pith. sign in

arxiv: 2602.20328 · v2 · submitted 2026-02-23 · 💻 cs.CV · eess.IV· math.OC

GSNR: Graph Smooth Null-Space Representation for Inverse Problems

Romario Gualdr\'on-Hurtado , Roman Jacome , Rafael S. Suarez , Henry Arguello This is my paper

Pith reviewed 2026-05-15 20:12 UTC · model grok-4.3

classification 💻 cs.CV eess.IVmath.OC
keywords inverse problemsnull spacegraph signal processingimage reconstructionplug-and-playdiffusion modelsspectral graph theory
0
0 comments X

The pith

A graph smoothness prior applied only to the null-space component improves reconstructions in ill-posed imaging problems.

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

Inverse problems in imaging admit infinitely many solutions because the sensing matrix has a non-trivial null space. Standard image priors act on the entire solution and therefore can distort the invisible component. GSNR instead builds a null-restricted graph Laplacian and projects the reconstruction onto its p smoothest eigenvectors, imposing structure exclusively where the measurements provide no constraint. The resulting low-dimensional regularizer is inserted into existing solvers such as plug-and-play, deep image prior, and diffusion models. Experiments on deblurring, compressed sensing, demosaicing, and super-resolution show consistent PSNR gains of up to 4.3 dB over the same solvers without the null-space term.

Core claim

GSNR constructs a null-restricted Laplacian from a chosen graph and extracts its p smoothest eigenmodes to form a projection matrix that regularizes only the null-space component of the reconstruction, thereby imposing smoothness where the sensing matrix provides no information.

What carries the argument

The null-restricted Laplacian, whose lowest-frequency eigenvectors supply a low-dimensional basis for the invisible part of the solution.

If this is right

  • A null-only regularizer accelerates convergence of iterative solvers by avoiding unnecessary penalties on the visible component.
  • A modest number p of spectral modes captures most of the variance in the null-space signal.
  • These modes remain predictable from the measurements alone, enabling the regularizer to be estimated without additional data.
  • The same mechanism yields measurable PSNR improvements when inserted into plug-and-play, deep image prior, and diffusion pipelines across four standard tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Decoupling visible and invisible constraints may extend the approach to other linear inverse problems outside imaging.
  • Learning the graph from data rather than fixing it by hand could adapt the method to domain-specific null-space structures.
  • The spectral restriction could be tested on mildly non-linear forward models to check whether the null-space basis remains effective.

Load-bearing premise

The chosen graph Laplacian accurately encodes pixel similarities that hold inside the null-space component and those smoothest modes can be reliably inferred from the measurements.

What would settle it

Replacing the projection matrix with a random subspace of equal dimension while keeping the same graph would eliminate the reported PSNR gains if the null-space smoothness claim is correct.

Figures

Figures reproduced from arXiv: 2602.20328 by Henry Arguello, Rafael S. Suarez, Roman Jacome, Romario Gualdr\'on-Hurtado.

