Pith. sign in

REVIEW 2 major objections 4 minor 44 references

PnP-IPA: A Provably Convergent Plug-and-Play Inexact Proximal Algorithm for Nonconvex Imaging Problems

T0 review · 2 major / 4 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A plug-and-play imaging algorithm that converges for any regularization strength and nonconvex data terms.

desk verdict Solid, usable advance on GS-denoiser PnP: removes the λ bound and fixed-step restriction while keeping a clean KL argument; experiments show the practical payoff. read the letter →

arxiv 2607.10223 v1 pith:27FVKK77 submitted 2026-07-11 math.NA cs.NAmath.OC

Cristiano Parenti , Silvia Bonettini , Marco Prato This is my paper
classification math.NAcs.NAmath.OC MSC 65K0590C2668U1094A08
keywords Plug-and-PlayinexactproximalalgorithmGradient-StepdenoiserKurdyka-Łojasiewicznonconveximaginglinesearchimagedeblurring
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Plug-and-play methods replace classical regularizers with deep denoisers and work well in practice, but most versions that come with proofs impose strict limits on the regularization parameter, freeze the step size, or require convex data fidelity. This paper introduces PnP-IPA, an inexact proximal algorithm that removes those limits. It splits the objective so that the proximal step of a scaled conjugate regularizer can be approximated by a short inner loop driven by the denoiser itself, then uses a surrogate merit function to run an Armijo line search without ever evaluating the true objective exactly. Under the Kurdyka-Łojasiewicz property the iterates are proved to converge globally to a stationary point for any positive regularization parameter and for nonconvex data terms. Experiments on deblurring with Gaussian and Cauchy noise show that the freedom to choose the optimal parameter yields higher restoration quality while still guaranteeing convergence.

What carries the argument

The novel splitting F = f_0 + f_1 with f_0 = f_data − (λ/2)∥x∥² and f_1 = λ h_σ^*, together with a surrogate merit function Φ(x,U) = f_0(x) + U that upper-bounds the nonsmooth term and drives the line search.

What would settle it

Run the algorithm on a deblurring problem with a data-fidelity gradient whose Lipschitz constant is deliberately made larger than any bound previously required by competing methods; if the iterates fail to approach a stationary point or the residual does not decrease, the unrestricted-convergence claim is false.

Watch

Extended reading notes

Core claim

PnP-IPA produces a sequence that converges globally to a stationary point of the nonconvex composite objective f_data + λ φ_σ for every λ > 0 and without requiring convexity of f_data, by combining an inexact proximal evaluation of the scaled conjugate regularizer with a surrogate-driven Armijo line search.

Load-bearing premise

The denoiser network must satisfy a global Lipschitz bound strictly less than one on the gradient of its residual, a condition that is only softly enforced during training and therefore holds only approximately for the networks actually used.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes PnP-IPA, a line-search-based inexact forward–backward algorithm for imaging inverse problems that replaces the proximity operator of a Gradient-Step (GS) regularizer φ_σ with an inexact evaluation obtained via Moreau decomposition and an inner gradient method on the strongly convex primal h_σ. A novel surrogate merit Φ(x_k,U_k)=f_0(x_k)+U_k, with U_k a convex upper envelope of f_1, drives an Armijo-like backtracking rule that does not require exact evaluations of the nonsmooth term. Under the Kurdyka–Łojasiewicz property the authors prove (Theorem 2) that the generated sequence converges to a stationary point of the nonconvex objective F=f_data+λφ_σ without any restriction on the regularization parameter λ and without convexity of the data-fidelity term. Numerical experiments on deblurring with Gaussian and Cauchy noise show that the method can exploit unconstrained λ and attains competitive or superior PSNR relative to Prox-PnP, alpha-Prox-PnP and a RED-style gradient method.

