Title: Degradation-Aware Metric Prompting for Hyperspectral Image Restoration

Abstract: Unified hyperspectral image (HSI) restoration seeks to address a variety of degradations within a single model. Yet, existing approaches frequently depend on explicit priors that are difficult to apply in practice or utilize opaque black-box representations that tend to overfit to training data, thereby limiting their ability to generalize to novel situations. To address these limitations, we introduce Degradation-Aware Metric Prompting (DAMP), a new framework that describes multi-dimensional degradations using interpretable spatial-spectral metrics. These metrics function as Degradation Prompts (DP), allowing the model to identify common traits across different tasks and adjust to previously unseen corruptions. The core of our approach is the Degradation-Adaptive Mixture-of-Experts (DAMoE) architecture. Within this structure, Spatial-Spectral Adaptive Modules (SSAMs) act as experts, employing learnable fusion coefficients to focus on specific levels of degradation severity. The DAMoE mechanism uses the DP as a gating router to dynamically select and activate the experts best suited for the particular degradation profile. Comprehensive evaluations on both natural and remote sensing HSI datasets show that DAMP delivers state-of-the-art results and demonstrates remarkable zero-shot generalization capabilities on restoration tasks involving unknown data. The code is open-source and can be accessed at \href{DAMP}{https://github.com/MiliLab/DAMP}.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC