Pith. sign in

REVIEW 4 major objections 6 minor 67 references

Moir\'e Zero: An Efficient and High-Performance Neural Architecture for Moir\'e Removal

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read MZNet claims that a convolutional network built from multi-scale dilated attention and multi-shape large-kernel convolutions removes moiré artifacts from 4K images more effectively than any prior demoiréing method, reporting PSNR gains of…

desk verdict A competent, honest architecture paper for demoiréing whose SOTA claim is undercut by an unshared TLC evaluation and a misaligned benchmark; worth reviewing but the numbers need a controlled rerun. read the letter →

arxiv 2507.22407 v1 pith:2WFRADXD submitted 2025-07-30 cs.CV eess.IV

classification cs.CVeess.IV
keywords moiréremovaldemoiréingimagerestorationlarge-kernelconvolutiondilateddualattentionfeaturefusionskipconnectionUHDbenchmarks
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 sets out to prove that a convolutional network can be the best practical tool for removing moiré artifacts from high-resolution photos, without needing a Transformer's cost. Its diagnosis is that prior CNN demoiréing methods fail mainly because small receptive fields cannot span the large, oriented, multi-scale bands moiré creates, and its proposed remedy is MZNet, a U-shaped network whose three components — multi-scale dilated convolution with dual attention, multi-shape large-kernel convolution at the bottleneck, and all-level skip-connection fusion — widen the receptive field cheaply. On the FHDMi and UHDM benchmarks, MZNet reports PSNR gains of 0.67 dB and 0.33 dB over the previous best method, with the best LPIPS on both, at about 1.19 T MACs for a 4K input. The paper also establishes that the UHDM ground truth is spatially misaligned, and that aligning it lifts every model's PSNR by roughly 0.9 dB, a caveat underneath the headline comparison. A reader comes away with a concrete picture: efficient demoiréing at 4K is reachable with a carefully widened CNN.

What carries the argument

The load-bearing device is a cheaply expanded receptive field. The Multi-Dilation Convolution Module runs four parallel $3\times 3$ depth-wise convolutions with dilation rates $\{1,4,7,9\}$, giving receptive fields of sizes $\{3,9,15,19\}$ so that one block sees moiré at several scales at once. The Multi-Shape Large Kernel Convolution Block sits at the lowest-resolution bottleneck and applies parallel depth-wise kernels of shapes $K\times K$, $K\times 1$, and $1\times K$, where $K$ is the largest odd number that fits the bottleneck feature (23 on UHDM), capturing elongated structures along different orientations. The Dual Attention Module combines Large Kernel Attention for spatial coverage with Simplified Channel Attention for channel reweighting, and the Feature Fusion-Based Skip Connection bilinearly interpolates every encoder feature to each decoder resolution and concatenates them, so no scale information is lost in the skip path. A Test-time Local Converter is applied at inference to reduce the mismatch between training crops and full-resolution test images. Each of these pieces is ablated, and each contributes to the final score.

What would settle it

Re-run the strongest baselines, P-BiC and ESDNet-L, under MZNet's exact settings (same crop sizes, losses, epochs, and the Test-time Local Converter at inference) and evaluate on the SIFT-aligned UHDM ground truth; if either baseline reaches or exceeds MZNet's reported scores, the claimed advantage is an evaluation artifact rather than an architectural one. A minimal version is to check whether P-BiC's FHDMi score moves when TLC is added.

Watch

Extended reading notes

Core claim

The central claim is that moiré removal is primarily a receptive-field problem, and that an efficient CNN can solve it. MZNet is a four-level U-shaped network in which every encoder and decoder stage uses a Multi-Scale Dual Attention Block — parallel $3\times 3$ depth-wise dilated convolutions with rates $\{1,4,7,9\}$, refined jointly by Large Kernel Attention and Simplified Channel Attention — and the bottleneck uses a Multi-Shape Large Kernel Convolution Block that runs parallel depth-wise kernels of shapes $K\times K$, $K\times 1$, and $1\times K$ to capture elongated moiré structures in different orientations. Feature Fusion-Based Skip Connections resize every encoder feature to each decoder resolution and concatenate them, so each decoder stage sees all scales at once. On FHDMi the paper reports 26.12 dB PSNR, 0.67 dB above the prior best P-BiC, and on UHDM 23.632 dB, 0.33 dB above P-BiC, with the best LPIPS on both datasets and 1.190 T MACs at 4K; on the low-resolution TIP2018 dataset it places second. The authors also establish that the UHDM dataset's input and ground truth are misaligned, and that SIFT-based realignment improves every method's scores, with MZNet gaining most (23.63 to 24.50 dB).

