Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Deep Plug-and-Play HIO Approach for Phase Retrieval

T0 review · 2 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Plug-and-Play HIO plugs a deep denoiser into HIO-type projections and reports state-of-the-art Fourier phase retrieval on 236 test images.

desk verdict Useful empirical study, but the paper's central HQS derivation does not hold; the method is a heuristic that works well in experiments. read the letter →

arxiv 2411.18967 v2 pith:RQU2CRBH submitted 2024-11-28 eess.IV cs.CVcs.LG

classification eess.IVcs.CVcs.LG
keywords Fourierphaseretrievalplug-and-playregularizationhalf-quadraticsplittinghybridinput-outputdeepdenoiserpriorimagereconstructionalternatingprojectionsnoisyintensitymeasurements
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

The paper develops Plug-and-Play HIO, a phase-retrieval algorithm that alternates half-quadratic-splitting update steps with a pre-trained deep denoiser, and claims it achieves state-of-the-art reconstruction quality for noisy Fourier phase retrieval. On a 236-image test set, the method reports higher average PSNR and SSIM than conventional HIO, PrDeep, and Iterative DNN-HIO at three Poisson noise levels, while running in about half the time of PrDeep. The paper argues that the success comes from treating the data-fidelity subproblem with closed-form measurement and image updates, and using HIO inner iterations rather than error-reduction projections when space-domain constraints are available. This matters because phase retrieval is nonlinear and ill-posed, and the method promises fast, flexible reconstruction without retraining when the measurement model changes.

What carries the argument

The load-bearing identity is the one-step fixed-point update derived from the necessary condition on the HQS data-fidelity subproblem: $$ x = U_k \,\mathrm{Re}\{$A^{{-1}}$(y \odot A z_k/|A z_k|)\} + (1-U_k)z_k. $$ The paper rewrites it as a convex measurement update $\tilde{y}_k = U_k y + (1-U_k)|A z_k|$ followed by an error-reduction projection using the current phase estimate, and then a deep denoiser for the $z$-subproblem. In Plug-and-Play HIO, the projection block is replaced by several hybrid input-output iterations, a relaxed alternating-projection scheme that updates only the components of the estimate that violate the space-domain constraints. The varying weight $U_k$ and denoiser noise level $\sigma_k$ decrease with iterations, so the algorithm increasingly trusts the learned prior and the evolving data model.

What would settle it

Compute the left-hand side of the necessary condition in Eq. (6) at the output $x$ produced by Eq. (7) for a set of test images; if the residual is far from zero and stays nonzero when the update is applied repeatedly, then the claim that Eq. (7) satisfies the optimality condition is false. A direct algorithmic version: replace the single update with several inner applications per outer iteration and see whether PSNR improves substantially; if it does, the one-step solution claim is not the reason the method works.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a plug-and-play variant of HIO, derived through half-quadratic splitting, can outperform both classical alternating projections and existing deep-learning phase-retrieval methods on Fourier intensity measurements. The derivation produces the measurement update $\tilde{y}_k = U_k y + (1 - U_k)|A z_k|$ and the image update $x_{k+1} = \mathrm{Re}\{A^{-1}(\tilde{y}_k \odot A z_k / |A z_k|)\}$, followed by the denoiser step $z_{k+1} = D(x_{k+1}, \sigma_k)$; Eq. (7) presents the combined update as satisfying the necessary optimality condition for the HQS data-fidelity subproblem. When the projection step is replaced by a few hybrid input-output iterations, the resulting Plug-and-Play HIO reports the best average PSNR and SSIM on the 236-image dataset, is less sensitive to initialization and noise, and generalizes to non-natural images even though the denoiser was trained only on natural images. The method inherits the flexibility of projection-based phase retrieval, so the same algorithm can be applied to other magnitude measurements, such as ptychography or coded diffraction patterns, without retraining.

Load-bearing premise

The load-bearing premise is that the single fixed-point update in Eq. (7) actually solves the necessary optimality condition in Eq. (6); one pass of a fixed-point map is not a fixed point in general, so if that step is only an approximation, the derivation's guarantee is not strict.

