Pith. sign in

REVIEW 2 major objections 6 minor 38 references

STAR-Net: An Interpretable Model-Aided Network for Remote Sensing Image Denoising

T0 review · 2 major / 6 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 STAR-Net claims a fully unrolled ADMM network with low-rank and sparse tensor priors beats state-of-the-art remote sensing image denoisers while staying interpretable.

desk verdict A well-executed empirical study whose central model-aided claim is undermined by an incorrect derivation in the Bi-block; worth sending to referees, but with a demand for correction or explicit approximation. read the letter →

arxiv 2505.24327 v1 pith:DT3JCMMM submitted 2025-05-30 cs.CV cs.LG

Jingjing Liu , Jiashun Jin , Xianchao Xiu , Jianhua Zhang , Wanquan Liu This is my paper
classification cs.CVcs.LG
keywords remotesensingimagedenoisinginterpretabilitysparsetensor-aidedrepresentationnetworkADMMdeepunrollinglow-rankpriornon-localself-similaritynoise
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 tries to establish that a remote sensing image denoiser can keep the structure of a model-based optimization problem while learning everything end-to-end. Its model represents the clean image in a spectral subspace, $X = \mathcal{G}\times_3 A$, and adds a low-rank prior on non-local patch groups of the representation coefficients to exploit non-local self-similarity. The sparse variant STAR-Net-S adds an explicit $S$ term with an $\ell_1$ prior to absorb non-Gaussian noise such as salt-and-pepper and dead lines. ADMM iterations for the model are unrolled into trainable layers, so the regularization parameters are learned rather than hand-tuned. If correct, the method offers interpretability and competitive denoising on both synthetic and real remote sensing data without tedious parameter tuning.

What carries the argument

The load-bearing object is the ADMM-guided unrolling, where each optimization subproblem becomes a named network block: LargNet implements linear updates for the subspace factor $G$ and basis $A$, ShrinkNet implements soft-thresholding of $B_i$ through ReLU, SvtNet implements singular value thresholding of the low-rank auxiliary variable $L_i$, and a linear layer updates the Lagrangian multiplier $P_i$. These blocks are chained for $K$ unrolled iterations, and the regularization parameters $\lambda, \gamma_1, \gamma_2, \beta, \mu, l$ are treated as learnable. This converts the fixed ADMM loop into an end-to-end trainable architecture while preserving the update structure of the original model.

What would settle it

Solve the coefficient subproblem in Eq. (12) exactly for the learned dictionaries with a generic convex solver and compare that solution to ShrinkNet's output at the same inputs; with any orthogonal non-identity dictionary the outputs will differ, which would show the network does not exactly reproduce the ADMM update.

Watch

Extended reading notes

Core claim

The paper's central claim is that the ADMM solution of a tensor subspace denoising model can be unrolled exactly into a deep network, so that the network inherits the model's interpretability while all regularization parameters become learnable. The low-rank prior is placed on the representation coefficient image through non-local patch groups, making the non-local self-similarity of remote sensing images an explicit part of the optimization. STAR-Net-S generalizes the model by introducing a sparse noise variable $S$ and an $\ell_1$ penalty, which is meant to handle non-Gaussian corruption. The authors report that both variants outperform ten existing model-based and deep-learning RSI denoising methods on the ICVL and PaviaU synthetic benchmarks and on real-world Beijing Capital Airport and Indian Pines images, with STAR-Net-S showing the most balanced performance under mixed Gaussian and non-Gaussian noise.

Load-bearing premise

The derivation that rewrites the coefficient update as a simple shrinkage layer holds exactly only when the learned dictionaries are identity transforms, not merely orthogonal, so the paper's exact-unrolling claim rests on that unstated simplification.