Significance. If the analysis holds, the work removes the two most restrictive theoretical bottlenecks of the GS-denoiser literature (bounds on λ and fixed unit step-size) while retaining global KL convergence for nonconvex data terms. The dual splitting (7)–(9), the surrogate merit Φ, and the adaptive inexactness rule ε_k=√(−Δ̃_k)/2 are concrete technical contributions that make the abstract inexact-FB framework of Bonettini et al. applicable to neural-network regularizers. The open-source implementation and the Cauchy-noise experiments (where existing provable PnP methods are severely constrained) supply concrete evidence that the theoretical flexibility translates into better restoration quality. These advances are of clear interest to the mathematical imaging and nonconvex optimization communities.

major comments (2)
  1. The global L_gσ-Lipschitz assumption with L_gσ<1 (Proposition 1) is only softly enforced at training (Remark 2). While the dual objective (7) remains well-defined, the identity D_σ=prox_φ_σ and the whole-space coincidence of (5) and (7) hold only approximately for the trained network actually used. A short quantitative check (e.g., empirical spectral-norm statistics of ∇²g_σ on the test images, or a comparison of the two objective values along the iterates) would strengthen the claim that the theory applies to the practical algorithm.
  2. Theorem 2 requires boundedness of {x_k}. The paper notes that coercivity of F would suffice, yet no argument or numerical evidence is given that the generated sequences remain bounded for the chosen λ and the nonconvex Cauchy fidelity. A brief remark or a plot of ||x_k|| would close this standard but load-bearing gap.
minor comments (4)
  1. In Algorithm 1 the line-search acceptance test mixes Φ and F(ỹ_k); the pseudocode would be clearer if the two cases of (33) were written as separate if-branches with explicit assignment of U_{k+1}.
  2. Table 1–3 report averages over only 10 images (CBSD10). A sentence on variance or a box-plot would help the reader assess statistical significance of the reported PSNR gains.
  3. The step-size schedule α_k (initially 10^6, reduced by 3 every N_α iterations) is motivated only heuristically in §5.5; a short sensitivity plot would make the practical recipe more reproducible.
  4. Typographical: “covex” → “convex” (Proposition 1(iii)); “the largest η” in Algorithm 1 should be “the largest η=δ^m that satisfies …”.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: convergence is obtained by verifying abstract KL conditions on a novel splitting and surrogate, not by reducing the claim to a fit or self-definition.

  1. self citation load bearing [§4.3, Theorem 2 and its proof (application of Theorem 3 from [9])]
    "The proof of Theorem 2 is given by showing that [S1]–[S3] are a special case of the abstract scheme defined in [9, Theorem 5(iii)] … Then, Theorem 3 applies and guarantees that the sequence (xk, ρk)k∈N converges to a stationary point of F."

    The final global-convergence statement is obtained by reducing the algorithm’s sequences to the hypotheses of an abstract theorem previously published by overlapping authors. The reduction itself is non-circular (the paper re-verifies every hypothesis), but the load-bearing general result is self-cited rather than re-proved; this is the only minor self-citation dependency and does not collapse the claim by construction.

full rationale

The central claim (global convergence of PnP-IPA iterates to a stationary point of the nonconvex F without restriction on λ) is derived by constructing sequences that satisfy the abstract hypotheses [S1]–[S3] and [H1]–[H5], then invoking a general KL theorem. The verification steps (inexactness control via (24)/(35), surrogate decrease of Φ via the Armijo rule (32)–(34), squared-norm estimates in Lemmas 2–4, finite termination of line-search in Lemma 5) are self-contained calculations that do not presuppose the conclusion. Self-citations to the authors’ earlier abstract frameworks ([9], [12], etc.) supply reusable lemmas whose assumptions are re-checked for the new dual splitting (7)–(9) and the GS-denoiser regularizer; they do not force the imaging-specific result by definition. Hyper-parameter choices of λ and σ appear only in the experimental section and are ordinary empirical tuning, not inputs renamed as theoretical predictions. The soft Lipschitz caveat of Remark 2 is openly stated and does not create a circular reduction. Hence the derivation chain is independent of its own target.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central claim rests on the GS-denoiser calculus (strong convexity of h_σ, weak convexity of φ_σ, D_σ = prox), the abstract inexact KL framework of [9], standard smoothness/convexity assumptions A1–A4, the KL assumption A5, and boundedness of the iterate sequence. Free parameters are the usual imaging knobs (λ, σ, step-size schedule, line-search constants) tuned for PSNR; they do not enter the proof. The only invented objects are the algorithmic constructs (surrogate Φ, envelope sequence U_k, adaptive ε_k rule).