Figure 1
Figure 1. Figure 1: For image SR task with SRF = 4, n = 3 · 1282 , we show ground-truth, adjoint reconstruction, NS projection, projec￾tion onto graph-smooth NS with L4nn and L8nn. The map Pnx ∗ isolates signal content invisible to the sen￾sor, while PnLPnx ∗ highlights where Lxn falls into the NS. These graph projections highlight the smoothest blind signal components, e.g., textures near edges. Then, by applying eigenvalue … view at source ↗
Figure 2
Figure 2. Figure 2: Coverage and spectral analysis for SR task. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-mode predictability for each case when [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: a) Coverage and b) Predictability vs. p for CS with CIFAR-10 dataset. In this case, m/n = 0.1 and p = 1 · · · n − m. 4.3. Graph-based regularizer Consider the PnP–PGD iteration solving the optimization problem (8) without the term ∥G∗ (y) − Sx∥ (for more de￾tailed analysis of this term, see [17, Theorem 1]), \vspace {-0.3cm} \mathbf {x}_{k+1}& = \mathrm {D}_\sigma \!\Big (\mathbf {x}_k-\alpha \big (\nabla … view at source ↗
Figure 5
Figure 5. Figure 5: Results of DM-based solvers (DPS [8] & DiffPIR [55]) for Baseline, NPN [17], and GSNR with L4nn and L8nn. Here, p = 0.1n [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results of PnP, NPN, and GSNR varying L and denois￾ers. Best results for each denoiser are in bold. Here, p = 0.1n. of GSNR is the requirement to perform EVD on the null￾restricted Laplacian T, which, at large image scales, neces￾sitates substantial computational resources. However, this computation is performed offline once per (H,L, n, p), and reused for learning the network G and the reconstruction step… view at source ↗
Figure 8
Figure 8. Figure 8: SR with different graph Laplacians. Left: ground truth [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (a) Variation of T normalized eigenvalues for Lθ with respect to their index in CS. (b) Coverage of S with Lθ in CS. lifts the Laplacian to multi–channel form with a Kronecker product \protect \mathbf {L}\leftarrow \mathbf {I}_C\otimes \mathbf {L} . For the numerical results of this work, we empirically set the dimension p. However, our framework provides a principled evaluation on how to select p based on… view at source ↗
Figure 10
Figure 10. Figure 10: PnP variant with null-only projector regularizer: illus [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Effect of the null-only graph regularizer on GSNR [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Effect of the null-only graph regularizer on GSNR-PnP [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
Figure 16
Figure 16. Figure 16: Spectral coverage curves C(p) for super-resolution, comparing different Laplacian choices in the GSNR construction: grid L4nn, grid L8nn, random-walk normalized Lrw, symmetric normalized Lsym, and the geometry-free baseline L = I. Cov￾erage C(p) is the fraction of null-space variance captured by the first p graph-smooth modes. A7. Coverage curve [PITH_FULL_IMAGE:figures/full_fig_p018_16.png] view at source ↗
Figure 15
Figure 15. Figure 15: Fixed-point convergence. A6.4. Minimax optimality bound To experimentally validate the theory of Theorem 2, two different operators S were tested with p = 0.1n and τ = 1 for [PITH_FULL_IMAGE:figures/full_fig_p018_15.png] view at source ↗
Figure 17
Figure 17. Figure 17: Results using latent-space diffusion models for SR. [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Convergence comparison with PnP, NPN, and GSNR [PITH_FULL_IMAGE:figures/full_fig_p020_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Evaluation of GSNR-PnP on a non-optical imaging ex [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Super-resolution results with GSNR-PnP for different graph Laplacians and denoisers. Left: PSNR versus iteration for GSNR [PITH_FULL_IMAGE:figures/full_fig_p022_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Super-resolution on CelebA (20 images) with Deep [PITH_FULL_IMAGE:figures/full_fig_p022_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Deblurring visual results comparing PGD-PnP, GSNR with [PITH_FULL_IMAGE:figures/full_fig_p023_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: PSNR with G ablation in Places for the deblurring task. covery still assumes the nominal H. This mismatch yields imperfect estimates of \protect \mathbf {P}_{n} and \protect \mathbf {S}, which in turn degrades the null-space representation used by GSNR. Despite these compounded imperfections, GSNR remains effective, im￾proving performance by approximately 1 dB and converg￾ing in fewer iterations [PITH_FU… view at source ↗
Figure 24
Figure 24. Figure 24: PSNR in deblurring with inexact forward operator. [PITH_FULL_IMAGE:figures/full_fig_p023_24.png] view at source ↗
read the original abstract

