Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Learning Multi-scale Spatial-frequency Features for Image Denoising

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Image denoising improves when a network is fed a pyramid of downscaled noisy images and learns to split its Fourier spectrum into high- and low-frequency bands with an adaptive mask; the proposed MADNet reports 39.78 dB on SIDD and 39.98…

desk verdict Plausible incremental denoising paper whose global fusion block is undefined as written; worth a referee's time if the resizing omission is fixed. read the letter →

arxiv 2506.16307 v1 pith:HV4KN7J5 submitted 2025-06-19 cs.CV cs.AIeess.IV

Xu Zhao , Chen Zhao , Xiantao Hu , Hongliang Zhang , Ying Tai , Jian Yang This is my paper · ORCID
classification cs.CVcs.AIeess.IV
keywords imagedenoisingmulti-scalearchitecturepyramidfrequencydomainFastFourierTransformadaptivemasktransposedself-attentionreal-worldnoise
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 image denoising improves when the network sees the noisy photo at several resolutions at once and treats high-frequency and low-frequency noise as separate problems. Its architecture, MADNet, feeds a downscaled image pyramid into a U-shaped network, uses a learnable mask in the Fourier domain to split features into high- and low-frequency branches that are refined with transposed self-attention, and fuses features from all scales in the skip connections. On real noisy photos, the paper reports a peak signal-to-noise ratio of 39.78 dB on the SIDD benchmark and 39.98 dB on the DND benchmark, ahead of the previous methods it compares against. If these numbers hold, the lesson is that scale and frequency structure are information, not just nuisance: explicitly exposing both to the network beats a single-input U-Net.

What carries the argument

The load-bearing object is the adaptive spatial-frequency learning unit (ASFU), which contains an adaptive frequency enhancement block (AFEB) and an adaptive spatial enhancement block (ASEB). AFEB applies a fast Fourier transform, multiplies the spectrum by a learnable rectangular mask whose dimensions scale with the input size, and uses the mask to split the signal into high- and low-frequency components; each branch is transformed back, refined by an MLP and by transposed self-attention, then added together. ASEB is that transposed self-attention, which models long-range channel dependencies. The global feature fusion block (GFFB) concatenates the four pyramid-scale features along the channel dimension and applies ASEB to produce a unified global feature. Together these units let the network adaptively choose how much of each frequency band and each scale to trust, which is the mechanism behind the claimed denoising gains.

What would settle it

Run the released implementation on a 256x256 dummy input and print the spatial shapes of the four features entering the global feature fusion block; if they are H, H/2, H/4, and H/8 and are concatenated without interpolation, the operation fails, showing the architecture is not as described. Independently, re-training with a fixed random mask instead of the learned adaptive mask would show whether the learned separation drives the reported PSNR gain.

Watch

Extended reading notes

Core claim

MADNet reaches state-of-the-art denoising performance on both synthetic and real noisy images by combining three mechanisms: multi-scale pyramid inputs that supply less-noisy low-resolution views of the same scene; an adaptive spatial-frequency learning unit (ASFU) whose learnable mask separates the Fourier spectrum into high- and low-frequency components, processes each band with MLPs and transposed self-attention, and recombines them; and a global feature fusion block (GFFB) that concatenates deep features from all four scales so coarse semantic structure can enhance fine detail. The paper reports PSNR 39.78 dB on SIDD and 39.98 dB on DND, and its ablation study attributes the gain to all three components rather than to any single one.

Load-bearing premise

The central claim rests on a concatenation of feature maps that have different resolutions in the global feature fusion block, and the paper does not specify the resizing that would make this concatenation possible.

Editorial extensions

If this is right

  • A single trained MADNet model should restore clean images at all four pyramid scales simultaneously, so the multi-scale outputs could be ensembled or used as extra supervision without extra inference cost.
  • The learnable frequency-separation mask means the network can adapt its high/low split to the noise level, so the same architecture should work across the tested sigma=15 to sigma=50 range and on real sensor noise.
  • If the reported benchmarks hold, MADNet becomes the reference point for real-image denoising on SIDD and DND, so future methods will need to beat 39.78 dB and 39.98 dB.
  • The ablation results imply that neither scale information nor frequency separation alone accounts for the gain; both are needed, so simpler single-domain or single-scale variants should underperform the full model.

