pith. machine review for the scientific record. sign in

arxiv: 2510.01608 · v2 · submitted 2025-10-02 · 💻 cs.CV · eess.SP· math.OC

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

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

Pith reviewed 2026-05-18 11:09 UTC · model grok-4.3

classification 💻 cs.CV eess.SPmath.OC
keywords null-space regularizationimaging inverse problemsplug-and-play methodsneural network priorscompressive sensingcomputed tomographymagnetic resonance imagingimage reconstruction
0
0 comments X

The pith

NPN uses a neural network to project solutions into a low-dimensional slice of the sensing matrix null-space, tightening reconstructions for ill-posed imaging problems.

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

The paper introduces Non-Linear Projections of the Null-Space (NPN) as a regularization strategy that operates on the null-space of the sensing operator rather than directly on image structure. By training a network to map solutions into a low-dimensional projection of that null-space, NPN supplies task-specific priors that capture information the sensor is fundamentally unable to see. The method is shown to be compatible with plug-and-play algorithms, unrolled networks, deep image priors, and diffusion models, and it carries convergence guarantees for the plug-and-play case. Experiments across compressive sensing, deblurring, super-resolution, CT, and MRI report consistent gains in reconstruction quality when NPN is added to existing frameworks.

Core claim

NPN promotes solutions that lie in a low-dimensional projection of the sensing matrix's null-space with a neural network, offering interpretability by capturing information orthogonal to the components that the sensing process cannot observe and flexibility as a modular complement to conventional image-domain priors.

What carries the argument

Non-Linear Projection of the Null-Space (NPN): a neural network that learns to enforce membership in a low-dimensional subspace drawn from the null-space of the forward operator.

If this is right

  • When inserted into plug-and-play iterations, NPN supplies convergence guarantees and higher reconstruction accuracy than image-domain priors alone.
  • NPN remains effective when paired with unrolling networks, deep image prior, and diffusion-based solvers.
  • The same null-space projection can be reused across compressive sensing, deblurring, super-resolution, CT, and MRI without redesign.
  • Because NPN acts orthogonally to image-domain constraints, it can be stacked with existing learned priors to tighten the solution set further.

Where Pith is reading between the lines

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

  • If the learned null-space projection generalizes, the same network might transfer to new forward operators whose null-spaces share similar low-dimensional structure.
  • Explicit control over projection dimension could be turned into an adaptive mechanism that shrinks or expands the subspace according to measured noise level.
  • Visualizing the directions the network keeps or discards might reveal which high-frequency patterns are lost for a given sensor geometry.
  • Extending NPN to nonlinear forward models would require replacing the linear null-space notion with a local tangent-space projection at each iterate.

Load-bearing premise

A neural network can learn a stable, low-dimensional projection of the null-space that remains useful and non-overfitting across different sensing matrices without requiring per-matrix retuning of dimension or capacity.

What would settle it

Reconstruction error or perceptual quality metrics show no statistically significant improvement, or even degrade, when NPN is inserted into standard plug-and-play or unrolled pipelines on multiple distinct sensing matrices.

Figures

Figures reproduced from arXiv: 2510.01608 by Henry Arguello, Leon Suarez, Roman Jacome, Romario Gualdr\'on-Hurtado.

