Abstract
Plug-and-Play methods constitute a class of iterative algorithms for imaging problems where regularization is performed by an off-the-shelf denoiser. Although Plug-and-Play methods can lead to tremendous visual performance for various image problems, the few existing convergence guarantees are based on unrealistic (or suboptimal) hypotheses on the denoiser, or limited to strongly convex data terms. In this work, we propose a new type of Plug-and-Play methods, based on half-quadratic splitting, for which the denoiser is realized as a gradient descent step on a functional parameterized by a deep neural network. Exploiting convergence results for proximal gradient descent algorithms in the non-convex setting, we show that the proposed Plug-and-Play algorithm is a convergent iterative scheme that targets stationary points of an explicit global functional. Besides, experiments show that it is possible to learn such a deep denoiser while not compromising the performance in comparison to other state-of-the-art deep denoisers used in Plug-and-Play schemes. We apply our proximal gradient algorithm to various ill-posed inverse problems, e.g. deblurring, super-resolution and inpainting. For all these applications, numerical results empirically confirm the convergence results. Experiments also show that this new algorithm reaches state-of-the-art performance, both quantitatively and qualitatively.
Introduction
Image restoration (IR) tasks like deblurring or super-resolution are "ill-posed" inverse problems, meaning there are many possible clean images that could have produced the same degraded observation. To solve them, we typically minimize an objective function consisting of a data-fidelity term (measuring how well the solution fits the observation) and a regularizer (incorporating prior knowledge about clean images).
Plug-and-Play (PnP) methods have emerged as a powerful framework that replaces the proximal operator of the regularizer with an off-the-shelf deep learning denoiser. While they produce visually stunning results, they often lack robust convergence guarantees. Most existing proofs require the denoiser to be "non-expansive" or are limited to specific data-fidelity terms. Without these guarantees, PnP algorithms can be unstable or diverge in practice.
Methodology: The Gradient Step Denoiser
The authors propose GS-PnP (Gradient Step PnP), a framework that ensures convergence to stationary points of an explicit global objective function. The core innovation lies in the structure of the denoiser $D_\sigma$:
- Denoiser as a Gradient Step: Instead of treating the denoiser as a black-box network, it is defined as a gradient descent step on a regularization functional $g_\sigma$: $D_\sigma = \text{Id} - \nabla g_\sigma$.
- Deep Parameterization: The functional $g_\sigma$ is parameterized using a neural network $N_\sigma$: $g_\sigma(x) = \frac{1}{2} \|x - N_\sigma(x)\|^2$. This ensures $g_\sigma$ is differentiable, allowing $D_\sigma$ to be implemented with standard differentiable architectures.
- Proximal Gradient Descent: By embedding this denoiser into a PnP scheme, the algorithm becomes a Proximal Gradient Descent (PGD) iteration. This allows the use of optimization theory to prove convergence to stationary points of the global functional $F(x) = f(x) + \lambda g_\sigma(x)$, even in non-convex settings.
- Backtracking: A backtracking strategy is employed to automatically tune the step size, ensuring stability without manually estimating the Lipschitz constant of the network's gradient.
Experimental Results
The denoiser was trained on a composite dataset of clean color images (CBSD, DIV2K, etc.) and evaluated on standard benchmarks like CBSD68 and Set3C. GS-PnP reaches or exceeds the performance of top-tier PnP methods like DPIR and IRCNN across deblurring, super-resolution, and inpainting tasks.
A critical advantage of GS-PnP is its explicit energy minimization. Unlike standard PnP, which often lacks a clear objective, GS-PnP targets a global energy function, providing better numerical control and interpretability. The authors provide clear evidence that the objective function values converge as predicted by their theory.
Quantitative gains are particularly notable in super-resolution, where GS-PnP consistently outperforms other convergent methods like RED. The method is also empirically shown to be robust to different starting images, likely due to the strong data-fidelity steps in the early iterations.
Conclusion
GS-PnP provides a principled way to integrate deep learning denoisers into iterative restoration algorithms with guaranteed convergence. By structuring the denoiser as a gradient step, the authors bridge the gap between high-performance deep learning and rigorous optimization theory. Future work may explore the representation power of this constrained architecture and its scalability to large-scale asynchronous optimization.
Summary updated