to-noise ratio and 4.22 in erreur relative globale adimensionnelle de synthese at most. Besides, we conduct ablation studies to demonstrate the effectiveness of the physically guided high-order degradation model and the MARSWIN cross-shot SR network architecture.