Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

SAR Despeckling via Region-Aware Sparse Representation and Statistical Noise Approximation

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A training-free sparse-representation estimator reports the best PSNR and SSIM at every tested speckle level and best edge preservation on real Sentinel-1 data.

desk verdict Training-free SAR despeckling combining Log-Yeo-Johnson with weighted sparse coding; good numbers, but w1/w2 estimation must be disclosed. read the letter →

arxiv 2412.18121 v2 pith:KLYDLVCY submitted 2024-12-24 cs.IT math.IT

Xuran Hu , Mingzhe Zhu , Djordje Stanković , Zhenpeng Feng , Yifang Ban , Ljubiša Stanković This is my paper · ORCID
classification cs.ITmath.IT
keywords SARdespecklingsparserepresentationcompressivesensingLog-Yeo-Johnsontransformationgammanoisenon-localpatchstackingweightedLassoSentinel-1
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 claims that a classical, training-free sparse-coding estimator can remove SAR speckle more effectively than supervised deep networks, provided the multiplicative gamma noise is first reshaped into near-Gaussian form. The method applies a Log-Yeo-Johnson transformation to log-intensity SAR data, stacks non-locally similar patches, and solves a weighted Lasso problem whose two diagonal weight matrices encode per-patch noise levels and per-atom sparsity. In synthetic experiments the method reports the highest PSNR and SSIM of all compared methods at 1, 2, 4, and 8 looks; on a real Sentinel-1 VV+VH scene it reports the highest edge-preservation indices and structural quality while keeping mean intensity close to the original. The intended significance is that a well-chosen noise model and sparsity prior can rival learned models without any training data.

What carries the argument

The load-bearing object is the MAP-derived weighted Lasso. Two diagonal auxiliary matrices carry the non-trivial information: $w_1$ weights each stacked patch's data-fidelity term by the inverse of its estimated noise standard deviation, and $w_2$ weights each sparse coefficient by the inverse of its singular-value scale, so that important dictionary atoms are penalized less. The non-local step groups the $k$ closest patches by Euclidean distance before the joint sparse coding, which is what lets the per-patch noise weights act on genuinely similar content. The dictionary itself is obtained from the SVD of the stacked patch matrix, making the whole pipeline closed-form except for the ADMM solution of the Lasso.

What would settle it

Take a heavily textured SAR-like image at ENL=1, apply the transform, and test the stacked-patch residuals for Gaussianity (for example with a Kolmogorov-Smirnov or kurtosis test); if the residuals depart strongly from Gaussian or the noise variance varies inside a patch, the reported PSNR/SSIM margin over the next-best method should shrink. A more direct check is to replace the Gaussian likelihood in the MAP objective with the exact gamma likelihood: if the gamma-likelihood version matches or beats the proposed method, the Gaussian approximation is not the source of the gains.

Watch

Extended reading notes

Core claim

After the logarithm turns multiplicative gamma speckle into additive noise, a Yeo-Johnson step with $\lambda$ chosen by minimizing kurtosis and skewness makes the residual approximately Gaussian. Under that approximation, the maximum-a-posteriori estimate of the sparse coefficients reduces to the weighted Lasso objective $\hat\alpha = \arg\min_\alpha \|(D\alpha - y)w_1\|_2^2 + c\|w_2\alpha\|_1$, where $w_1 = \mathrm{diag}(\sigma_1^{-1},\dots,\sigma_K^{-1})$ rescales the residual of each stacked patch by its estimated noise level and $w_2$ scales each dictionary atom by its singular-value importance. The paper reports that, with the dictionary $D$ taken from the SVD of the stacked similar-patch matrix and the regularization constant $c=1.5$ set by hand, this formulation achieves the best quantitative scores in its comparisons: for example, 23.31 dB PSNR and 66.08 SSIM at 1-look versus 21.05 dB and 54.49 for the next-best method, and on real data an edge-preservation index of 0.733 versus 0.719 for the best baseline. The construction has no learned parameters, so the same estimator applies to any scene or noise level without retraining.