Reading between the lines

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

  • Because the pyramid input works by presenting the network with progressively less noisy versions of the same scene, the same trick could transfer to other restoration tasks such as deblurring or super-resolution, where low-resolution views also carry cleaner global information.
  • The learned frequency mask is a potential interpretability probe: extracting the trained mask and plotting its shape across noise levels would show whether the network settles on a fixed high/low split or adapts per image, a test the paper does not run.
  • The reported gaps over the nearest competitors are small (about 0.2 to 0.3 dB on SIDD and DND), and the paper does not report variance across training runs; a multi-seed rerun would indicate whether the improvement is systematic or within noise.
  • If the global feature fusion concatenation requires an interpolation step that the paper omits, the fix is straightforward, but the description in Eq. (9) needs that step spelled out before the architecture can be reproduced as written.
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 / 5 minor

Summary. The paper proposes MADNet, a multi-scale adaptive dual-domain network for image denoising. The architecture uses image pyramid inputs, a dual-domain modulation layer with an adaptive spatial-frequency learning unit (ASFU) that separates high- and low-frequency components via a learnable mask, and a global feature fusion block (GFFB) in skip connections. The authors report experiments on synthetic (CBSD68, Set12) and real (SIDD, DND) datasets, claiming state-of-the-art PSNR/SSIM results. The central claim is that the combination of multi-scale inputs, frequency separation, and global feature fusion improves denoising performance.

Significance. If the architecture is implementable as described, the paper makes a plausible empirical contribution by combining multi-scale pyramid inputs, adaptive frequency separation, and cross-scale feature fusion, and the reported gains on SIDD and DND are nonzero. The ablation study attempts to attribute improvements to individual components, and the paper includes both real and synthetic noise benchmarks. However, the method description contains a fundamental tensor-dimension inconsistency in the GFFB, and the ablation table has internal inconsistencies; until these are resolved, the central architectural innovation and its claimed benefit cannot be verified. The SOTA claim is also weakened by the absence of several widely used recent baselines and the lack of error bars. The work is not circular: metrics come from held-out test sets and the learnable mask is a trained parameter.

