Image Deblurring Based on Convex Non-Convex Sparse Regularization and Plug-and-Play Algorithm

Yi Wang, Yating Xu, Tianjian Li, Tao Zhang, Jian ZouAlgorithms 2023 · DOI

TL;DR

This paper introduces a novel image deblurring method that combines a "convex non-convex" (CNC) sparse regularization technique with a "plug-and-play" (PnP) algorithm. This approach aims to overcome the limitations of traditional sparse regularization methods, leading to improved deblurring performance and robustness to noise.

What Problem Does This Paper Address?

Image deblurring is a crucial task in computer vision, aiming to recover a sharp image from a blurred one. A common approach involves sparse regularization, which leverages prior knowledge about image sparsity. However, existing methods face challenges: convex sparse regularization (like L1-norm) can lead to biased estimations, particularly for high-magnitude image components, while non-convex methods (like Lp-norm with p<1) are difficult to solve and can get stuck in local optima. Furthermore, traditional iterative deblurring algorithms can be slow and computationally expensive. This paper seeks to address these issues by proposing a more effective regularization strategy and a more efficient algorithmic framework.

Methods

The core of this work lies in two main innovations:

  1. Convex Non-Convex (CNC) Sparse Regularization: The authors introduce a CNC sparse regularization term, $\phi_b(x) = ||x||_1 - s_b(x)$, where $s_b(x)$ is the Moreau envelope of the L1-norm. This formulation aims to mitigate the estimation bias of convex methods while ensuring the overall convexity of the deblurring objective function by carefully selecting the non-convexity parameter $b$. The paper proves conditions under which the objective function remains convex, a critical step for algorithmic convergence.
  2. Plug-and-Play Forward-Backward Splitting (PnP-FBS) Algorithm: To efficiently solve the deblurring problem with the new regularization, the authors adapt the Forward-Backward Splitting (FBS) algorithm. The key idea of PnP is to replace the proximal operator (a standard component in optimization algorithms for non-smooth functions) with a pre-trained image denoiser. This allows leveraging the power of deep learning-based denoisers within a principled iterative optimization framework. Specifically, they develop an FBS-CNC algorithm and then a PnP-FBS-CNC algorithm by substituting the proximal operators with two distinct denoisers, $D_{\sigma_1}$ and $D_{\sigma_2}$, to handle different noise levels.

Datasets

The paper evaluates its proposed method on the CBSD68 dataset, a standard benchmark for image restoration tasks, which contains 68 color images. They also use eight real-world camera shake kernels from Levin et al. [17]. Gaussian noise is added at three levels: $\nu \in \{0.01, 0.03, 0.05\}$.

Key Insights and Ideas

  • Bridging the Gap: The CNC regularization elegantly bridges the gap between convex and non-convex approaches, aiming for the best of both worlds: reduced bias and solvable optimization problems.
  • The Power of PnP: Replacing proximal operators with powerful deep learning denoisers is a highly effective strategy for enhancing iterative optimization algorithms, leading to improved performance without sacrificing theoretical convergence guarantees.
  • Convexity is Key: Proving the conditions for the overall convexity of the objective function is crucial for ensuring that the proposed PnP-FBS-CNC algorithm converges to a meaningful solution.
  • Denoiser Choice Matters: The performance of the PnP algorithm is directly tied to the quality of the denoiser used. The paper highlights the use of state-of-the-art denoisers.

Key Results

The proposed PnP-FBS-CNC algorithm demonstrates superior performance compared to other state-of-the-art methods like IRCNN, DPIR, and GS-PnP.

  • Quantitative: The algorithm consistently achieves the highest average Peak Signal-to-Noise Ratio (PSNR) across various noise levels and blur kernels. For instance, at a noise level of 0.01, PnP-FBS-CNC achieved the highest PSNR values for all blur kernels, with values reaching up to 34.69 dB. The CPU time is comparable to other methods, indicating a good balance between performance and efficiency.
  • Qualitative: Visual comparisons show that PnP-FBS-CNC produces images that are more closely aligned with the clean image, exhibiting superior clarity and finer texture restoration, especially in challenging regions like water or sky.

Limitations and Future Work

  • Denoiser Dependency: The performance of the PnP-FBS-CNC algorithm is heavily reliant on the quality and characteristics of the denoisers used. The paper notes that the computational complexity can vary significantly depending on the type of denoiser (TV, BM3D, or DNN).
  • Hyperparameter Tuning: While the paper sets uniform parameters for its method, optimal performance might require fine-tuning of hyperparameters like $\alpha$, $\lambda$, and $b$ for different image types and noise conditions.
  • Future Work: The paper suggests that the PnP framework, combined with CNC regularization, offers a powerful direction for future research in image restoration tasks. Further exploration into different types of denoisers and more complex regularization schemes could yield even better results.

Summary updated

Abstract

Image deblurring based on sparse regularization has garnered significant attention, but there are still certain limitations that need to be addressed. For instance, convex sparse regularization tends to exhibit biased estimation, which can adversely impact the deblurring performance, while non-convex sparse regularization poses challenges in terms of solving techniques. Furthermore, the performance of the traditional iterative algorithm also needs to be improved. In this paper, we propose an image deblurring method based on convex non-convex (CNC) sparse regularization and a plug-and-play (PnP) algorithm. The utilization of CNC sparse regularization not only mitigates estimation bias but also guarantees the overall convexity of the image deblurring model. The PnP algorithm is an advanced learning-based optimization algorithm that surpasses traditional optimization algorithms in terms of efficiency and performance by utilizing the state-of-the-art denoiser to replace the proximal operator. Numerical experiments verify the performance of our proposed algorithm in image deblurring.
BibTeX
@inproceedings{wang2023image,
  author = {Yi Wang and Yating Xu and Tianjian Li and Tao Zhang and Jian Zou},
  title = {{Image Deblurring Based on Convex Non-Convex Sparse Regularization and Plug-and-Play Algorithm}},
  booktitle = {Algorithms},
  year = {2023},
  doi = {10.3390/a16120574},
}
Figure