free parameters (4)
  • regularization parameter λ (or 1/λ in the dual convention) = 3 / 10 / 1/300 depending on regime
    Tuned per noise regime by grid search to maximize PSNR; values 3 (high Gaussian), 10 (low Gaussian), 1/300 (Cauchy). Central practical claim depends on being free to choose these values.
  • denoiser noise level σ = 0.05 / 0.02 / 0.03
    Chosen relative to measurement noise (0.05, 0.02, 0.03) by grid search; affects both quality and the implicit regularizer.
  • forward step-size schedule α_k = α_0=1e6, factor 3, N_α=10 or 25
    Hand-designed: start at 10^6, divide by 3 every N_α iterations until α_k ≤ 10^2. Acts as a practical knob controlling how close the inexact prox is to a pure PnP step.
  • line-search constants δ, ω = δ=0.5, ω=1e-4
    Fixed to δ=0.5, ω=10^{-4}; standard Armijo parameters that affect the proven η_min but are not data-fitted.
assumptions (5)
  • domain assumption ∇g_σ is globally L_gσ-Lipschitz with L_gσ < 1, so D_σ = prox_φ_σ and Im(D_σ)=R^n (Proposition 1, Remark 1–2).
    Inherited from the GS-denoiser literature; only softly enforced at training. Load-bearing for the whole-space objective identity and for strong convexity constants used in the error bounds.
  • standard math F(x,ρ)=F(x)+½ρ² is a KL function (Assumption A5).
    Standard for semi-algebraic / analytic objectives; justified because f1=λ h_σ^* is real-analytic and f0 is smooth. Needed for global (not just subsequential) convergence.
  • domain assumption The iterate sequence {x_k} is bounded (Theorem 2).
    Common nonconvex assumption; paper notes it holds if F is coercive but does not prove coercivity for the GS regularizer.
  • domain assumption f_data is bounded below, C¹, with L_fdata-Lipschitz gradient (Assumption 1 / A1–A4).
    Standard for FB methods; holds for least-squares and for the Cauchy log-likelihood on a compact domain after clipping.
  • standard math Abstract inexact KL convergence framework of Bonettini et al. [9, Thm 5(iii)] applies once [H1]–[H5] are verified.
    External theorem used as a black box; the paper’s contribution is the verification for the PnP surrogate sequences.
invented entities (2)
  • Surrogate merit function Φ(x_k, U_k) = f_0(x_k) + U_k with convex upper envelope U_k
    purpose: Drives Armijo line search when f1 cannot be evaluated at intermediate points x_k + η d_k.
    Algorithmic construct introduced in §3.3; no independent physical meaning, but fully defined and used only inside the method.
  • Adaptive inexactness rule ε_k = √(−Δ̃_k)/2 (realized via inner gradient descent on G_k)
    purpose: Guarantees that the approximate prox yields a descent direction and that the line-search acceptance condition is eventually satisfied.
    New relative to fixed-tolerance inexact prox schemes; defined purely for the algorithm.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PnP-IPA: A Provably Convergent Plug-and-Play Inexact Proximal Algorithm for Nonconvex Imaging Problems." pith.science (2026). https://pith.science/paper/27FVKK77

@misc{pith2026260710223,
  author       = {Pith},
  title        = {Pith review of: PnP-IPA: A Provably Convergent Plug-and-Play Inexact Proximal Algorithm for Nonconvex Imaging Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27FVKK77}},
  note         = {Machine review of arXiv:2607.10223}
}
read the original abstract