Inverse problems in imaging are ill-posed, leading to infinitely many solutions consistent with the measurements due to the non-trivial null-space of the sensing matrix. Common image priors promote solutions on the general image manifold, such as sparsity, smoothness, or score function. However, as these priors do not constrain the null-space component, they can bias the reconstruction. Thus, we aim to incorporate meaningful null-space information in the reconstruction framework. Inspired by smooth image representation on graphs, we propose Graph-Smooth Null-Space Representation (GSNR), a mechanism that imposes structure only into the invisible component. Particularly, given a graph Laplacian, we construct a null-restricted Laplacian that encodes similarity between neighboring pixels in the null-space signal, and we design a low-dimensional projection matrix from the $p$-smoothest spectral graph modes (lowest graph frequencies). This approach has strong theoretical and practical implications: i) improved convergence via a null-only graph regularizer, ii) better coverage, how much null-space variance is captured by $p$ modes, and iii) high predictability, how well these modes can be inferred from the measurements. GSNR is incorporated into well-known inverse problem solvers, e.g., PnP, DIP, and diffusion solvers, in four scenarios: image deblurring, compressed sensing, demosaicing, and image super-resolution, providing consistent improvement of up to 4.3 dB over baseline formulations and up to 1 dB compared with end-to-end learned models in terms of PSNR.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes Graph-Smooth Null-Space Representation (GSNR) to address the null-space component in ill-posed inverse imaging problems. Given a graph Laplacian, it constructs a null-restricted Laplacian and projects onto its p lowest-eigenvalue (smoothest) modes to impose structure only on the invisible subspace. GSNR is plugged into PnP, DIP, and diffusion solvers for deblurring, compressed sensing, demosaicing, and super-resolution, reporting PSNR gains of up to 4.3 dB over baselines and 1 dB over end-to-end learned models.

Significance. If the null-space-specific regularization can be shown to be stably estimable from measurements, the approach would offer a principled way to avoid biasing reconstructions toward the visible range while still leveraging graph smoothness. The reported consistent gains across four tasks and multiple solvers indicate potential practical value, but the absence of derivations for the claimed convergence/coverage/predictability properties and lack of error bounds on mode recovery limit the current significance.

major comments (3)
  1. [Abstract] Abstract: the three theoretical implications (improved convergence via null-only regularizer, better coverage of null-space variance by p modes, high predictability from measurements) are stated without any derivation, quantitative bound, or error analysis linking measurement consistency to mode-recovery error; this is load-bearing because the PSNR gains are attributed specifically to null-space regularization.
  2. [Abstract] The construction of the null-restricted Laplacian and the projection matrix onto the p-smoothest modes (lowest graph frequencies) treats p and the underlying graph as free parameters; no analysis shows that the recovered modes align with the true null component of the ground-truth image or that the projection is stably estimable from y = Ax.
  3. [Experiments (implied by results)] The reported PSNR improvements (up to 4.3 dB) are given without confidence intervals, ablation on graph-construction choices, or sensitivity analysis on p, leaving open whether the gains are robust or arise from post-hoc parameter selection.
minor comments (1)
  1. [Abstract] Notation for the null-restricted Laplacian and the projection matrix P_p should be introduced with explicit definitions and an equation showing how the low-dimensional representation is formed.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below, clarifying the existing theoretical and experimental support while committing to targeted revisions that strengthen the claims without altering the core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the three theoretical implications (improved convergence via null-only regularizer, better coverage of null-space variance by p modes, high predictability from measurements) are stated without any derivation, quantitative bound, or error analysis linking measurement consistency to mode-recovery error; this is load-bearing because the PSNR gains are attributed specifically to null-space regularization.

    Authors: We agree the abstract is concise and does not preview the supporting derivations. Section 3.1 derives improved convergence by showing that the null-only regularizer leaves the range-space solution unchanged while contracting the null-space component. Coverage is quantified in Section 3.2 via the cumulative variance captured by the p lowest-eigenvalue modes of the null-restricted Laplacian. Predictability is supported empirically in Section 4.2 by measuring the error between coefficients recovered from y and those of the ground-truth null component. We acknowledge the absence of a formal error bound; the revision will add a proposition providing a quantitative bound on mode-recovery error in terms of measurement noise and the conditioning of A. revision: partial

  2. Referee: [Abstract] The construction of the null-restricted Laplacian and the projection matrix onto the p-smoothest modes (lowest graph frequencies) treats p and the underlying graph as free parameters; no analysis shows that the recovered modes align with the true null component of the ground-truth image or that the projection is stably estimable from y = Ax.

