Plug and Play Splitting Techniques for Poisson Image Restoration
Alessandro Benfenati · 2025
Read Plug and Play Splitting Techniques for Poisson Image Restoration by Alessandro Benfenati (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-19378.txt.
Introduction
Alessandro Benfenati, a solo author, extends a popular image-restoration strategy to a harder noise model in this 2025 paper. Plug and play methods restore images by alternating a deblurring step with an off-the-shelf denoiser, and they shine when the noise is Gaussian. Poisson noise, common in photon-limited imaging, breaks the theory: the fit term has no closed-form step and its slope is not tame.
Benfenati's method, PnPSplit+, rebuilds the scheme around the alternating direction method of multipliers. That change yields a closed-form deblurring step, so no inner iterative solver is needed. Convergence is guaranteed as long as the denoiser used is firmly non-expansive, a mild stability property. On a range of Poisson restoration problems, the method performed strongly even with heavy noise and severe blur. The author presents it as a practical route to plug and play restoration beyond the Gaussian world.
Catalog record
| CREATOR | Alessandro Benfenati |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-19378
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Plug and Play Splitting Techniques for Poisson Image Restoration Alessandro Benfenati Abstract Plug and Play (PnP) methods achieve remarkable results in the framework of image restoration problems for Gaussian data. Nonetheless, the theory available for the Gaussian case cannot be extended to the Poisson case, due to the non-Lipschitz gradient of the fidelity function, the Kullback-Leibler functional, or the absence of closed-form solution for the proximal operator of such term, leading to employ iterative solvers for the inner subproblem. In this work we extend the idea of PIDSplit+ algorithm, exploiting the Alternating Direction Method of Multipliers, to PnP scheme: this allows to provide a closed form solution for the deblurring step, with no need for iterative solvers. The convergence of the method is assured by employing a firmly non expansive denoiser. The proposed method, namely PnPSplit+, is tested on different Poisson image restoration problems, showing remarkable performance even in presence of high noise level and severe blurring conditions. 1 Introduction Imaging problems arise in several scientific applications, such as Medicine [ 41 , 33 ] , Astronomy [ 2 , 24 , 4 ] , Microscopy [ 13 , 63 , 15 ] . The mathematical model underlying the physics process is shared among all these disciplines [ 7 ] , and it reads as {\mathbf{g}}=\mathcal{N}\left({\mathbf{H}}\,{\mathbf{x}}^{\star}+b\right), (1) where {\mathbf{x}}^{\star}\in\mathbb{R}^{n} the ground truth image, {\mathbf{H}}\in\mathbb{R}^{m\times n} is a linear operator perturbing the data, b\in\mathbb{R}^{+} is a known background parameter, {\mathbf{g}}\in\mathbb{R}^{m} the recorded image and \mathcal{N} denotes the statistical noise on recorded data. The operator {\mathbf{H}} is also called Point Spread Function (PSF), since its representation is the registered image of a point source; classical hypotheses, abiding to real life systems properties, are that {\mathbf{H}}^{\top}\bm{1}=\bm{1} and \sum_{ij}H_{ij}=1 . The aim of image restoration problems is to recover an estimation of {\mathbf{x}}^{\star} given the registered data {\mathbf{g}} and the operator {\mathbf{H}} . When the recorded data {\mathbf{g}} is affected by Poisson noise, under a Bayesian framework and adopting a maximum a posteriori approach [ 8 , 55 , 31 ] , one is led to solve the optimization problem depicted below: \underset{\begin{subarray}{c}{{\mathbf{x}}\geq\bm{0}}\end{subarray}}{\amin}\,\,KL({\mathbf{H}}\,{\mathbf{x}}+b,{\mathbf{g}})+\beta\,R({\mathbf{x}}), (2) where KL is the generalized Kullback-Leibler functional KL({\mathbf{H}}\,{\mathbf{x}}+b,{\mathbf{g}})={\mathbf{g}}\log\left(\frac{{\mathbf{g}}}{{\mathbf{H}}{\mathbf{x}}+b}\right)+{\mathbf{H}}{\mathbf{x}}+b-{\mathbf{g}}. The operations are intended component wise and one assumes 0\log(0)=0 . In particular, KL(\cdot,{\mathbf{g}}) is a proper, convex and differentiable functional. The function R is the regularization functional and its role is to preserve the desired characteristics on the estimated solution, such as sharp edges or sparseness, and to control the influence of the noise on the estimated solution. Common choice for R consists of proper, lower semi-continuous (l.s.c.) convex function, such as \ell_{2} regularization, which