Editorial extensions

If this is right

  • Plug-and-Play HIO lifts average reconstruction quality on the full test set: at the lowest tested noise it reports 26.35 dB PSNR and 0.751 SSIM, versus 25.78 dB and 0.663 for PrDeep and 25.63 dB and 0.737 for Iterative DNN-HIO.
  • Reconstruction is fast: about 247 seconds per image versus 475 seconds for PrDeep, making the learned-prior quality gain available at a practical runtime.
  • Because the data-fidelity step is projection-based and the prior enters only as a denoiser, the algorithm can switch between measurement models (Fourier magnitude, ptychography, coded diffraction patterns) without retraining.
  • The HIO inner loop is what earns the improvement: Plug-and-Play PR, the error-reduction variant, is clearly worse in PSNR and SSIM, so the relaxation matters.
  • Parameter robustness: performance remains state-of-the-art at noise levels other than the one used for parameter tuning, and with a single random initialization the method can be made robust by increasing inner and outer iterations.

Reading between the lines

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

  • Inference: if the paper's derivation is taken literally, the natural stress test is to run several inner fixed-point updates in place of Eq. (7) and check whether PSNR improves; a material improvement would show the single-step update is only an approximation, not the claimed solution of subproblem (5a).
  • Inference: the same HQS-plus-denoiser recipe should transfer to ptychography or coded-diffraction phase retrieval by swapping the forward operator, since the method already accommodates non-invertible $A$ via the pseudoinverse.
  • Inference: for complex-valued images, applying a real-valued denoiser separately to amplitude and phase is a heuristic; a denoiser trained on complex-valued data or on the target image distribution would likely narrow the gap to natural-image performance.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper proposes two phase-retrieval algorithms, Plug-and-Play PR and Plug-and-Play HIO, that combine half-quadratic splitting (HQS) with a pre-trained CNN denoiser. The authors claim to derive analytical measurement and image update steps from HQS and to combine them with learning-based denoising. The experimental section compares the proposed methods against HIO, PrDeep, and Iterative DNN-HIO on 236 test images at three Poisson noise levels, using 5 Monte Carlo runs, and reports average PSNR, SSIM, and runtime. The results show that Plug-and-Play HIO achieves the best average PSNR/SSIM for natural images at all tested noise levels, with runtime roughly half that of PrDeep, and exhibits improved robustness to initialization and noise.

Significance. The empirical evaluation is a clear strength: 236 test images, 5 Monte Carlo runs per setting, comparisons against the authors' implementations of competing learning-based methods, and separate sensitivity analyses for initialization and noise. If the empirical results are correct, Plug-and-Play HIO is practically useful because it combines the computational efficiency of classical projection methods with the prior power of deep denoisers and appears less initialization-sensitive than the compared alternatives. However, the paper's central mathematical claim—that the update steps are derived from half-quadratic splitting—is not supported: Eq. (7) is a single fixed-point iterate rather than a solution of the stated optimality condition, and the HIO inner iterations in Algorithm 2 are introduced heuristically without an associated objective. The theoretical contribution therefore fails as written, and the paper should be reframed as an empirical study of a well-motivated heuristic or accompanied by a corrected derivation.

