Proximal Denoiser for Convergent Plug-and-Play Optimization with Nonconvex Regularization

Samuel Hurault, Arthur Leclaire, Nicolas Papadakis

Introduction: The Plug-and-Play Dilemma

In the world of image restoration—encompassing tasks like deblurring, super-resolution, and inpainting—researchers often frame the challenge as an inverse problem. The goal is to recover a clean image $x$ from a degraded observation $y$. Mathematically, this is typically solved by minimizing a "data-fidelity" term (how well $x$ fits the observed data) alongside a "regularization" term (a prior dictating how much $x$ resembles a natural image).

Plug-and-Play (PnP) methods revolutionized this field by replacing the complex, hand-crafted mathematics of the regularization term with a pre-trained deep neural network denoiser. While these methods produce stunning visual results, they are notoriously difficult to analyze mathematically. Because deep denoisers operate as "black boxes," it is often unclear if the optimization algorithm will actually converge, or what specific objective function it is implicitly trying to minimize. Previous attempts to prove convergence often required forcing the denoiser to be "nonexpansive"—a strict mathematical constraint that significantly degrades the final image quality.

This paper provides a solid mathematical foundation for PnP image restoration by proving that specific "gradient-step" deep denoisers are actually proximal operators of nonconvex functions. This breakthrough allows the authors to guarantee that PnP versions of popular optimization algorithms will reliably converge to stationary points of an explicit objective function without sacrificing image quality.

The Mathematical Bridge: Gradient-Step Denoisers

The authors bridge the gap between empirical success and theoretical rigor by focusing on a specific architecture called a Gradient-Step (GS) Denoiser. Instead of using a generic, unconstrained network, a GS denoiser is defined as $D_\sigma = \text{Id} - \nabla g_\sigma$, where $g_\sigma$ is a scalar function parameterized by a neural network.

The paper's key technical contribution (Proposition 3.1) proves that if the residual of this denoiser is contractive (meaning it doesn't "stretch" the mathematical space too much), then the denoiser is mathematically equivalent to a proximal operator of a (possibly nonconvex) function $\phi_\sigma$.

By establishing this link, the authors can apply powerful tools from nonconvex optimization theory—specifically the Kurdyka-Lojasiewicz (KL) property—to prove convergence for three major PnP frameworks:

  1. PnP-PGD (Proximal Gradient Descent)
  2. PnP-ADMM (Alternating Direction Method of Multipliers)
  3. PnP-DRS (Douglas-Rachford Splitting)

To ensure their deep denoiser (based on the DRUNet architecture) meets these necessary mathematical conditions, they fine-tune it using a custom loss function that penalizes the spectral norm of the Hessian of $g_\sigma$.

Key Insights

  • From Black Box to Explicit Prior: A deep denoiser doesn't have to be a heuristic trick. If structured as a gradient step, it corresponds to an explicit (though nonconvex) prior. This transforms PnP into a rigorous optimization method.
  • Constraining the Residual: Previous work tried to make the denoiser itself nonexpansive, which is overly restrictive and blurs details. This paper shows it is much more effective to constrain the residual (the difference between the input and the denoised output). This preserves high-frequency details while still providing convergence guarantees.
  • Predictable Behavior: Because the denoiser acts as a proximal operator, the PnP algorithm is guaranteed to minimize a specific, explicit functional $F(x) = \lambda f(x) + \phi_\sigma(x)$. This makes the algorithm's behavior predictable and monitorable.

Empirical Validation and Results

The GS denoiser was trained on a large database of clean images and evaluated on the standard CBSD68 benchmark dataset for deblurring, super-resolution, and inpainting tasks.

Visual and quantitative results comparing IRCNN and the proposed Prox-PnP methods on a deblurring task, demonstrating both high image quality and steady objective convergence.

The authors empirically demonstrated that while popular methods like DPIR can fail to converge (with residual errors oscillating wildly), their Prox-PnP methods converge steadily to a stationary point.

Furthermore, this mathematical rigor does not come at the cost of performance. On CBSD68 deblurring tasks, their method achieved a competitive PSNR of 31.54 dB, outperforming the well-known IRCNN (31.42 dB) and performing comparably to the state-of-the-art DPIR (31.93 dB)—but with the crucial added benefit of guaranteed stability. The PnP-DRS version proved particularly robust to initialization, consistently finding high-quality solutions even when started from heavily degraded images.

Limitations and Future Directions

While a significant step forward, the authors note a few limitations:

  • Convexity Assumptions: One of the stronger convergence theorems requires the image of the denoiser to be a convex set, which is difficult to verify for complex neural networks in practice.
  • Parameter Tuning: The theory provides bounds for the regularization parameter $\lambda$, but finding the exact "sweet spot" for optimal visual quality still requires manual tuning or automatic search.
  • Geometry of Denoisers: The authors suggest that future work should investigate the geometry of the set of images produced by deep denoisers to better understand exactly why these methods work so well on natural images.

Summary updated

Abstract

Plug-and-Play (PnP) methods solve ill-posed inverse problems through iterative proximal algorithms by replacing a proximal operator by a denoising operation. When applied with deep neural network d...
BibTeX
@inproceedings{hurault2022proximal,
  author = {Samuel Hurault and Arthur Leclaire and Nicolas Papadakis},
  title = {{Proximal Denoiser for Convergent Plug-and-Play Optimization with Nonconvex Regularization}},
  booktitle = {International Conference on Machine Learning},
  year = {2022},
  url = {https://openreview.net/forum?id=wUMFyjWlS5},
}
Figure