Pith. sign in

REVIEW 3 major objections 5 minor 43 references

A TRPCA-Inspired Deep Unfolding Network for Hyperspectral Image Denoising via Thresholded t-SVD and Top-K Sparse Transformer

T0 review · 3 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 This paper aims to show that a deep unfolding network strictly alternating a thresholded t-SVD low-rank module and a Top-K sparse transformer module outperforms existing hyperspectral denoisers under severe mixed noise while staying…

desk verdict Strong empirical mixed-noise results, but the TRPCA-unfolding identity is mathematically overstated and should be fixed before publication. read the letter →

arxiv 2506.02364 v1 pith:VEQRI4AC submitted 2025-06-03 cs.CV

Liang Li , Jianli Zhao , Sheng Fang , Siyu Chen , Hui Sun This is my paper
classification cs.CV
keywords hyperspectralimagedenoisingdeepunfoldingnetworktensorrobustprincipalcomponentanalysisthresholdedt-SVDTop-Ksparsetransformermixednoiselow-rankandpriorsspectral-spatial
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes DU-TRPCA, a deep unfolding network that turns the two alternating updates of tensor robust principal component analysis (TRPCA) into trainable network stages for hyperspectral image denoising. In each stage, a differentiable hard-thresholded t-SVD layer enforces a low-rank prior, and then a transformer with a Top-K sparsity constraint in its deepest encoder layer enforces a sparse prior, mimicking the $\ell^1$ update in TRPCA. The central claim is that this strict stage-wise alternation, rather than extra model capacity, is what lets the network outperform existing methods under severe mixed noise while using only about 1.05M parameters. If true, the paper offers a template for hybrid denoisers that keep an optimization algorithm's structure and interpretability without giving up deep representation power.

What carries the argument

The load-bearing object is the unfolded TRPCA update: each network stage computes $L^{k+1}=X-r_L^k(X-P_{\mathrm{t-SVD},r}(X-S^k))$ followed by $S^{k+1}=X-r_S^k(X-P_{\mathrm{TopK}}(X-L^{k+1}))$. The low-rank operator $P_{\mathrm{t-SVD},r}$ is a rank-$r$ truncated tensor SVD computed per frontal slice in the frequency domain, which is the closed-form proximal operator of the tensor nuclear norm in classical TRPCA; because hard truncation is non-differentiable, the paper uses a custom backward pass that propagates gradients only through the top-$r$ singular components. The sparse operator $P_{\mathrm{TopK}}$ is a transformer encoder-decoder built on a hybrid spectral denoising transformer, with a learnable Top-K sparsity constraint at the deepest encoder layer, so that only the most significant features survive; this is the neural counterpart of the $\ell^1$ soft-threshold proximal update. Learnable residual weights $r_L^k$, $r_S^k$ and the Top-K rate are trained end-to-end with a stage-wise $\ell^2$ loss over all unfolded stages, and a 'first stage independent, later stages shared' parameter scheme keeps the model at about 1.05M parameters.

What would settle it

Replace the hard-thresholded t-SVD backward pass with a differentiable soft-thresholded SVD or a straight-through estimator and retrain under the same mixed-noise protocol; if the custom pseudo-gradient version does not outperform it on impulse and deadline noise, the claim that the hard-threshold proximal step is what carries the TRPCA prior is not supported. Similarly, if removing the Top-K module entirely while keeping the unfolded t-SVD stages does not hurt performance on sparse impulse noise, the sparse-prior claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that explicit, stage-by-stage alternation between a low-rank proximal step and a sparse proximal step, matching the structure of the TRPCA iterations, can be realized as a trainable network that beats both pure model-driven TRPCA solvers and pure learned denoisers on mixed-noise hyperspectral restoration. The low-rank step is implemented by rank-truncated t-SVD in the Fourier domain, using hard thresholding in the forward pass and a pseudo-gradient in the backward pass; the sparse step is implemented by a transformer whose deepest encoder features are Top-K sparsified. The paper reports that on synthetic benchmarks under stripe, deadline, impulse, non-i.i.d. Gaussian, and mixture noise, DU-TRPCA achieves the best PSNR, SSIM, and SAM in most settings, and that it transfers across datasets better than prior methods. The authors interpret this as evidence that the TRPCA alternating structure is the source of robustness, with the t-SVD module capturing global spatial-spectral correlation and the Top-K module removing localized outliers.