major comments (2)
  1. [Sec. 3, Eqs. (6)-(7)] The derivation of the analytical update step is incorrect. The necessary condition for the x-subproblem in Eq. (5a) is a nonlinear equation in x; for general invertible A it contains an additional term involving A^H A x, and even for unitary A the phase term Ax/|Ax| depends on x. The paper evaluates the phase at the previous iterate z_N and then solves a linear equation for x, obtaining Eq. (7). This is one iteration of a fixed-point map, not a point satisfying the stated necessary condition. A concrete scalar example illustrates the failure: with A=2, y=2, z=0.1, Q=1, Eq. (7) gives x=0.55, while the exact minimizer of Eq. (5a) is x=0.82. Thus the claim that Eq. (7) 'provides the following solution that satisfies the above necessary condition for optimality' is false, and the paper's stated contribution of deriving the update steps from half-quadratic splitting is not established.
  2. [Algorithm 2, steps 3-12] The paper presents Plug-and-Play HIO by replacing the error-reduction projection steps of Algorithm 1 with a few HIO inner iterations, but this substitution is not derived from HQS or from any stated objective function. The inner HIO loop is introduced as a heuristic improvement (justified by the known behavior of HIO for nonconvex constraints), and no argument is given that alternating the measurement update, HIO iterations, and the denoiser corresponds to minimizing a common cost. Consequently, the 'mathematical development' announced in the abstract and introduction covers only the simplified PnP-PR update, not the algorithm that actually achieves the reported state-of-the-art results. The authors should either provide a corrected derivation for the HIO variant or explicitly characterize it as an empirically validated heuristic.
minor comments (6)
  1. [Sec. 2, Eq. (1)] The noise model uses the symbol L for the normal distribution; this should be N, and the relation between the scaling parameter M and the SNR should be defined more explicitly.
  2. [Sec. 3, Eqs. (4)-(5)] The factor 1/2 appears inconsistently between the objective in Eq. (4) and the subproblem in Eq. (5a); this is cosmetic but should be made consistent.
  3. [Sec. 4, Table 1] The plus-or-minus values in Table 1 are not identified as standard deviations or standard errors; please specify the measure of variability.
  4. [Algorithm 2] The pseudocode uses the inner loop variable Z but refers to iterates as v_{P+1} and v_P; this inconsistent indexing should be corrected to avoid confusion.
  5. [Sec. 4, parameter selection] The parameters S_max=40, S_min=5, X=5, and W=200 are described as optimized for the M=3 case only; the paper should state explicitly whether the same schedule is used for M=2 and M=4 or whether it was retuned for those noise levels.
  6. [Data availability] Because the quantitative conclusions depend on the exact denoiser-switching schedule and parameter schedules, providing the reconstruction code or a fully specified pseudo-code would substantially improve reproducibility; the current data availability statement does not mention code.

Circularity Check

0 steps flagged · score 1.0 of 10

No construction-level circularity: Eq. (7) is an unsupported one-step approximation rather than a solution equivalent to Eq. (6), and the self-citations are provenance-only.

full rationale

The paper's derivation chain runs from the HQS objective (Eqs. 3-5) through the necessary condition (Eq. 6) to the update (Eq. 7). This step is not circular: Eq. (7) is not obtained by substituting Eq. (7) into Eq. (6), and for generic inputs it does not satisfy the necessary condition. That is a mathematical-justification gap, not an equivalence-by-construction or a fitted quantity renamed as a prediction. The numerical parameters S_MAX and S_MIN are selected by numerical search; even if tuned for the M=3 noise level, the M=2 and M=4 results are evaluated under different noise conditions, so the robustness claim is not statistically forced by construction. The self-citations [1,2] are used only to attribute the earlier conference presentation and provenance of the algorithm ('The results of this manuscript were obtained in 2019-2020 [1,2]' and 'first presented in [2]'); the present paper restates the full algorithm in Algorithms 1-2 and reruns the comparisons, so those citations are not load-bearing. No uniqueness theorem from the same authors is imported, and no known result is merely renamed: the contribution is an empirical combination of HIO-type projections with an external pretrained denoiser [25]. The Data availability statement's lack of public code and data is a reproducibility limitation, not a circularity. Overall, the central empirical claim is self-contained against external baselines, and the circularity burden is very low.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on hyperparameters tuned on one noise level and on the unproven one-step optimality of the update. No new physical entities are introduced.

