Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Plug-and-Play Half-Quadratic Splitting for Ptychography

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Plug-and-play denoisers beat classical ptychography at low overlap.

desk verdict A clean PnP-HQS extension for ptychography with a correct analytic data-fidelity step, solid simulated evidence, but an overstated abstract and an unvalidated complex-denoiser heuristic. read the letter →

arxiv 2412.02548 v1 pith:JBIFIFXP submitted 2024-12-03 eess.IV cs.LG

Alexander Denker , Johannes Hertrich , Zeljko Kereta , Silvia Cipiccia , Ecem Erin , Simon Arridge This is my paper · ORCID
classification eess.IVcs.LG
keywords ptychographyphaseretrievalhalf-quadraticsplittingplug-and-playimagedenoisingcomplex-valuedimaginglowoverlapregularisation
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

This paper extends half-quadratic splitting, a standard variational-reconstruction scheme, to ptychography, and shows that replacing the regularisation step with a plug-and-play denoiser yields reconstructions that beat classical ptychographic algorithms. The authors derive a closed-form solution for the data-consistency subproblem, so each iteration is cheap, and they handle complex-valued images by denoising the real and imaginary parts separately with off-the-shelf greyscale denoisers. In simulated experiments on natural images and a brain phantom, the approach keeps quality high where classical methods degrade: at only 38% probe overlap and under strong shot noise. The practical payoff is that scanning time could be cut while keeping reconstruction quality.

What carries the argument

The load-bearing object is the half-quadratic splitting formulation with multiple auxiliary variables $z_\ell = A_\ell x$, one per probe position. It reduces each data-consistency update to the same closed-form Fourier phase retrieval step: in Fourier space, set the amplitude of $\hat{z}$ to a convex combination of the measured amplitude and the current amplitude, and keep the current phase. The regularisation update is a Gaussian denoising problem with a spatially varying noise level controlled by the accumulated probe intensity $D^2 = \sum_\ell |A_\ell|^2$; under a regularity assumption this reduces to a standard denoiser applied to the probe-weighted average of the auxiliary images. Complex values are handled by splitting into real and imaginary parts, shifting each by a constant equal to the maximum amplitude so that pretrained non-negative greyscale denoisers apply.

What would settle it

Run the algorithm on a real ptychography dataset with an independently known ground truth, such as a test pattern reconstructed at very high overlap with ePIE, and compare amplitude and phase PSNR against SimPIE at 38% overlap. If the reported performance gap disappears or visible phase artifacts appear, the complex-denoiser heuristic would be the likely culprit.

Watch

Extended reading notes

Core claim

The central claim is that plug-and-play half-quadratic splitting (PnP-HQS) provides a stable, flexible reconstruction method for ptychography that outperforms unregularised PIE variants, particularly at low overlap and high noise. In the paper's formulation, the ptychographic forward model is split into one auxiliary variable per probe position; the resulting z-subproblems are identical to Fourier phase retrieval and admit an explicit closed-form update using the measured amplitudes and the phase of the current iterate. The x-subproblem becomes a spatially varying Gaussian denoising task, which can be solved either classically (TV, weakly convex ridge regulariser) or by a pretrained deep denoiser applied independently to real and imaginary channels. Tables 1 and 2 report consistent PSNR gains over Simultaneous PIE and Sequential PIE, with the gap widening as overlap drops from 68% to 38% and as noise level alpha rises from 10 to 40. The paper concludes that this is a first step toward shorter acquisition times.

Load-bearing premise

The method assumes that denoising the real and imaginary parts independently, after shifting them to be non-negative, is a good stand-in for a true complex-valued denoiser, and the paper gives no experimental data validating this on real samples.