major comments (4)
  1. [Section 3.4, Eq. (9)] The global feature fusion block concatenates features from different pyramid scales along the channel dimension, but Section 3.1 defines these features as F_i ∈ R^{H/n × W/n × nC} with n = 1, 2, 4, 8. Concatenation along channels requires identical spatial dimensions; no upsampling, interpolation, or pooling operation is described anywhere in the paper. Since Table 5 attributes a 0.063 dB improvement (39.777 vs 39.714 dB) to GFFB, this is a load-bearing definitional gap. Please specify how the resolution mismatch is resolved (e.g., bilinear upsampling to the finest scale) and update Figure 3 and the ablation description accordingly.
  2. [Section 3.3.1, Eq. (4) and Section 3.4, Eq. (10)] The mask operation in Eq. (4), F_high, F_low = Mask(F_freq), is not defined: the paper does not state how the binary mask is constructed, how the learnable coefficient controls the mask dimensions, or whether the separation is an element-wise multiplication in the Fourier domain. In addition, Eq. (10) uses ASEB(F_g, F_i) as a two-argument operation, while Eq. (7) defines transposed self-attention for a single input feature. The intended mechanism by which the global feature F_g modulates each scale-specific feature is unspecified. These details are needed to reproduce the core ASFU and GFFB components.
  3. [Table 5] The ablation table contains internally inconsistent checkmark patterns. For example, the rows labeled 'AFEB w/o Sep' and 'AFEB w/o Enh' have ASEB=✗ and AFEB=✗, meaning both the spatial and frequency enhancement blocks are removed, which does not match the row labels. The reported values 39.625 and 39.621 dB therefore cannot be interpreted as isolating the effect of 'Sep' or 'Enh' alone. The text in Section 4.6.2 also says 'Removal of frequency separation (39.670 dB)' but that row (ASEB+AFEB w/o Sep) has a different configuration. Please correct the table so the ablation conditions match the components being removed, or clarify the intended experimental design.
  4. [Tables 2-4] The claim of state-of-the-art performance is not adequately supported by the chosen baselines. The comparisons omit several widely used recent denoisers, including Restormer, Uformer, and other top-performing methods on SIDD and DND, and no error bars or statistical significance tests are provided. The reported margins over the strongest included baseline (DRANet) are small (e.g., 0.24 dB on SIDD, 0.35 dB on DND), so the SOTA claim should be either backed by comparisons against those methods or rephrased as competitive performance against the included baselines.
minor comments (5)
  1. [Section 4.6.1, Table 5] The text 'removing both results in a more noticeable 0.04 dB reduction (39.695 dB vs 39.735 dB)' does not match the table: the value 39.735 dB does not appear in Table 5, and the actual drop from the full model (39.777 dB) to 'w/o MSI & GFFB' (39.695 dB) is 0.082 dB. Please correct the numbers.
  2. [Section 4.4, Table 3] The text states that on Set12 the method 'maintains leading PSNR values of 33.05 dB at σ=15', but Table 3 lists MADNet's Set12 PSNR at σ=15 as 33.00 dB. Please align the text with the table.
  3. [Section 4.2 and Table 2] Section 4.2 says evaluation is performed at noise levels 15, 25, 30, and 50, but the caption and text in Section 4.3 refer to 'three noise levels (σ=15, 25, 50)' while the table actually includes four columns (σ=15, 25, 30, 50). Please correct the inconsistent wording.
  4. [Table 2] The DudeNet row is duplicated in the CBSD68 block.
  5. [Section 4.7] Section 4.7 is titled 'Limitation' but it reports FLOPs and parameter counts; the actual limitation (the large parameter count) is discussed in Section 5. Please consider retitling or reorganizing this section.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical benchmark results with independently evaluated held-out test sets, and the cited self-works are related-work background rather than load-bearing evidence.

full rationale

MADNet is an empirical architecture paper. The central claims are PSNR/SSIM numbers on CBSD68, Set12, SIDD, and DND, which are external benchmarks evaluated on held-out test sets. No parameter is fitted to the test data and then reported as a prediction; the learnable frequency mask and network weights are optimized on training splits (Flickr2K, SIDD Medium) only. The multi-scale pyramid inputs, ASFU, and GFFB are architectural components whose contributions are assessed by ablations, not by construction that forces the reported improvements. The loss functions in Eqs. (11)-(13) supervise training with ground-truth targets, and the final metrics are computed independently on test benchmarks. The paper does cite works by overlapping authors (e.g., [19], [30], [33], [34]), but these appear in related-work and frequency-learning context, not as load-bearing justification of MADNet's design or performance; none is invoked as a uniqueness theorem or as a substitute for experimental evidence. The strongest concern in the manuscript is that Eq. (9) concatenates features of different spatial resolutions along the channel dimension without describing a resizing operation; this is an implementability and reproducibility defect, not a circularity defect, because it does not make the claimed result equivalent to an input by construction. Accordingly, no circular step meeting the required evidence standard is present.

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

The paper introduces no physical or conceptual entities beyond network modules. The frequency mask and fusion block are architecture components validated only by the paper's own ablations, so no independent evidence exists outside the paper.

free parameters (3)
  • Number of pyramid scales = 4
    Chosen by hand; used to define the multi-scale input and output; not ablated beyond a single configuration.
  • Loss scale weights alpha_s = not reported
    Weights in the multi-scale Charbonnier and frequency losses (Eq. 11-12); critical for training but no values given.
  • Learnable frequency mask coefficient = not reported
    Controls the rectangular high/low frequency split in ASFU (Section 3.3.1); learned during training but its scale or final value is not reported.
assumptions (3)
  • standard math FFT and IFFT are defined as in Eq. (2) and can be used inside a differentiable network.
    Section 3.3 relies on FFT for frequency separation; this is a standard mathematical tool, but the paper does not discuss finite-size or boundary effects.
  • domain assumption Low-resolution versions of a noisy image contain less noise and provide cleaner contextual information.
    Used to justify pyramid inputs (Section 1, Figure 1, Table 1); this is an empirical observation, not proven, and the effect depends on the downsampling method.
  • ad hoc to paper The adaptive binary frequency mask is trainable by gradient descent.
    Section 3.3.1 uses a binary mask with a learnable coefficient but does not describe a differentiable relaxation; without a soft mask, the gradient would be zero almost everywhere.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Multi-scale Spatial-frequency Features for Image Denoising." pith.science (2026). https://pith.science/paper/HV4KN7J5