free parameters (5)
  • S_max (initial denoiser noise level) = 40
    Selected by numerical search for the M=3 noise case; controls the starting point of the denoiser noise schedule.
  • S_min (final denoiser noise level) = 5
    Selected by numerical search for the M=3 noise case; controls the ending point of the denoiser noise schedule.
  • X (inner HIO iterations) = 5
    Chosen as generally sufficient for convergence; not optimized per image.
  • W (outer iterations) = 200
    Chosen as generally sufficient for convergence; not optimized per image.
  • beta (HIO feedback parameter) = 0.9
    Uses the common default value from Fienup's HIO algorithm.
assumptions (3)
  • domain assumption HQS splitting is applicable to the non-convex phase retrieval objective in Eq. (2).
    The data fidelity term is non-convex, yet the paper applies HQS without proving convergence to a stationary point.
  • domain assumption A pre-trained Gaussian denoiser can serve as the proximal operator / prior for the z-subproblem in Eq. (5b).
    Standard plug-and-play assumption; the paper uses a CNN denoiser trained for Gaussian noise.
  • ad hoc to paper The one-step fixed-point update in Eq. (7) solves the necessary condition in Eq. (6) for the x-subproblem.
    This is claimed but not proven and is generally false; only one fixed-point iteration is performed, not a convergent fixed point.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Plug-and-Play HIO Approach for Phase Retrieval." pith.science (2026). https://pith.science/paper/RQU2CRBH

@misc{pith2026241118967,
  author       = {Pith},
  title        = {Pith review of: Deep Plug-and-Play HIO Approach for Phase Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQU2CRBH}},
  note         = {Machine review of arXiv:2411.18967}
}
read the original abstract

In the phase retrieval problem, the aim is the recovery of an unknown image from intensity-only measurements such as Fourier intensity. Although there are several solution approaches, solving this problem is challenging due to its nonlinear and ill-posed nature. Recently, learning-based approaches have emerged as powerful alternatives to the analytical methods for several inverse problems. In the context of phase retrieval, a novel plug-and-play approach that exploits learning-based prior and efficient update steps has been presented at the Computational Optical Sensing and Imaging topical meeting, with demonstrated state-of-the-art performance. The key idea was to incorporate learning-based prior to the Gerchberg-Saxton type algorithms through plug-and-play regularization. In this paper, we present the mathematical development of the method including the derivation of its analytical update steps based on half-quadratic splitting and comparatively evaluate its performance through extensive simulations on a large test dataset. The results show the effectiveness of the method in terms of both image quality, computational efficiency, and robustness to initialization and noise.

Figures

Figures reproduced from arXiv: 2411.18967 by the authors.

Figure 1
Figure 1. Reconstructions obtained with di"erent methods for three test images and 𝑀=3 case [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Comparison of PnP PR and PnP HIO 𝑋 = 50 and 𝑊 = 400, a reconstruction with a quality similar to the earlier result in [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Normalized di"erence (↘z𝑁+1 ↗ z𝑁 ↘2/↘z𝑁 ↘2) and PSNR versus iteration count for randomly chosen five test images to demonstrate the convergence of PnP HIO [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The histograms of PSNR (left column) and SSIM (right column) for the [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Reconstructions of PnP HIO with a single random initialization for di [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

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. DDRM-PR: Fourier Phase Retrieval using Denoising Diffusion Restoration Models

    eess.IV 2025-01 conditional novelty 6.0 of 10

    A training-free method (DDRM-PR) that combines pretrained diffusion priors with alternating projection improves Fourier phase retrieval over HIO alone, but with only modest simulated gains and a heuristic theoretical basis.

Reference graph

Works this paper leans on

61 extracted references · 59 canonical work pages · cited by 1 Pith paper

  1. [1]

    Development of novel analysis and reconstruction techniques for coherent optical imaging systems,

    Isil, “Development of novel analysis and reconstruction techniques for coherent optical imaging systems,” Master’s thesis, Middle East Technical University (2019)

  2. [2]

    Model-based phase retrieval with deep denoiser prior,

    Ç. I#ıl and F. S. Oktem, “Model-based phase retrieval with deep denoiser prior,” in Imaging and Applied Optics Congress, (Optica Publishing Group, 2020), p. CF2C.5

  3. [3]