    Authors: The null-restricted Laplacian is defined in Equation (5) as L_null = P_null L P_null, where P_null projects onto the null-space of A, and the projection matrix V_p comprises the p eigenvectors of L_null with smallest eigenvalues. Alignment with the true null component is demonstrated in Section 4.1 through correlation metrics and visualizations on held-out images. Direct estimation from y is impossible because the null-space is invisible, so we select p and the graph via a measurement-consistency criterion. The revision will include a stability proposition showing that small perturbations in the estimated null-space induce bounded changes in the selected modes, together with additional alignment statistics across datasets. revision: partial

  3. Referee: [Experiments (implied by results)] The reported PSNR improvements (up to 4.3 dB) are given without confidence intervals, ablation on graph-construction choices, or sensitivity analysis on p, leaving open whether the gains are robust or arise from post-hoc parameter selection.

    Authors: We accept this criticism. The revised experimental section will report 95% confidence intervals computed over five independent runs per method and task. We will add ablations comparing three graph constructions (8-connected grid, k-NN with k=8, and adaptive k-NN) and sensitivity curves for p ranging from 5 to 100. These plots will show that PSNR gains remain above 2 dB for p in [15,60] across all four tasks, confirming robustness rather than post-hoc tuning. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; parameter choices introduce mild dependency

full rationale

The GSNR construction starts from an externally supplied graph Laplacian and deterministically extracts the p lowest-eigenvalue modes to form the projection matrix P_p. No equation in the abstract reduces a claimed prediction or null-space coverage result back to the same fitted quantities by construction. The reported PSNR gains are measured on standard benchmarks against baselines, providing external falsifiability. The only potential weakness is that selection of the graph and integer p is not shown to be independent of test data, which is a standard hyper-parameter issue rather than a definitional loop. This warrants a low score of 2 rather than zero.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The method rests on the standard properties of the graph Laplacian and the assumption that smoothness on the graph transfers to the null-space signal. No new physical entities are postulated.

free parameters (2)
  • p
    Number of smoothest spectral modes retained in the low-dimensional projection; chosen to balance coverage and predictability.
  • graph construction parameters
    Edge weights and connectivity rule for the pixel graph; not specified in the abstract.
axioms (2)
  • domain assumption The graph Laplacian encodes meaningful similarity between neighboring pixels that remains valid inside the null-space component.
    Invoked when constructing the null-restricted Laplacian.
  • domain assumption The lowest-frequency eigenvectors of the null-restricted Laplacian form a useful basis for the invisible signal.
    Basis for the low-dimensional projection matrix.

pith-pipeline@v0.9.0 · 5590 in / 1575 out tokens · 18059 ms · 2026-05-15T20:12:49.855427+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. DIPA: Distilled Preconditioned Algorithms for Solving Imaging Inverse Problems

    eess.IV 2026-05 unverdicted novelty 7.0

    DIPA learns preconditioning operators via distillation from a teacher with a better sensing matrix to improve reconstruction quality for the student's physically constrained matrix in imaging inverse problems.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ntire 2017 challenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. InThe IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR) Workshops, 2017. 11

    work page 2017

  2. [2]

    Computed tomography reconstruction using deep image prior and learned reconstruction methods.Inverse Problems, 36(9):094004, 2020

    Daniel Otero Baguer, Johannes Leuschner, and Maximilian Schmidt. Computed tomography reconstruction using deep image prior and learned reconstruction methods.Inverse Problems, 36(9):094004, 2020. 1

    work page 2020

  3. [3]

    Color imaging array, 1976

    Bryce E Bayer. Color imaging array, 1976. US Patent 3,971,065. 6

    work page 1976

  4. [4]

    Modern regularization methods for inverse problems.Acta numerica, 27:1–111,

    Martin Benning and Martin Burger. Modern regularization methods for inverse problems.Acta numerica, 27:1–111,

    work page

  5. [5]

    E. J. Candes and M. B. Wakin. An introduction to com- pressive sampling.IEEE Signal Processing Magazine, 25 (2):21–30, 2008. 1, 6

    work page 2008

  6. [6]

