Hyperspectral Image Super-resolution via Subspace-Based Low Tensor Multi-Rank Regularization.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral image (HSI) enhancement is crucial for detailed analysis.
• Combining low spatial resolution HSI (LR-HSI) with high spatial resolution multispectral images (HR-MSI) is a common approach to achieve high spatial resolution HSI (HR-HSI).
• Existing methods often struggle to fully exploit spectral correlations and non-local similarities.

Purpose of the Study:

• To propose a novel subspace-based low tensor multi-rank regularization method for fusing LR-HSI and HR-MSI.
• To enhance the spatial resolution of HSI by effectively leveraging spectral correlations and non-local similarities.
• To improve the accuracy and detail in the resulting HR-HSI.

Main Methods:

• Approximating the HR-HSI using a spectral subspace and coefficients.
• Learning the spectral subspace from LR-HSI via singular value decomposition (SVD).
• Estimating coefficients using a low tensor multi-rank prior, grouping patches based on HR-MSI cluster structure, and applying regularization to grouped tensors.
• Solving coefficient optimization using the alternating direction method of multipliers (ADMM).

Main Results:

• The proposed method effectively exploits spectral correlations and non-local self-similarities in HR-HSI.
• Experimental results on two public HSI datasets demonstrate the advantages of the proposed fusion technique.
• The method achieves superior performance in enhancing HSI spatial resolution compared to existing approaches.

Conclusions:

• The subspace-based low tensor multi-rank regularization method is a powerful tool for HSI enhancement.
• This approach offers a significant improvement in fusing LR-HSI and HR-MSI for generating HR-HSI.
• The findings highlight the potential of tensor multi-rank priors for modeling complex image structures in remote sensing applications.