Load-bearing premise

The whole approach rests on the assumption that the truncated t-SVD layer with its custom pseudo-gradient is a faithful, trainable approximation of the TRPCA low-rank proximal update, so the trained network really is learning the intended alternating optimization rather than a generic denoiser with a low-rank-looking module.

Editorial extensions

If this is right

  • Under the paper's synthetic benchmark protocol, DU-TRPCA reports the best PSNR, SSIM, and SAM in most mixed-noise settings on both benchmark datasets, including gains over the transformer backbone it builds on.
  • Trained on one dataset and tested on another, DU-TRPCA reports higher PSNR and SSIM than the compared methods, indicating that the explicit priors aid cross-domain transfer.
  • When moving from Gaussian to impulse noise, DU-TRPCA reports the smallest relative PSNR drop among compared methods, consistent with the claim that the sparse module specifically handles sparse, high-magnitude corruptions.
  • Ablation without the Top-K module performs slightly better under pure Gaussian noise, while adding it improves impulse and mixed-noise robustness; the paper interprets this as expected behavior for an explicit sparsity prior.
  • The parameter count is about 1.05M, roughly half that of the larger transformer baseline, which the paper uses to argue that the gains come from the alternating low-rank and sparse structure rather than from model capacity.

Reading between the lines

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

  • A strict reading suggests the TRPCA guarantees are inherited structurally, not as formal convergence theorems, because the learned transformer module and the pseudo-gradient replace exact proximal maps with approximations.
  • The same alternating low-rank/sparse unfolding could be applied to other tensor restoration problems, such as multi-spectral inpainting or video denoising, by keeping the t-SVD proximal step and swapping the transformer branch for a task-appropriate network.
  • The paper applies Top-K sparsity to feature representations rather than to attention maps; a direct testable extension is to compare feature-level and attention-level sparsity within the same unfolding framework to see which better matches the TRPCA sparse prior.
  • One could construct a synthetic noise mix with a tunable ratio of dense Gaussian to sparse impulse corruption; the model's PSNR curve across that ratio would test the claim that alternation is most valuable when both noise types are present.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes DU-TRPCA, a deep unfolding network for hyperspectral image denoising that alternates a low-rank module based on thresholded tensor SVD (t-SVD) with a Top-K sparse transformer module, inspired by tensor robust principal component analysis (TRPCA). The manuscript claims that this architecture strictly unfolds the TRPCA alternating updates, thereby inheriting TRPCA's interpretability and theoretical guarantees, and reports state-of-the-art results under severe mixed noise on ICVL and CAVE, strong cross-dataset generalization, and a compact 1.05M parameter budget. Experiments include synthetic mixed-noise and Gaussian-noise benchmarks, real Urban and Indian Pines data, and ablations over stage count and module composition.

Significance. If the empirical results hold up, DU-TRPCA is a useful hybrid architecture for HSI denoising. The paper includes extensive comparisons, a clear ablation study, code release, and a thoughtful observation that Top-K sparsity helps with impulse/mixed noise but can slightly hurt under pure Gaussian noise. The main weakness is conceptual: the low-rank module is presented as the exact proximal update of the tensor nuclear norm problem in Eq. (3), but it is actually a hard-thresholded t-SVD projection, not the soft-thresholding operator that solves that problem. This undercuts the 'strict unfolding' and 'inherited theoretical guarantees' claims, although it does not necessarily invalidate the empirical mixed-noise results, which are separable from the theoretical framing.

