Abstract
In this paper, we propose a novel nonconvex approach to robust principal component analysis (RPCA) for hyperspectral image (HSI) denoising. Our method focuses on simultaneously developing more accurate approximations to both rank and column-wise sparsity for the low-rank and sparse components, respectively. Specifically, we adopt a log-determinant rank approximation and a novel $\ell_{2,\log}$ norm to restrict local low-rank and column-wise sparse properties. For the $\ell_{2,\log}$-regularized shrinkage problem, we develop an efficient, closed-form solution named the $\ell_{2,\log}$-shrinkage operator. Furthermore, we impose spatial-spectral total variation (SSTV) regularization to enhance global piece-wise smoothness and spectral consistency. Extensive experiments on simulated and real HSIs demonstrate the effectiveness of the proposed method.
Introduction
Hyperspectral images (HSIs) capture rich spectral information across hundreds of narrow bands, but they are frequently corrupted by complex "cocktail" noise, including Gaussian noise, impulse noise, deadlines, and stripes. Traditional denoising methods often rely on Robust Principal Component Analysis (RPCA), which decomposes an image into a low-rank clean component and a sparse noise component. However, standard RPCA models typically use convex relaxations—the nuclear norm for rank and the $\ell_{2,1}$ norm for sparsity—which can be "too loose," leading to over-smoothed results and biased estimates.
The $L^3S^3TV$ Framework
To overcome the limitations of convex models, we propose the $L^3S^3TV$ (Log-based Local Low-rank and Sparse separation with Spatial-Spectral Total Variation) model. This approach introduces several key innovations:
- Non-convex Log-based Approximations: We replace standard convex norms with log-determinant functions for rank and a novel $\ell_{2,\log}$ norm for column-wise sparsity. These non-convex functions more accurately mimic the true rank and sparsity by penalizing small values (noise) while preserving large signal components.
- Local Patch Processing: The HSI is divided into overlapping 3D patches to exploit local low-rankness, capturing fine-grained spectral signatures that global models might miss.
- Spatial-Spectral Total Variation (SSTV): To ensure spatial and spectral consistency across the recovered image, we integrate a global SSTV term that enforces piecewise smoothness.
A major mathematical contribution of this work is the derivation of a closed-form solution for the $\ell_{2,\log}$ shrinkage problem. This operator allows the non-convex problem to be solved with computational efficiency comparable to simpler convex models.
Experimental Evaluation
We evaluated the $L^3S^3TV$ method against seven state-of-the-art baselines, including BM4D, LRMR, and LRTDTV, using both simulated and real-world datasets such as Indian Pines and HYDICE Urban data.
Quantitative Results
In simulated experiments across six noise cases, our method consistently achieved the highest Mean Peak Signal-to-Noise Ratio (MPSNR) and Mean Structural Similarity Index (MSSIM). For example, in Case 1 (Gaussian noise), $L^3S^3TV$ improved MPSNR by over 2 dB compared to the next best method (LRTDTV).
Qualitative Results
Visually, the proposed method excels at removing complex artifacts like stripes and deadlines while preserving sharp edges and fine textures. Unlike traditional methods that may leave visible artifacts or blur structural details, $L^3S^3TV$ maintains high spectral integrity, which is vital for downstream applications like mineral identification.
Conclusion
The $L^3S^3TV$ model significantly advances HSI denoising by combining accurate non-convex approximations with local and global structural regularizations. The efficiency of the new $\ell_{2,\log}$-shrinkage operator makes this approach highly practical for complex image restoration tasks and potentially applicable to broader fields like subspace clustering and multi-task learning.
Summary updated