Few-shot domain-adaptive hyperspectral image denoising via dynamic quantile-guided learning.

Area of Science:

• Optics and Photonics
• Computer Vision
• Data Science

Background:

• Hyperspectral imaging (HSI) offers rich spectral information but is hampered by noise.
• Existing deep learning denoising methods require extensive paired datasets, which are difficult to obtain for HSI.
• Cross-system variability and data scarcity limit the application of HSI.

Purpose of the Study:

• To develop a few-shot, domain-adaptive hyperspectral denoising framework (FT-DSES) to overcome data limitations.
• To enable accurate hyperspectral denoising with minimal paired samples for diverse optical imaging systems.
• To improve the fidelity and efficiency of hyperspectral image reconstruction.

Main Methods:

• FT-DSES integrates a spectral-enhanced spatial attention module and adaptive dynamic quantile pooling for noise-aware encoding.
• A two-stage adaptation scheme allows rapid transfer to new instruments using 5-8 paired samples.
• The framework is trained on synthetic noise models and fine-tuned on target domain data.

Main Results:

• FT-DSES achieves high-fidelity reconstruction (up to ~36.47 dB in PSRN, ~0.15 in SAM) with significantly reduced parameters and computation time.
• Demonstrates robust cross-source, cross-domain, and cross-modal generalization in photon-limited and data-limited scenarios.
• Shows significant gains in Peak Signal-to-Noise Ratio (PSNR) and Spectral Angle Mapper (SAM) across remote sensing, reflectance imaging, and fluorescence microscopy.

Conclusions:

• FT-DSES effectively addresses data scarcity and cross-system variability in hyperspectral denoising.
• The framework offers a computationally efficient and highly generalizable solution for diverse HSI applications.
• Enables high-quality hyperspectral data reconstruction even in challenging, low-data environments.