Load-bearing premise

The numbers reported for the earlier methods were produced under conditions comparable to MZNet's, so the PSNR and LPIPS gains come from the architecture itself and not from evaluation choices such as the Test-time Local Converter applied to MZNet at inference or how the misaligned UHDM ground truth was handled.

Editorial extensions

If this is right

  • High-resolution demoiréing no longer requires Transformer-scale compute: MZNet runs at 1.190 T MACs for a 4K input, the low end among top methods, and a lightweight variant at 0.596 T MACs keeps most of the gain.
  • The UHDM benchmark carries an alignment confound: realigning the ground truth improves every model by roughly 0.7–0.9 dB PSNR, so future comparisons should report scores on the aligned set or control for alignment.
  • Display inspection can drop the defocusing step: MZNet removes moiré while preserving the relative brightness variations between adjacent pixels that mark defects.
  • The ablation study supports the orientation-coverage story: removing Large Kernel Attention costs 0.48 dB PSNR, and removing the horizontal and vertical stripe kernels hurts more than removing the square kernel.
  • Because the architecture achieves its wide receptive field through depth-wise dilated and large kernels, it demonstrates that a CNN can capture long-range moiré structure without self-attention.

Reading between the lines

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

  • Editorial inference: if the receptive-field diagnosis is correct, the same component recipe — multi-scale dilation, multi-shape large kernels, and all-level feature fusion — should transfer to other periodic or aliasing artifacts such as screen-content compression or sensor-CFA aliasing, a claim the paper does not make and does not test.
  • Editorial inference: the paper does not report whether the Test-time Local Converter changes the baselines' scores; that single experiment is the cheapest way to test whether the headline gains are architectural or procedural.
  • Editorial inference: the alignment table suggests part of the UHDM ranking reflects how models cope with misaligned targets rather than pure removal skill; an independently aligned re-benchmark could reshuffle the leaderboard.
  • Editorial inference: because MSLKB's kernel size $K$ is hand-tuned per dataset resolution, learning or adaptively selecting the kernel shapes would test whether orientation coverage, rather than the tuning, drives the gain.
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 / 6 minor

Summary. The paper proposes MZNet, a U-shaped CNN for moiré removal that combines three components: Multi-Scale Dual Attention Block (MSDAB), Multi-Shape Large Kernel Convolution Block (MSLKB), and Feature Fusion-Based Skip Connection (FFSC). The model is evaluated on TIP2018, FHDMi, and UHDM against prior demoiréing methods, with the authors reporting state-of-the-art PSNR/SSIM/LPIPS on FHDMi and UHDM while keeping MACs low (1.190 T at 4K input). The paper also includes component ablations, an analysis of ground-truth misalignment in UHDM, and a demonstration on real-world display inspection images.

Significance. If the head-to-head results hold, MZNet would be a genuinely efficient and effective high-resolution demoiréing architecture: the FHDMi and UHDM gains over prior work are sizable, the components are clearly described, and the ablation study is informative. The paper deserves credit for explicitly disclosing the UHDM alignment problem and quantifying its effect in Table 4, and for stating limitations in the supplementary material. The main significance claim is, however, currently weakened by the asymmetric test-time protocol used for MZNet only and by the unresolved benchmark-alignment issue; these need to be addressed before the state-of-the-art claim is fully supported.

