Tensor Completion via Nonlocal Low-Rank Regularization.

Area of Science:

• Computer Vision
• Data Science
• Remote Sensing

Background:

• Tensor completion (TC) is crucial for recovering degraded hyperspectral images (HSIs).
• Existing TC methods often use convex trace norm penalties, leading to biased solutions and failing to capture complex data structures.
• The assumption of global low-rank for HSIs can limit the recovery of detailed information.

Purpose of the Study:

• To propose a novel nonlocal low-rank regularization-based tensor completion (NLRR-TC) method for HSIs.
• To improve the accuracy and detail recovery in HSI data completion.
• To address the limitations of existing TC methods in handling complex HSI structures.

Main Methods:

• Developed a two-step NLRR-TC approach for HSI data.
• Step 1: Introduced a low-rank regularization-based TC (LRR-TC) model combining determinant logarithm and tensor trace norm for initial completion.
• Step 2: Integrated nonlocal spatial-spectral similarity by grouping similar HSI cubes and applying LRR-TC to each group.

Main Results:

• The proposed LRR-TC model adaptively tunes logarithm function values for better tensor rank approximation.
• Grouping nonlocal similar cubes and applying LRR-TC to each group effectively leverages local low-rank properties.
• Experimental results show the NLRR-TC method significantly outperforms state-of-the-art HSIs completion techniques.

Conclusions:

• The NLRR-TC method provides a superior approach for HSI tensor completion.
• Leveraging nonlocal spatial-spectral similarities enhances the recovery of detailed information in HSIs.
• The proposed method offers a more effective solution compared to existing TC techniques for HSI data.