major comments (3)
  1. [Section III-B.a, Eqs. (3)-(5)] The claim that Eq. (4) is the closed-form solution of Eq. (3) is mathematically incorrect. The proximal operator of the tensor nuclear norm induced by t-SVD is tensor singular-value soft-thresholding, performed slice-wise in the Fourier domain with threshold lambda_L, not hard rank-r truncation. The cited [14, Algorithm 2] also uses tensor singular-value thresholding, not hard truncation. Consequently, even when the learnable residual weight r_L^k equals 1, Eq. (4) is a hard projection, not the minimizer of Eq. (3). This invalidates the 'strict unfolding' claim and the statement in Section III-D that DU-TRPCA 'inherits the theoretical guarantees, convergence behavior, and interpretability of the original optimization framework.'
  2. [Section III-B.b, Eqs. (6)-(8)] The sparse module is described as 'directly matching' the TRPCA sparse update and as 'strictly corresponding' to it, but Eq. (7) with P_Top-K is not the element-wise soft-thresholding update of Eq. (6); it is a learned transformer mapping with feature sparsity. This is a heuristic analogy rather than a derivation. The paper should reframe the Top-K sparse transformer as a trainable analogue of the sparse prior, not as an exact realization of the TRPCA proximal step, or provide a formal argument for the correspondence.
  3. [Section III-B.b, paragraph on learnable parameters] The text states that 'All modules and parameters, including the Top-K rate, are optimized end-to-end,' but no mechanism is described for differentiating through the discrete Top-K selection or through the rank parameter k. If the Top-K rate is a hard integer hyperparameter, it is not end-to-end learnable in the standard sense; if a differentiable relaxation is used (e.g., soft-top-k or straight-through estimation), it should be specified. This is a reproducibility gap for one of the paper's two central modules.
minor comments (5)
  1. [Section II.B] There is a duplicated sentence: 'Notably, unlike existing approaches where Top-K sparsity is applied to the attention maps, our method imposes Top-K sparsity directly on the feature representations. Notably, unlike previous approaches that apply Top-K sparsity to attention maps, our method imposes Top-K sparsity directly on feature representations.' One of the two should be removed.
  2. [Section IV.B.1] The text says 'Figure 4 presents a visual comparison on the mixture-noise scenario' but Figure 4 is the reflectance-spectra plot; the visual denoising comparison is Figure 3. The cross-reference should be corrected.
  3. [Section III.C, Eq. (9)] There is a notation conflict: in Eqs. (1)-(8), X denotes the observed degraded tensor, while in Section III.C and Eq. (9), X denotes the ground-truth image and Y the degraded measurement. Using different symbols (e.g., X_gt and X_obs) would avoid confusion.
  4. [Tables I, II, VII] The noise type 'deadline' appears repeatedly; this should likely be 'dead line' or 'dead-pixel line' to match the description in the text. The current spelling reads as a typo.
  5. [Section IV.D.1, Table VII] The text says performance 'improves initially but gradually saturates,' but the table shows a non-monotonic pattern: D5 and D6 are sometimes worse than D4 (e.g., mix PSNR 37.920 and 38.279 vs. 38.759). The wording should reflect a peak at four stages rather than a saturation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the denoising results are held-out evaluations, and the TRPCA-unfolding claim, while mathematically loose in its proximal-operator identification, is not a self-referential or fitted-to-target derivation.

full rationale

The paper's empirical claims are self-contained against external benchmarks: the network is trained end-to-end with a stage-wise supervised loss (Eq. 9) and evaluated on held-out ICVL, CAVE, Urban, and Indian Pines data, following the standard QRNN3D protocol. No parameter is fitted to the reported test metric and then presented as a prediction. The low-rank and sparse modules contain trainable residual weights and a custom pseudo-gradient, but these are trained parameters, not quantities derived from the evaluation targets. There are no load-bearing self-citations and no uniqueness theorem imported from the authors' own prior work; reference [14] is an external TRPCA paper. The arXiv note and the method section do contain a real mathematical-fidelity concern: Eq. (4) calls rank-r hard-thresholded t-SVD the closed-form solution of Eq. (3), while the proximal operator for the t-SVD tensor nuclear norm is tensor singular-value soft-thresholding. That is an incorrect-identity or overclaiming issue about whether the network truly inherits TRPCA's guarantees, but it is not a circularity: the network's performance is not derived from that identity, and the empirical comparison does not reduce to the paper's own assumptions. Therefore no circular step can be quoted, and the appropriate score is 0.