@misc{pith2026250616307,
  author       = {Pith},
  title        = {Pith review of: Learning Multi-scale Spatial-frequency Features for Image Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HV4KN7J5}},
  note         = {Machine review of arXiv:2506.16307}
}
read the original abstract

Recent advancements in multi-scale architectures have demonstrated exceptional performance in image denoising tasks. However, existing architectures mainly depends on a fixed single-input single-output Unet architecture, ignoring the multi-scale representations of pixel level. In addition, previous methods treat the frequency domain uniformly, ignoring the different characteristics of high-frequency and low-frequency noise. In this paper, we propose a novel multi-scale adaptive dual-domain network (MADNet) for image denoising. We use image pyramid inputs to restore noise-free results from low-resolution images. In order to realize the interaction of high-frequency and low-frequency information, we design an adaptive spatial-frequency learning unit (ASFU), where a learnable mask is used to separate the information into high-frequency and low-frequency components. In the skip connections, we design a global feature fusion block to enhance the features at different scales. Extensive experiments on both synthetic and real noisy image datasets verify the effectiveness of MADNet compared with current state-of-the-art denoising approaches.

Figures

Figures reproduced from arXiv: 2506.16307 by the authors.

Figure 1
Figure 1. Comparison of degradation magnitude of low-resolution images. The observed degradation in low [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of degradation magnitude of low-resolution images. The degradation of the image is [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The structure of the MADNet. A Unet structure processes multi-scale pyramid inputs with dual [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Detailed architecture of each block in MADNet. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Feature visualization across fusion layers: (a) Noisy image (b)-(e) Fusion Layer feature heatmaps. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on AWGN color images from the CBSD68 dataset. From left to right, we show [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results on AWGN gray images from the CBSD68 dataset. From left to right, we show [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on real noisy images from the SIDD validation dataset. From left to right, we [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Zheng, J

    Y . Zheng, J. Zhan, S. He, J. Dong, Y . Du, Curricular contrastive regularization for physics-aware single image dehazing, in: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, 2023, pp. 5785–5794

    work page 2023

  2. [2]

    J. Bao, X. Tao, Y . Zhou, An emotion recognition method based on eye movement and audiovisual features in mooc learning environment, IEEE Transactions on Computational Social Systems 11 (1) (2022) 171–183

    work page 2022

  3. [3]

    Dabov, A

    K. Dabov, A. Foi, V . Katkovnik, K. Egiazarian, Image denoising with block- matching and 3d filtering, in: Image processing: algorithms and systems, neural networks, and machine learning, V ol. 6064, SPIE, 2006, pp. 354–365

    work page 2006

  4. [4]

    X. Xie, Y . Wu, H. Ni, C. He, Node-imgnet: a pde-informed effective and robust model for image denoising, Pattern Recognition 148 (2024) 110176. 19

    work page 2024

  5. [5]

    E. Ning, Y . Wang, C. Wang, H. Zhang, X. Ning, Enhancement, integration, ex- pansion: Activating representation of detailed features for occluded person re- identification, Neural Networks 169 (2024) 532–541

    work page 2024

  6. [6]

    X. Ji, Y . Cao, Y . Tai, C. Wang, J. Li, F. Huang, Real-world super-resolution via kernel estimation and noise injection, in: proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition workshops, 2020, pp. 466–467

    work page 2020

  7. [7]

    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 (7) (2017) 3142–3155

    work page 2017

  8. [8]

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, Restormer: Efficient transformer for high-resolution image restoration, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