Plug-and-Play (PnP) methods have emerged as a highly effective paradigm for solving imaging inverse problems by replacing traditional proximity operators of regularization terms with highly expressive deep denoisers. While empirically successful, establishing rigorous convergence guarantees for PnP algorithms remains a major challenge. Existing provable approaches based on the Gradient-Step (GS) denoiser suffer from theoretical and practical limitations, such as restrictive bounds on the regularization parameter, rigid step-size rules, and the inability to handle nonconvex data-fidelity terms. In this paper, we introduce PnP-IPA (Plug-and-Play Inexact Proximal Algorithm), a novel optimization scheme that overcomes these bottlenecks. We propose a new splitting strategy that evaluates the proximal operator of the scaled implicit regularizer inexactly. To enable adaptive step-size selection without exact objective evaluations, we design a novel surrogate merit function that successfully drives an Armijo-like backtracking line-search. Relying on the Kurdyka-Lojasiewicz property, we establish global convergence to a stationary point of the nonconvex objective without imposing any assumption on the regularization parameter. Extensive numerical experiments on image deblurring under both Gaussian and Cauchy noise demonstrate the practical advantages of PnP-IPA. By effectively lifting previous theoretical constraints, our method allows for optimal parameter tuning, yielding state-of-the-art restoration quality and robust convergence even in nonconvex regimes.

Figures

Figures reproduced from arXiv: 2607.10223 by the authors.

Figure 1
Figure 1. Reconstruction results and convergence comparison under aggressive Gaussian noise [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Reconstruction results and convergence comparison under low Gaussian noise ( [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. Reconstruction results and convergence comparison under Cauchy noise ( [PITH_FULL_IMAGE:figures/full_fig_p030_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 2 linked inside Pith

  1. [9]

    Bonettini, P

    S. Bonettini, P. Ochs, M. Prato, and S. Rebegoldi. An abstract convergence framework with application to inertial inexact forward–backward methods.Comput. Optim. Appl., 84(2):319–362, 2023

    2023

  2. [1]

    Arridge, P

    S. Arridge, P. Maass, O. Öktem, and C.-B. Schönlieb. Solving inverse problems using data- driven models.Acta Numer., 28:1–174, 2019

    2019

  3. [2]

    Attouch, J

    H. Attouch, J. Bolte, and B. F. Svaiter. Convergence of descent methods for semi-algebraic and tame problems: Proximal algorithms, forward–backward splitting, and regularized Gauss–Seidel methods.Math. Program., 137(1–2):91–129, 2013. 26 Parenti C., Bonettini S. and Prato M. A PnP Inexact Proximal Algorithm for Nonconvex Imaging Table 4: Hyperparameters o...

    2013

  4. [3]

    H. H. Bauschke and P. L. Combettes.Convex Analysis and Monotone Operator Theory in Hilbert Spaces. Springer, New York, 2011

    2011

  5. [4]

    Bertero, P

    M. Bertero, P. Boccacci, and C. De Mol.Introduction to Inverse Problems in Imaging. CRC press, Boca Raton, 2nd edition, 2022

    2022

  6. [5]

    Bertocchi, E

    C. Bertocchi, E. Chouzenoux, M.-C. Corbineau, J.-C. Pesquet, and M. Prato. Deep unfold- ing of a proximal interior point method for image restoration.Inverse Probl., 36(3):034005, 2020

    2020

  7. [6]

    M. I. H. Bhuiyan, M. O. Ahmad, and M. N. S. Swamy. Spatially adaptive wavelet-based method using the Cauchy prior for denoising the SAR images.IEEE Trans. Circuits Syst. Video Technol., 17(4):500–507, 2007

    2007

  8. [7]

    Bolte, S

    J. Bolte, S. Sabach, and M. Teboulle. Proximal alternating linearized minimization for nonconvex and nonsmooth problems.Math. Program., 146(1–2):459–494, 2014

    2014

Show all 44 references
  1. [8]

    Bonettini, I

    S. Bonettini, I. Loris, F. Porta, and M. Prato. Variable metric inexact line–search-based methods for nonsmooth optimization.SIAM J. Optim., 26(2):891–921, 2016

    2016

  2. [10]