Editorial extensions

If this is right

  • At 38% overlap, PnP-HQS with DRUNet or WCRR reconstructs amplitude and phase with several dB higher PSNR than SimPIE, so the required scan overlap could be relaxed.
  • The closed-form z-update keeps per-iteration cost low enough that 600 iterations suffice, versus 2000 iterations for the PIE baselines.
  • The method inherits the behaviour of any chosen denoiser, so better datasets or denoisers directly improve reconstruction without changing the algorithm.
  • Because the regularisation step is generic, the same framework covers TV, learned convex regularisers, and deep denoisers within one code path.
  • Since the probe is assumed known, the authors point to simultaneous probe and object reconstruction as the next step toward realistic ptychography.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The real/imaginary decomposition with a constant shift is not validated on experimentally measured complex samples; testing a true complex-valued proximal operator instead would isolate whether this heuristic limits performance.
  • The same HQS derivation should transfer to other non-linear Fourier measurements, such as Fourier ptychography with coded illuminations, because the data-consistency step only relies on the magnitude-Fourier structure.
  • Low-overlap robustness suggests PnP-HQS could act as an initialisation or regulariser inside joint probe-object estimation, potentially reducing the risk of local minima in that harder problem.
  • All experiments are simulated with a known binary circular probe; testing on real beam profiles, where probe uncertainty and position errors enter, is the natural next comparison.
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

3 major / 5 minor

Summary. This paper proposes a plug-and-play half-quadratic splitting (PnP-HQS) algorithm for ptychographic phase retrieval. The authors derive an explicit solution for the data-fidelity subproblem in Fourier phase retrieval, extend it to ptychography via multiple auxiliary variables, and use the resulting update with classical (TV, WCRR) and neural (DRUNet) denoisers. For complex-valued images, they construct the denoiser step by applying real-valued denoisers separately to the real and imaginary parts after a constant shift. Experiments on synthetic complex images from BSD500 and on a brain phantom show consistent PSNR improvements over unregularized sequential and simultaneous PIE, particularly at low overlap and high noise.

Significance. If the results hold, the paper makes a useful contribution by showing that PnP-HQS, with its closed-form data-consistency updates, is a practical and flexible framework for ptychographic reconstruction with reduced overlap. The derivation in Section 2 is algebraically sound, and the evaluation covers overlap and noise sweeps with multiple denoisers; the use of equivariant evaluation for DRUNet is a positive methodological detail. The significance is tempered by the fact that the quantitative evidence is entirely simulation-based and by two heuristics (the real/imaginary-part complex denoiser and the D²≈γI approximation for DRUNet) that are not validated or quantified. The paper would be strengthened by tests against a true complex-valued proximal/denoiser and by a clearer statement of the simulation-only scope of the low-overlap claim.

major comments (3)
  1. [Section 2.3] The complex-valued denoiser is implemented by adding a constant equal to the maximum amplitude to the real and imaginary parts, applying a real-valued pre-trained denoiser to each part independently, and (implicitly) subtracting the constant afterward. This is a heuristic proxy for the complex proximal operator, and it is load-bearing: every PnP result in Tables 1–3 uses this construction. The paper provides no evidence that this proxy behaves like a true complex-valued proximal mapping, and the text does not specify the subtraction step explicitly. Please either validate the heuristic (for instance, by comparing with an explicitly complex proximal operation for a known regularizer, or with a magnitude-phase decomposition) or soften the claims that depend on it.
  2. [Section 2.2, Remark 1] For the DRUNet experiments, the spatially varying operator D is replaced by γI through the approximation D²≈γI, but the approximation error is never quantified. For the binary circular probe used in Section 3.1, D is not a multiple of the identity: it takes different values in the interior, the partially illuminated boundary, and the corners of the reconstructed field. Since all DRUNet results in Tables 1 and 2 rely on this approximation, the paper should quantify its effect (e.g., by comparing with the exact spatially varying update used for TV and WCRR on a subset of images) or justify that the chosen τ_k schedule makes the reconstruction insensitive to it.
  3. [Abstract, Section 3.2, Section 4] The abstract states that the method is evaluated on 'real test objects,' and the conclusion states that PnP-HQS 'was able to produce good reconstructions while using lower probe overlap than classical methods.' However, all quantitative experiments in Section 3 use simulated measurements: the BSD500 images are synthetic complex images with a simulated circular probe, and the brain phantom in Section 3.2 also uses simulated measurements ('We simulate measurements using a circular probe'). No real ptychographic diffraction data are presented. Please either add real-data experiments or revise the abstract/conclusion to state explicitly that the low-overlap claim is demonstrated only in simulation.
minor comments (5)
  1. [Section 2.1, after Eq. (5)] The formula 'c = n/(n+µσ²)' contains an undefined σ²; it should be c = n/(n+µ) to match Algorithm 1 and the derivation.
  2. [Section 3.1, 'Influence of Noise'] The text says 'The results are in Table 1,' but the noise-sweep results are in Table 2; please correct the cross-reference.
  3. [Section 3.1] The paper does not state how many BSD500 images are used for testing or how the validation set is split; please specify the number of test images and random seeds so that the standard deviations in Tables 1 and 2 are interpretable.
  4. [Remark 1] The claim that the algorithm recovers simultaneous PIE when R is the indicator of the nonnegative orthant is imprecise: SimPIE's update is a step-size-weighted averaging update, whereas the HQS update with R=ι_{[0,∞)ⁿ} is a projection of the auxiliary least-squares solution; the two coincide only under additional assumptions, which should be stated.
  5. [Section 2.3] The denoising procedure for complex images is not fully specified: after adding the maximum-amplitude constant, applying the real-valued denoiser to each component, and recombining, the constant must be subtracted; the paper should state this explicitly to ensure reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the HQS derivation is self-contained and the empirical claims are benchmarked against external algorithms and data.