major comments (4)
  1. [Supplementary 'Dataset and Experimental Setup'; Table 1] The state-of-the-art comparison is run under an asymmetric inference protocol. The supplementary states that 'we employed the Test-time Local Converter (TLC) ... to ensure consistency by adapting local features during inference,' and it is applied only to MZNet; the baselines in Table 1 are evaluated without TLC. Because the claimed margins over P-BiC are only 0.67 dB on FHDMi and 0.33 dB on UHDM, a generic test-time adaptation applied to the baselines could close or even reverse these margins. If TLC is instead a trainable component of MZNet, its parameters and inference cost need to be included in the reported 14.82 M parameters and 1.190 T MACs. Please re-evaluate all Table 1 competitors under a matched protocol (TLC applied to all models, or to none) and report the resulting numbers.
  2. [Section 'Moiré Dataset Alignment Issues'; Figure 6; Table 4] The UHDM comparison rests on a ground truth that the authors themselves show to be misaligned. Table 4 reports that applying SIFT-based alignment changes MZNet's PSNR from 23.63 to 24.50 dB and ESDNet-L's from 22.42 to 23.09 dB; both shifts are larger than MZNet's claimed 0.33 dB advantage over P-BiC in Table 1. Yet the aligned evaluation covers only MZNet and ESDNet-L, not P-BiC or the other Table 1 baselines, so it cannot establish how the ranking changes on a correctly aligned UHDM test set. The main UHDM results should be reported on a consistently aligned ground truth for all methods, or the authors should justify why the original misaligned ground truth is the appropriate benchmark.
  3. [Tables 1 and 2; 'Experimental settings'] No measure of repeated-run variability is reported for any table. The central margins (0.33 dB on UHDM, 0.67 dB on FHDMi) are small relative to the alignment-induced shifts in Table 4, and the ablation differences in Table 2 (e.g., 23.16 vs 23.04 vs 22.97 dB) are within the range that could change with training seed. Please report results over multiple training runs, at least for the main comparison and the key ablations, with means and standard deviations, or otherwise justify that the differences are statistically meaningful.
  4. [Table 1; 'Comparison with State-of-the-Art Methods'] The provenance of the prior-model numbers in Table 1 is not made explicit; they appear to be taken from the original papers rather than re-run under MZNet's training set-up (crop sizes, epochs, hardware, optimizer settings). Since training conditions differ across methods, the comparison is not fully controlled. Please state the source of each baseline number in the table or its caption, and, where feasible, re-run the closest competitors under the same training and inference protocol.
minor comments (6)
  1. [Figure 3 caption] The caption begins with 'llustration of the core components'; this is a typo for 'Illustration'.
  2. [Equation (3)] The definition of MSCM is ambiguous: the SCA term appears to be added to the sum of four depth-wise convolutions, but it is not clear whether SCA is applied to the input feature or to the fused output of the parallel branches. Please clarify the intended data flow.
  3. [Supplementary 'Dataset and Experimental Setup'] The main model is described as using four dilation rates {1,4,7,9} in MDCM, while the lightweight variants in the supplementary use three rates {1,4,9}. Since Eq. (1) is written for a fixed four-branch sum, please clarify whether the number of branches is variable or whether the lightweight variants use a different module.
  4. [Supplementary 'Dataset and Experimental Setup'] The statement that the MSLKB kernel size K 'is set to the largest odd number that does not exceed the size of the bottleneck feature' is not consistent with the reported values K=7, 15, and 23 for the stated input resolutions. Please give an operational definition of how K is chosen, including the effect of pixel unshuffle and downsampling on the bottleneck resolution.
  5. [Supplementary 'Limitations and Future work'] The limitations section correctly notes that downstream defect-detection metrics were not measured. The industrial-application claim in the main text should be softened accordingly, or quantitative inspection metrics (e.g., defect detection accuracy before and after demoiréing) should be provided.
  6. [Table 1 caption] The caption says 'MACs are for our configuration on the UHDM dataset,' while the supplementary reports dataset-specific MACs in Table 5. Please clarify whether all listed MACs use the UHDM configuration or provide per-dataset MACs for all models.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MZNet is an empirical benchmark paper whose claims rest on external benchmark metrics, not on fitted constants or a self-citation chain.

full rationale

MZNet's central claims are architectural and empirical: the contributions are the MSDAB, MSLKB, and FFSC modules, and the evidence is PSNR/SSIM/LPIPS on TIP2018, FHDMi, and UHDM. None of these quantities is defined in terms of the model's outputs by construction, and no parameter is fitted to the test benchmarks and then reported as a prediction. The dilation rates and MSLKB kernel size are disclosed as hand-tuned choices, but they do not define the target metrics. The ablation studies compare independently trained variants on held-out benchmark data, which is standard empirical evidence rather than circular reasoning. The most notable evaluation concerns are the use of Test-time Local Converter (TLC) at inference for MZNet without applying it to baselines, and the UHDM ground-truth alignment analysis; these are protocol-fairness and benchmark-validity issues, not circularity, because TLC is an external published technique and the alignment experiment does not define the paper's performance metric. The paper's references to NAFNet, LKA, SCA, and TLC are external works, and there is no load-bearing self-citation chain. No derivation in the paper reduces to its own inputs, so the appropriate finding is no circularity.

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

The method introduces no new physical entities or theoretical constructs. Its load-bearing choices are hyperparameters and standard empirical assumptions about benchmark validity and comparability.

free parameters (4)
  • MDCM dilation rates = {1, 4, 7, 9}
    Reported as empirically optimal in the supplementary; affects receptive field coverage.
  • MSLKB kernel size K = 7 (TIP2018), 15 (FHDMi), 23 (UHDM)
    Manually set per dataset based on bottleneck feature size; authors acknowledge dataset-specific tuning in Limitations.
  • Loss weight lambda = 1.0
    Empirically set to balance L1 and LPIPS losses in the Method section.
  • Stage block counts = Encoder 4,4,6,8; decoder 4,4,6,6
    Architectural capacity hyperparameters chosen by hand, not derived.
assumptions (4)
  • domain assumption Benchmark ground-truth images are sufficiently aligned for PSNR and SSIM to be meaningful.
    The paper itself shows substantial misalignment in UHDM (Fig. 6) and that alignment changes PSNR by about 0.9 dB (Table 4), so the reported benchmark numbers depend on this assumption.
  • domain assumption Baseline results cited from prior papers were obtained under training and inference settings comparable to MZNet's.
    No baselines are retrained; Table 1 numbers are mostly from prior publications, and MZNet uses TLC at inference while baselines may not.
  • domain assumption PSNR, SSIM, and LPIPS adequately capture moiré removal quality.
    Standard assumption in image restoration; no user study or task-specific metric is used.
  • domain assumption Real-world display inspection images share the degradation distribution of the training data.
    Only qualitative results are shown for inspection images (Figs. 7 and 8); no downstream defect-detection metric is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Moir\'e Zero: An Efficient and High-Performance Neural Architecture for Moir\'e Removal." pith.science (2026). https://pith.science/paper/2WFRADXD