Load-bearing premise

The load-bearing premise is that, after the Log-Yeo-Johnson transform, the speckle residual in each stacked patch is close enough to independent Gaussian noise with a single per-patch variance that the weighted-Lasso likelihood is trustworthy; the paper itself notes the transform only approximates Gaussianity and never quantifies the approximation error.

Editorial extensions

If this is right

  • At every tested ENL level (1, 2, 4, 8 looks), the proposed method reports the best PSNR and SSIM among the nine compared methods, with the largest margin at 1-look.
  • On the real Sentinel-1 scene, it reports the highest EPI (0.733), EPD (14.33 horizontal, 14.32 vertical), and SQI (1.595) among all compared methods, indicating that edges and structure survive despeckling.
  • The ablation demonstrates that the two auxiliary matrices, not the transform alone, drive most of the gain: adding them raises 4-look PSNR from 22.62 to 25.60 and real-scene EPI from 0.571 to 0.658 before the transform adds further improvement.
  • Because the method is training-free, the same estimator can be applied directly to a new scene or a new noise level without collecting training data or retraining a network.
  • The method preserves fine details such as striped clothing patterns under 1-look speckle, where the compared supervised networks tend to over-smooth.

Reading between the lines

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

  • A natural extension, not pursued in the paper, is to apply the same Log-Yeo-Johnson plus weighted-Lasso pipeline to other coherent imaging modalities with gamma-like speckle, such as ultrasound or optical coherence tomography.
  • The MAP derivation suggests that per-patch regularization weights could be set automatically from local noise estimates in other sparse-coding denoisers, reducing the need for hand-tuned constants; this is an implicit consequence of the paper's construction.
  • A sharper comparison would retrain all deep baselines on the identical synthetic gamma-noise dataset and ENL conditions; the paper retrains only two of the eight baselines, so part of the reported margin could reflect checkpoint mismatch rather than method superiority.
  • The high mean intensity retained on the real scene (44.52 versus lower values for most baselines) hints that the method may preserve radiometric calibration; a dedicated radiometric-accuracy study could make that explicit.
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

4 major / 4 minor

Summary. The manuscript proposes a training-free SAR despeckling method combining a Log-Yeo-Johnson transformation with sparse representation. The method transforms multiplicative gamma noise into an approximately Gaussian additive noise, groups similar patches, and solves a weighted Lasso problem with two auxiliary matrices intended to capture per-patch noise levels and per-atom sparsity scales. Experiments on synthetic gamma-noised images and real Sentinel-1 data compare the method with eight existing despeckling approaches and report the best PSNR/SSIM on synthetic data and best EPI/EPD on real data. An ablation study attributes part of the gain to the transformation and the auxiliary matrices.

Significance. If the reported results are reproducible, the method would be a useful contribution: it is a classical, training-free estimator that claims to outperform supervised deep networks on the tested SAR despeckling tasks, and the ablation study gives some evidence that the Log-Yeo-Johnson transform and the two weighting matrices contribute to the final performance. The paper also compares against a broad set of baselines on both synthetic and real data, which is valuable. However, the central claim is currently conditional on details that are not in the manuscript: the extra regularization constant c is not derived from the stated MAP model, and the estimation of the noise weights and sparsity scales is not specified. Because the quantitative superiority claim rests precisely on these components, the significance cannot be fully assessed until the derivation and implementation details are supplied.

