Weighted Schatten $p$-Norm Minimization for Image Denoising with Local and Nonlocal Regularization

Yuan Xie

TL;DR

This paper introduces a novel image denoising method that combines a weighted Schatten $p$-norm minimization for low-rank matrix approximation with local and nonlocal regularization. This approach effectively removes noise while simultaneously reducing artifacts, outperforming existing state-of-the-art methods both visually and quantitatively.

What Problem Does This Paper Address?

Digital images often suffer from noise introduced during acquisition and transmission, making them unsuitable for various applications. Image denoising aims to suppress this noise while preserving crucial image details. Traditional patch-based denoising methods often rely on low-rank matrix approximation, typically using nuclear norm minimization (NNM). However, NNM has limitations: it can suppress important singular values, leading to data shrinkage, and treats all singular values equally, ignoring their physical significance. Furthermore, the aggregation of denoised patches in these methods can introduce artifacts. This paper addresses these issues by proposing a more robust low-rank approximation and a sophisticated regularization strategy to mitigate artifacts.

Methods

The paper introduces two main innovations:

  1. Weighted Schatten $p$-Norm Minimization (WSNM): This extends the traditional nuclear norm minimization by incorporating weights for singular values and allowing for a non-convex $p$-norm ($0 < p < 1$). This allows for more flexible approximation of the low-rank assumption and better consideration of the physical meaning of different data components. An efficient algorithm is developed to solve this WSNM-based low-rank matrix approximation problem.
  2. Local and Nonlocal Regularization: To combat artifacts arising from patch aggregation, the authors propose a novel regularization framework that combines Steering Total Variation (STV) and Nonlocal Total Variation (NLTV). STV captures local image correlations, while NLTV enforces nonlocal consistency. This combined approach is integrated into a variational model, and a fast, efficient optimization algorithm is developed to solve it, avoiding complex partial differential equations.

These two components are integrated into an iterative restoration framework to produce the final denoised image.

Datasets

The paper evaluates the proposed method on standard test images, including "House," "Monarch," "Barbara," "Boat," and "Butterfly." The authors do not introduce new datasets or benchmarks.

Key Insights and Ideas

  • Beyond Nuclear Norm: The paper demonstrates that the standard nuclear norm minimization can lead to suppression of important singular values and data shrinkage. The proposed WSNM, by incorporating weights and a $p$-norm, offers a more nuanced approach to low-rank approximation.
  • The Power of $p$: The choice of the power $p$ in WSNM is shown to be crucial and dependent on the noise level. Lower $p$ values are better for lower noise levels, while higher $p$ values are preferred for stronger noise.
  • Regularization for Artifacts: The combination of local (STV) and nonlocal (NLTV) regularization is essential for effectively removing artifacts that plague simpler patch-based methods. This dual approach ensures both local detail preservation and global consistency.
  • Efficient Solvers: Developing fast and efficient algorithms for both the WSNM optimization and the variational regularization problem is key to the practical applicability of the proposed method. The paper presents algorithms that are significantly faster than existing approaches.

Key Results

  • Quantitative Improvement: The proposed method, termed CWSNM (Constrained Weighted Schatten $p$-Norm Minimization), consistently achieves higher Peak Signal-to-Noise Ratio (PSNR) values compared to state-of-the-art methods like BM3D, PBNO, SAIST, and WNNM across various noise levels. For instance, at higher noise levels (ση = 50 and 100), CWSNM shows notable PSNR gains over WNNM (0.1-0.12dB on average) and BM3D (0.32-0.52dB on average).
  • Qualitative Improvement: Visual comparisons (e.g., in Figures 7, 8, and 9) demonstrate that CWSNM produces visually superior denoised images with fewer artifacts and better preservation of image textures and structures, especially in challenging scenarios with high noise.
  • Computational Efficiency: The proposed fast algorithm for the variational model is significantly faster than the split Bregman method, offering a reduction of up to 37.44% in CPU time for larger image sizes. The method also converges faster.

Limitations and Future Work

  • Parameter Tuning: The performance of the proposed method relies on the careful selection of several parameters, including the power $p$ for WSNM, regularization parameters $\mu_1$ and $\mu_2$, and patch sizes. While the paper provides guidance, optimal parameter selection can still be challenging.
  • Low-Rank Tensor Recovery (LRTR): The authors suggest extending the WSNM framework to low-rank tensor recovery (LRTR) as a future direction. This would involve adapting the methods to handle multi-dimensional data structures, which could benefit applications beyond 2D image denoising.
  • Broader Applications: The paper highlights the potential of WSNM and the proposed regularization techniques for various computer vision applications, such as background modeling, structure from motion, and MRI data recovery.

Summary updated

Abstract

This paper presents a patch-wise low-rank based image denoising method with constrained variational model involving local and nonlocal regularization. On one hand, recent patch-wise methods can be represented as a low-rank matrix approximation problem whose convex relaxation usually depends on nuclear norm minimization (NNM). Here, we extend the NNM to the nonconvex schatten p-norm minimization with additional weights assigned to different singular values, which is referred to as the Weighted Schatten p-Norm Minimization (WSNM). An efficient algorithm is also proposed to solve the WSNM problem. The proposed WSNM not only gives better approximation to the original low-rank assumption, but also considers physical meanings of different data components. On the other hand, due to the naive aggregation schema which integrates all the denoised patches into a whole image, current patch-wise denoising methods always produce various degree of artifacts in denoised results. Therefore, to further reduce artifacts, a data-driven regularizer called Steering Total Variation (STV) combined with nonlocal TV is derived for a variational model, which imposes local and nonlocal consistency constraints on the patch-wise denoised image. A highly simple but efficient algorithm is proposed to solve this variational model with convergence guarantee. Both WSNM and local \&amp; nonlocal consistent regularization are integrated into an iterative restoration framework to produce final results. Extensive experimental testing shows, both qualitatively and quantitatively, that the proposed method can effectively remove noise, as well as reduce artifacts compared with state-of-the-art methods.
BibTeX
@misc{xie2015weighted,
  author = {Yuan Xie},
  title = {{Weighted Schatten $p$-Norm Minimization for Image Denoising with Local and Nonlocal Regularization}},
  year = {2015},
  eprint = {1501.01372},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1501.01372},
}
Figure