    Plug- and-play admm for image restoration: Fixed-point conver- gence and applications.IEEE Transactions on Computa- tional Imaging, 3(1):84–98, 2016

    Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug- and-play admm for image restoration: Fixed-point conver- gence and applications.IEEE Transactions on Computa- tional Imaging, 3(1):84–98, 2016. 1, 7

    work page 2016

  7. [7]

    Deep decomposition learning for inverse imaging problems

    Dongdong Chen and Mike E Davies. Deep decomposition learning for inverse imaging problems. InComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, Au- gust 23–28, 2020, Proceedings, Part XXVIII 16, pages 510–

    work page 2020

  8. [8]

    Springer, 2020. 2, 7

    work page 2020

  9. [9]

    Diffusion pos- terior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mc- cann, Marc Louis Klasky, and Jong Chul Ye. Diffusion pos- terior sampling for general noisy inverse problems. InThe Eleventh International Conference on Learning Representa- tions, 2023. 6, 7, 1

    work page 2023

  10. [10]

    Regu- larization by denoising via fixed-point projection (red-pro)

    Regev Cohen, Michael Elad, and Peyman Milanfar. Regu- larization by denoising via fixed-point projection (red-pro). SIAM Journal on Imaging Sciences, 14(3):1374–1406, 2021. 1

    work page 2021

  11. [11]

    AdaIR: Adap- tive all-in-one image restoration via frequency mining and modulation

    Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. AdaIR: Adap- tive all-in-one image restoration via frequency mining and modulation. InThe Thirteenth International Conference on Learning Representations, 2025. 11

    work page 2025

  12. [12]

    Sar image despeckling by deep neural networks: from a pre-trained model to an end-to-end training strategy.Remote Sensing, 12(16), 2020

    Emanuele Dalsasso, Xiangli Yang, Lo ¨ıc Denis, Florence Tupin, and Wen Yang. Sar image despeckling by deep neural networks: from a pre-trained model to an end-to-end training strategy.Remote Sensing, 12(16), 2020. 6, 7

    work page 2020

  13. [13]

    D.L. Donoho. De-noising by soft-thresholding.IEEE Trans- actions on Information Theory, 41(3):613–627, 1995. 7

    work page 1995

  14. [14]

    Ideal spatial adapta- tion by wavelet shrinkage.Biometrika, 81(3):425–455, 1994

    David L Donoho and Iain M Johnstone. Ideal spatial adapta- tion by wavelet shrinkage.Biometrika, 81(3):425–455, 1994. 2

    work page 1994

  15. [15]

    Pseudo- invertible neural networks, 2026

    Yamit Ehrlich, Nimrod Berman, and Assaf Shocher. Pseudo- invertible neural networks, 2026. 8

    work page 2026

  16. [16]

    Manifold preserving guided diffusion

    Yutong He, Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Dongjun Kim, Wei-Hsiang Liao, Yuki Mitsufuji, J Zico Kolter, Ruslan Salakhutdinov, and Ste- fano Ermon. Manifold preserving guided diffusion. InThe Twelfth International Conference on Learning Representa- tions, 2024. 7, 1, 8

    work page 2024

  17. [17]

    A relaxed proximal gradient descent algorithm for convergent plug-and-play with proximal de- noiser

    Samuel Hurault, Antonin Chambolle, Arthur Leclaire, and Nicolas Papadakis. A relaxed proximal gradient descent algorithm for convergent plug-and-play with proximal de- noiser. InInternational Conference on Scale Space and Variational Methods in Computer Vision, pages 379–392. Springer, 2023. 7

    work page 2023

  18. [18]

    NPN: Non-Linear Projec- tions of the Null-Space for Imaging Inverse Problems

    Roman Jacome, Romario Gualdr ´on-Hurtado, Le ´on Su ´arez- Rodr´ıguez, and Henry Arguello. NPN: Non-Linear Projec- tions of the Null-Space for Imaging Inverse Problems. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 2, 6, 7, 8, 9, 11

    work page 2025

  19. [19]