Figure 1
Figure 1. Figure 1: Geometric comparison of subspace–prior learning versus direct reconstruction in a R 3 toy example. (a) In the low–dimensional projection space, the learned mapping G ∗ (y) trained on points inside the unit circle, closely matches the true null–space projection Sx∗ for both training (solid) and test (semi-transparent) inputs, whereas the direct–reconstruction estimate x˜0 projected into S is significantly i… view at source ↗
Figure 2
Figure 2. Figure 2: PnP-FISTA convergence analysis in CS. (a) Reconstruction error. (b) Null-space prediction error for (red) Initialization S˜ = QR(H) from Algorithm 1, and (blue) Designed S with Eq. (3) and m/n = p/n = 0.1. In this case, the CIZ from Definition 3 is highlighted in light red and light blue. (c) Acceleration ratio of signal convergence; here, the CIZ is defined as the empirical convergence ratio of the propos… view at source ↗
Figure 4
Figure 4. Figure 4: Deblurring and MRI reconstruction results for PnP and PnP-NPN using a DnCNN prior, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance of DIP and NPN-DIP for different [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of γ in NPN on PSNR (dB) for MRI reconstruction, with α = 1 × 10−4 . The maximum PSNR of 33.67 dB is achieved when γ = 6 × 10−3 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effect of γ on PSNR (dB) in SPC reconstruction, with α = 8 × 10−4 . The maximum PSNR of 21.17 dB is achieved when γ = 1.04. Method Restormer DnCNN DnCNN-Lipschitz DRUNet Sparsity Prior Baseline 29.86 29.55 30.36 29.68 28.75 NPN 32.62 32.12 32.35 32.07 29.75 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of the low-dimensional subspace dimension [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MRI masks Sampling Type ∥Sx − G(y)∥ ∥Sx∥ ∥G(y)∥ ∥Sx∥ PSNR(G(y), Sx) [dB] Disjoint Sampling 1.00 0.0197 1.72 30.7 Adjacent Sampling 0.40 98.7 112 59.5 [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
read the original abstract

Imaging inverse problems aim to recover high-dimensional signals from undersampled, noisy measurements, a fundamentally ill-posed task with infinite solutions in the null-space of the sensing operator. To resolve this ambiguity, prior information is typically incorporated through handcrafted regularizers or learned models that constrain the solution space. However, these priors typically ignore the task-specific structure of that null-space. In this work, we propose Non-Linear Projections of the Null-Space (NPN), a novel class of regularization that, instead of enforcing structural constraints in the image domain, promotes solutions that lie in a low-dimensional projection of the sensing matrix's null-space with a neural network. Our approach has two key advantages: (1) Interpretability: by focusing on the structure of the null-space, we design sensing-matrix-specific priors that capture information orthogonal to the signal components that are fundamentally blind to the sensing process. (2) Flexibility: NPN is adaptable to various inverse problems, compatible with existing reconstruction frameworks, and complementary to conventional image-domain priors. We provide theoretical guarantees on convergence and reconstruction accuracy when used within plug-and-play methods. Empirical results across diverse sensing matrices demonstrate that NPN priors consistently enhance reconstruction fidelity in various imaging inverse problems, such as compressive sensing, deblurring, super-resolution, computed tomography, and magnetic resonance imaging, with plug-and-play methods, unrolling networks, deep image prior, and diffusion models.

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

2 major / 2 minor

Summary. The paper introduces Non-Linear Projections of the Null-Space (NPN), a regularization technique for imaging inverse problems. Rather than applying priors in the image domain, NPN employs a neural network to promote reconstructed solutions that lie in a low-dimensional projection of the null-space of the sensing matrix. The approach is positioned as interpretable (sensing-matrix-specific) and flexible (plug-and-play compatible with PnP, unrolling networks, deep image prior, and diffusion models). The manuscript claims theoretical guarantees on convergence and reconstruction accuracy for PnP methods, along with consistent empirical improvements across compressive sensing, deblurring, super-resolution, CT, and MRI tasks.