    Wide-field, high-resolution Fourier ptychographic microscopy,

    G. Zheng, R. Horstmeyer, and C. Y ang, “Wide-field, high-resolution Fourier ptychographic microscopy,” Nat. Photonics 7, 739 (2013)

  4. [4]

    Phase recovery and holographic image reconstruction using deep learning in neural networks,

    Y . Rivenson, Y . Zhang, H. Günaydın, D. Teng, and A. Ozcan, “Phase recovery and holographic image reconstruction using deep learning in neural networks,” Light. Sci. & Appl. 7, 17141 (2018)

  5. [5]

    Phase retrieval in crystallography and optics,

    R. P . Millane, “Phase retrieval in crystallography and optics,” J. Opt. Soc. Am. A 7, 394–411 (1990)

  6. [6]

    The question of phase retrieval in optics,

    A. Walther, “The question of phase retrieval in optics,” Opt. Acta: Int. J. Opt. 10, 41–49 (1963)

  7. [7]

    Image recovery from correlations,

    T. J. Schulz and D. L. Snyder, “Image recovery from correlations,” J. Opt. Soc. Am. A 9, 1266–1272 (1992)

  8. [8]

    Phase retrieval and image reconstruction for astronomy,

    J. C. Dainty and J. Fienup, “Phase retrieval and image reconstruction for astronomy,” Image Recover. Theory Appl 13, 231–275 (1987)

Show all 61 references
  1. [9]

    The reconstruction of a multidimensional sequence from the phase or magnitude of its Fourier transform,

    M. Hayes, “The reconstruction of a multidimensional sequence from the phase or magnitude of its Fourier transform,” IEEE Transactions on Acoust. Speech, Signal Process. 30, 140–154 (1982)

  2. [10]

    Phase retrieval with application to optical imaging: a contemporary overview,

    Y . Shechtman, Y . C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and M. Segev, “Phase retrieval with application to optical imaging: a contemporary overview,” IEEE Signal Process. Mag. 32, 87–109 (2015)

  3. [11]

    Phase retrieval algorithms: a comparison,

    J. R. Fienup, “Phase retrieval algorithms: a comparison,” Appl. Opt. 21, 2758–2769 (1982)

  4. [12]

    A practical algorithm for the determination of phase from image and di

    R. W . Gerchberg and W . O. Saxton, “A practical algorithm for the determination of phase from image and di"raction plane pictures,” Optik 35, 237–250 (1972)

  5. [13]

    Reconstruction of an object from the modulus of its Fourier transform,

    J. R. Fienup, “Reconstruction of an object from the modulus of its Fourier transform,” Opt. letters 3, 27–29 (1978)

  6. [14]

    The numerics of phase retrieval,

    A. Fannjiang and T. Strohmer, “The numerics of phase retrieval,” Acta Numer. 29, 125–228 (2020)

  7. [15]

    Dolphin—dictionary learning for phase retrieval,

    A. M. Tillmann, Y . C. Eldar, and J. Mairal, “Dolphin—dictionary learning for phase retrieval,” IEEE Transactions on Signal Process. 64, 6485–6500

  8. [16]

    Iterative projection meets sparsity regularization: towards practical single-shot quantitative phase imaging with in-line holography,

    Y . Gao and L. Cao, “Iterative projection meets sparsity regularization: towards practical single-shot quantitative phase imaging with in-line holography,” Light. Adv. Manuf. 4, 37–53 (2023)

  9. [17]

    Deep learning techniques for inverse problems in imaging,

    G. Ongie, A. Jalal, C. A. Metzler, R. G. Baraniuk, A. G. Dimakis, and R. Willett, “Deep learning techniques for inverse problems in imaging,” IEEE J. on Sel. Areas Inf. Theory 1, 39–56 (2020)

  10. [18]

    prDeep: Robust phase retrieval with a flexible deep network,

    C. Metzler, P . Schniter, A. Veeraraghavan, and R. Baraniuk, “prDeep: Robust phase retrieval with a flexible deep network,” in Proceedings of the 35th International Conference on Machine Learning, (JMLR.org, 2018), pp. 3498–3507