major comments (4)
  1. [Section II-C, Eq. (16) and Eq. (17)] The manuscript derives the MAP objective in Eq. (16) from a Gaussian likelihood and a Laplace prior, and that objective contains no free regularization parameter. The actual solved objective in Eq. (17) inserts an extra coefficient c, set to 1.5 in Section III-A, that appears neither in Eq. (16) nor in the stated priors. As written, the implemented Lasso problem is not the MAP estimator derived from the model, so the 'sparsity-guided posterior estimation' framing does not justify the reported objective.
  2. [Section II-C, Eqs. (12)-(14) and (17)] The method is underspecified because w1 and w2 are never estimated. The text says 'we apply maximum a posteriori (MAP) estimation to determine w1 and w2,' but the subsequent derivation only solves for α and treats w1 and w2 as fixed. No procedure or formula is given for estimating σ_k in Eq. (12) or S_i in Eq. (13); the statement that S_i values are 'derived from singular value decomposition' does not define the mapping from singular values to Laplacian scales, and the matrix Σ in w2 = Σ^{-1} is never defined. Without this estimation procedure, a third party cannot reproduce the method, the ablation in Table III, or the comparisons in Tables I and II.
  3. [Table I] The SSIM values in Table I exceed the theoretical maximum of 1.0: for example, the reported SSIM values are 27.86 and 40.04 for ANLM and 66.08 for the proposed method at 1-look. If these are SSIM values multiplied by 100, the scaling must be stated explicitly in the table header or caption; otherwise the quantitative comparison is uninterpretable and the claim of the highest SSIM cannot be verified.
  4. [Section I and Section III-D] The paper itself concedes in Section I that Log-Yeo-Johnson and similar transformations 'only approximate Gaussian distribution, undermining the reliability of Gaussian-based models,' but the validity of the likelihood in Eq. (12) is never quantified. The ablation in Section III-D shows histograms of residual noise, but no numerical test of Gaussianity is reported across ENL levels or image textures. Since the method's model is misspecified when the transformed noise is substantially non-Gaussian or signal-dependent, the reported gains may not generalize beyond the test conditions.
minor comments (4)
  1. [Eq. (4) and Eq. (5)] The branch conditions in the Yeo-Johnson transformation appear to contain typos: the λ=2 branch of Eq. (4) is written with the condition 'x ≥ 0' but the expression applies to x < 0, and Eq. (5) has the same issue in its final branch.
  2. [Section III-A] The text says 'After Section 3.3, we revert the transformed image back to the spatial domain,' but the relevant sections are in the methodology, not Section 3.3, and the cross-reference is unclear.
  3. [Abstract] The phrase 'to noise sparse assumption' in the abstract is ungrammatical and should be reworded, for example to 'to satisfy a sparse-noise assumption.'
  4. [Section II-B] The notation in Eq. (8) uses 'Da' while the dictionary is denoted D and the sparse coefficient matrix α; the typo 'a' instead of 'α' appears in several equations and should be fixed throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: Eq. (17) is an explicitly constructed weighted-Lasso objective, and the claimed SOTA results are evaluated against external test benchmarks, not predicted from the fitted parameters.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The MAP objective in Eq. (16) follows from the stated Gaussian likelihood (12) and Laplace prior (13)-(14); Eq. (17) is the matrix form of that objective, with the extra multiplicative factor c set to 1.5 in Section III-A as an explicit heuristic regularization weight rather than a quantity derived from the data. The auxiliary matrices w1 and w2 are data-adaptive: w1 encodes per-patch noise standard deviations and w2 encodes inverse singular values from the SVD dictionary, both obtained from the observed noisy image. This is standard empirical modeling (analogous to empirical Bayes) and does not make the target denoised image an input to its own derivation; the method does not fit the evaluation metrics or the ground-truth images. The only self-citation, reference [5] (authors' RDDPM), is used as a comparison baseline and is not load-bearing for the proposed derivation. The acknowledged limitation that Log-Yeo-Johnson only approximates Gaussianity is a modeling-risk statement, not a circular step. No equation in the paper reduces to its inputs by construction, and no fitted parameter is renamed as a prediction. Therefore the appropriate circularity score is 0.

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

The method rests on standard SAR noise modeling and several data-fitted components. No new physical entities are introduced. The main burden is that the Gaussianity assumption, the dictionary, and the sparsity scales are all derived from or fit to the observed noisy image, and the paper does not specify how sigma_k is obtained.

