Pith. sign in

REVIEW 5 major objections 5 minor 43 references

TRUST -- Transformer-Driven U-Net for Sparse Target Recovery

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TRUST, a transformer-driven U-Net, recovers sparse images better than classical compressed sensing and standalone U-Net models, with fewer hallucinated structures.

desk verdict A correct but simple theory and a workable hybrid architecture undermined by PSNR numbers that don't match the stated metric definition and an unfair Restormer baseline. read the letter →

arxiv 2506.01112 v1 pith:ERZ2UPDX submitted 2025-06-01 eess.IV

Di An , Dylan Poppert , Jiayue Li , Mark Foster , Trac D. Tran This is my paper
classification eess.IV
keywords inverseproblemssparserecoverytransformerU-Netattentionmechanismrestrictedisometrypropertyimagereconstructionhallucinationsuppression
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 proposes TRUST, a hybrid network that pairs a pretrained Vision Transformer encoder with a U-Net-style decoder, to solve sparse-signal recovery problems in which the sensing operator is unknown and only measurement–target pairs are available. The authors argue that self-attention computed directly on the measurements can stand in for attention on the true signal when the sensing matrix satisfies the Restricted Isometry Property, and that this global context helps a U-Net decoder avoid the hallucinated structures that pure convolutional reconstructions produce. On a custom multicore-fiber microendoscope dataset and on the fastMRI knee benchmark, TRUST is reported to beat classical sparse-recovery baselines and several deep-learning alternatives on PSNR, SSIM, and a hallucination metric. The design is meant for cross-domain inverse problems, notably optical imaging where the forward physics is nonlinear or hard to model.

What carries the argument

The central identity is the RIP inner-product preservation bound $|y^{\top}y' - x^{\top}x'| \le \delta_{2k}$, which the paper derives from the polarization identity and the Restricted Isometry Property. This justifies using scaled dot-product self-attention on the measurement $y$ as a proxy for attention on the true sparse signal $x$. The architecture then couples this Transformer encoder to an adaptive pooling layer and a U-Net-style upsampling decoder with skip connections, so global attention features guide multiscale local refinement.

What would settle it

Retrain the all-transformer baseline for the same number of epochs as TRUST on the same FastMRI split, using the same loss and hyperparameter budget, and measure whether the PSNR and SSIM gap remains; additionally, estimate the RIP constant of the actual microendoscope sensing operator to check whether the stated attention bound applies.

Watch

Extended reading notes

Core claim

The central claim is that a Transformer encoder can learn the sparse support structure directly from measurements, and that feeding the resulting attention features through skip connections into a U-Net decoder yields reconstructions with higher SSIM and PSNR and fewer false-positive regions than U-Net alone, TransUNet, Restormer, or Orthogonal Matching Pursuit. The theoretical anchor is a bound showing that, when the sensing matrix satisfies the Restricted Isometry Property with constant $\delta_{2k}$, the inner product of two measurements is within $\delta_{2k}$ of the inner product of the corresponding ground-truth signals, so attention maps computed in the measurement domain approximate attention maps in the target domain. This lets the encoder act as a support-estimation prior while the decoder refines local detail across scales.

Load-bearing premise

The reported superiority over the all-transformer baseline rests on the assumption that all models were compared after roughly equal training effort; if the baseline was undertrained, the performance gap may shrink or reverse.

Editorial extensions

If this is right

  • Attention maps extracted from measurement space can serve as a sparse-support prior for inverse problems whose sensing matrices obey RIP, potentially removing the need to estimate the sensing operator explicitly.
  • A U-Net decoder with skip connections preserves high-frequency detail that a transformer bottleneck alone flattens; removing any single skip connection measurably degrades SSIM and increases false-positive regions.
  • Transfer learning from a general-purpose vision transformer accelerates convergence and improves reconstruction quality, which matters when domain-specific training pairs are scarce.
  • On the microendoscope task, TRUST lowers the false-discovery rate by roughly a factor of five relative to U-Net and by roughly a factor of six relative to OMP, meaning recovered images contain fewer spurious structures.

