Equivariant Bayesian Hyperspectral Imaging via Mosaiced and PAN Image Fusion

Area of Science:

• Computational Imaging
• Machine Learning
• Remote Sensing

Background:

• High-resolution hyperspectral imaging (HSI) is crucial for various applications.
• Current fusion methods for HSI often introduce spatial or spectral distortions.
• Achieving high-resolution HSI at video rates requires efficient fusion techniques.

Purpose of the Study:

• To develop an advanced fusion framework for high-resolution hyperspectral imaging (HSI).
• To mitigate spatial and spectral distortions inherent in existing HSI fusion methods.
• To enable accurate HSI reconstruction without ground truth data.

Main Methods:

• Proposed an equivariant Bayesian variational inference framework for HSI fusion.
• Decomposed HSI into principal component and sparsity residual, modeled as latent variables.
• Integrated equivariant imaging (EI) prior and learnable degradation functions (PSF, SRF) within a deep neural network (DNN).

Main Results:

• The framework effectively fuses low-resolution (LR) and high-resolution (HR) images, reducing distortions.
• Achieved enhanced parameter efficiency and reconstruction accuracy by leveraging shared spatial structures.
• Demonstrated effectiveness on both simulated and real-world datasets, validating the approach.

Conclusions:

• The proposed equivariant Bayesian variational inference framework significantly improves HSI fusion.
• The method enables accurate HSI reconstruction even without ground truth data.
• This work advances the field of HSI computational imaging, offering a robust solution for real-world applications.