free parameters (5)
  • Log-Yeo-Johnson shape parameter lambda = not reported (exhaustive search)
    Chosen per image by minimizing kurtosis and skewness of the transformed data; a data-fitted parameter whose range and criterion are not specified.
  • Regularization coefficient c = 1.5
    Manually set in Eq (17) and Implementation Specifics; it is not derived from the Laplace prior in Eq (13).
  • Per-patch noise standard deviations sigma_k = not reported
    Used to build w1 in Eq (17); the paper never states how sigma_k is estimated from the noisy image.
  • Sparsity scales S_i (w2=Sigma^{-1}) = singular values of the noisy patch matrix
    Derived from SVD of the noisy stacked patches in Eq (9) and used as the Laplace prior scale in Eqs (13)-(17); fitted to the same data being denoised.
  • Patch size and stack count = p=16, K=10
    Chosen by hand in Implementation Specifics; results may depend on these choices.
assumptions (5)
  • domain assumption SAR intensity y equals clean image x times multiplicative gamma noise n with pdf p(n)=L^L n^{L-1} e^{-Ln}/Gamma(L).
    Standard SAR speckle model, invoked in Section II-A Eqs (1)-(2).
  • domain assumption After Log-Yeo-Johnson transformation, the noise is approximately i.i.d. Gaussian.
    Central to the Gaussian likelihood Eq (12); the paper acknowledges in Section I that this is only approximate.
  • domain assumption Similar patches found by Euclidean distance in the noisy image share the same sparse structure.
    Nonlocal grouping in Section II-B assumes noisy-domain similarity is a reliable proxy for clean similarity.
  • domain assumption The SVD dictionary D=U and singular values Sigma from the noisy patch matrix are valid for sparse coding the clean image.
    Eq (9) defines the dictionary from noisy data; this assumes the noise does not corrupt the dictionary structure enough to mislead the estimate.
  • domain assumption Noise in each patch is independent with patch-specific variance sigma_k.
    Likelihood Eq (12) assumes independence across patches and known variances, but sigma_k estimation is unspecified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAR Despeckling via Region-Aware Sparse Representation and Statistical Noise Approximation." pith.science (2026). https://pith.science/paper/KLYDLVCY

@misc{pith2026241218121,
  author       = {Pith},
  title        = {Pith review of: SAR Despeckling via Region-Aware Sparse Representation and Statistical Noise Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KLYDLVCY}},
  note         = {Machine review of arXiv:2412.18121}
}
read the original abstract

Synthetic Aperture Radar (SAR) imagery are widely utilized in remote sensing due to their all-weather, all-day imaging capabilities. However, SAR images are highly susceptible to noise, particularly speckle noise, caused by the coherent imaging process, which severely degrades image quality. This has driven increasing research interest in SAR despeckling. Sparse representation-based methods have been extensively applied in natural image processing, yet SAR despeckling requires addressing non-Gaussian assumption and ensuring sparsity in the transform domain. In this work, we propose a simple, intuitive, and efficient SAR despeckling approach grounded in compressive sensing theory. By applying Log-Yeo-Johnson transformation, we convert gamma-distributed noise into an approximate Gaussian distribution to noise sparse assumption. The method incorporates noise and sparsity priors, leveraging a non-local sparse representation through auxiliary matrices: one capturing varying noise characteristics across regions and the other encoding adaptive sparsity information. Extensive experiments validate the effectiveness of our method.

Figures

Figures reproduced from arXiv: 2412.18121 by the authors.

Figure 1
Figure 1. Synthetic image despeckling, row from top to bottom: noisy image, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Real image despeckling, column from right to left: noisy image, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Images and Noise distributions before and after Log-Yeo-Johnson [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation study, column from right to left: noisy image, denoised image [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. $\gamma$-Bridge: A Look-Parametric Diffusion Bridge

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A single diffusion model trained on synthetic single-look Gamma noise restores SAR images over the full (input look, output look) grid and transfers zero-shot to six real sensors, by making bridge time equal the physi...

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [8]

    Despeckling sar images with log-yeo-johnson transformation and conditional diffusion models,

    Y . Ma, P. Ke, H. Aghababaei, L. Chang, and J. Wei, “Despeckling sar images with log-yeo-johnson transformation and conditional diffusion models,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    work page 2024

  2. [1]

    Speckle removal using dictionary learn- ing and pnp-based fast iterative shrinkage threshold algorithm,

    S. Baraha and A. K. Sahoo, “Speckle removal using dictionary learn- ing and pnp-based fast iterative shrinkage threshold algorithm,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2023

    work page 2023

  3. [2]

    Sar image despeckling using a convolutional neural network,

    P. Wang, H. Zhang, and V . M. Patel, “Sar image despeckling using a convolutional neural network,” IEEE Signal Processing Letters , vol. 24, no. 12, pp. 1763–1767, 2017

    work page 2017

  4. [3]