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

The core building blocks are existing mathematical and architectural tools. The paper's contribution is in their combination and training scheme, not in inventing a new physical entity. The main assumptions are domain-level (low-rank + sparse model) plus two ad hoc approximations that are not fully derived.

free parameters (4)
  • low-rank truncation rank r = not specified
    The rank r in Eq. (4)-(5) controls the t-SVD truncation and is a critical hyperparameter, but no value or sensitivity study is reported.
  • Top-K rate k = learnable
    The paper says the Top-K rate is optimized end-to-end, but no final value is given.
  • learnable residual weights r_L^k and r_S^k = learnable per stage
    These weights scale the residual connections in Eqs. (4), (7), and (10) and are trained from data.
  • number of unfolded stages K = 4
    Stage count is chosen via Table VII; the paper selects stage4 as the best trade-off.
assumptions (4)
  • domain assumption HSI data can be modeled as a low-rank tensor plus a sparse tensor, as in TRPCA.
    This is the basis for Eq. (1) and the entire unfolding design, and it may not hold for all real scenes.
  • standard math Tensor nuclear norm based on t-SVD is a valid convex surrogate for tensor low-rankness.
    Used in Eq. (1) and (3), following Lu et al. [14].
  • ad hoc to paper Hard-thresholded t-SVD with the custom pseudo-gradient is a trainable proxy for the proximal operator.
    The paper introduces this approximation in Section III-B.a without proving that it preserves convergence or optimization fidelity.
  • ad hoc to paper Top-K sparse transformer features correspond to the sparse noise component in TRPCA.
    The mapping is argued by analogy in Section III-B.b, but no formal equivalence with l1 soft-thresholding is established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A TRPCA-Inspired Deep Unfolding Network for Hyperspectral Image Denoising via Thresholded t-SVD and Top-K Sparse Transformer." pith.science (2026). https://pith.science/paper/VEQRI4AC

@misc{pith2026250602364,
  author       = {Pith},
  title        = {Pith review of: A TRPCA-Inspired Deep Unfolding Network for Hyperspectral Image Denoising via Thresholded t-SVD and Top-K Sparse Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VEQRI4AC}},
  note         = {Machine review of arXiv:2506.02364}
}
read the original abstract

Hyperspectral images (HSIs) are often degraded by complex mixed noise during acquisition and transmission, making effective denoising essential for subsequent analysis. Recent hybrid approaches that bridge model-driven and data-driven paradigms have shown great promise. However, most of these approaches lack effective alternation between different priors or modules, resulting in loosely coupled regularization and insufficient exploitation of their complementary strengths. Inspired by tensor robust principal component analysis (TRPCA), we propose a novel deep unfolding network (DU-TRPCA) that enforces stage-wise alternation between two tightly integrated modules: low-rank and sparse. The low-rank module employs thresholded tensor singular value decomposition (t-SVD), providing a widely adopted convex surrogate for tensor low-rankness and has been demonstrated to effectively capture the global spatial-spectral structure of HSIs. The Top-K sparse transformer module adaptively imposes sparse constraints, directly matching the sparse regularization in TRPCA and enabling effective removal of localized outliers and complex noise. This tightly coupled architecture preserves the stage-wise alternation between low-rank approximation and sparse refinement inherent in TRPCA, while enhancing representational capacity through attention mechanisms. Extensive experiments on synthetic and real-world HSIs demonstrate that DU-TRPCA surpasses state-of-the-art methods under severe mixed noise, while offering interpretability benefits and stable denoising dynamics inspired by iterative optimization. Code is available at https://github.com/liangli97/TRPCA-Deep-Unfolding-HSI-Denoising.

Figures

Figures reproduced from arXiv: 2506.02364 by the authors.

Figure 1
Figure 1. Overview of the DU-TRPCA network. (a) The complete architecture, consisting of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the Guided Spectral Self-Attention (GSSA) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Denoising comparison on the mixture-noise scenario of ICVL’s [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of reflectance spectra at pixel [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Denoising comparison on the real scenario of [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Denoising comparison on the real scenario of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [14]

    Tensor robust principal component analysis with a new tensor nuclear norm,

    C. Lu, J. Feng, Y . Chen, W. Liu, Z. Lin, and S. Yan, “Tensor robust principal component analysis with a new tensor nuclear norm,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 4, pp. 925–938, 2019

    2019

  2. [1]

    Diverse Region-Based CNN for Hyperspectral Image Classification,

    M. Zhang, W. Li, and Q. Du, “Diverse Region-Based CNN for Hyperspectral Image Classification,”IEEE Transactions on Image Processing, vol. 27, no. 6, pp. 2623–2634, Jun. 2018

    work page 2018

  3. [2]

    A review of change detection in multitemporal hyperspectral images: Current techniques, applications, and challenges,