    Ian T Jolliffe and Jorge Cadima. Principal component analy- sis: a review and recent developments.Philosophical trans- actions of the royal society A: Mathematical, Physical and Engineering Sciences, 374(2065):20150202, 2016. 4

    work page 2065

  20. [20]

    How to learn a graph from smooth sig- nals

    Vassilis Kalofolias. How to learn a graph from smooth sig- nals. InProceedings of the 19th International Conference on Artificial Intelligence and Statistics, pages 920–929, Cadiz, Spain, 2016. PMLR. 2

    work page 2016

  21. [21]

    Kamilov, Charles A

    Ulugbek S. Kamilov, Charles A. Bouman, Gregery T. Buz- zard, and Brendt Wohlberg. Plug-and-play methods for inte- grating physical and learned models in computational imag- ing: Theory, algorithms, and applications.IEEE Signal Pro- cessing Magazine, 40(1):85–97, 2023. 1, 2, 7

    work page 2023

  22. [22]

    Adam: a method for stochastic opti- mization, 2014

    DP Kingma and J Ba. Adam: a method for stochastic opti- mization, 2014. 6

    work page 2014

  23. [23]

    Efficient visual state space model for image deblurring

    Lingshun Kong, Jiangxin Dong, Jinhui Tang, Ming-Hsuan Yang, and Jinshan Pan. Efficient visual state space model for image deblurring. InProceedings of the computer vision and pattern recognition conference, pages 12710–12719, 2025. 11

    work page 2025

  24. [24]

    Krizhevsky

    A. Krizhevsky. Learning multiple layers of features from tiny images.Master’s thesis, University of Tront, page 1, 2009. 6

    work page 2009

  25. [25]

    ARPACK users’ guide: solution of large-scale eigenvalue problems with implicitly restarted Arnoldi methods

    Richard B Lehoucq, Danny C Sorensen, and Chao Yang. ARPACK users’ guide: solution of large-scale eigenvalue problems with implicitly restarted Arnoldi methods. SIAM,

    work page

  26. [26]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. InProceedings of International Conference on Computer Vision (ICCV), 2015. 6, 7

    work page 2015

  27. [27]

    Compressed sensing mri.IEEE signal pro- cessing magazine, 25(2):72–82, 2008

    Michael Lustig, David L Donoho, Juan M Santos, and John M Pauly. Compressed sensing mri.IEEE signal pro- cessing magazine, 25(2):72–82, 2008. 1

    work page 2008

  28. [28]

    Denoising: a powerful building block for imaging, inverse problems and machine learning.Philosophical Transactions A, 383(2299): 20240326, 2025

    Peyman Milanfar and Mauricio Delbracio. Denoising: a powerful building block for imaging, inverse problems and machine learning.Philosophical Transactions A, 383(2299): 20240326, 2025. 1

    work page 2025

  29. [29]

    Graph-wavelet filterbanks for edge-aware image processing

    Sunil K Narang, Yung Hsuan Chao, and Antonio Ortega. Graph-wavelet filterbanks for edge-aware image processing. In2012 IEEE Statistical Signal Processing Workshop (SSP), pages 141–144, 2012. 2

    work page 2012

  30. [30]

    Graph signal processing: Overview, challenges, and applications.Proceedings of the IEEE, 106(5):808–828, 2018

    Antonio Ortega, Pascal Frossard, Jelena Kovaˇcevi´c, Jos´e MF Moura, and Pierre Vandergheynst. Graph signal processing: Overview, challenges, and applications.Proceedings of the IEEE, 106(5):808–828, 2018. 2, 3, 4

    work page 2018

  31. [31]

    Proximal algorithms.Foun- dations and Trends in Optimization, 1(3):127–239, 2014

    Neal Parikh and Stephen Boyd. Proximal algorithms.Foun- dations and Trends in Optimization, 1(3):127–239, 2014. 2

    work page 2014

  32. [32]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Rai- son, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-per...

    work page 2019

  33. [33]

