Hyperspectral Image Denoising via Weighted Multidirectional Low-Rank Tensor Recovery

Area of Science:

• Remote Sensing
• Image Processing
• Signal Processing

Background:

• Low-rank tensor recovery methods are popular for hyperspectral image (HSI) denoising.
• Existing methods often treat different structural dimensions equally, limiting denoising effectiveness.
• Subspace representation techniques analyze HSI data but can be improved by considering mode-specific properties.

Purpose of the Study:

• To develop an advanced HSI denoising method by incorporating multidirectional low-rank regularization.
• To address the limitations of existing methods that indiscriminately analyze structural information in HSI data.
• To improve the accuracy and efficiency of hyperspectral image mixed noise removal.

Main Methods:

• Investigated low-rank properties within the subspace of HSI data.
• Proved stronger structure correlation in the nonlocal self-similarity mode compared to spatial sparsity and spectral correlation modes.
• Introduced a novel multidirectional low-rank regularization assigning distinct weights to each mode for tensor rank estimation.
• Developed an optimization model for HSI mixed noise removal using the proposed regularization within a subspace-based tensor recovery framework.
• Employed the alternating minimization algorithm for efficient model optimization.

Main Results:

• The proposed multidirectional low-rank regularization effectively characterizes the contribution of each mode to tensor rank estimation.
• The developed optimization model demonstrated superior performance in removing mixed noise from hyperspectral images.
• Extensive experiments on synthetic and real HSI data confirmed significant improvements over state-of-the-art denoising methods.
• The method showed enhanced capability in preserving crucial structural information during the denoising process.

Conclusions:

• The proposed method, leveraging weighted multidirectional low-rank regularization, offers a significant advancement in hyperspectral image denoising.
• The findings highlight the importance of considering mode-specific structural correlations for effective tensor recovery in HSI.
• The developed approach provides a robust and efficient solution for hyperspectral image mixed noise removal, outperforming existing techniques.