Unsupervised Spectral Demosaicing With Lightweight Spectral Attention Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Hyperspectral Imaging

Background:

• Supervised deep learning methods for spectral demosaicing often fail on real-world data, especially with increasing spectral bands.
• Existing unsupervised methods lack robustness and struggle with spatial distortion and spectral fidelity.

Purpose of the Study:

• To develop a comprehensive unsupervised spectral demosaicing (USD) framework that overcomes limitations of supervised methods.
• To enhance the dynamic modeling of spectral correlations within a compact parameter space.

Main Methods:

• A novel unsupervised deep learning framework for spectral demosaicing.
• Reduced complexity spectral attention module using spatial and channel dimension decomposition.
• Introduction of Mosaic25, a real-world 25-band hyperspectral mosaic dataset.

Main Results:

• The proposed USD method outperforms conventional unsupervised techniques on synthetic and real-world datasets.
• Demonstrated improvements in spatial distortion suppression, spectral fidelity, robustness, and computational efficiency.
• The Mosaic25 dataset provides a valuable benchmark for hyperspectral image processing.

Conclusions:

• The developed unsupervised spectral demosaicing framework offers a robust and efficient solution for hyperspectral imaging.
• The novel approach and dataset advance the field of spectral demosaicing, particularly for real-world applications.
• Public availability of code and dataset facilitates further research and development.