  11. [19]

    Plug-and-play priors for model based reconstruction,

    S. V . Venkatakrishnan, C. A. Bouman, and B. Wohlberg, “Plug-and-play priors for model based reconstruction,” in IEEE Global Conference on Signal and Information Processing, (IEEE, 2013), pp. 945–948

  12. [20]

    T urning a denoiser into a super-resolver using plug and play priors,

    A. Brifman, Y . Romano, and M. Elad, “T urning a denoiser into a super-resolver using plug and play priors,” in 2016 IEEE International Conference on Image Processing (ICIP), (2016), pp. 1404–1408

  13. [21]

    Image restoration and reconstruction using variable splitting and class-adapted image priors,

    A. M. Teodoro, J. M. Bioucas-Dias, and M. A. T. Figueiredo, “Image restoration and reconstruction using variable splitting and class-adapted image priors,” in 2016 IEEE International Conference on Image Processing (ICIP), (2016), pp. 3518–3522

  14. [22]

    Poisson inverse problems by the plug-and-play scheme,

    A. Rond, R. Giryes, and M. Elad, “Poisson inverse problems by the plug-and-play scheme,” J. Vis. Commun. Image Represent. 41, 96–108 (2016)

  15. [23]

    Nonlinear image recovery with half-quadratic regularization,

    D. Geman and C. Y ang, “Nonlinear image recovery with half-quadratic regularization,” IEEE Transactions on Image Process. 4, 932–946 (1995)

  16. [24]

    From learning models of natural image patches to whole image restoration,

    D. Zoran and Y . Weiss, “From learning models of natural image patches to whole image restoration,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), (IEEE, 2011), pp. 479–486

  17. [25]

    Learning deep CNN denoiser prior for image restoration,

    K. Zhang, W. Zuo, S. Gu, and L. Zhang, “Learning deep CNN denoiser prior for image restoration,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (IEEE, 2017), pp. 3929–3938

  18. [26]

    Distributed optimization and statistical learning via the alternating direction method of multipliers,

    S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein et al., “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Foundations Trends Mach. learning 3, 1–122 (2011)

  19. [27]

    Flexisp: A flexible camera image processing framework,

    F. Heide, M. Steinberger, Y .-T. Tsai, M. Rouf, D. Pajak, D. Reddy, O. Gallo, J. Liu, W . Heidrich, K. Egiazarian et al., “Flexisp: A flexible camera image processing framework,” ACM Transactions on Graph. (ToG) 33, 1–13 (2014)

  20. [28]

    Proximal: E!cient image optimization using proximal algorithms,

    F. Heide, S. Diamond, M. Nießner, J. Ragan-Kelley, W. Heidrich, and G. Wetzstein, “Proximal: E!cient image optimization using proximal algorithms,” ACM Transactions on Graph. 35, 84 (2016)

  21. [29]

    Multi-scale weighted nuclear norm image restoration,

    N. Y air and T. Michaeli, “Multi-scale weighted nuclear norm image restoration,” in IEEE Conference on Computer Vision and Pattern recognition, (2018), pp. 3165–3174

  22. [30]

    Learning proximal operators: Using denoising networks for regularizing inverse imaging problems,

    T. Meinhardt, M. Moller, C. Hazirbas, and D. Cremers, “Learning proximal operators: Using denoising networks for regularizing inverse imaging problems,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), (IEEE, 2017), pp. 1781–1790

  23. [31]

    One network to solve them all—solving linear inverse problems using deep projection models,

    J. R. Chang, C. L. Li, B. Poczos, and B. V . Kumar, “One network to solve them all—solving linear inverse problems using deep projection models,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), (IEEE, 2017), pp. 5889–5898

  24. [32]

    MoDL: Model-based deep learning architecture for inverse problems,

    H. K. Aggarwal, M. P . Mani, and M. Jacob, “MoDL: Model-based deep learning architecture for inverse problems,” IEEE Transactions on Med. Imaging 38, 394–405 (2019)