    2022

Show all 46 references
  1. [9]

    Fan, T.-J

    C.-M. Fan, T.-J. Liu, K.-H. Liu, Sunet: Swin transformer unet for image denois- ing, in: 2022 IEEE International Symposium on Circuits and Systems (ISCAS), IEEE, 2022, pp. 2333–2337

    2022

  2. [10]

    E. Ning, W. Li, J. Fang, J. Yuan, Q. Duan, G. Wang, 3d-guided multi-feature semantic enhancement network for person re-id, Information Fusion 117 (2025) 102863

    2025

  3. [11]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10012–10022

    2021

  4. [12]

    Zhang, M

    K. Zhang, M. Long, J. Chen, M. Liu, J. Li, Cfpnet: A denoising network for complex frequency band signal processing, IEEE Transactions on Multimedia 25 (2023) 8212–8224

    2023

  5. [13]

    M. Yan, J. Qian, R. Wang, S. Gao, J. Yang, Dynamic group difference coding based on thermal infrared face image for fever screening, IEEE Transactions on Instrumentation and Measurement 72 (2023) 1–13. 20

    2023

  6. [14]

    R. Ma, S. Li, B. Zhang, Z. Li, Generative adaptive convolutions for real-world noisy image denoising, in: Proceedings of the AAAI conference on artificial in- telligence, V ol. 36, 2022, pp. 1935–1943

    2022

  7. [15]

    M. Zhou, J. Huang, C.-L. Guo, C. Li, Fourmer: An efficient global modeling paradigm for image restoration, in: International conference on machine learning, PMLR, 2023, pp. 42589–42601

    2023

  8. [16]

    X. Hu, B. Zhong, Q. Liang, S. Zhang, N. Li, X. Li, Towards modalities corre- lation for rgb-t tracking, IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  9. [17]

    E. Ning, C. Wang, H. Zhang, X. Ning, P. Tiwari, Occluded person re- identification with deep learning: a survey and perspectives, Expert systems with applications 239 (2024) 122419

    2024

  10. [18]

    S. Chen, Y . Chen, Y . Zheng, Z.-X. Yang, E. Wu, A transformer-based adaptive prototype matching network for few-shot semantic segmentation, in: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024, pp. 659–667

    2024

  11. [19]

    X. Hu, Y . Tai, X. Zhao, C. Zhao, Z. Zhang, J. Li, B. Zhong, J. Yang, Exploiting multimodal spatial-temporal patterns for video object tracking, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 39, 2025, pp. 3581–3589

    2025

  12. [20]

    L. Wang, S. Li, F. Yang, J. Wang, Z. Zhang, Y . Liu, Y . Wang, J. Yang, Not all parameters matter: Masking diffusion models for enhancing generation ability, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12880–12890

    2025

  13. [21]

    Z. Wang, Z. Yan, J. Pan, G. Gao, K. Zhang, J. Yang, Dornet: A degradation ori- ented and regularized network for blind depth super-resolution, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 15813– 15822. 21

    2025

  14. [22]

    Y . Wu, Z. Yan, Z. Wang, X. Li, L. Hui, J. Yang, Deep height decoupling for precise vision-based 3d occupancy prediction, arXiv preprint arXiv:2409.07972 (2024)

    2024 arXiv

  15. [23]

    Zhang, S

    H. Zhang, S. Chen, L. Luo, J. Yang, Few-shot learning with long-tailed labels, Pattern Recognition 156 (2024) 110806

    2024

  16. [24]

    C. Ren, X. He, C. Wang, Z. Zhao, Adaptive consistency prior based deep network for image denoising, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8596–8606

    2021

  17. [25]

    R. K. Thakur, S. K. Maji, Multi scale pixel attention and feature extraction based neural network for image denoising, Pattern recognition 141 (2023) 109603

    2023

  18. [26]

    W. Wu, S. Liu, Y . Xia, Y . Zhang, Dual residual attention network for image de- noising, Pattern Recognition 149 (2024) 110291

    2024

  19. [27]