Significance. If the claimed theoretical guarantees hold under standard PnP assumptions and the empirical gains prove robust to ablations and statistical testing, NPN could supply a useful complementary prior that exploits null-space structure orthogonal to conventional image-domain models. The stated compatibility with multiple reconstruction frameworks would be a practical strength for the field.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Theoretical Analysis): The central claim of 'theoretical guarantees on convergence and reconstruction accuracy' when NPN is used inside plug-and-play methods is load-bearing. Standard PnP convergence arguments (fixed-point theorems, ADMM, or Krasnosel'skii-Mann iterations) require the regularizer/denoiser to be non-expansive or firmly non-expansive. The manuscript does not indicate that the learned non-linear neural-network projection operator is constrained (e.g., via spectral normalization, contractive layers, or post-training verification) to satisfy these conditions. Without such verification the cited guarantees do not necessarily transfer.
  2. [§3 and §5] §3 (Method) and §5 (Experiments): The projection dimension is treated as a free hyper-parameter. The manuscript does not report cross-validation procedures, sensitivity analysis, or explicit checks that this dimension was not tuned on the same test sensing matrices used for final evaluation. This directly affects the weakest assumption that the low-dimensional null-space projection remains generalizable and non-overfitting.
minor comments (2)
  1. [§4] Add explicit statements of all assumptions (including operator properties) in the theorem statements of §4 so that readers can immediately see which PnP convergence results are being invoked.
  2. [§5] Include error bars, dataset splits, and at least one ablation on projection dimension in the main experimental tables or figures to allow quantitative assessment of the reported fidelity gains.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. The comments identify important points regarding the theoretical claims and experimental validation that we will address in the revision. Below we respond point by point to the major comments.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Theoretical Analysis): The central claim of 'theoretical guarantees on convergence and reconstruction accuracy' when NPN is used inside plug-and-play methods is load-bearing. Standard PnP convergence arguments (fixed-point theorems, ADMM, or Krasnosel'skii-Mann iterations) require the regularizer/denoiser to be non-expansive or firmly non-expansive. The manuscript does not indicate that the learned non-linear neural-network projection operator is constrained (e.g., via spectral normalization, contractive layers, or post-training verification) to satisfy these conditions. Without such verification the cited guarantees do not necessarily transfer.

    Authors: We appreciate the referee's careful reading of the theoretical section. The guarantees presented in §4 are stated under the standard assumption that the NPN operator satisfies non-expansiveness (or firm non-expansiveness) so that existing PnP fixed-point results apply. However, the current manuscript does not explicitly describe how this property is enforced for the learned neural-network projection (e.g., through architectural constraints or verification). In the revised manuscript we will add a dedicated paragraph in §4 explaining the use of spectral normalization during training together with post-training Lipschitz-constant checks on the learned operator. These additions will make the applicability of the cited convergence theorems explicit and verifiable. revision: yes

  2. Referee: [§3 and §5] §3 (Method) and §5 (Experiments): The projection dimension is treated as a free hyper-parameter. The manuscript does not report cross-validation procedures, sensitivity analysis, or explicit checks that this dimension was not tuned on the same test sensing matrices used for final evaluation. This directly affects the weakest assumption that the low-dimensional null-space projection remains generalizable and non-overfitting.

    Authors: The referee correctly notes that the manuscript lacks a full account of hyper-parameter selection for the projection dimension. In our experiments the dimension was chosen via cross-validation on separate validation sets for each task and sensing matrix, but this procedure and the associated sensitivity results were not reported. We will revise §5 to include (i) a sensitivity plot showing PSNR/SSIM versus projection dimension across representative tasks and (ii) an explicit statement that all tuning was performed exclusively on validation data disjoint from the reported test sensing matrices. These changes will strengthen the claim that the chosen dimension generalizes. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces NPN as a novel regularization approach that uses a neural network to project onto a low-dimensional subspace of the sensing matrix null-space, claiming interpretability, flexibility, and compatibility with existing frameworks like PnP. Theoretical guarantees on convergence and accuracy are asserted for PnP integration, but the provided text shows no reduction of these guarantees to fitted parameters on evaluation data, no self-definitional equations where outputs are inputs by construction, and no load-bearing self-citations that substitute for independent derivation. Empirical results across multiple inverse problems are presented as validation rather than forced by the method definition itself. The central claims remain independent of the inputs under the examined patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the existence of exploitable low-dimensional structure in the null-space of arbitrary sensing matrices and on the ability of a neural network to learn a useful projection operator without introducing circular dependence on the reconstruction target.