Editorial extensions

If this is right

  • Regularization parameters in the underlying optimization model become trainable, so the same ADMM structure adapts to different noise levels without manual tuning.
  • The low-rank prior on non-local patch groups gives the network a mechanism for preserving edges and fine details while maintaining spectral consistency.
  • STAR-Net-S absorbs mixed Gaussian and non-Gaussian noise in a single architecture, removing the need for a separate salt-and-pepper or dead-line preprocessing step.
  • Stage-by-stage visualizations show each unrolled iteration addressing different corruption, supporting the claim that the network's internals correspond to identifiable optimization steps.

Reading between the lines

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

  • Because the step from Eq. (12) to Eq. (13) appears to require identity dictionaries while the model only assumes orthogonality, the exact-unrolling claim is likely approximate for the learned DCT dictionaries; the network may still perform well if the learned layers compensate for this mismatch.
  • The same ADMM-unrolling recipe could be carried to other remote-sensing inverse problems, such as fusion, super-resolution, or destriping, where non-local self-similarity and mixed noise are both present.
  • A direct comparison between the ShrinkNet $B_i$ layer and an exact solver for Eq. (12) would show whether the identity-dictionary simplification is empirically harmless or whether it changes the denoising result.
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

2 major / 6 minor

Summary. The paper proposes STAR-Net and STAR-Net-S, two deep unrolling networks for remote sensing image denoising. The method starts from a tensor subspace model with non-local self-similarity encoded through a low-rank prior on patch dictionaries, adds a sparse noise term in the STAR-Net-S variant, and unrolls an ADMM iteration into a trainable network with learnable regularization parameters. The authors report extensive experiments on two synthetic datasets (ICVL, PaviaU) and two real-world datasets (Beijing Capital Airport, Indian Pines), comparing against ten model-based and deep learning-based methods, and include parameter counts, runtime, ablations, a Friedman/Nemenyi statistical analysis, and a downstream classification experiment on Indian Pines.

Significance. If the central derivation were sound, the paper would make a useful contribution: it attacks a real problem (interpretable, parameter-light RSI denoising) and the empirical study is unusually thorough, with cross-dataset tests, real-world data, a downstream task, ablations on the unrolling depth and dictionary size, and a statistical significance analysis. The parameter efficiency of STAR-Net is also a strength. However, the exact-unrolling claim is the core intellectual selling point, and the mathematical step that connects the ADMM subproblem to the ShrinkNet layer is not valid as written. The empirical results may still stand as a strong learned architecture, but the interpretability/model-aided narrative currently rests on an incorrect algebraic equivalence.

