Degradation-Aware Metric Prompting for Hyperspectral Image Restoration

TL;DR

DAMP addresses unified hyperspectral image restoration without explicit degradation priors by introducing Degradation Prompts (DP)—a compact set of spatial-spectral metrics—and a Spatial-Spectral Adaptive Module (SSAM) within a Degradation-Adaptive MoE (DAMoE). DP provides degradation-aware conditioning while SSAM enables dynamic, task-relevant feature fusion, allowing the model to adapt to diverse, mixed, or unseen degradations. The approach yields state-of-the-art results on natural and remote-sensing HSIs and demonstrates strong zero-shot generalization to unseen degradations, with minimal runtime overhead. The work offers a practical, interpretable pathway to robust, single-model HSIR across varied real-world degradation scenarios, and provides code for reproducibility.

Abstract

Unified hyperspectral image (HSI) restoration aims to recover various degraded HSIs using a single model, offering great practical value. However, existing methods often depend on explicit degradation priors (e.g., degradation labels) as prompts to guide restoration, which are difficult to obtain due to complex and mixed degradations in real-world scenarios. To address this challenge, we propose a Degradation-Aware Metric Prompting (DAMP) framework. Instead of relying on predefined degradation priors, we design spatial-spectral degradation metrics to continuously quantify multi-dimensional degradations, serving as Degradation Prompts (DP). These DP enable the model to capture cross-task similarities in degradation distributions and enhance shared feature learning. Furthermore, we introduce a Spatial-Spectral Adaptive Module (SSAM) that dynamically modulates spatial and spectral feature extraction through learnable parameters. By integrating SSAM as experts within a Mixture-of-Experts architecture, and using DP as the gating router, the framework enables adaptive, efficient, and robust restoration under diverse, mixed, or unseen degradations. Extensive experiments on natural and remote sensing HSI datasets show that DAMP achieves state-of-the-art performance and demonstrates exceptional generalization capability. Code is publicly available at https://github.com/MiliLab/DAMP.

Figures (9)

• Figure 1: PSNR comparison with the state-of-the-art all-in-one methods: Inpainting, Super Resolution, Gaussian Deblurring, and Gaussian Denoising results are evaluated on the ARAD dataset after unified training, while Poisson Denoising and Motion Deblurring are reported as zero-shot results on the CAVE dataset. $[\cdot]$ denotes the range of PSNR values across different methods.
• Figure 2: (a) Comparison between explicit prompt-based methods and degradation-aware metric prompting approaches. (b) Confusion matrix for classifying five degradation types based on HFER, STU and SCM. (c) Distribution of different degradation types across the HFER, STU and SCM.
• Figure 3: (a) The architecture of the proposed DAMP framework. (b) Transformer blocks. (c) The Degradation-Adaptive MoE.
• Figure 4: Visual comparison of HSI recovery performance on tasks with known degradation types. From top to bottom: super-resolution on the ARAD dataset arad2022NITRE, denoising on the ICVL dataset arad2016ICVL, deblurring on the PaviaU dataset huang2009PaviaU, and inpainting on the Xiong’an dataset yi2020xiongan. The content within the small red boxes in each image is magnified, with the left side showing the error map compared to the ground truth (GT), and the right side displaying the magnified result images.
• Figure 5: Visual comparison of Poisson denoising on the CAVE Dataset CAVE. The first row shows the restoration results, and the second row displays the error maps.