free parameters (1)
  • projection dimension
    The dimensionality of the low-dimensional null-space projection must be selected or learned; this choice directly affects the regularization strength.
axioms (1)
  • domain assumption The null-space of the sensing operator contains task-specific structure that is orthogonal to recoverable signal components and can be captured by a neural network.
    Invoked when the abstract states that NPN captures 'information orthogonal to the signal components that are fundamentally blind to the sensing process'.

pith-pipeline@v0.9.0 · 5802 in / 1484 out tokens · 46292 ms · 2026-05-18T11:09:38.323451+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

70 extracted references · 70 canonical work pages · 3 internal anchors

  1. [1]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems

    Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2(1):183–202, 2009

    work page 2009

  2. [2]

    Non-Local Means Denoising

    Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. Non-Local Means Denoising. Image Processing On Line, 1:208–212, 2011

    work page 2011

  3. [3]

    Image denoising: Can plain neural networks compete with bm3d? In 2012 IEEE conference on computer vision and pattern recognition, pages 2392–2399

    Harold C Burger, Christian J Schuler, and Stefan Harmeling. Image denoising: Can plain neural networks compete with bm3d? In 2012 IEEE conference on computer vision and pattern recognition, pages 2392–2399. IEEE, 2012

    work page 2012

  4. [4]

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

    work page 2008

  5. [5]

    Deep physics-guided unrolling generalization for compressed sensing.International Journal of Computer Vision, 131(11):2864– 2887, 2023

    Bin Chen, Jiechong Song, Jingfen Xie, and Jian Zhang. Deep physics-guided unrolling generalization for compressed sensing.International Journal of Computer Vision, 131(11):2864– 2887, 2023

    work page 2023

  6. [6]

    Deep decomposition learning for inverse imaging problems

    Dongdong Chen and Mike E Davies. Deep decomposition learning for inverse imaging problems. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII 16, pages 510–526. Springer, 2020

    work page 2020

  7. [7]

    Equivariant imaging: Learning beyond the range space

    Dongdong Chen, Julián Tachella, and Mike E Davies. Equivariant imaging: Learning beyond the range space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4379–4388, 2021

    work page 2021

  8. [8]

    Learning to optimize: A primer and a benchmark.Journal of Machine Learning Research, 23(189):1–59, 2022

    Tianlong Chen, Xiaohan Chen, Wuyang Chen, Howard Heaton, Jialin Liu, Zhangyang Wang, and Wotao Yin. Learning to optimize: A primer and a benchmark.Journal of Machine Learning Research, 23(189):1–59, 2022

    work page 2022

  9. [9]

    Theoretical linear convergence of unfolded ista and its practical weights and thresholds

    Xiaohan Chen, Jialin Liu, Zhangyang Wang, and Wotao Yin. Theoretical linear convergence of unfolded ista and its practical weights and thresholds. Advances in Neural Information Processing Systems, 31, 2018

    work page 2018

  10. [10]

    Unsupervised range-nullspace learning prior for multispectral images reconstruction

    Yurong Chen, Yaonan Wang, and Hui Zhang. Unsupervised range-nullspace learning prior for multispectral images reconstruction. IEEE Transactions on Image Processing, 2025

    work page 2025

  11. [11]

    Null-space diffusion sampling for zero-shot point cloud completion

    Xinhua Cheng, Nan Zhang, Jiwen Yu, Yinhuai Wang, Ge Li, and Jian Zhang. Null-space diffusion sampling for zero-shot point cloud completion. In IJCAI, pages 618–626, 2023

    work page 2023

  12. [12]

    Diffusion Posterior Sampling for General Noisy Inverse Problems

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffu- sion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  13. [13]

    An analysis of single-layer networks in unsuper- vised feature learning

    Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsuper- vised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215–223. JMLR Workshop and Conference Proceedings, 2011

    work page 2011

  14. [14]

    Image denoising with block-matching and 3d filtering

    Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising with block-matching and 3d filtering. In Image processing: algorithms and systems, neural networks, and machine learning, volume 6064, pages 354–365. SPIE, 2006. 11

    work page 2006

  15. [15]