Reading between the lines

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

  • If the RIP bound transfers to other sensing operators, any global-attention encoder, not only the pretrained ViT used here, could provide support estimates, making TRUST a special case of a broader attention-as-prior principle for inverse problems.
  • The FastMRI comparison trains the all-transformer baseline for about 8 epochs versus 50 epochs for TRUST under the same hardware budget; equalizing the optimization budget could narrow the reported gap, and that test would separate architectural benefit from training benefit.
  • The optical test set comes from a single unseen slide at a fixed depth; a natural stress test is varying depth, tissue type, and illumination to see whether the learned inverse map generalizes beyond the specific diffraction geometry.
  • The reported PSNR on the optics task is near 72 dB, far above typical image-reconstruction ranges, suggesting the task is nearly noiseless; the method's advantage may be more pronounced in lower-signal-to-noise regimes where hallucination suppression matters most.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes TRUST, a hybrid architecture that combines a pretrained Vision Transformer (ViT) encoder with a U-Net-style decoder, connected by an adaptive pooling layer and skip connections, for sparse target recovery from linear measurements y = Ax when A is unknown. The authors provide a theoretical bound showing that, under the Restricted Isometry Property, inner products between sparse vectors are approximately preserved under A, which they use to motivate attention applied directly to measurements. They evaluate TRUST on a custom optical microendoscope dataset and on FastMRI, comparing against OMP, U-Net, TransUNet, and Restormer, and report that TRUST achieves superior PSNR, SSIM, and reduced hallucination artifacts.

Significance. If the empirical claims were fully supported, the paper would offer a simple and effective architectural recipe for learned inverse problems with unknown forward operators. The theoretical derivation in Appendix A is self-contained and parameter-free, which is a strength. However, the experimental support is currently undermined by internal inconsistencies in the reported metrics, an undertrained baseline, and missing details. The paper also provides no code or data, limiting reproducibility. The contribution is otherwise incremental, combining existing components (ViT, U-Net, adaptive pooling) without a clear new principle beyond the inner-product bound, whose connection to the actual attention maps is not fully established.

major comments (5)
  1. [Table 1 and Appendix B] The reported PSNR values are inconsistent with the reported MSE values under the metric definition in Appendix B. For TRUST on the optics dataset, MSE=0.00431 implies RMSE=0.0656, giving PSNR=20*log10(1/0.0656)=23.66 dB, not the reported 71.99 dB; matching 71.99 dB requires MAX≈255, contradicting the stated assumption that MAX=1.0 after normalization. On FastMRI (Table 2), TRUST's MSE=0.0613 gives 12.13 dB with MAX=1 and 60.3 dB with MAX=255, neither matching the reported 24.81 dB; the implied MAX is approximately 4.3. This same inconsistency appears for all methods and in Tables 3 and 4, so the quantitative comparisons in the paper cannot be taken at face value.
  2. [Appendix D and Table 2] Appendix D states that under the same hardware, TRUST, TransUNet, and U-Net complete 50 epochs in 24 hours, while Restormer completes only 8 epochs. Table 2 therefore compares TRUST, U-Net, and TransUNet against an undertrained Restormer baseline. The claimed superiority of TRUST on FastMRI may reflect a training-budget imbalance rather than an architectural advantage. The authors should either train Restormer to a comparable epoch count or explicitly frame the comparison as restricted to limited training budgets.
  3. [Appendix B and Tables 1-5] The hallucination metric is defined in Appendix B as a 'False Positive Region (FPR) score' that depends on two thresholds, thigh and tlow, but these threshold values are never reported. Moreover, the main text and all tables label the column 'FDR' (False Discovery Rate), which is a different statistical quantity. Without threshold values and a consistent definition, the FDR columns cannot be interpreted or reproduced.
  4. [Section 3.3 and Appendix A] The theoretical derivation bounds the difference |x^T A^T A x' - x^T x'| for unit-norm k-sparse vectors, but the paper's claim that the attention map computed from y 'closely resembles' the attention map from x requires a bound on the output of the softmax function, not merely on the input inner products. The softmax nonlinearity and normalization across tokens can amplify or suppress inner-product differences. In addition, the unit-norm k-sparse assumption is not verified on the experimental data, so the connection between the theorem and the observed attention-map similarity in Figure 3 is not established.
  5. [Tables 1-2] No statistical significance tests are reported, and the standard deviations overlap on several key metrics. For example, in Table 2 the SSIM values are 0.717 ± 0.085 for TRUST and 0.698 ± 0.095 for Restormer, and in Table 1 the FDR values are 0.901 ± 0.22 for TRUST and 0.907 ± 0.36 for Restormer. The abstract's claim that TRUST 'significantly outperforms' the baselines is therefore not supported by the evidence presented.
minor comments (5)
  1. [Figure 7 caption] The word 'Restomer' should be spelled 'Restormer'.
  2. [Section 4.1] The text says that 'full details on preprocessing/sampling masks above are provided in the Appendix,' but Appendix B only defines evaluation metrics; the FastMRI undersampling pattern, acceleration factor, and normalization procedure are not described.
  3. [Section 4.2] Section 4.1 states that the test set contains 16,000 pairs, but Section 4.2 reports results on '5,000 randomly selected optical samples'; the relationship between these two numbers should be clarified.
  4. [Section 4.3] The loss weights for the combined ℓ2 + SSIM objective are not reported, so the loss-function ablation is not fully reproducible.
  5. [Table 3] The standard deviation for the ℓ2 + ℓ1 MSE value (0.0101 ± 0.18) is much larger than the mean, which suggests a typo or a highly skewed distribution; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the attention-bound derivation is self-contained and the empirical claims are not derived from the theory by construction.