major comments (2)
  1. [Section 3.1.2, Eqs. (12)–(16)] Equations (12) and (13) are not equivalent. Writing T(B)=B×1D1×2D2×3D3 and C=L_i^k+P_i^k/β, the smooth part of Eq. (12) has normal equations (βI+λT*T)B=λT*(R_iG^{k+1})+βC; when the dictionaries are orthogonal, T*T=I and the proximal update is soft_{λγ1/(λ+β)}((λT*(R_iG^{k+1})+βC)/(λ+β)). Equation (13), however, minimizes ½||(βI+λT)B-(λR_iG^{k+1}+βC)||², whose gradient contains the cross term βλ(T+T*)B; that term is absent from the true normal equations unless T is the identity. The DCT dictionaries used in Section 4.1.4 are orthogonal but not the identity, so the ShrinkNet update in Eqs. (14)–(16) does not solve the B_i-subproblem stated in Eq. (12). Because this step underlies the claim in the abstract and in Section 3 that the network fully unrolls the ADMM model, the derivation must be corrected, or Eq. (13) must be explicitly presented as an approximation and the exact-unrolling claims adjusted accordingly.
  2. [Section 3.1.5, Eq. (24), and Algorithms 1–2] Section 3.1.5 states Eq. (23) as P_i^{k+1}=P_i^k+β(L_i^{k+1}-B_i^{k+1}), but the sentence below Eq. (24) says Θ_i is calculated as P_i^k+β(L_i^{k+1}+B_i^{k+1}). The plus sign contradicts the minus sign in Eq. (23). Since the multiplier update affects the subsequent G- and B_i-blocks in the ADMM recursion, this sign inconsistency must be fixed in the text and in the implementations summarized in Algorithms 1 and 2.
minor comments (6)
  1. [Section 4.1.3, Eq. (33)] The notation in Eq. (33) is confusing: the text says Y denotes the denoised RSI generated by the network, while Y was defined in Eq. (1) as the noisy input tensor; please use distinct symbols for the network input and output.
  2. [Section 4.1.4 and Section 4.4.3] Please state explicitly whether the dictionaries D_1, D_2, D_3 are learned during training or remain fixed after DCT initialization; the phrase 'initialized using a DCT basis' is ambiguous, and this matters for the 'all parameters are learned' claim.
  3. [Tables 1–3] The quantitative comparisons report single numbers with no standard deviations or number of repeated runs; please clarify whether one training run is reported and, if so, consider adding variance information or multiple-seed results.
  4. [Section 4.4.8] The 'Convergence Analysis' heading is misleading: Figure 19 only shows training loss versus epochs, which demonstrates training convergence, not convergence of the ADMM recursion to a minimizer of the model in Eq. (5); please rename the section or add ADMM iteration curves.
  5. [Section 4.4.6] For the Friedman test, report the number of blocks used for ranking (four noise levels times two datasets) and the test statistic value, not only the p-value, so the reader can judge the strength of the evidence.
  6. [Throughout] There are several typographical errors: 'sparaity' in Section 3.2, 'Gernerally' in the Introduction, 'temn' in the Figure 17 caption, and many 'e ffectively' spacing artifacts; please proofread the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: STAR-Net's unrolling is trained on external labels and tested on held-out data; the flagged Eq. (12)-(13) issue is an algebraic-correctness concern, not a circular derivation.

full rationale

STAR-Net's central claim—that ADMM unrolling with low-rank and sparse priors outperforms state-of-the-art RSI denoising methods—is not circular. The objective in Eq. (5) is an independent variational model built from standard low-rank/nuclear-norm and l1 priors, and the network is trained on the ICVL training set against ground-truth labels via Eq. (33), then evaluated on held-out ICVL images, PaviaU, and real-world datasets. No fitted parameter from the test set is renamed as a prediction; the learnable regularization parameters are trained from labeled data rather than derived from the outputs. The paper contains no load-bearing self-citations: the compared baselines and cited techniques (ISTA, SVT, Procrustes rotation, subspace representation) are all external prior work. The reviewer-identified issue in Section 3.1.2—that the algebraic reduction from Eq. (12) to Eq. (13) requires D_j to be the identity, whereas the paper uses orthogonal DCT dictionaries—concerns the correctness of the "exact unrolling" claim and is a derivational error, not circular reasoning. Under the stated review rules, that concern belongs to correctness risk rather than circularity, since the reported PSNR/SSIM results would still be an externally trained empirical outcome even if the unrolling derivation needs correction. Therefore no circular step is exhibited and the circularity score is 0.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a low-rank subspace model for the clean image, a low-rank/sparse prior on non-local patch coefficients, and a sequence of standard optimization results. The main fragility is the unstated orthogonality (or identity) assumption in the B_i-block, plus the unstated subspace dimension n4.