    A Survey on Diffusion Models for Inverse Problems

    Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Eduardo S. L. Gastal and Manuel M. Oliveira. Domain transform for edge-aware image and video processing. In ACM SIGGRAPH 2011 Papers, SIGGRAPH ’11, New York, NY , USA,

    work page 2011

  17. [17]

    Association for Computing Machinery

    work page

  18. [18]

    Joint sparse and low rank recovery algorithm for compressive hyperspectral imaging

    Tatiana Gelvez, Hoover Rueda, and Henry Arguello. Joint sparse and low rank recovery algorithm for compressive hyperspectral imaging. Appl. Opt., 56(24):6785–6795, Aug 2017

    work page 2017

  19. [19]

    Deep equilibrium architectures for inverse problems in imaging

    Davis Gilton, Gregory Ongie, and Rebecca Willett. Deep equilibrium architectures for inverse problems in imaging. IEEE Transactions on Computational Imaging, 7:1123–1133, 2021

    work page 2021

  20. [20]

    Tikhonov regularization and total least squares

    Gene H Golub, Per Christian Hansen, and Dianne P O’Leary. Tikhonov regularization and total least squares. SIAM journal on matrix analysis and applications, 21(1):185–194, 1999

    work page 1999

  21. [21]

    Improving compressive imaging recovery via measurement augmentation

    Romario Gualdrón-Hurtado, Roman Jacome, Leon Suarez, Emmanuel Martinez, and Henry Arguello. Improving compressive imaging recovery via measurement augmentation. InICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2025

    work page 2025

  22. [22]

    Image restoration

    Bahadir Gunturk and Xin Li. Image restoration. CRC Press, 2018

    work page 2018

  23. [23]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors,Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc., 2020

    work page 2020

  24. [24]

    A restoration network as an implicit prior

    Yuyang Hu, Mauricio Delbracio, Peyman Milanfar, and Ulugbek S Kamilov. A restoration network as an implicit prior. arXiv preprint arXiv:2310.01391, 2023

    work page arXiv 2023

  25. [25]

    Stochastic deep restoration priors for imaging inverse problems

    Yuyang Hu, Albert Peng, Weijie Gan, Peyman Milanfar, Mauricio Delbracio, and Ulugbek S Kamilov. Stochastic deep restoration priors for imaging inverse problems. arXiv preprint arXiv:2410.02057, 2024

    work page arXiv 2024

  26. [26]

    Plug-and-play methods for integrating physical and learned models in compu- tational imaging: Theory, algorithms, and applications

    Kamilov et al. Plug-and-play methods for integrating physical and learned models in compu- tational imaging: Theory, algorithms, and applications. IEEE Sig. Proc. Mag., 40(1):85–97, 2023

    work page 2023

  27. [27]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  28. [28]

    FastMRI: A publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning

    Florian Knoll, Jure Zbontar, Anuroop Sriram, Matthew J Muckley, Mary Bruno, Aaron Defazio, Marc Parente, Krzysztof J Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdzalv, Adriana Romero, Michael Rabbat, Pascal Vincent, James Pinkerton, Duo Wang, Nafissa Yakubova, Erich Owens, C Lawrence Zitnick, Michael P Recht, Daniel K Sodickson, and ...

    work page 2020

  29. [29]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    work page 2009

  30. [30]

    Learning a projection operator onto the null space of a linear imaging operator

    Joseph Kuo, Jason Granstedt, Umberto Villa, and Mark A Anastasio. Learning a projection operator onto the null space of a linear imaging operator. In Medical Imaging 2021: Physics of Medical Imaging, volume 11595, pages 1019–1025. SPIE, 2021

    work page 2021

  31. [31]

    LoDoPaB- CT, a benchmark dataset for low-dose computed tomography reconstruction.Sci

    Johannes Leuschner, Maximilian Schmidt, Daniel Otero Baguer, and Peter Maass. LoDoPaB- CT, a benchmark dataset for low-dose computed tomography reconstruction.Sci. Data, 8(1):109, April 2021

    work page 2021

  32. [32]