    S. Liu, D. Marinelli, L. Bruzzone, and F. Bovolo, “A review of change detection in multitemporal hyperspectral images: Current techniques, applications, and challenges,”IEEE Geoscience and Remote Sensing Magazine, vol. 7, no. 2, pp. 140–158, Jun. 2019

    work page 2019

  4. [3]

    Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources,

    X. X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, and F. Fraundorfer, “Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources,”IEEE Geoscience and Remote Sensing Magazine, vol. 5, no. 4, pp. 8–36, Dec. 2017

    work page 2017

  5. [4]

    Noise re- duction in hyperspectral imagery: Overview and application,

    B. Rasti, P. Scheunders, P. Ghamisi, G. Licciardi, and J. Chanussot, “Noise re- duction in hyperspectral imagery: Overview and application,”Remote Sensing, vol. 10, no. 3, p. 482, 2018

    work page 2018

  6. [5]

    Hyperspectral image denoising with realistic data,

    T. Zhang, Y . Fu, and C. Li, “Hyperspectral image denoising with realistic data,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2248–2257

    2021

  7. [6]

    Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven,

    Q. Zhang, Y . Zheng, Q. Yuan, M. Song, H. Yu, and Y . Xiao, “Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 10, pp. 13 143–13 163, Oct. 2024

    work page 2024

  8. [7]

    Flex-DLD: Deep low-rank decomposition model with flexible priors for hyperspectral image denoising and restoration,

    Y . Chen, H. Zhang, Y . Wang, Y . Yang, and J. Wu, “Flex-DLD: Deep low-rank decomposition model with flexible priors for hyperspectral image denoising and restoration,”IEEE Transactions on Image Processing, vol. 33, pp. 1211– 1226, 2024

    work page 2024

Show all 43 references
  1. [8]

    Hyper-Laplacian Regularized Unidirectional Low-Rank Tensor Recovery for Multispectral Image Denoising,

    Y . Chang, L. Yan, and S. Zhong, “Hyper-Laplacian Regularized Unidirectional Low-Rank Tensor Recovery for Multispectral Image Denoising,” in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, Jul. 2017, pp. 5901–5909

    2017

  2. [9]

    Weighted Low-Rank Tensor Recovery for Hyperspectral Image Restoration,

    Y . Chang, L. Yan, X.-L. Zhao, H. Fang, Z. Zhang, and S. Zhong, “Weighted Low-Rank Tensor Recovery for Hyperspectral Image Restoration,”IEEE Transactions on Cybernetics, vol. 50, no. 11, pp. 4558–4572, Nov. 2020

    2020

  3. [10]

    Total-variation-regularized low-rank matrix factorization for hyperspectral image restoration,

    W. He, H. Zhang, L. Zhang, and H. Shen, “Total-variation-regularized low-rank matrix factorization for hyperspectral image restoration,”IEEE transactions on geoscience and remote sensing, vol. 54, no. 1, pp. 178–188, 2015

    2015

  4. [11]

    Hyperspectral Image Denoising Using Local Low-Rank Matrix Recovery and Global Spatial–Spectral Total Variation,

    W. He, H. Zhang, H. Shen, and L. Zhang, “Hyperspectral Image Denoising Using Local Low-Rank Matrix Recovery and Global Spatial–Spectral Total Variation,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 11, no. 3, pp. 713–729, Mar. 2018

    2018

  5. [12]

    Non-local meets global: An iterative paradigm for hyperspectral image restoration,

    W. He, Q. Yao, C. Li, N. Yokoya, Q. Zhao, H. Zhang, and L. Zhang, “Non-local meets global: An iterative paradigm for hyperspectral image restoration,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 4, pp. 2089–2107, 2020

    2020

  6. [13]

    Hy- perspectral image denoising based on nonlocal low-rank and tv regularization,

    X. Kong, Y . Zhao, J. Xue, J. C.-W. Chan, Z. Ren, H. Huang, and J. Zang, “Hy- perspectral image denoising based on nonlocal low-rank and tv regularization,” Remote Sensing, vol. 12, no. 12, p. 1956, 2020

    1956

  7. [15]

    Hyperspectral image denoising employing a spatial–spectral deep residual convolutional neural network,

    Q. Yuan, Q. Zhang, J. Li, H. Shen, and L. Zhang, “Hyperspectral image denoising employing a spatial–spectral deep residual convolutional neural network,”IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 2, pp. 1205–1218, 2018