full rationale

The paper's only formal derivation is Appendix A, which bounds |x^T A^T A x' - x^T x'| using the RIP assumption, the polarization identity, and the triangle inequality. This derivation assumes y = Ax, normalized k-sparse vectors, and RIP of order 2k; it contains no fitted parameters, no hidden appeal to the target result, and no self-citation chain. The bound is used only as motivation for the attention-based encoder, not as a definition of TRUST or as an input to the architecture's construction. The experimental comparisons are standard supervised training and evaluation: models are trained with explicit loss functions and then evaluated on held-out data, so reporting PSNR and SSIM for a model trained with an L2+SSIM loss is metric reporting rather than a circular reduction. The numerical inconsistency between reported MSE and PSNR values (Appendix B defines PSNR with MAX=1.0, which does not reproduce Tables 1-3) and the unequal training budget for Restormer (Appendix D reports 8 epochs vs. 50 epochs in 24 hours) are correctness and fairness concerns, not circularity. No load-bearing self-citation was found: the microendoscope references are contextual, and the pretrained ViT is an external checkpoint. Therefore the derivation chain is self-contained and the score is 0.

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

The paper introduces no new physical entities or mediators. Its central empirical claims depend on the free parameters above and on domain assumptions about linearity, RIP, and transferability of pretrained features. The theoretical section is self-contained but narrow.

free parameters (3)
  • SSIM loss weight in combined ℓ2+SSIM objective
    Section 4.3 reports best results with this combination but never states the weighting coefficient; it appears to be hand-tuned and affects the reported numbers.
  • hallucination threshold parameters t_high and t_low
    Appendix B defines a hallucinated pixel by xhat > t_high and xtrue <= t_low but does not provide the threshold values used in Tables 1-2, so the FDR numbers depend on unreported choices.
  • adaptive pooling output size
    Section 3.4 introduces adaptive pooling to standardize feature maps to a fixed size, but the chosen output dimensions are not reported and affect the decoder capacity.
assumptions (4)
  • domain assumption The measurement model is linear, y = Ax + w, with a known additive noise term.
    Assumed in Section 2 and used throughout; the optical microendoscope's diffraction process may be nonlinear, which would break the theory in Section 3.3.
  • domain assumption The sensing matrix A satisfies the Restricted Isometry Property of order 2k with a small constant, or has orthonormal columns.
    Used in Section 3.3 and Appendix A to derive the attention inner-product bound; the actual optical system is never verified to satisfy RIP.
  • ad hoc to paper A ViT pretrained on ImageNet (google/vit-base-patch16-224) transfers useful features to non-natural diffraction-pattern measurements.
    Section 4.5 shows pretraining helps empirically, but there is no theoretical or domain justification for why ImageNet features should apply to optical measurements.
  • domain assumption The test set from a third neuron slide is representative of the target distribution and does not overlap with training data.
    Section 4.1 states this separation, but no analysis of distribution shift, sample diversity, or potential leakage is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TRUST -- Transformer-Driven U-Net for Sparse Target Recovery." pith.science (2026). https://pith.science/paper/ERZ2UPDX

@misc{pith2026250601112,
  author       = {Pith},
  title        = {Pith review of: TRUST -- Transformer-Driven U-Net for Sparse Target Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERZ2UPDX}},
  note         = {Machine review of arXiv:2506.01112}
}
abstract