full rationale

The derivation chain is not circular. The variational problem in Eq. (2) is split by half-quadratic splitting into z- and x-subproblems. The z-update is solved analytically using the Fourier isometry and separability of the objective (Eqs. (3)-(5)), with no fitted parameters entering the derivation. The ptychographic extension introduces auxiliary variables and reduces the x-update to a weighted average of probe patches followed by a denoising step (Eqs. (6)-(9)); again, no parameter is fitted as part of the derivation. Hyperparameters are tuned on a small validation set at one overlap/noise setting and then fixed for all reported test settings, which is standard practice rather than fitting the reported results. The comparisons are against external classical algorithms (SimPIE, SeqPIE) and an external pre-trained denoiser (DRUNet), and the datasets are external (BSD500, brain phantom), so the central empirical claim is not forced by construction. The only self-citation is reference [1] in the introduction, used as background for patch priors; it is not load-bearing for any claim in the paper. The paper's phrase 'real test objects' is an overstatement because the experiments use simulated measurements, and the real/imaginary-part complex denoiser is a heuristic proxy for a complex proximal operator; these are correctness or transfer risks, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central algorithm relies on standard variational splitting and a few modeling heuristics. No new physical entities are introduced. The main free parameters are tuning knobs for regularisation strength and denoiser noise level, all selected on a validation set.

free parameters (3)
  • lambda (regularisation strength) = 0.0001 for TV and DRUNet, 0.01 for WCRR
    Chosen on a small validation set to maximize amplitude PSNR at 38% overlap and alpha=20, as described in Section 3.1.
  • tau (denoising strength) = Ranges: TV [6.0, 30.0], WCRR [200, 1500], DRUNet [5.0, 30.0]
    Selected with lambda on the same validation set; controls the implicit prior strength in the plug-and-play update.
  • Offset constant for complex denoiser = Maximum amplitude of the current image
    Added to real and imaginary parts before applying denoisers to force non-negativity; a hand-chosen preprocessing constant, Section 2.3.
assumptions (4)
  • domain assumption The probe P and scan positions are known exactly.
    Used in the forward model Eq. (1) and throughout the algorithm; joint probe-object reconstruction is deferred to future work in Section 4.
  • domain assumption The shot noise model can be approximated by a Gaussian L2 data fidelity term.
    Stated in Section 1, following [38]; the loss 1/2 ||y - G(z)||^2 is used even though the noise model is Poisson-like.
  • domain assumption The accumulated probe intensity satisfies D² ≈ γI for some constant γ, so that the ptychography x-step reduces to a standard denoiser.
    Remark 1 in Section 2.2; used for the DRUNet experiments, and the error of this approximation is not quantified.
  • ad hoc to paper Applying a real-valued denoiser separately to the real and imaginary parts (after an offset) is a valid complex-valued denoiser.
    Section 2.3; this heuristic is required to use pre-trained greyscale denoisers and is not theoretically justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Plug-and-Play Half-Quadratic Splitting for Ptychography." pith.science (2026). https://pith.science/paper/JBIFIFXP