goes also under the name Tikhonov regularization [ 54 , 32 ] or Ridge Regression in other frameworks [ 1 ] , \ell_{1} norm for promoting sparsity on the solution [ 30 ] , a convex combination of \ell_{2} and \ell_{1} norms, commonly referenced to as Elastic Net [ 3 ] . Another popular choice is the Total Variation functional [ 17 ] , for promoting sharp edges, and its offsprings [ 42 , 39 ] . The parameter \beta\in\mathbb{R}^{+} is the regularization parameter and balances the trade–off between the KL and R . A common requirement in imaging problems is that the solution’s components are non-negative, since they represents pixels’ intensity: therefore, the estimated solution is required to belong to the non negative orthant. The literature presents a plethora of variational methods to solve this particular instance of restoration problem: among them one can find gradient approaches [ 18 ] and the related variation [ 23 , 61 , 50 ] , Bregman iterative methods [ 40 , 6 , 48 ] , proximal approaches [ 20 ] . The Alternating Direction Method of Multipliers (ADMM) has gained a predominant role in image restoration problem [ 29 , 11 , 26 ] , showing particularly interesting results in managing optimization problems with linear constraints. The seminal work [ 56 ] introduced a novel approach, called Plug and Play (PnP) technique. This strategy consists of solving optimization problems, whose objective functional encompasses two terms. Employing splitting techniques as ADMM, the authors in [ 56 ] observed that the update for one of the primal variables reads actually as a Gaussian denoising step: therefore, they propose to substitute such updating step with an off-the–shelf denoiser D , such as Block-Matching and 3D Filtering (BM3D) [ 25 ] , Nonlocal Mean Filter (NLM) [ 12 ] . Modern approaches encompass also the usage of deep neural networks, tailored for Gaussian denoising [ 62 ] . The main hypothesis is that such denoiser is the proximal operator of some function R : the numerical experience showed the remarkable results of this approach. The research interest then moved to search for the theoretical hypothesis to have on the denoiser for assuring the convergence of PnP: indeed, fixed point theory tells us that such denoiser needs to be firmly non expansive [ 49 , 53 ] , but unfortunately most of the employed denoisers do not fulfill this requirement [ 22 ] , despite their impressive performance results. Even classical neural networks, that show remarkable performances in Gaussian denoising tasks, cannot satisfy this requirement, unless properly trained with tailored loss function [ 45 ] . The scientific research explored the control of the Lipschitz constant of the neural network [ 27 , 35 , 59 ] , but the quality of such control is not strong enough to ensure the convergence property and moreover the computational cost is rather high. In [ 5 ] the PnP framework has been addressed by considering it as a constrained problem under an ADMM approach, where a discrepancy principle is used in reformulating the problem. This approach allows to automatically chose the regularization parameter. Different techniques have been explored to assure convergence of PnP method: bounded denoisers assure fixed point convergence [ 19 ] ; in [ 52 ] an incremental version of PnP with explicit requirements on the denoiser, namely its firmly non expansiveness, assures the convergence while maintaining scalability in terms of speed and memory. In [ 43 ] under the hypotheses of the denoiser being averaged and the convexity of the data fidelity term the PnP scheme converges, and moreover it is shown that some of the employed denoisers are indeed the proximal operator of particular functions, e.g. , the NLM is the prox of a quadratic convex function. One has to mention alternative approaches to PnP, which try to address the theoretical issues posed by PnP. The Regularizaton by Denoising (RED) method [ 46 ] is among them, it tries to overcome the PnP limitations by requiring the denoiser to have a symmetric Jacobian and to be locally homogeneous: unfortunately, although the theoretical framework is very rich and interesting, the majority of the employed denoiser do not satisfy this requirements. RED has been then investigated from different points of view: it has been reformulated [ 21 ] as a constrained optimization problem (RED-PRO), where the least square minimum is projected on the fixed-point sets of demicontractive denoisers, which reveal to be convex sets. In [ 14 ] the RED-PRO has been reversed following a discrepancy principle, leading to a constrained RED