    X. An, L. Zhao, C. Gong, N. Wang, D. Wang, J. Yang, SHaRPose: Sparse High- Resolution Representation for Human Pose Estimation, Proceedings of the AAAI Conference on Artificial Intelligence 38 (2) (2024) 691–699

    2024

  20. [28]

    E. Ning, C. Zhang, C. Wang, X. Ning, H. Chen, X. Bai, Pedestrian re-id based on feature consistency and contrast enhancement, Displays 79 (2023) 102467

    2023

  21. [29]

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, H. Li, Uformer: A general u-shaped transformer for image restoration, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17683–17693

    2022

  22. [30]

    C. Zhao, W. Cai, C. Hu, Z. Yuan, Cycle contrastive adversarial learning with structural consistency for unsupervised high-quality image deraining transformer, Neural Networks (2024) 106428

    2024

  23. [31]

    X. Hu, B. Zhong, Q. Liang, S. Zhang, N. Li, X. Li, R. Ji, Transformer track- ing via frequency fusion, IEEE Transactions on Circuits and Systems for Video Technology 34 (2) (2023) 1020–1031. 22

    2023

  24. [32]

    Fuoli, L

    D. Fuoli, L. Van Gool, R. Timofte, Fourier space losses for efficient perceptual image super-resolution, in: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, 2021, pp. 2360–2369

    2021

  25. [33]

    C. Zhao, W. Cai, C. Dong, Z. Zeng, Toward sufficient spatial-frequency inter- action for gradient-aware underwater image enhancement, in: ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2024, pp. 3220–3224

    2024

  26. [34]

    C. Zhao, W. Cai, C. Dong, C. Hu, Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 8281–8291

    2024

  27. [35]

    Lai, J.-B

    W.-S. Lai, J.-B. Huang, N. Ahuja, M.-H. Yang, Fast and accurate image super- resolution with deep laplacian pyramid networks, IEEE transactions on pattern analysis and machine intelligence 41 (11) (2018) 2599–2613

    2018

  28. [36]

    B. Lim, S. Son, H. Kim, S. Nah, K. Mu Lee, Enhanced deep residual networks for single image super-resolution, in: Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 2017, pp. 136–144

    2017

  29. [37]

    S. Roth, M. J. Black, Fields of experts: A framework for learning image priors, in: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), V ol. 2, IEEE, 2005, pp. 860–867

    2005

  30. [38]

    Abdelhamed, S

    A. Abdelhamed, S. Lin, M. S. Brown, A high-quality denoising dataset for smart- phone cameras, in: IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018

    2018

  31. [39]

    Plotz, S

    T. Plotz, S. Roth, Benchmarking denoising algorithms with real photographs, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1586–1595

    2017

  32. [40]

    C. Tian, Y . Xu, W. Zuo, B. Du, C.-W. Lin, D. Zhang, Designing and training of a dual cnn for image denoising, Knowledge-Based Systems 226 (2021) 106949. 23

    2021

  33. [41]

    Liang, J

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, R. Timofte, Swinir: Image restoration using swin transformer, in: Proceedings of the IEEE/CVF interna- tional conference on computer vision, 2021, pp. 1833–1844

    2021

  34. [42]

    Y . Gou, P. Hu, J. Lv, J. T. Zhou, X. Peng, Multi-scale adaptive network for single image denoising, Advances in Neural Information Processing Systems 35 (2022) 14099–14112

    2022

  35. [43]

    B. Li, X. Liu, P. Hu, Z. Wu, J. Lv, X. Peng, All-in-one image restoration for unknown corruption, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17452–17462

    2022

  36. [44]

    C. Tian, M. Zheng, W. Zuo, B. Zhang, Y . Zhang, D. Zhang, Multi-stage image denoising with the wavelet transform, Pattern Recognition 134 (2023) 109050

    2023

  37. [45]

    J. W. Soh, N. I. Cho, Variational deep image restoration, IEEE Transactions on Image Processing 31 (2022) 4363–4376

    2022

  38. [46]

    J. Du, X. Qiao, Z. Yan, H. Zhang, W. Zuo, Flexible image denoising model with multi-layer conditional feature modulation, Pattern Recognition 152 (2024) 110372. 24

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.