free parameters (9)
  • λ (regularization weight) = learned (init 0.02)
    Balances data fidelity and the prior terms in the STAR-Net objective; trained from data.
  • γ1 (sparsity weight) = learned (init 0.02)
    Controls the ℓ1 penalty on coefficient tensors B_i; trained from data.
  • γ2 (low-rank weight) = learned (init 0.02)
    Controls the tensor nuclear norm penalty on L_i in STAR-Net; trained from data.
  • μ (sparse noise weight) = learned (init 0.02)
    Controls the ℓ1 penalty on sparse noise S in STAR-Net-S; trained from data.
  • β (ADMM penalty) = learned (init 0.02)
    ADMM penalty parameter, treated as learnable in the unrolled network.
  • l (Lipschitz constant) = learned (init 0.02)
    Step size in the ISTA-based B_i update; learned from data.
  • K (number of unrolling iterations) = 9
    Chosen from a validation study (Figures 12-13) balancing PSNR, SSIM, SAM, and parameters.
  • Dictionary size = 9
    DCT dictionary size [9,9,9] selected from a validation study (Figures 14-15).
  • Subspace dimension n4 = not specified
    The model defines G of size n1×n2×n4 with n4 << n3, but the paper never states the value used; a key modeling choice left unspecified.
assumptions (5)
  • domain assumption Spectral low-rank subspace representation: X = G ×_3 A with orthogonal A.
    Assumes the clean RSI lies in a low-dimensional spectral subspace; standard in HSI processing but an approximation.
  • domain assumption Non-local self-similarity is captured by low-rankness of coefficient tensors B_i.
    This is the paper's core modeling assumption; no derivation is given.
  • domain assumption Sparse noise (salt-and-pepper, dead lines) is modeled by an additive ℓ1 term S.
    For STAR-Net-S; the noise type is specific to the synthetic experiments.
  • ad hoc to paper Dictionaries D_j are orthogonal (DCT initialized).
    The simplification in Eq. (13) is only correct if D_j^T D_j = I or the operator H is defined differently; this is not stated.
  • standard math SVT and Procrustes solutions are valid for the L_i and A blocks.
    Standard results from matrix/tensor optimization (Cai et al. 2010; Zou et al. 2006).

how reviews work

0 comments
Cite this review

Pith. "Pith review of STAR-Net: An Interpretable Model-Aided Network for Remote Sensing Image Denoising." pith.science (2026). https://pith.science/paper/DT3JCMMM

@misc{pith2026250524327,
  author       = {Pith},
  title        = {Pith review of: STAR-Net: An Interpretable Model-Aided Network for Remote Sensing Image Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DT3JCMMM}},
  note         = {Machine review of arXiv:2505.24327}
}
read the original abstract

Remote sensing image (RSI) denoising is an important topic in the field of remote sensing. Despite the impressive denoising performance of RSI denoising methods, most current deep learning-based approaches function as black boxes and lack integration with physical information models, leading to limited interpretability. Additionally, many methods may struggle with insufficient attention to non-local self-similarity in RSI and require tedious tuning of regularization parameters to achieve optimal performance, particularly in conventional iterative optimization approaches. In this paper, we first propose a novel RSI denoising method named sparse tensor-aided representation network (STAR-Net), which leverages a low-rank prior to effectively capture the non-local self-similarity within RSI. Furthermore, we extend STAR-Net to a sparse variant called STAR-Net-S to deal with the interference caused by non-Gaussian noise in original RSI for the purpose of improving robustness. Different from conventional iterative optimization, we develop an alternating direction method of multipliers (ADMM)-guided deep unrolling network, in which all regularization parameters can be automatically learned, thus inheriting the advantages of both model-based and deep learning-based approaches and successfully addressing the above-mentioned shortcomings. Comprehensive experiments on synthetic and real-world datasets demonstrate that STAR-Net and STAR-Net-S outperform state-of-the-art RSI denoising methods.

Figures

Figures reproduced from arXiv: 2505.24327 by the authors.