  25. [33]

    Plug-and-play regularization on magnitude with deep priors for 3D near-field MIMO imaging,

    O. Oral and F. S. Oktem, “Plug-and-play regularization on magnitude with deep priors for 3D near-field MIMO imaging,” IEEE Transactions on Comput. Imaging (2024)

  26. [34]

    Phase retrieval via spatial light modulator phase modulation in 4f optical setup: numerical inverse imaging with sparse regularization for phase and amplitude,

    V . Katkovnik and J. Astola, “Phase retrieval via spatial light modulator phase modulation in 4f optical setup: numerical inverse imaging with sparse regularization for phase and amplitude,” J. Opt. Soc. Am. A 29, 105–116 (2012)

  27. [35]

    Phase retrieval from noisy data based on sparse approximation of object phase and amplitude,

    V . Katkovnik, “Phase retrieval from noisy data based on sparse approximation of object phase and amplitude,” arXiv preprint arXiv:1709.01071 (2017)

  28. [36]

    Tuning-free plug-and-play proximal algorithm for inverse imaging problems,

    K. Wei, A. Aviles-Rivero, J. Liang, Y . Fu, C.-B. Schönlieb, and H. Huang, “Tuning-free plug-and-play proximal algorithm for inverse imaging problems,” in International Conference on Machine Learning, (PMLR, 2020), pp. 10158–10169

  29. [37]

    Constrained phase retrieval: when alternating projection meets regularization,

    B. Shi, Q. Lian, X. Huang, and N. An, “Constrained phase retrieval: when alternating projection meets regularization,” JOSA B 35, 1271–1281 (2018)

  30. [38]

    BM3D-PRGAMP: Compressive phase retrieval based on BM3D denoising,

    C. A. Metzler, A. Maleki, and R. G. Baraniuk, “BM3D-PRGAMP: Compressive phase retrieval based on BM3D denoising,” in IEEE International Conference on Image Processing (ICIP), (IEEE, 2016), pp. 2504–2508

  31. [39]

    The little engine that could: Regularization by denoising (RED),

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

  32. [40]

    Online regularization by denoising with applications to phase retrieval,

    Z. Wu, Y . Sun, J. Liu, and U. Kamilov, “Online regularization by denoising with applications to phase retrieval,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, (2019), pp. 0–0

  33. [41]

    PhaseGAN: a deep-learning phase-retrieval approach for unpaired datasets,

    Y . Zhang, M. A. Noack, P . Vagovic, K. Fezzaa, F. Garcia-Moreno, T. Ritschel, and P . Villanueva-Perez, “PhaseGAN: a deep-learning phase-retrieval approach for unpaired datasets,” Opt. Express 29, 19593–19604 (2021)

  34. [42]

    SiSPRNet: end-to-end learning for single-shot phase retrieval,

    Q. Y e, L.-W . Wang, and D. P . Lun, “SiSPRNet: end-to-end learning for single-shot phase retrieval,” Opt. Express30, 31937–31958 (2022)

  35. [43]

    Deep iterative reconstruction for phase retrieval,

    Ç. I#ıl, F. S. Oktem, and A. Koç, “Deep iterative reconstruction for phase retrieval,” Appl. Opt. 58, 5422–5431 (2019)

  36. [44]

    Di"usion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Y e, “Di"usion posterior sampling for general noisy inverse problems,” in The 11th International Conference on Learning Representations, (2023)

  37. [45]

    Phase retrieval using conditional generative adversarial networks,

    T. Uelwer, A. Oberstraß, and S. Harmeling, “Phase retrieval using conditional generative adversarial networks,” in 2020 25th International Conference on Pattern Recognition (ICPR), (IEEE, 2021), pp. 731–738

  38. [46]

    DOLPH: Di

    S. Shoushtari, J. Liu, and U. S. Kamilov, “DOLPH: Di"usion models for phase retrieval,” arXiv preprint arXiv:2211.00529 (2022)