In the context of inverse problems $\bf y = Ax$, sparse recovery offers a powerful paradigm shift by enabling the stable solution of ill-posed or underdetermined systems through the exploitation of structure, particularly sparsity. Sparse regularization techniques via $\ell_0$- or $\ell_1$-norm minimization encourage solutions $\bf x$ that are both consistent with observations $\bf y$ and parsimonious in representation, often yielding physically meaningful interpretations. In this work, we address the classical inverse problem under the challenging condition where the sensing operator $\bf A$ is unknown and only a limited set of observation-target pairs $\{ \bf x,\bf y \}$ is available. We propose a novel neural architecture, TRUST, that integrates the attention mechanism of Transformers with the decoder pathway of a UNet to simultaneously learn the sensing operator and reconstruct the sparse signal. The TRUST model incorporates a Transformer-based encoding branch to capture long-range dependencies and estimate sparse support, which then guides a U-Net-style decoder to refine reconstruction through multiscale feature integration. The skip connections between the transformer stages and the decoder not only enhance image quality but also enable the decoder to access image features at different levels of abstraction. This hybrid architecture enables more accurate and robust recovery by combining global context with local details. Experimental results demonstrate that TRUST significantly outperforms traditional sparse recovery methods and standalone U-Net models, achieving superior performance in SSIM and PSNR metrics while effectively suppressing hallucination artifacts that commonly plague deep learning-based inverse solvers.

Figures

Figures reproduced from arXiv: 2506.01112 by the authors.

Figure 1
Figure 1. A multicore fiber coded aperture microendoscope. The fiber bundle contains around 6000 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. TRUST Architecture – Transformer-Driven U-Net for Sparse Target Recovery [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overlaying attention map of a sample collected from the microendoscope in Figure 1. From [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Adaptive pooling layer function’s effect on a typical attention map. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Different stages of decoding. From left to right: response [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Example of reconstruction results with corresponding SSIM and PSNR values. From left to right: response y, target x, OMP {0.301, 68.723dB}, U-Net {0.779, 71.691dB}, TRUST {0.862, 72.744dB} [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Two examples of FastMRI reconstruction results. From left to right: undersampled aliased image, true target x, U-Net reconstruction, Restormer reconstruction, and TRUST reconstruction [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Example of optical reconstruction with different loss functions with SSIM and PSNR(dB) value. From left to right: y, ℓ2{0.137, 48.756}, ℓ2 + ℓ1{0.251, 67.693}, ℓ2+ SSIM {0.798, 73.012}, and x. 4.4 Ablation Study: The Role of Skip Connections We investigate here the imp…
Figure 9
Figure 9. Figure 9: , removing even a single skip connection results in a noticeable drop in performance across all evaluation metrics. The degradation is particularly pronounced in high-frequency regions and structural boundaries, where spatial detail is most critical. These findings rea…
Figure 10
Figure 10. Figure 10: Pretrained-vs-Not reconstruction results with SSIM and PSNR(dB) value. From left to [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Simulation of similarity between attention on [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Example of reconstruction results with corresponding SSIM and PSNR values. Top row, from left to right: response y, target x, OMP {0.301, 68.723dB}, and U-Net {0.779, 71.691dB}. Bottom row, from left to right: TransUnet {0.672, 67.236dB}, Restormer {0.752, 71.762dB}, …
Figure 13
Figure 13. Figure 13: Example of reconstruction results with corresponding SSIM and PSNR values. Top row, from left to right: response y, target x, OMP {0.325, 63.071dB}, and U-Net {0.636, 66.712dB}. Bottom row, from left to right: TransUnet {0.553, 66.351dB}, Restormer {0.625, 66.583dB}, …
Figure 14
Figure 14. Figure 14: Example of reconstruction results with corresponding SSIM and PSNR values. Top row, from left to right: response y, target x, OMP {0.244, 58.232dB}, and U-Net {0.513, 62.105dB}. Bottom row, from left to right: TransUnet {0.409, 61.812dB}, Restormer {0.542, 62.503dB}, …
Figure 15
Figure 15. Figure 15: Example of reconstruction results with corresponding SSIM and PSNR values. Top row, from left to right: undersampled input y, target x, OMP {0.173, 15.682dB}, U-Net {0.610, 21.623dB}. Bottom row, from left to right: TransUnet {0.614, 21.956dB}, Restormer {0.623, 22.63…
Figure 16
Figure 16. Figure 16: Example of reconstruction results with corresponding SSIM and PSNR values. Top row, from left to right: undersampled input y, target x, OMP {0.2430, 12.812dB}, U-Net {0.612, 18.844dB}. Bot￾tom row, from left to right:: TransUnet {0.635, 19.593dB}, Restormer {0.636, 20…
Figure 17
Figure 17. Figure 17: Example of reconstruction results with corresponding SSIM and PSNR values. Top row, from left to right: undersampled input y, target x, OMP {0.5230, 19.083dB}, U-Net {0.586, 21.693dB}, TransUnet {0.871, 22.631dB}, Restormer {0.877, 26.568dB}, and TRUST {0.889, 30.602d…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [1]