Figure 1
Figure 1. Illustration of the proposed STAR-Net and STAR-Net-S with [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustration of tensor low-rank decomposition. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Denoising results on gavyam 0823-0933 with the noise variance of 50. The false-color images are generated by combining bands 5, 18, and 25. (a) Clean, (b) Noisy, (c) BM4D, (d) LLRT, (e) LRTDTV, (f) NGMeet, (g) NLSSR, (h) FastHyMix, (i) HSI-SDeCNN, (j) SMDS-Net, (k) Eigen-CNN, (l) RCILD, (m) STAR-Net, (n) STAR-Net-S. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) (k) (l) (m) (n) [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Denoising results of pixel (400, 100) on gavyam [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Denoising results on PaviaU with the noise variance of 50. The false-color images are generated [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: The denoising results of pixel (90, 130) on PaviaU with the noise variance of 50. (a) Clean, (b) [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Denoising results on Beijing Capital Airport. The false-color images are generated by combin [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Denoising results of pixel (160, 185) on Beijing Capital Airport. (a) Noisy, (b) BM4D, (c) LLRT, [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Denoising results on Indian Pines. The false-color images are generated by combining bands 1, [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: The denoising results of pixel (100, 100) on Indian Pines. (a) Noisy, (b) BM4D, (c) LLRT, (d) [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Classification results obtained by all methods on Indian Pines. (a) Ground truth, (b) Noisy, (c) [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Impact of unrolling iteration K of STAR-Net. LLRT and STAR-Net. This further validates the effectiveness of our proposed STAR￾Net and STAR-Net-S on downstream tasks. 4.4. Discussion 4.4.1. Number of Parameters The number of parameters for each deep learning-based meth…
Figure 13
Figure 13. Figure 13: Impact of unrolling iteration K of STAR-Net-S. (a) PSNR (b) SSIM (c) SAM (d) #Parameters [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: Impact of the dictionary number of STAR-Net. [PITH_FULL_IMAGE:figures/full_fig_p026_14.png]
Figure 15
Figure 15. Figure 15: Impact of the dictionary number of STAR-Net-S. [PITH_FULL_IMAGE:figures/full_fig_p026_15.png]
Figure 16
Figure 16. Figure 16: Stage-by-stage visualization process of STAR-Net-S. [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: Post-hoc Nemenyi test in temns of PSNR. all four indexes. Therefore, we initialize the learnable parameters γ1, γ2, l, λ, µ, β to 0.02 in this paper. 4.4.6. Statistical Tests The Friedman test analyzes the rankings of measurements across multiple condi￾tions to determ…
Figure 18
Figure 18. Figure 18: Performance and runtime comparison on the PaviaU dataset. The size of each circle represents [PITH_FULL_IMAGE:figures/full_fig_p029_18.png]
Figure 19
Figure 19. Figure 19: Convergence analysis of the loss function. [PITH_FULL_IMAGE:figures/full_fig_p030_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    Z. Wang, X. Wang, K. Tan, B. Han, J. Ding, Z. Liu, Hyperspectral anomaly detection based on variational background inference and generative adversarial network, Pattern Recognition 143 (2023) 109795

    work page 2023

  2. [2]

    Zhang, Y

    Q. Zhang, Y . Zheng, Q. Yuan, M. Song, H. Yu, Y . Xiao, Hyperspectral image 31 denoising: From model-driven, data-driven, to model-data-driven, IEEE Trans- actions on Neural Networks and Learning Systems (2023) 1–21

    work page 2023

  3. [3]

    C. Wang, W. Zheng, X. Sun, J. Zhou, J. Lu, Probabilistic deep metric learning for hyperspectral image classification, Pattern Recognition 157 (2025) 110878

    work page 2025

  4. [4]

    J. Chen, M. Zhao, X. Wang, C. Richard, S. Rahardja, Integration of physics-based and data-driven models for hyperspectral image unmixing: A summary of current methods, IEEE Signal Processing Magazine 40 (2023) 61–74

    work page 2023

  5. [5]

    Y . Su, H. Zhu, K.-C. Wong, Y . Chang, X. Li, Hyperspectral image denoising via weighted multidirectional low-rank tensor recovery, IEEE Transactions on Cybernetics 53 (2023) 2753–2766

    work page 2023

  6. [6]

    Dabov, A

    K. Dabov, A. Foi, V . Katkovnik, K. Egiazarian, Image denoising by sparse 3-D transform-domain collaborative filtering, IEEE Transactions on Image Processing 16 (2007) 2080–2095

    work page 2007

  7. [7]

    Maggioni, V

    M. Maggioni, V . Katkovnik, K. Egiazarian, A. Foi, Nonlocal transform-domain filter for volumetric data denoising and reconstruction, IEEE Transactions on Image Processing 22 (2013) 119–133

    work page 2013

  8. [8]

