Plug-and-Play Denoisers in Image Restoration

Updated 22 April 2026

Plug-and-play denoisers are advanced frameworks that replace conventional proximal operators with modular denoising methods in image restoration.
They leverage various techniques including deep CNNs, GMM-based models, and pixel-adaptive networks to address noise in diverse imaging applications.
Their design integrates into algorithms like ADMM and FBS, supported by theoretical guarantees based on nonexpansive and contractive conditions.

Plug-and-play (PnP) denoisers are a class of algorithms and frameworks for imaging inverse problems that leverage state-of-the-art denoisers—frequently based on deep neural networks, nonlocal patch models, or adaptive statistical filters—by integrating them as modular priors within iterative optimization schemes. Rather than specifying explicit regularization terms with tractable closed-form proximal operators, PnP frameworks substitute the proximal mapping step with a black-box denoising operation. This strategy enables domain-adaptive, learned, or scene-adapted denoisers to be seamlessly incorporated into classical algorithms such as ADMM, HQS, forward–backward splitting, and their many variants. The mathematical, algorithmic, and theoretical foundations of PnP have been established through a variety of fixed-point, operator-theoretic, and variational analyses, enabling guarantees under suitable conditions. PnP denoisers are now a widely used paradigm in generic image restoration (deblurring, inpainting, super-resolution), computed imaging (MRI, CT, PET), microscopy, seismic inversion, and beyond.

1. Mathematical Framework and MAP Connections

The canonical inverse problem formulation is to recover an unknown image $x \in \mathbb{R}^n$ from measurements $y = Kx + \eta$ , with $K$ a linear operator and $\eta \sim \mathcal{N}(0, \sigma^2 I)$ . A standard Maximum a Posteriori (MAP) estimator seeks

$\hat{x} = \arg\min_x \frac{1}{2\sigma^2}\|Kx-y\|_2^2 + \phi(x),$

where $\phi(x) = -\log p(x)$ encodes the prior. Classic proximal algorithms such as ADMM and HQS decouple the data and prior terms by splitting and iteratively solving quadratic subproblems and proximal-mapping steps: $\begin{aligned} x^{t+1} &\leftarrow \arg\min_x \frac{1}{2\sigma^2}\|Kx-y\|_2^2 + \frac{\rho}{2}\|x-z^t+\lambda^t\|_2^2, \ z^{t+1} &\leftarrow \mathrm{prox}_{\phi/\rho}(x^{t+1}+\lambda^t), \ \lambda^{t+1} &\leftarrow \lambda^t + (x^{t+1} - z^{t+1}). \end{aligned}$ Plug-and-play replaces $\mathrm{prox}_{\phi/\rho}$ with a generic denoiser $D^*$ trained to map noisy images toward clean ones, usually at a fixed (or variable) noise level. When the denoiser is the exact proximal operator of some (possibly learned) regularizer, classical convergence guarantees apply. More generally, the PnP formulation yields a flexible, modular framework for incorporating learned priors (Bigdeli et al., 2019).

2. Denoiser Construction, Training, and Adaptation

A wide range of denoiser constructions are compatible with PnP:

Deep CNN Denoisers: DnCNN, DRUNet, and their variants are frequently adopted as the backbone for plug-and-play methods. These networks are trained on large image corpora with additive Gaussian noise at a prescribed standard deviation to minimize empirical risk measures such as mean-square error (MMSE). In advanced architectures, the denoiser is trained via a two-stage process to directly minimize the loss associated with the MAP denoising objective, ensuring that the output approximates the true proximal mapping (Bigdeli et al., 2019).
GMM-based Priors: Scene-adapted denoisers constructed using Gaussian mixture models (GMMs), with mixture weights learned from the target or a related image and fixed during inference, can be cast as linear proximal operators of convex quadratic functionals (Teodoro et al., 2017).
Pixel-Wise Adaptive Networks: For applications with spatially inhomogeneous noise, like demosaicing or missing data, PnP frameworks have been extended to leverage denoisers trained to accept full maps of per-pixel noise standard deviations, enabled by a preconditioning modification in the ADMM scheme (Pendu et al., 2021).

Denoisers may also be constructed in transformed domains (e.g., wavelet or gradient), or even on a per-instance basis ("single-shot") (Cheng et al., 2023), with implicit neural representations fit to a single corrupted observation.

3. Algorithmic Variants and Restoration Workflows

Key PnP iterative schemes include:

PnP-ADMM (Alternating Direction Method of Multipliers): Alternates model inversion steps with denoising. The denoiser operates as a modular proximal mapping.
PnP-FBS (Forward–Backward Splitting): Alternates gradient-descent on the data-fidelity term with denoiser steps. If the denoiser is MMSE or MAP-trained and contractive, convergence to stationary points is ensured (Hauptmann et al., 2023).
Gradient-step and Proximal Denoiser Algorithms: Rather than using generic denoisers, these methods train networks to be the exact gradient or proximal step of an explicit or weakly convex functional, yielding transparent convergence properties (Herfeld et al., 11 Sep 2025).
Analysis (Transform-Domain) PnP: Applies denoisers in domains such as the image gradient, extending classical TV regularization to learned functionals within half-quadratic splitting or ADMM frameworks (Chandler et al., 18 Sep 2025).
Expectation Consistent (EC) PnP and AMP-type Algorithms: For forward models with structured operators (e.g., MRI), expectation-consistent PnP tracks variances block-wise and presents a denoiser with nearly white, stationary noise, enabling optimal training and state-evolution tracking (Shastri et al., 2022).
Single-Shot PnP: Fits an implicit neural prior directly to one instance, integrating it into ADMM for data-limited scenarios without reliance on large training datasets (Cheng et al., 2023).