    Bonettini, M

    S. Bonettini, M. Prato, and S. Rebegoldi. Convergence of inexact forward–backward algo- rithms using the forward–backward envelope.SIAM J. Optim., 30(4):3069–3097, 2020

    2020

  3. [11]

    Bonettini, M

    S. Bonettini, M. Prato, and S. Rebegoldi. New convergence results for the inexact variable metric forward–backward method.Appl. Math. Comput., 392:125719, 2021

    2021

  4. [12]

    Bonettini, M

    S. Bonettini, M. Prato, and S. Rebegoldi. A new proximal heavy ball inexact line–search algorithm.Comput. Optim. Appl., 88(2):525–565, 2024

    2024

  5. [13]

    T. A. Bubba.Data–driven Models in Inverse Problems. De Gruyter, Berlin, 2025

    2025

  6. [14]

    Cohen, Y

    R. Cohen, Y. Blau, D. Freedman, and E. Rivlin. It has potential: Gradient-driven denoisers for convergent solutions to inverse problems. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, vo...

    2021

  7. [15]

    P. L. Combettes and J.-C. Pesquet. Proximal splitting methods in signal processing. In H. H. Bauschke, R. S. Burachik, P. L. Combettes, V. Elser, D. R. Luke, and H. Wolkowicz, editors,Fixed-point algorithms for inverse problems in science and engineering, Springer Optimization...

    2011

  8. [16]

    P. L. Combettes and V. R. Wajs. Signal recovery by proximal forward–backward splitting. Multiscale Model. Simul., 4(4):1168–1200, 2005

    2005

  9. [17]

    File:dolomitas cinque torri @vilanchelo.jpg — wikimedia commons, the free media repository, 2026

    Wikimedia Commons. File:dolomitas cinque torri @vilanchelo.jpg — wikimedia commons, the free media repository, 2026. [Online; accessed 26-May-2026]

    2026

  10. [18]

    H. W. Engl, M. Hanke, and A. Neubauer.Regularization of Inverse Problems. Kluwer, Dordrecht, 1996

    1996

  11. [19]

    Frankel, G

    P. Frankel, G. Garrigos, and J. Peypouquet. Splitting methods with variable metric for Kurdyka–Łojasiewicz functions and general convergence rates.J. Opt. Theory Appl., 165:874–900, 2015

    2015

  12. [20]

    Gregor and Y

    K. Gregor and Y. LeCun. Learning fast approximations of sparse coding. InProceedings of the 27th International Conference on Machine Learning, pages 399–406, 2010

    2010

  13. [21]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman.The Elements of Statistical Learning. Springer, New York, 2nd edition, 2009

    2009

  14. [22]

    Hurault, A

    S. Hurault, A. Chambolle, A. Leclaire, and N. Papadakis. Convergent plug-and-play with proximal denoiser and unconstrained regularization parameter.J. Math. Imaging Vis., 66(4):616–638, 2024

    2024

  15. [23]

    Hurault, A

    S. Hurault, A. Leclaire, and N. Papadakis. Gradient step denoiser for convergent plug-and- play. InInternational Conference on Learning Representations, 2022

    2022

  16. [24]

    Hurault, A

    S. Hurault, A. Leclaire, and N. Papadakis. Proximal denoiser for convergent plug-and- play optimization with nonconvex regularization. InInternational Conference on Machine Learning, pages 9483–9505. PMLR, 2022

    2022

  17. [25]

    Idan and J

    M. Idan and J. L. Speyer. Cauchy estimation for linear scalar systems.IEEE Trans. Autom. Control, 55(6):1329–1342, 2010

    2010

  18. [26]

    U. S. Kamilov, C. A. Bouman, G. T. Buzzard, and B. Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications.IEEE Signal Process. Mag., 40(1):85–97, 2023

    2023

  19. [27]

    Levin, Y

    A. Levin, Y. Weiss, F. Durand, and W. T. Freeman. Understanding and evaluating blind deconvolutionalgorithms. In2009 IEEE Conference on Computer Vision and Pattern Recog- nition, pages 1964–1971, 2009

    1964

  20. [28]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InProceedings of the 8th IEEE International Conference on Computer Vision. ICCV 2001, volume...