    Recovery analysis for plug-and-play priors using the restricted eigenvalue condition

    Jiaming Liu, Salman Asif, Brendt Wohlberg, and Ulugbek Kamilov. Recovery analysis for plug-and-play priors using the restricted eigenvalue condition. Advances in Neural Information Processing Systems, 34:5921–5933, 2021. 12

    work page 2021

  33. [33]

    Accelerated MRI with CIRcular cartesian UnderSampling (CIR- CUS): a variable density cartesian sampling strategy for compressed sensing and parallel imaging

    Jing Liu and David Saloner. Accelerated MRI with CIRcular cartesian UnderSampling (CIR- CUS): a variable density cartesian sampling strategy for compressed sensing and parallel imaging. Quant. Imaging Med. Surg., 4(1):57–67, February 2014

    work page 2014

  34. [34]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976–11986, 2022

    work page 2022

  35. [35]

    Deep learning face attributes in the wild

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

    work page 2015

  36. [36]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

    work page 2019

  37. [37]

    Compressed sensing mri

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

    work page 2008

  38. [38]

    Improving medical image denoising via a lightweight plug-and-play module

    Lei Ma, Hulin Kuang, Jin Liu, Chengchao Shen, and Jianxin Wang. Improving medical image denoising via a lightweight plug-and-play module. In 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 1350–1355. IEEE, 2023

    work page 2023

  39. [39]

    Compressive sensing with augmented measurements via generative self-distillation

    Emmanuel Martinez, Roman Jacome, Romario Gualdrón-Hurtado, Iñaki Esnaola, and Henry Arguello. Compressive sensing with augmented measurements via generative self-distillation. In 2025 IEEE Statistical Signal Processing Workshop (SSP), pages 31–35, 2025

    work page 2025

  40. [40]

    Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing

    Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18–44, 2021

    work page 2021

  41. [41]

    Proximal algorithms

    Neal Parikh and Stephen Boyd. Proximal algorithms. Foundations and Trends in optimization, 1(3):127–239, 2014

    work page 2014

  42. [42]

    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

    work page 2017

  43. [43]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234–241. Springer, 2015

    work page 2015

  44. [44]

    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. In International Conference on Machine Learning, pages 5546–5557. PMLR, 2019

    work page 2019

  45. [45]

    Theoretical perspectives on deep learning methods in inverse problems

    Jonathan Scarlett, Reinhard Heckel, Miguel RD Rodrigues, Paul Hand, and Yonina C Eldar. Theoretical perspectives on deep learning methods in inverse problems. IEEE journal on selected areas in information theory, 3(3):433–453, 2023

    work page 2023

  46. [46]

    Deep null space learning for inverse problems: convergence analysis and rates

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

    work page 2019

  47. [47]

    Iterative projections for signal identification on manifolds: Global recovery guarantees

    Parikshit Shah and Venkat Chandrasekaran. Iterative projections for signal identification on manifolds: Global recovery guarantees. In 2011 49th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 760–767. IEEE, 2011

    work page 2011

  48. [48]

    Deep unsu- pervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsu- pervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 2256–2265, Lille, France, 07–09 Jul 2015. PMLR

    work page 2015

  49. [49]

    Siqi Wan, Yehao Li, Jingwen Chen, Yingwei Pan, Ting Yao, Yang Cao, and Tao Mei

    Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solv- ing inverse problems with latent diffusion models via hard data consistency. arXiv preprint arXiv:2307.08123, 2023. 13

    work page arXiv 2023

  50. [50]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021

    work page 2021

  51. [51]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021

    work page 2021

  52. [52]

    Image restoration and reconstruction using targeted plug-and-play priors

    Afonso M Teodoro, José M Bioucas-Dias, and Mário AT Figueiredo. Image restoration and reconstruction using targeted plug-and-play priors. IEEE Transactions on Computational Imaging, 5(4):675–686, 2019

    work page 2019

  53. [53]