    Non-local regularization of inverse problems

    Gabriel Peyr ´e, S ´ebastien Bougleux, and Laurent Cohen. Non-local regularization of inverse problems. InComputer Vision – ECCV 2008, pages 57–68, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2

    work page 2008

  34. [34]

    Plug and play augmented hqs: Convergence analysis and its application in mri reconstruction.Neuro- computing, 518:1–14, 2023

    Arash Rasti-Meymandi, Aboozar Ghaffari, and Emad Fatemizadeh. Plug and play augmented hqs: Convergence analysis and its application in mri reconstruction.Neuro- computing, 518:1–14, 2023. 7

    work page 2023

  35. [35]

    Regularization by denoising: Clarifications and new interpretations.IEEE transactions on computational imaging, 5(1):52–67, 2018

    Edward T Reehorst and Philip Schniter. Regularization by denoising: Clarifications and new interpretations.IEEE transactions on computational imaging, 5(1):52–67, 2018. 1

    work page 2018

  36. [36]

    The little engine that could: Regularization by denoising (red)

    Yaniv Romano, Michael Elad, and Peyman Milanfar. The little engine that could: Regularization by denoising (red). SIAM journal on imaging sciences, 10(4):1804–1844, 2017. 1, 2

    work page 2017

  37. [37]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 8

    work page 2022

  38. [38]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 6

    work page 2015

  39. [39]

    Rudin, Stanley Osher, and Emad Fatemi

    Leonid I. Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms.Physica D: Nonlinear Phenomena, 60(1):259–268, 1992. 2

    work page 1992

  40. [40]

    Chapman and Hall/CRC,

    Havard Rue and Leonhard Held.Gaussian Markov random fields: theory and applications. Chapman and Hall/CRC,

    work page

  41. [41]

    Plug-and-play methods provably converge with properly trained denoisers

    Ernest Ryu, Jialin Liu, Sicheng Wang, Xiaohan Chen, Zhangyang Wang, and Wotao Yin. Plug-and-play methods provably converge with properly trained denoisers. InIn- ternational Conference on Machine Learning, pages 5546–

    work page

  42. [42]

    Deep null space learning for inverse problems: convergence analysis and rates.Inverse Problems, 35(2): 025008, 2019

    Johannes Schwab, Stephan Antholzer, and Markus Halt- meier. Deep null space learning for inverse problems: convergence analysis and rates.Inverse Problems, 35(2): 025008, 2019. 2, 7

    work page 2019

  43. [43]

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of sig- nal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains.IEEE sig- nal processing magazine, 30(3):83–98, 2013. 3, 4

    work page 2013

  44. [44]

    Liaudat, Maxime Song, Johannes Hertrich, Se- bastian Neumayer, and Georg Schramm

    Juli ´an Tachella, Matthieu Terris, Samuel Hurault, Andrew Wang, Leo Davy, J ´er´emy Scanvic, Victor Sechaud, Ro- main V o, Thomas Moreau, Thomas Davies, Dongdong Chen, Nils Laurent, Brayan Monroy, Jonathan Dong, Zhiyuan Hu, Minh-Hai Nguyen, Florian Sarron, Pierre Weiss, Paul Es- cande, Mathurin Massias, Thibaut Modrzyk, Brett Levac, Tob´ıas I. Liaudat, M...

    work page 2025

  45. [45]

    Solv- ing imaging inverse problems using plug-and-play denoisers: Regularization and optimization perspectives, 2025

    Hong Ye Tan, Subhadip Mukherjee, and Junqi Tang. Solv- ing imaging inverse problems using plug-and-play denoisers: Regularization and optimization perspectives, 2025. 1

    work page 2025

  46. [46]

    Equivariant plug-and-play image reconstruc- tion

    Matthieu Terris, Thomas Moreau, Nelly Pustelnik, and Ju- lian Tachella. Equivariant plug-and-play image reconstruc- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 25255–25264,

    work page

  47. [47]

    A survey on super-resolution imaging.Signal, Image and Video Processing, 5:329–342,