    Regression shrinkage and selection via the lasso

    Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology , 58:267–288, 1996

    work page 1996

  2. [2]

    Curtis R. V ogel. Computational Methods for Inverse Problems . Society for Industrial and Applied Mathematics, 2002

    work page 2002

  3. [3]

    Inverse Problem Theory and Methods for Model Parameter Estimation

    Albert Tarantola. Inverse Problem Theory and Methods for Model Parameter Estimation . Society for Industrial and Applied Mathematics, 2005

    work page 2005

  4. [4]

    Linear inverse problems in imaging

    Alejandro Ribes and Francis Schmitt. Linear inverse problems in imaging. IEEE Signal Processing Magazine, 25:84–99, 2008

    work page 2008

  5. [5]

    Candès, Justin K

    Emmanuel J. Candès, Justin K. Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences , 59(8):1207–1223, 2006

    work page 2006

  6. [6]

    Compressed sensing

    David L Donoho. Compressed sensing. IEEE Transactions on Information Theory, 52(4):1289– 1306, 2006

    2006

  7. [7]

    Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information

    Emmanuel J Candès, Justin Romberg, and Terence Tao. Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information. IEEE Transactions on Information Theory, 52(2):489–509, 2006

    work page 2006

  8. [8]

    Sparse and Redundant Representations: from Theory to Applications in Signal and Image Processing

    Michael Elad. Sparse and Redundant Representations: from Theory to Applications in Signal and Image Processing. Springer Science & Business Media, 2010

    work page 2010

Show all 43 references
  1. [9]

    U-net: convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) , pages 234–241. Springer, 2015

    2015

  2. [10]

    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

    2017

  3. [11]

    FFDNet: Toward a fast and flexible solution for CNN-based image denoising

    Kai Zhang, Wangmeng Zuo, and Lei Zhang. FFDNet: Toward a fast and flexible solution for CNN-based image denoising. IEEE Transactions on Image Processing, 27(9):4608–4622, 2018

    2018

  4. [12]

    Photo-realistic single image super-resolution using a generative adversarial network

    Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE C...

    2017

  5. [13]

    Patel, and Richard G

    Ali Mousavi, Ankit B. Patel, and Richard G. Baraniuk. A deep learning approach to struc- tured signal recovery. In 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 1336–1343. IEEE, 2015

    2015

  6. [14]

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

    Yair Rivenson, Yibo Zhang, Harun Günaydın, Da Teng, and Aydogan Ozcan. Phase recovery and holographic image reconstruction using deep learning in neural networks. Light: Science & Applications, 7(2):17141–17141, 2018

    2018

  7. [15]

    Deep convolu- tional neural network for inverse problems in imaging

    Kyong Hwan Jin, Michael T McCann, Emmanuel Froustey, and Michael Unser. Deep convolu- tional neural network for inverse problems in imaging. IEEE Transactions on Image Processing, 26(9):4509–4522, 2017

    2017

  8. [16]

    An image is worth 16x16 words: transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  9. [17]

    Yuille, and Yuyin Zhou

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L. Yuille, and Yuyin Zhou. TransUNet: transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 10

    2021 arXiv

  10. [18]

    K-svd: An algorithm for designing overcomplete dictionaries for sparse representation

    Michal Aharon, Michael Elad, and Alfred Bruckstein. K-svd: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Transactions on Signal Processing, 54(11):4311–4322, 2006

    2006

  11. [19]

    Signal recovery from random measurements via orthogonal matching pursuit

    Joel A Tropp and Anna C Gilbert. Signal recovery from random measurements via orthogonal matching pursuit. IEEE Transactions on Information Theory , 53(12):4655–4666, 2007

    2007

  12. [20]

    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

    2009

  13. [21]