    A total variation recursive space- variant filter for image denoising,

    M. V . Afonso and J. M. Sanches, “A total variation recursive space- variant filter for image denoising,” Digital Signal Processing , vol. 40, pp. 101–116, 2015

    work page 2015

  5. [4]

    A new nonconvex approach for image restoration with gamma noise,

    L. Bai, “A new nonconvex approach for image restoration with gamma noise,” Computers & Mathematics with Applications , vol. 77, no. 10, pp. 2627–2639, 2019

    work page 2019

  6. [5]

    Sar despeckling via regional denoising diffusion probabilistic model,

    X. Hu, Z. Xu, Z. Chen, Z. Feng, M. Zhu, and L. Stankovi ´c, “Sar despeckling via regional denoising diffusion probabilistic model,” in IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024, pp. 7226–7230

    work page 2024

  7. [6]

    Edge preserved low-rank sar image despeckling via hierarchical prior knowl- edge regulation,

    Z. Xu, X. Feng, S. Tian, X.-J. Shen, H. Zhang, and C. Wang, “Edge preserved low-rank sar image despeckling via hierarchical prior knowl- edge regulation,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–17, 2023

    work page 2023

  8. [7]

    Image denoising using ransac and compressive sensing,

    I. Stankovi ´c, M. Brajovi ´c, J. Lerga, M. Dakovi ´c, and L. Stankovi ´c, “Image denoising using ransac and compressive sensing,” Multimedia tools and applications , vol. 81, no. 30, pp. 44 311–44 333, 2022

    work page 2022

Show all 19 references
  1. [9]

    Image denoising via learned dictionaries and sparse representation,

    M. Elad and M. Aharon, “Image denoising via learned dictionaries and sparse representation,” in 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06) , vol. 1. IEEE, 2006, pp. 895–900

    2006

  2. [10]

    Stochastic alternating direction method of multipliers,

    H. Ouyang, N. He, L. Tran, and A. Gray, “Stochastic alternating direction method of multipliers,” in International conference on machine learning. PMLR, 2013, pp. 80–88

    2013

  3. [11]

    A trilateral weighted sparse coding scheme for real-world image denoising,

    J. Xu, L. Zhang, and D. Zhang, “A trilateral weighted sparse coding scheme for real-world image denoising,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 20–36

    2018

  4. [12]

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

    M. Aharon, M. Elad, and A. Bruckstein, “K-svd: An algorithm for designing overcomplete dictionaries for sparse representation,” IEEE Transactions on signal processing, vol. 54, no. 11, pp. 4311–4322, 2006

    2006

  5. [13]

    Asymptotic non-local means image denoising algorithm,

    X. Xiao-Xiao, W. Hai-Long, L. Jian, and Z. Xuan-De, “Asymptotic non-local means image denoising algorithm,” Acta Automatica Sinica , vol. 46, no. 9, pp. 1952–1960, 2020

    1952

  6. [14]

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

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,” IEEE transactions on image processing , vol. 26, no. 7, pp. 3142–3155, 2017

    2017

  7. [15]

    Sar image despeckling using continuous attention module,

    J. Ko and S. Lee, “Sar image despeckling using continuous attention module,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 15, pp. 3–19, 2022

    2022

  8. [16]

    Sar2sar: A semi-supervised despeckling algorithm for sar images,

    E. Dalsasso, L. Denis, and F. Tupin, “Sar2sar: A semi-supervised despeckling algorithm for sar images,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 14, pp. 4321– 4329, 2021

    2021

  9. [17]

    Agsdnet: Attention and gradient-based sar denoising network,

    R. K. Thakur and S. K. Maji, “Agsdnet: Attention and gradient-based sar denoising network,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2022

    2022

  10. [18]

    Sifsdnet: Sharp image feature based sar denoising network,

    ——, “Sifsdnet: Sharp image feature based sar denoising network,” in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2022, pp. 3428–3431

    2022

  11. [19]

    Analysis on the building of training dataset for deep learning sar despeckling,

    S. Vitale, G. Ferraioli, and V . Pascazio, “Analysis on the building of training dataset for deep learning sar despeckling,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2021

    2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.