@misc{pith2026241202548,
  author       = {Pith},
  title        = {Pith review of: Plug-and-Play Half-Quadratic Splitting for Ptychography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBIFIFXP}},
  note         = {Machine review of arXiv:2412.02548}
}
read the original abstract

Ptychography is a coherent diffraction imaging method that uses phase retrieval techniques to reconstruct complex-valued images. It achieves this by sequentially illuminating overlapping regions of a sample with a coherent beam and recording the diffraction pattern. Although this addresses traditional imaging system challenges, it is computationally intensive and highly sensitive to noise, especially with reduced illumination overlap. Data-driven regularisation techniques have been applied in phase retrieval to improve reconstruction quality. In particular, plug-and-play (PnP) offers flexibility by integrating data-driven denoisers as implicit priors. In this work, we propose a half-quadratic splitting framework for using PnP and other data-driven priors for ptychography. We evaluate our method both on natural images and real test objects to validate its effectiveness for ptychographic image reconstruction.

Figures

Figures reproduced from arXiv: 2412.02548 by the authors.

Figure 1
Figure 1. The measurement model for ptychography. We first extract a patch from [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Reconstructions for the 7 × 7 probe setting with α = 20.0. We only show the center of the image where PSNR is calculated. where MSEϕ is the mean squared error on the sphere defined as MSEϕ(ϕreco, ϕgt) = 1 N X N i=1 ((ϕreco,i − ϕgt,i + π) mod 2π − π) 2 , and ϕreco and ϕgt are the phases of the reconstruction and ground truth image. Because measurements are sparse at the border, it cannot be reconstructed. Therefore, … view at source ↗
Figure 3
Figure 3. Reconstruction of the phase of the brain phantom. The ground truth is [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LoRePIE: $\ell_0$ Regularised Extended Ptychographical Iterative Engine for Low-dose and Fast Electron Ptychography

    physics.app-ph 2024-11 conditional novelty 4.0 of 10

    LoRePIE, an l0-regularized ePIE with DCT-domain hard thresholding, reconstructs useful electron ptychography images from probe positions with as low as 56% overlap.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Inverse Problems 39(6), 064006 (2023) Plug-and-Play Half-Quadratic Splitting for Ptychography 11

    Altekrüger, F., Denker, A., Hagemann, P., Hertrich, J., Maass, P., Steidl, G.: PatchNR: learning from very few images by patch normalizing flow regularization. Inverse Problems 39(6), 064006 (2023) Plug-and-Play Half-Quadratic Splitting for Ptychography 11

    work page 2023

  2. [2]

    Acta Numerica 27, 1–111 (2018)

    Benning, M., Burger, M.: Modern regularization methods for inverse problems. Acta Numerica 27, 1–111 (2018)

    work page 2018

  3. [3]

    Buzzard, G.T., Chan, S.H., Sreehari, S., Bouman, C.A.: Plug-and-play unplugged: Optimization-free reconstruction using consensus equilibrium. SIAM J. Imaging Sci. 11(3), 2001–2020 (2018)

    work page 2018

  4. [4]

    Candes, E.J., Li, X., Soltanolkotabi, M.: Phase retrieval from coded diffraction patterns. Appl. Comput. Harmon A.39(2), 277–299 (2015)

    work page 2015

  5. [5]

    IEEE Trans

    Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play ADMM for image restoration: Fixed-point convergence and applications. IEEE Trans. Comput. Imaging3(1), 84– 98 (2016)

    work page 2016

  6. [6]

    IEEE Signal Processing Magazine 40(1), 45–57 (2023)

    Dong, J., Valzania, L., Maillard, A., Pham, T.a., Gigan, S., Unser, M.: Phase re- trieval: From computational imaging to machine learning. IEEE Signal Processing Magazine 40(1), 45–57 (2023)

    work page 2023

  7. [7]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 472(2196), 20160640 (2016)