  39. [47]

    Compressed sensing-based robust phase retrieval via deep generative priors,

    F. Shamshad and A. Ahmed, “Compressed sensing-based robust phase retrieval via deep generative priors,” IEEE Sensors J. 21, 2286–2298 (2021)

  40. [48]

    Phase retrieval using alternating minimization,

    P . Netrapalli, P . Jain, and S. Sanghavi, “Phase retrieval using alternating minimization,” Adv. Neural Inf. Process. Syst. 26 (2013)

  41. [49]

    Coherent inverse scattering via transmission matrices: E!cient phase retrieval algorithms and a public dataset,

    C. A. Metzler, M. K. Sharma, S. Nagesh, R. G. Baraniuk, O. Cossairt, and A. Veeraraghavan, “Coherent inverse scattering via transmission matrices: E!cient phase retrieval algorithms and a public dataset,” in 2017 IEEE International Conference on Computational Photography (ICCP...

  42. [50]

    Phasepack: A phase retrieval library,

    R. Chandra, T. Goldstein, and C. Studer, “Phasepack: A phase retrieval library,” in2019 13th International conference on Sampling Theory and Applications (SampTA), (IEEE, 2019), pp. 1–5

  43. [51]

    Complex-valued neural networks: A comprehensive survey,

    C. Lee, H. Hasegawa, and S. Gao, “Complex-valued neural networks: A comprehensive survey,” IEEE/CAA J. Autom. Sinica 9, 1406–1426 (2022)

  44. [52]

    Phase retrieval and image reconstruction for astronomy,

    J. Dainty and C. Fienup, “Phase retrieval and image reconstruction for astronomy,” Image recovery: theory application 231, 275 (1987)

  45. [53]

    Imaging correlography with sparse arrays of detectors,

    J. R. Fienup and P . S. Idell, “Imaging correlography with sparse arrays of detectors,” Opt. Eng. 27, 778–0 (1988)

  46. [54]

    Non-invasive single-shot imaging through scattering layers and around corners via speckle correlations,

    O. Katz, P . Heidmann, M. Fink, and S. Gigan, “Non-invasive single-shot imaging through scattering layers and around corners via speckle correlations,” Nat. photonics 8, 784–790 (2014)

  47. [55]

    True color kodak images,

    R. W . Franzen, “True color kodak images,” http://r0k.us/graphics/kodak (2013)

  48. [56]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P . Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Process. 13, 600–612 (2004)

  49. [57]

    MatConvNet: Convolutional neural networks for MATLAB,

    A. Vedaldi and K. Lenc, “MatConvNet: Convolutional neural networks for MATLAB,” in Proceedings of the 23rd ACM International Conference on Multimedia, (ACM, 2015), pp. 689–692

  50. [58]

    Using deep neural networks for inverse problems in imaging: Beyond analytical methods,

    A. Lucas, M. Iliadis, R. Molina, and A. K. Katsaggelos, “Using deep neural networks for inverse problems in imaging: Beyond analytical methods,” IEEE Signal Process. Mag. 35, 20–36 (2018)

  51. [59]

    Phase imaging with an untrained neural network,

    F. Wang, Y . Bian, H. Wang, M. Lyu, G. Pedrini, W. Osten, G. Barbastathis, and G. Situ, “Phase imaging with an untrained neural network,” Light. Sci. & Appl. 9, 77 (2020)

  52. [60]

    Deep phase decoder: self-calibrating phase microscopy with an untrained deep neural network,

    E. Bostan, R. Heckel, M. Chen, M. Kellman, and L. Waller, “Deep phase decoder: self-calibrating phase microscopy with an untrained deep neural network,” Optica 7, 559–562 (2020)

  53. [61]

    DNF: di"ractive neural field for lensless microscopic imaging,

    H. Zhu, Z. Liu, Y . Zhou, Z. Ma, and X. Cao, “DNF: di"ractive neural field for lensless microscopic imaging,” Opt. express 30, 18168–18178 (2022)

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.