    2018

  8. [16]

    3-D quasi-recurrent neural network for hyperspectral image denoising,

    K. Wei, Y . Fu, and H. Huang, “3-D quasi-recurrent neural network for hyperspectral image denoising,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 1, pp. 363–375, Jan. 2021

    2021

  9. [17]

    Hybrid spectral denoising transformer with guided attention,

    Z. Lai, C. Yan, and Y . Fu, “Hybrid spectral denoising transformer with guided attention,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV). Paris, France: IEEE, Oct. 2023, pp. 13 019–13 029

    2023

  10. [18]

    SMDS-net: Model guided spectral-spatial network for hyperspectral image denoising,

    F. Xiong, J. Zhou, S. Tao, J. Lu, J. Zhou, and Y . Qian, “SMDS-net: Model guided spectral-spatial network for hyperspectral image denoising,”IEEE Transactions on Image Processing, vol. 31, pp. 5469–5483, 2022

    2022

  11. [19]

    Deep plug-and-play prior for hyperspectral image restoration,

    Z. Lai, K. Wei, and Y . Fu, “Deep plug-and-play prior for hyperspectral image restoration,”Neurocomputing, vol. 481, pp. 281–293, Apr. 2022

    2022

  12. [20]

    A trainable spectral- spatial sparse coding model for hyperspectral image restoration,

    T. Bodrito, A. Zouaoui, J. Chanussot, and J. Mairal, “A trainable spectral- spatial sparse coding model for hyperspectral image restoration,”Advances in Neural Information Processing Systems, vol. 34, pp. 5430–5442, 2021

    2021

  13. [21]

    LR-Net: Low-Rank Spatial- JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 11 Spectral Network for Hyperspectral Image Denoising,

    H. Zhang, H. Chen, G. Yang, and L. Zhang, “LR-Net: Low-Rank Spatial- JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 11 Spectral Network for Hyperspectral Image Denoising,”IEEE Transactions on Image Processing, vol. 30, pp. 8743–8758, 2021

    2014

  14. [22]

    MAC-Net: Model-Aided Non- local Neural Network for Hyperspectral Image Denoising,

    F. Xiong, J. Zhou, Q. Zhao, J. Lu, and Y . Qian, “MAC-Net: Model-Aided Non- local Neural Network for Hyperspectral Image Denoising,”IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2022

    2022

  15. [23]

    Low-rank prompt-guided trans- former for hyperspectral image denoising,

    X. Tan, M. Shao, Y . Qiao, T. Liu, and X. Cao, “Low-rank prompt-guided trans- former for hyperspectral image denoising,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–15, 2024

    2024

  16. [24]

    ISTA-net: Interpretable optimization-inspired deep network for image compressive sensing,

    J. Zhang and B. Ghanem, “ISTA-net: Interpretable optimization-inspired deep network for image compressive sensing,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, Jun. 2018, pp. 1828–1837

    2018

  17. [25]

    Recursions are all you need: Towards efficient deep unfolding networks,

    R. Alhejaili, M. Alfarraj, H. Luqman, and A. Al-Shaikhi, “Recursions are all you need: Towards efficient deep unfolding networks,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Vancouver, BC, Canada: IEEE, Jun. 2023, pp. 4705–4714

    2023

  18. [26]

    Deep unfolding network for image desnowing with snow shape prior,

    X. Guo, X. Wang, X. Fu, and Z.-J. Zha, “Deep unfolding network for image desnowing with snow shape prior,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  19. [27]

    Deep unfolding transformers for sparse recovery of video,

    B. De Weerdt, Y . C. Eldar, and N. Deligiannis, “Deep unfolding transformers for sparse recovery of video,”IEEE Transactions on Signal Processing, 2024

    2024

  20. [28]

    Deep generalized unfolding networks for image restoration,

    C. Mou, Q. Wang, and J. Zhang, “Deep generalized unfolding networks for image restoration,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, LA, USA: IEEE, Jun. 2022, pp. 17 378–17 389

    2022

  21. [29]

    Iterative low-rank network for hyperspectral image denoising,

    J. Ye, F. Xiong, J. Zhou, and Y . Qian, “Iterative low-rank network for hyperspectral image denoising,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–15, 2024

    2024

  22. [30]

    Degradation-noise-aware deep unfolding transformer for hyperspectral image denoising,