    Enders,B.,Thibault,P.:Acomputationalframeworkforptychographicreconstruc- tions. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 472(2196), 20160640 (2016)

    work page 2016

  8. [8]

    Applied Optics21(15), 2758–2769 (1982)

    Fienup, J.R.: Phase retrieval algorithms: a comparison. Applied Optics21(15), 2758–2769 (1982)

    work page 1982

Show all 41 references
  1. [9]

    IEEE Transactions on Image Processing4(7), 932–946 (1995)

    Geman, D., Yang, C.: Nonlinear image recovery with half-quadratic regularization. IEEE Transactions on Image Processing4(7), 932–946 (1995)

    1995

  2. [10]

    IEEE Transactions on Computa- tional Imaging 9, 781–795 (2023)

    Goujon, A., Neumayer, S., Bohra, P., Ducotterd, S., Unser, M.: A neural-network- based convex regularizer for inverse problems. IEEE Transactions on Computa- tional Imaging 9, 781–795 (2023)

    2023

  3. [11]

    Goujon, A., Neumayer, S., Unser, M.: Learning weakly convex regularizers for convergent image-reconstruction algorithms. SIAM J. Imaging Sci.17(1), 91–115 (2024)

    2024

  4. [12]

    NeurIPS 35, 14715–14728 (2022)

    Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. NeurIPS 35, 14715–14728 (2022)

    2022

  5. [13]

    Optics express16, 7264–78 (06 2008)

    Guizar-Sicairos, M., Fienup, J.: Phase retrieval with transverse translation diver- sity: A nonlinear optimization approach. Optics express16, 7264–78 (06 2008)

    2008

  6. [14]

    GAMM-Mitteilungen p

    Habring, A., Holler, M.: Neural-network-based regularization methods for inverse problems in imaging. GAMM-Mitteilungen p. e202470004 (2024)

    2024

  7. [15]

    Linear Algebra Appl.631, 203–234 (2021)

    Hertrich, J., Neumayer, S., Steidl, G.: Convolutional proximal neural networks and plug-and-play algorithms. Linear Algebra Appl.631, 203–234 (2021)

    2021

  8. [16]

    In: ICLR (2022)

    Hurault, S., Leclaire, A., Papadakis, N.: Gradient step denoiser for convergent plug-and-play. In: ICLR (2022)

    2022

  9. [17]

    In: ICML

    Hurault, S., Leclaire, A., Papadakis, N.: Proximal denoiser for convergent plug- and-play optimization with nonconvex regularization. In: ICML. pp. 9483–9505 (2022)

    2022

  10. [18]

    Ultrami- croscopy 171, 43–54 (2016)

    Konijnenberg, A.P., Coene, W.M.J., Pereira, S.F., Urbach, H.P.: Combining pty- chographical algorithms with the hybrid input-output (HIO) algorithm. Ultrami- croscopy 171, 43–54 (2016)

    2016

  11. [19]

    Ultramicroscopy109 10, 1256–62 (2009)

    Maiden, A., Rodenburg, J.M.: An improved ptychographical phase retrieval algo- rithm for diffractive imaging. Ultramicroscopy109 10, 1256–62 (2009)

    2009

  12. [20]

    In: IEEE ICCV

    Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: IEEE ICCV. vol. 2, pp. 416–423 (2001)

    2001

  13. [21]

    Martin, S., Gagneux, A., Hagemann, P., Steidl, G.: PnP-flow: Plug-and-play image restoration with flow matching (2024),arXiv:2410.02423 12 Denker et al

    2024 arXiv

  14. [22]

    In: ICCV

    Meinhardt, T.,Moller, M., Hazirbas, C., Cremers, D.: Learning proximal operators: Using denoising networks for regularizing inverse imaging problems. In: ICCV. pp. 1781–1790 (2017)

    2017

  15. [23]

    In: ICML

    Metzler, C., Schniter, P., Veeraraghavan, A., Baraniuk, R.: prdeep: Robust phase retrieval with a flexible deep network. In: ICML. pp. 3501–3510. PMLR (2018)