    2001

  21. [29]

    Martin, A

    S. Martin, A. Gagneux, P. Hagemann, and G. Steidl. PnP-Flow: Plug-and-play image restoration with flow matching. InInternational Conference on Learning Representations, volume 2025, pages 45466–45492, 2025. 28 Parenti C., Bonettini S. and Prato M. A PnP Inexact Proximal Algori...

    2025

  22. [30]

    Moreau, M

    T. Moreau, M. Massias, A. Gramfort, et al. Benchopt: Reproducible, efficient and col- laborative optimization benchmarks. InProceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, pages 25404–25421, Red Hook, 2022. Curran Associates Inc

    2022

  23. [31]

    Learningmaximallymonotoneoperators for image recovery.SIAM J

    J.-C.Pesquet, A.Repetti, M.Terris, andY.Wiaux. Learningmaximallymonotoneoperators for image recovery.SIAM J. Imaging Sci., 14(3):1206–1237, 2021

    2021

  24. [32]

    E. T. Reehorst and P. Schniter. Regularization by denoising: Clarifications and new inter- pretations.IEEE Trans. Comput. imaging, 5(1):52–67, 2018

    2018

  25. [33]

    Renaud, J

    M. Renaud, J. Prost, A. Leclaire, and N. Papadakis. Plug-and-play image restoration with stochastic denoising regularization. InProceedings of the 41st International Conference on Machine Learning, ICML’24, pages 42484–42520, 2024

    2024

  26. [34]

    R. T. Rockafellar, R. J.-B. Wets, and M. Wets.Variational Analysis, volume 317 of Grundlehren der Mathematischen Wissenschaften. Springer, Berlin, 1998

    1998

  27. [35]

    Romano, M

    Y. Romano, M. Elad, and P. Milanfar. The little engine that could: Regularization by denoising (RED).SIAM J. Imaging Sci., 10(4):1804–1844, 2017

    2017

  28. [36]

    L. I. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based noise removal algo- rithms.J. Phys. D., 60(1–4):259–268, 1992

    1992

  29. [37]

    E. K. Ryu, J. Liu, S. Wang, X. Chen, Z. Wang, and W. Yin. Plug-and-play methods provably converge with properly trained denoisers. InInternational Conference on Machine Learning, pages 5546–5557. PMLR, 2019

    2019

  30. [38]

    Schmidt, N

    M. Schmidt, N. Roux, and F. Bach. Convergence rates of inexact proximal-gradient meth- ods for convex optimization. In J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger, editors,Advances in Neural Information Processing Systems, volume 24. Curran Associates...

    2011

  31. [39]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score- based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  32. [40]

    Tachella, M

    J. Tachella, M. Terris, S. Hurault, et al. Deepinverse: A Python package for solving imaging inverse problems with deep learning.arXiv preprint arXiv:2505.20160, 2025

    2025 arXiv

  33. [41]

    G. A. Tsihrintzis and C. L. Nikias. Evaluation of fractional, lower-order statistics-based detection algorithms on real radar sea-clutter data.IEE Proceedings - Radar, Sonar and Navigation, 144(1):29–37, 1997

    1997

  34. [42]

    S. V. Venkatakrishnan, C. A. Bouman, and B. Wohlberg. Plug-and-play priors for model based reconstruction. In2013 IEEE Global Conference on Signal and Information Process- ing, pages 945–948. IEEE, 2013

    2013

  35. [43]

    Villa, S

    S. Villa, S. Salzo, L. Baldassarre, and A. Verri. Accelerated and inexact forward–backward algorithms.SIAM J. Optim., 23(3):1607–1633, 2013

    2013

  36. [44]

    Zhang, Y

    K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte. Plug-and-play image restoration with deep denoiser prior.IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6360– 6376, 2021. 29 Parenti C., Bonettini S. and Prato M. A PnP Inexact Proximal Algorithm for Nonconvex Im...

    2021

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.