Major acceleration strategies include preconditioning (to decouple per-pixel variances), quasi-Newton updates for faster convergence, and policy-learning for tuning-free parameter selection using reinforcement learning (Wei et al., 2020).

4. Theoretical Guarantees and Convergence Analysis

The convergence of PnP methods depends critically on the regularity properties of the denoiser and the splitting scheme:

Exact Proximal Operators: If the denoiser is the exact proximal mapping of a convex function, convergence to the unique minimizer or fixed point is guaranteed under mild conditions; this holds for symmetrized nonlocal means, fixed-weight GMMs, and some deep networks explicitly trained for this property (Teodoro et al., 2017, Herfeld et al., 11 Sep 2025).
Lipschitz and Nonexpansive Denoisers: When the denoiser is close to the identity (i.e., the difference is strictly contractive or nonexpansive in the appropriate norm), fixed-point convergence of the iterates is obtained for both ADMM and FBS variants, with global rates determined by the contraction factor (Ryu et al., 2019, Hauptmann et al., 2023, Gavaskar et al., 2020, Sinha et al., 21 May 2025).
Gradient and Proximal Denoiser Training: Training denoisers to be explicit gradient steps (Id – ∇g_σ) or proximal operators (with L_g < 1) ensures that the resulting PnP scheme can be interpreted as energy descent on an explicit global functional $F(x) = f(x) + g_\sigma(x)$ , and that all limit points are critical points (Hurault et al., 2021, Herfeld et al., 11 Sep 2025).
Regularization and Limiting Behavior: In the limit of vanishing measurement noise, linear PnP schemes with spectral filtering of the denoiser correspond to convergent regularization methods in the classical inverse problem sense (Hauptmann et al., 2023).
Quasi-Newton and Viscosity Stabilization: Advanced iterative schemes can combine curvature information (L-BFGS) for faster energy descent (Tan et al., 2023) or stabilize unstable fixed points via viscosity control by averaging the PnP operator with a contractive anchor mapping to suppress oscillations and divergence (Sinha et al., 2 Aug 2025).

Assumptions underlying these guarantees include contractivity, spectral properties of the kernel denoiser, or weak convexity of the associated (possibly learned) potential.

5. Empirical Performance and Applications

PnP methods have been extensively validated across a range of generic and application-specific imaging tasks:

Image Deblurring and Inpainting: CNN MAP PnP achieves PSNR within 0.1–0.3 dB of other state-of-the-art methods but is 70× faster, with only one denoiser network and a single noise-level (Bigdeli et al., 2019).
MRI and Tomography: Expectation consistent PnP with blockwise, noise-adaptive denoisers achieves state-of-the-art PSNR/SSIM in single and multi-coil MRI, outperforming both classical and other PnP schemes (Shastri et al., 2022).
Microscopy: PnP with variance-trained DRUNet models efficiently reconstructs super-resolved or fluctuation-based images, outperforming hand-crafted sparsity regularization by wide margins in Jaccard index and PSNR (Stergiopoulou et al., 2023).
Seismic Inversion: PnP with a pretrained 2D DRUNet as the regularizing denoiser, extended via multi-agent consensus (MACE) to 3D, yields higher-resolution impedance models than classical TV regularization in both synthetic and real seismic volumes (Luiken et al., 2024).
Particle Beam Micrograph Denoising: PnP with total-variation, BM3D, or DnCNN denoisers, and physically accurate non-Gaussian data-fidelity terms, reduces RMSE by a factor of 2–4 over conventional approaches in SEM and HIM simulation studies (Peng et al., 2022).
Tuning-Free and Adaptive Algorithms: Policy learning via reinforcement approaches achieves state-of-the-art reconstruction in challenging scenarios such as compressed-sensing MRI and phase retrieval while dramatically reducing the number of iterations needed (Wei et al., 2020).

Performance in denoising, deblurring, super-resolution, and inpainting consistently matches or surpasses classical methods and other deep restoration pipelines, with added interpretability and modularity.

6. Extensions, Generalizations, and Open Directions

Recent work extends PnP denoisers to diverse modalities and operational regimes:

Gradient/Analysis-Domain PnP: Denoisers trained on the image-gradient domain serve as implicit learned TV priors, and PnP-ADMM or HQS with these operators achieve performance comparable to standard image-domain PnP at similar complexity (Chandler et al., 18 Sep 2025).
Single-Shot and Instance-Specific Denoisers: PnP-ADMM with an implicit neural representation fitted to a single corrupted instance demonstrates strong gains in data-limited regimes, outperforming both TV and generic pre-trained CNN denoisers (Cheng et al., 2023).
Viscosity and Stability: Data-driven stabilization ensures robust peak performance for PnP even with highly nonexpansive (and hence unstable) deep denoisers by adaptively mixing with contractive operators, precluding divergence or oscillatory failure (Sinha et al., 2 Aug 2025).
Linear and Kernel Denoiser Analysis: Theoretical frameworks now encompass affine and kernel-denoiser variants, providing explicit bounds on contraction and convergence rates for symmetric and nonsymmetric denoisers over a range of inverse problems, and clarifying the trade-offs between denoising strength and convergence speed (Gavaskar et al., 2020, Sinha et al., 21 May 2025).

Active directions include: extending PnP to non-Gaussian and compound-noise likelihoods, automated hyperparameter tuning or bilevel learning of denoising schedules, joint learning of transform and denoiser, application to video and 3D/tensor data, and developing a unified theory for nonconvex/instability-prone denoisers and black-box plug-ins. Formal theoretical guarantees for all PnP settings with generic black-box denoisers remain an open and continually evolving topic.

References (by arXiv id):