    2018

  16. [24]

    Pendu, M.L., Guillemot, C.: Preconditioned plug-and-play ADMM with locally adjustable denoiser for image restoration. SIAM J. Imaging Sci.16(1), 393–422 (2023)

    2023

  17. [25]

    Pesquet, J.C., Repetti, A., Terris, M., Wiaux, Y.: Learning maximally monotone operators for image recovery. SIAM J. Imaging Sci.14(3), 1206–1237 (2021)

    2021

  18. [26]

    Rodenburg, J., Maiden, A.: Ptychography, p. 819–904. Springer Handbooks, Springer International Publishing, Cham (2019)

    2019

  19. [27]

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

    2017

  20. [28]

    Physica D: Nonlinear Phenomena60(1-4), 259–268 (1992)

    Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena60(1-4), 259–268 (1992)

    1992

  21. [29]

    In: ICML

    Ryu, E., Liu, J., Wang, S., Chen, X., Wang, Z., Yin, W.: Plug-and-play methods provably converge with properly trained denoisers. In: ICML. pp. 5546–5557 (2019)

    2019

  22. [30]

    Acta Crystallographica 5(6), 843–843 (1952)

    Sayre, D.: Some implications of a theorem due to Shannon. Acta Crystallographica 5(6), 843–843 (1952)

    1952

  23. [31]

    Shastri, S.K., Schniter, P.: Fast and robust phase retrieval via deep expectation- consistent approximation (2024),arxiv:2407.09687

    2024 arXiv

  24. [32]

    Sommerhoff, H., Kolb, A., Moeller, M.: Energy dissipation with plug-and-play pri- ors.In:NeurIPSWorkshoponSolvingInverseProblemswithDeepNetworks(2019)

    2019

  25. [33]

    In: IEEE ICASSP

    Sun, Y., Xu, S., Li, Y., Tian, L., Wohlberg, B., Kamilov, U.S.: Regularized Fourier ptychography using an online plug-and-play algorithm. In: IEEE ICASSP. pp. 7665–7669. IEEE (2019)

    2019

  26. [34]

    In: CVPR

    Terris,M.,Moreau,T.,Pustelnik,N.,Tachella,J.:Equivariantplug-and-playimage reconstruction. In: CVPR. pp. 25255–25264 (2024)

    2024

  27. [35]

    In: IEEE GlobalSIP

    Venkatakrishnan, S.V., Bouman, C.A., Wohlberg, B.: Plug-and-play priors for model based reconstruction. In: IEEE GlobalSIP. pp. 945–948 (2013)

    2013

  28. [36]

    Wakonig, K., Stadler, H.C., Odstrčil, M., Tsai, E.H.R., Diaz, A., Holler, M., Usov, I., Raabe, J., Menzel, A., Guizar-Sicairos, M.: It ptychoshelves, a versatile high- level framework for high-performance analysis of ptychographic data. J. Appl. Crystallogr. 53(2), 574–586 (2020)

    2020

  29. [37]

    In: ICML

    Wang, Y., Sun, X., Fleischer, J.: When deep denoising meets iterative phase re- trieval. In: ICML. pp. 10007–10017. PMLR (2020)

    2020

  30. [38]

    Yeh, L.H., Dong, J., Zhong, J., Tian, L., Chen, M., Tang, G., Soltanolkotabi, M., Waller, L.: Experimental robustness of Fourier ptychography phase retrieval algorithms. Opt. Express23(26), 33214–33240 (2015)

    2015

  31. [39]

    Optica4(10), 1306–1313 (2017)

    Zhang, J., Pégard, N., Zhong, J., Adesnik, H., Waller, L.: 3D computer-generated holography by non-convex optimization. Optica4(10), 1306–1313 (2017)

    2017

  32. [40]

    Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play imagerestorationwithdeepdenoiserprior.IEEETPAMI 44(10),6360–6376(2021)

    2021

  33. [41]

    In: CVPR

    Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep CNN denoiser prior for image restoration. In: CVPR. pp. 3929–3938 (2017)

    2017

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.