    J. Peng, W. Sun, H.-C. Li, W. Li, X. Meng, C. Ge, Q. Du, Low-rank and sparse representation for hyperspectral image processing: A review, IEEE Geoscience and Remote Sensing Magazine 10 (2022) 10–43

    work page 2022

Show all 38 references
  1. [9]

    Zhang, W

    H. Zhang, W. He, L. Zhang, H. Shen, Q. Yuan, Hyperspectral image restoration using low-rank matrix recovery, IEEE Transactions on Geoscience and Remote Sensing 52 (2014) 4729–4743

    2014

  2. [10]

    F. Xu, Y . Chen, C. Peng, Y . Wang, X. Liu, G. He, Denoising of hyperspectral im- age using low-rank matrix factorization, IEEE Geoscience and Remote Sensing Letters 14 (2017) 1141–1145

    2017

  3. [11]

    Chang, L

    Y . Chang, L. Yan, S. Zhong, Hyper-laplacian regularized unidirectional low-rank tensor recovery for multispectral image denoising, in: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 5901–5909. 32

    2017

  4. [12]

    Y . Wang, J. Peng, Q. Zhao, Y . Leung, X.-L. Zhao, D. Meng, Hyperspectral image restoration via total variation regularized low-rank tensor decomposition, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 11 (2018) 1227–1243

    2018

  5. [13]

    W. He, Q. Yao, C. Li, N. Yokoya, Q. Zhao, Non-local meets global: An inte- grated paradigm for hyperspectral denoising, in: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 6861–6870

    2019

  6. [14]

    Z. Zha, B. Wen, X. Yuan, J. Zhang, J. Zhou, Y . Lu, C. Zhu, Nonlocal struc- tured sparsity regularization modeling for hyperspectral image denoising, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–16

    2023

  7. [15]

    E. Pan, Y . Ma, X. Mei, J. Huang, Q. Chen, J. Ma, Hyperspectral image destriping and denoising from a task decomposition view, Pattern Recognition 144 (2023) 109832

    2023

  8. [16]

    Q. Yuan, Q. Zhang, J. Li, H. Shen, L. Zhang, Hyperspectral image denoising employing a spatial–spectral deep residual convolutional neural network, IEEE Transactions on Geoscience and Remote Sensing 57 (2019) 1205–1218

    2019

  9. [17]

    K. Wei, Y . Fu, H. Huang, 3D quasi-recurrent neural network for hyperspectral image denoising, IEEE Transactions on Neural Networks and Learning Systems 32 (2021) 363–375

    2021

  10. [18]

    Ma ffei, J

    A. Ma ffei, J. M. Haut, M. E. Paoletti, J. Plaza, L. Bruzzone, A. Plaza, A sin- gle model CNN for hyperspectral image denoising, IEEE Transactions on Geo- science and Remote Sensing 58 (2020) 2516–2529

    2020

  11. [19]

    Zhuang, M

    L. Zhuang, M. K. Ng, L. Gao, Z. Wang, Eigen-CNN: Eigenimages plus eigen- noise level maps guided network for hyperspectral image denoising, IEEE Trans- actions on Geoscience and Remote Sensing 62 (2024) 1–18

    2024

  12. [20]

    Q. Xie, M. Zhou, Q. Zhao, Z. Xu, D. Meng, MHF-Net: An interpretable deep network for multispectral and hyperspectral image fusion, IEEE Transactions on Pattern Analysis and Machine Intelligence 44 (2022) 1457–1473. 33

    2022

  13. [21]

    Zhuang, M

    L. Zhuang, M. K. Ng, FastHyMix: Fast and parameter-free hyperspectral im- age mixed noise removal, IEEE Transactions on Neural Networks and Learning Systems 34 (2023) 4702–4716

    2023

  14. [22]

    Zhang, W

    K. Zhang, W. Zuo, L. Zhang, FFDNet: Toward a fast and flexible solution for cnn-based image denoising, IEEE Transactions on Image Processing 27 (2018) 4608–4622

