Tensor Cascaded-Rank Minimization in Subspace: A Unified Regime for Hyperspectral Image Low-Level Vision.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Low-rank tensor representation is vital for hyperspectral image (HSI) analysis.
• Previous methods often fail to fully exploit HSI's low-rank properties across different modes.
• Existing approaches typically address only a single HSI low-level vision task.

Purpose of the Study:

• To propose a unified framework for HSI low-level vision tasks.
• To develop a novel tensor cascaded rank minimization method (STCR).
• To comprehensively exploit spatial and spectral correlations in HSI data.

Main Methods:

• Projecting high-dimensional HSI into a low-dimensional tensor subspace.
• Employing a novel tensor low-cascaded-rank decomposition to exploit spatial, nonlocal, and spectral modes.
• Utilizing difference continuity-regularization for endmember approximation.

Main Results:

• The proposed STCR regime effectively integrates low-rankness across different HSI domains.
• Demonstrated high effectiveness and robustness on eight datasets for denoising, compressive sensing reconstruction, inpainting, and destriping.
• Achieved state-of-the-art (SOTA) performance compared to existing methods.

Conclusions:

• STCR offers a comprehensive approach to HSI tensor analysis.
• The cascaded rank minimization effectively captures complex HSI correlations.
• The unified framework is versatile for multiple HSI low-level vision applications.