    H. Zeng, K. Feng, X. Zhao, J. Cao, S. Huang, H. Luong, and W. Philips, “Degradation-noise-aware deep unfolding transformer for hyperspectral image denoising,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–12, 2025

    2025

  23. [31]

    Learning a sparse transformer network for effective image deraining,

    X. Chen, H. Li, M. Li, and J. Pan, “Learning a sparse transformer network for effective image deraining,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 5896–5905

    2023

  24. [32]

    Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration,

    S. Zhou, D. Chen, J. Pan, J. Shi, and J. Yang, “Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 2952–2963

    2024

  25. [33]

    Denoising of hyperspectral image using low-rank matrix factorization,

    F. Xu, Y . Chen, C. Peng, Y . Wang, X. Liu, and G. He, “Denoising of hyperspectral image using low-rank matrix factorization,”IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 7, pp. 1141–1145, 2017

    2017

  26. [34]

    Hyperspectral image denoising using nonconvex local low- rank and sparse separation with spatial–spectral total variation regularization,

    C. Peng, Y . Liu, K. Kang, Y . Chen, X. Wu, A. Cheng, Z. Kang, C. Chen, and Q. Cheng, “Hyperspectral image denoising using nonconvex local low- rank and sparse separation with spatial–spectral total variation regularization,” IEEE Transactions on Geoscience and Remote Sensing,...

    2022

  27. [35]

    Deep hyperspectral prior: Single-image denoising, inpainting, super-resolution,

    O. Sidorov and J. Yngve Hardeberg, “Deep hyperspectral prior: Single-image denoising, inpainting, super-resolution,” inProceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019, pp. 0–0

    2019

  28. [36]

    Spatial-spectral transformer for hyperspectral im- age denoising,

    M. Li, Y . Fu, and Y . Zhang, “Spatial-spectral transformer for hyperspectral im- age denoising,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, pp. 1368–1376, Jun. 2023

    2023

  29. [37]

    Spectral enhanced rectangle transformer for hyperspectral image denoising

    M. Li, J. Liu, Y . Fu, Y . Zhang, and D. Dou, “Spectral enhanced rectangle transformer for hyperspectral image denoising.” arXiv, Apr. 2023

    2023

  30. [38]

    Matrix backpropagation for deep networks with structured layers,

    C. Ionescu, O. Vantzos, and C. Sminchisescu, “Matrix backpropagation for deep networks with structured layers,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 2965–2973

    2015

  31. [39]

    Double- factor-regularized low-rank tensor factorization for mixed noise removal in hyperspectral image,

    Y .-B. Zheng, T.-Z. Huang, X.-L. Zhao, Y . Chen, and W. He, “Double- factor-regularized low-rank tensor factorization for mixed noise removal in hyperspectral image,”IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 12, pp. 8450–8464, 2020

    2020

  32. [40]

    Fasthymix: Fast and parameter-free hyperspectral image mixed noise removal,

    L. Zhuang and M. K. Ng, “Fasthymix: Fast and parameter-free hyperspectral image mixed noise removal,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 8, pp. 4702–4716, 2021

    2021

  33. [41]

    Trq3dnet: A 3d quasi-recurrent and transformer based network for hyperspectral image denoising,

    L. Pang, W. Gu, and X. Cao, “Trq3dnet: A 3d quasi-recurrent and transformer based network for hyperspectral image denoising,”Remote Sensing, vol. 14, no. 18, p. 4598, 2022

    2022

  34. [42]

    Sparse recovery of hyperspectral signal from natural rgb images,

    B. Arad and O. Ben-Shahar, “Sparse recovery of hyperspectral signal from natural rgb images,” inComputer Vision–ECCV 2016: 14th European Confer- ence, Amsterdam, the Netherlands, October 11–14, 2016, Proceedings, Part VII 14. Springer, 2016, pp. 19–34

    2016

  35. [43]

    Generalized assorted pixel camera: postcapture control of resolution, dynamic range, and spectrum,

    F. Yasuma, T. Mitsunaga, D. Iso, and S. K. Nayar, “Generalized assorted pixel camera: postcapture control of resolution, dynamic range, and spectrum,”IEEE transactions on image processing, vol. 19, no. 9, pp. 2241–2253, 2010. Liang Lireceived his M.E. degree in Computer Scienc...

    2010

Pith tools

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