    2018

  15. [23]

    Xiong, J

    F. Xiong, J. Zhou, Q. Zhao, J. Lu, Y . Qian, MAC-Net: Model-aided nonlocal neu- ral network for hyperspectral image denoising, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1–14

    2022

  16. [24]

    Xiong, J

    F. Xiong, J. Zhou, S. Tao, J. Lu, J. Zhou, Y . Qian, SMDS-Net: Model guided spectral-spatial network for hyperspectral image denoising, IEEE Transactions on Image Processing 31 (2022) 5469–5483

    2022

  17. [25]

    J. Peng, H. Wang, X. Cao, Q. Zhao, J. Yao, H. Zhang, D. Meng, Learnable repre- sentative coefficient image denoiser for hyperspectral image, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–16

    2024

  18. [26]

    Zhang, W

    K. Zhang, W. Zuo, Y . Chen, D. Meng, L. Zhang, Beyond a gaussian denoiser: Residual learning of deep CNN for image denoising, IEEE Transactions on Image Processing 26 (2017) 3142–3155

    2017

  19. [27]

    T. Xie, S. Li, J. Lai, Adaptive rank and structured sparsity corrections for hyper- spectral image restoration, IEEE Transactions on Cybernetics 52 (2022) 8729– 8740

    2022

  20. [28]

    Monga, Y

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

    2021

  21. [29]

    Z. Zha, B. Wen, X. Yuan, J. Zhang, J. Zhou, X. Jiang, C. Zhu, Multiple comple- mentary priors for multispectral image compressive sensing reconstruction, IEEE Transactions on Cybernetics 54 (2024) 3338–3351. 34

    2024

  22. [30]

    J. M. Bioucas-Dias, J. M. P. Nascimento, Hyperspectral subspace identification, IEEE Transactions on Geoscience and Remote Sensing 46 (2008) 2435–2445

    2008

  23. [31]

    M. Wang, D. Hong, Z. Han, J. Li, J. Yao, L. Gao, B. Zhang, J. Chanussot, Ten- sor decompositions for hyperspectral data processing in remote sensing: A com- prehensive review, IEEE Geoscience and Remote Sensing Magazine 11 (2023) 26–72

    2023

  24. [32]

    C. Li, B. Zhang, D. Hong, J. Yao, J. Chanussot, LRR-Net: An interpretable deep unfolding network for hyperspectral anomaly detection, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–12

    2023

  25. [33]

    Y . Yang, J. Sun, H. Li, Z. Xu, ADMM-CSNet: A deep learning approach for image compressive sensing, IEEE Transactions on Pattern Analysis and Machine Intelligence 42 (2020) 521–538

    2020

  26. [34]

    X. Li, Y . Yang, H. Zheng, Z. Xu, ISP-IRLNet: Joint optimization of interpretable sampler and implicit regularization learning network for accerlerated MRI, Pat- tern Recognition 151 (2024) 110412

    2024

  27. [35]

    Zhang, B

    J. Zhang, B. Ghanem, ISTA-Net: Interpretable optimization-inspired deep net- work for image compressive sensing, in: 2018 IEEE /CVF Conference on Com- puter Vision and Pattern Recognition, 2018, pp. 1828–1837

    2018

  28. [36]

    J.-F. Cai, E. J. Cand `es, Z. Shen, A singular value thresholding algorithm for matrix completion, SIAM Journal on Optimization 20 (2010) 1956–1982

    2010

  29. [37]

    H. Zou, T. J. Hastie, R. Tibshirani, Sparse principal component analysis, Journal of Computational and Graphical Statistics 15 (2006) 265 – 286

    2006

  30. [38]

    Y . Chen, J. Zeng, W. He, X.-L. Zhao, T.-X. Jiang, Q. Huang, Fast large-scale hyperspectral image denoising via noniterative low-rank subspace representation, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–14. 35

    2024

Pith tools

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