    Inverse Problem Theory and Methods for Model Parameter Estimation

    Albert Tarantola. Inverse Problem Theory and Methods for Model Parameter Estimation. SIAM, 2005

    2005

  14. [22]

    Curtis R. V ogel. Computational Methods for Inverse Problems. SIAM, 2002

    2002

  15. [23]

    Deep generative adversarial neural networks for compressive sensing mri

    Morteza Mardani, Enhao Gong, Joseph Cheng, and et al. Deep generative adversarial neural networks for compressive sensing mri. IEEE Transactions on Medical Imaging , 38(1):167–179, 2019

    2019

  16. [24]

    Rebecca Willett, Roummel Marcia, and Justin M. Nichols. Coded aperture imaging: principles, progress, and prospects. IEEE Signal Processing Magazine, 25(1):61–70, 2007

    2007

  17. [25]

    Deep tissue fluorescence microscopy with a multimode fiber

    S Farahi, Y Guan, K Wagner, and et al. Deep tissue fluorescence microscopy with a multimode fiber. Optics Express, 21(20):24566–24575, 2013

    2013

  18. [26]

    ISTA-Net: interpretable optimization-inspired deep network for image compressive sensing

    Jian Zhang and Bernard Ghanem. ISTA-Net: interpretable optimization-inspired deep network for image compressive sensing. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1828–1837, 2018

    2018

  19. [27]

    Deep ADMM-Net for compressive sensing MRI

    Jian Sun, Huibin Li Li, and Zongben Xu. Deep ADMM-Net for compressive sensing MRI. Advances in Neural Information Processing Systems , 29, 2016

    2016

  20. [28]

    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

    2021

  21. [29]

    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, pages 234–241. Springer, 2015

    2015

  22. [30]

    Restormer: efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: efficient transformer for high-resolution image restoration. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5728–5739, 2022

    2022

  23. [31]

    Decoding by linear programming

    Emmanuel J Candès and Terence Tao. Decoding by linear programming. IEEE Transactions on Information Theory, 51(12):4203–4215, 2005

    2005

  24. [32]

    Spatial pyramid pooling in deep convolutional networks for visual recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(9):1904–1916, 2015

    1904

  25. [33]

    FastMRI: an open dataset and benchmarks for accelerated MRI

    Jure Zbontar, Florian Knoll, Anuroop Sriram, Tim Murrell, Zizhao Huang, Matthew J Muckley, Aaron Defazio, Rachel Stern, Patricia Johnson, Michael Bruno, et al. FastMRI: an open dataset and benchmarks for accelerated MRI. In Advances in Neural Information Processing Systems (Ne...

    2018

  26. [34]

    Sparse MRI: the application of compressed sensing for rapid MR imaging

    Michael Lustig, David Donoho, and John M Pauly. Sparse MRI: the application of compressed sensing for rapid MR imaging. Magnetic Resonance in Medicine , 58(6):1182–1195, 2007

    2007

  27. [35]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600– 612, 2004. 11

    2004

  28. [36]

    Digital Image Processing

    Rafael C Gonzalez and Richard E Woods. Digital Image Processing. Prentice Hall, 2nd edition, 2002

    2002

  29. [37]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS) , volume 33, pages 6840–6851, 2020

    2020

  30. [38]

    Loss functions for image restoration with neural networks

    Hang Zhao, Orazio Gallo, Iuri Frosio, and Jan Kautz. Loss functions for image restoration with neural networks. IEEE Transactions on Computational Imaging, 3(1):47–57, 2016

    2016

  31. [39]

    Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections

    Xinbo Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016

    2016

  32. [40]

    Deep residual learning for im- age recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    2016

  33. [41]

    Med3d: Transfer learning for 3d medical image analysis

    Sheng Chen, Kai Ma, and Yefeng Zheng. Med3d: Transfer learning for 3d medical image analysis. arXiv preprint arXiv:1904.00625, 2019

    1904 arXiv

  34. [42]

    Training data-efficient image transformers & distillation through attention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning , pages 10347–10357. PMLR, 2021

    2021

  35. [43]

    MobileViT: light-weight, general-purpose, and mobile-friendly vision transformer

    Sachin Mehta and Mohammad Rastegari. MobileViT: light-weight, general-purpose, and mobile-friendly vision transformer. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 12 Appendix A Error Bound for the Attention Mechanism We assume that ...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.