    Jing Tian and Kai-Kuang Ma. A survey on super-resolution imaging.Signal, Image and Video Processing, 5:329–342,

    work page

  48. [48]

    Degradation-aware feature perturbation for all-in- one image restoration

    Xiangpeng Tian, Xiangyu Liao, Xiao Liu, Meng Li, and Chao Ren. Degradation-aware feature perturbation for all-in- one image restoration. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 28165–28175, 2025. 11

    work page 2025

  49. [49]

    Venkatakrishnan, Charles A

    Singanallur V . Venkatakrishnan, Charles A. Bouman, and Brendt Wohlberg. Plug-and-play priors for model based re- construction. In2013 IEEE Global Conference on Signal and Information Processing, pages 945–948, 2013. 1

    work page 2013

  50. [50]

    Hyperspectral image reconstruction using a deep spatial-spectral prior

    Lizhi Wang, Chen Sun, Ying Fu, Min H Kim, and Hua Huang. Hyperspectral image reconstruction using a deep spatial-spectral prior. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 8032–8041, 2019. 8

    work page 2019

  51. [51]

    Fista-net: Learning a fast iterative shrinkage thresholding network for inverse problems in imaging.IEEE Transactions on Medical Imaging, 40(5):1329–1339, 2021

    Jinxi Xiang, Yonggui Dong, and Yunjie Yang. Fista-net: Learning a fast iterative shrinkage thresholding network for inverse problems in imaging.IEEE Transactions on Medical Imaging, 40(5):1329–1339, 2021. 7

    work page 2021

  52. [52]

    Generative image inpainting with con- textual attention

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with con- textual attention. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5505–5514,

    work page

  53. [53]

    Plug-and-play image restora- tion with deep denoiser prior.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 44(10):6360–6376,

    Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restora- tion with deep denoiser prior.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 44(10):6360–6376,

    work page

  54. [54]

    Deep image deblurring: A survey.International Journal of Com- puter Vision, 130(9):2103–2130, 2022

    Kaihao Zhang, Wenqi Ren, Wenhan Luo, Wei-Sheng Lai, Bj¨orn Stenger, Ming-Hsuan Yang, and Hongdong Li. Deep image deblurring: A survey.International Journal of Com- puter Vision, 130(9):2103–2130, 2022. 1

    work page 2022

  55. [55]

    Places: A 10 million image database for scene recognition.IEEE Transactions on Pattern Analy- sis and Machine Intelligence, 2017

    Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition.IEEE Transactions on Pattern Analy- sis and Machine Intelligence, 2017. 6

    work page 2017

  56. [56]

    Denoising dif- fusion models for plug-and-play image restoration

    Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bi- han Wen, Radu Timofte, and Luc Van Gool. Denoising dif- fusion models for plug-and-play image restoration. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1219– 1229, 2023. 6, 7, 1, 8 GSNR: Graph Smooth Null-Space Representation for Inver...

    work page 2023

  57. [57]

    Consider measurements from(1)and denote Cy :=HC rrH⊤ +σ 2Im

    Define thej-th null coefficienta j :=v ⊤ j xn where xn :=P nx. Consider measurements from(1)and denote Cy :=HC rrH⊤ +σ 2Im. Then the populationR 2 of the optimal linear predictor ofa j fromysatisfies ρ2 j := Cov(aj,y) ⊤ C−1 y Cov(y, aj) Var(aj) ≤ cj cj +µ j , cj :=v ⊤ j QrnCrrQnr vj. Equality holds in the ideal caseH=Iandσ 2 = 0. SinceQ≻0andQ nn ≻0, the b...

    work page

  58. [58]

    Algorithm 3 shows the integration of GSNR in DPS. A8.2. DiffPIR [55]. σn >0denotes the standard deviation of the measure- ment noise, andη >0is the data–proximal penalty that trades off data fidelity and the denoiser prior inside the sub- problem;ρ i ≜η σ 2 n/˜σ2 i is the iteration-dependent weight used in the proximal objective at stepi; ˜x(i) 0 is the s...

    work page