    Fire: Fixed-points of restoration priors for solving inverse problems

    Matthieu Terris, Ulugbek S Kamilov, and Thomas Moreau. Fire: Fixed-points of restoration priors for solving inverse problems. arXiv preprint arXiv:2411.18970, 2024

    work page arXiv 2024

  54. [54]

    Equivariant plug-and- play image reconstruction

    Matthieu Terris, Thomas Moreau, Nelly Pustelnik, and Julian Tachella. Equivariant plug-and- play image reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25255–25264, 2024

    work page 2024

  55. [55]

    Deep image prior

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 9446–9454, 2018

    work page 2018

  56. [56]

    Venkatakrishnan, Charles A

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

    work page 2013

  57. [57]

    Gan prior based null-space learning for consistent super-resolution

    Yinhuai Wang, Yujie Hu, Jiwen Yu, and Jian Zhang. Gan prior based null-space learning for consistent super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2724–2732, 2023

    work page 2023

  58. [58]

    Unlimited-size diffusion restoration

    Yinhuai Wang, Jiwen Yu, Runyi Yu, and Jian Zhang. Unlimited-size diffusion restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1160–1167, 2023

    work page 2023

  59. [59]

    Zero-shot image restoration using denoising diffusion null-space model

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022

    work page arXiv 2022

  60. [60]

    Iterative reconstruction techniques for computed tomography part 1: technical principles

    Martin J Willemink, Pim A de Jong, Tim Leiner, Linda M de Heer, Rutger AJ Nievelstein, Ricardo PJ Budde, and Arnold MR Schilham. Iterative reconstruction techniques for computed tomography part 1: technical principles. European radiology, 23:1623–1631, 2013

    work page 2013

  61. [61]

    Restormer: Efficient transformer for high-resolution image restoration

    Zamir et al. Restormer: Efficient transformer for high-resolution image restoration. In IEEE/CVF CVPR, pages 5728–5739, 2022

    work page 2022

  62. [62]

    Learning nonlocal sparse and low-rank models for image compressive sensing: Nonlocal sparse and low-rank modeling

    Zhiyuan Zha, Bihan Wen, Xin Yuan, Saiprasad Ravishankar, Jiantao Zhou, and Ce Zhu. Learning nonlocal sparse and low-rank models for image compressive sensing: Nonlocal sparse and low-rank modeling. IEEE Signal Processing Magazine, 40(1):32–44, 2023

    work page 2023

  63. [63]

    Plug-and- play image restoration with deep denoiser prior

    Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and- play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360–6376, 2021

    work page 2021

  64. [64]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142–3155, 2017

    work page 2017

  65. [65]

    Places: A 10 million image database for scene recognition

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

    work page 2017

  66. [66]

    Denoising diffusion models for plug-and-play image restoration

    Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1219–1229, June 2023. 14 A Technical Appendices A.1 Settings for experiment in Figure 1 ...

    work page 2023

  67. [67]

    Deep null space network (DNSN)[45]: ˆx=H †y+ (I−H †H)R(H†y)

    work page

  68. [68]

    Deep decomposition network cascade (DDN-C) [ 6]: ˆx=H †y+ P r(Rr(H†y)) + Pn(Rn(H†y+ P r(Rr(H†y))))

    work page

  69. [69]

    • Network R and Rn: is a lightweight version of the U-Net architecture for image segmenta- tion

    Deep decomposition network independent (DDN-I) [ 6]: ˆx=H †y+ P r(Rr(H†y)) + Pn(Rn(H†y)) We used the source code 2 of [6] to implement the modelsR,R r andR n. • Network R and Rn: is a lightweight version of the U-Net architecture for image segmenta- tion. It features an encoder-decoder structure with skip connections. The encoder consists of five convolut...

    work page

  70. [70]

    (3) on each dataset further improves it

    The results show that the QR- initialized matrix provides good invertibility across all datasets, and that optimizing S via Eq. (3) on each dataset further improves it. Thus, the data distribution can enhance the data-driven invertibility of an already orthogonal matrix. Testing S on in-distribution samples yields the best invertibility (bold), while dive...

    work page