RGB denoisers repurposed for hyperspectral image restoration
A new method leverages pretrained RGB denoisers for hyperspectral image restoration, addressing challenges like limited training data and high spectral dimensionality. The approach uses a lightweight adapter that projects hyperspectral data into a low-dimensional space, applies frozen RGB denoisers, and reconstructs the cube via linear aggregation. Experiments on denoising, deblurring, and super-resolution show consistent improvements over hyperspectral-specific baselines, demonstrating strong transferability of large-scale RGB priors.
Key facts
- Hyperspectral image restoration faces limited training data, sensor specificity, and high spectral dimensionality.
- Proposed method repurposes frozen pretrained RGB denoisers via a projection mapping.
- Method denoises low-dimensional spectral projections and reconstructs via constrained linear aggregation.
- Plug-and-play compatibility and stability properties of underlying RGB denoiser are preserved.
- Experiments on denoising, deblurring, and super-resolution across multiple datasets.
- Consistent improvements over hyperspectral-specific baselines.
- Shows strong transferability of large-scale RGB priors.
- Method is minimally trained and lightweight.
Entities
—