@misc{pith2026250722407,
  author       = {Pith},
  title        = {Pith review of: Moir\'e Zero: An Efficient and High-Performance Neural Architecture for Moir\'e Removal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2WFRADXD}},
  note         = {Machine review of arXiv:2507.22407}
}
read the original abstract

Moir\'e patterns, caused by frequency aliasing between fine repetitive structures and a camera sensor's sampling process, have been a significant obstacle in various real-world applications, such as consumer photography and industrial defect inspection. With the advancements in deep learning algorithms, numerous studies-predominantly based on convolutional neural networks-have suggested various solutions to address this issue. Despite these efforts, existing approaches still struggle to effectively eliminate artifacts due to the diverse scales, orientations, and color shifts of moir\'e patterns, primarily because the constrained receptive field of CNN-based architectures limits their ability to capture the complex characteristics of moir\'e patterns. In this paper, we propose MZNet, a U-shaped network designed to bring images closer to a 'Moire-Zero' state by effectively removing moir\'e patterns. It integrates three specialized components: Multi-Scale Dual Attention Block (MSDAB) for extracting and refining multi-scale features, Multi-Shape Large Kernel Convolution Block (MSLKB) for capturing diverse moir\'e structures, and Feature Fusion-Based Skip Connection for enhancing information flow. Together, these components enhance local texture restoration and large-scale artifact suppression. Experiments on benchmark datasets demonstrate that MZNet achieves state-of-the-art performance on high-resolution datasets and delivers competitive results on lower-resolution dataset, while maintaining a low computational cost, suggesting that it is an efficient and practical solution for real-world applications. Project page: https://sngryonglee.github.io/MoireZero

Figures

Figures reproduced from arXiv: 2507.22407 by the authors.

Figure 1
Figure 1. Visualization of moire patterns from the UHDM ´ dataset. (A) shows the full image with moire patterns, while ´ (B), (C), and (D) present magnified regions. They illustrate the varying directions and scales of moire within a single ´ image. (E) is extracted from the ground truth corresponding to the same region as (D), highlighting the color distortions caused by moire patterns. ´ the receptive field of the network w… view at source ↗
Figure 2
Figure 2. Overview of MZNet. The model follows a U-Net architecture with four levels of encoders and decoders. MSDAB are [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. llustration of the core components in MZNet. Sim [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons with state-of-the-art methods on the UHDM dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons of ablation study. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: (a) Overlay of the moire image and ground truth ´ from the original UHDM dataset, showing noticeable mis￾alignment. (b) Overlay after additional alignment. Dataset MZNet (Ours) ESDNet-L PSNR↑ SSIM↑ LPIPS↓ PSNR↑ SSIM↑ LPIPS↓ Original 23.63 0.810 0.223 22.42 0.799 0.245 …
Figure 8
Figure 8. Figure 8: Challenging inspection scenario with synthetic [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparisons with state-of-the-art methods on the UHDM dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparisons with state-of-the-art methods on the UHDM dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparisons with state-of-the-art methods on the UHDM dataset. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparisons with state-of-the-art methods on the UHDM dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Qualitative comparisons with state-of-the-art methods on the UHDM dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparisons with state-of-the-art methods on the FHDMi dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Examples of moire artifact removal in display inspection images using MZNet. The left image shows the input with ´ visible moire patterns, while the right image presents the result after moir ´ e removal. ´ [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 45 canonical work pages

  1. [1]

    Anwar, S.; and Barnes, N. 2020. Densely residual laplacian super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3): 1192--1204

  2. [2]

    Chen, H.; Wang, Y.; Guo, T.; Xu, C.; Deng, Y.; Liu, Z.; Ma, S.; Xu, C.; Xu, C.; and Gao, W. 2021. Pre-Trained Image Processing Transformer. arXiv:2012.00364

  3. [3]

    Chen, L.; Chu, X.; Zhang, X.; and Sun, J. 2022. Simple baselines for image restoration. In European conference on computer vision, 17--33. Springer

  4. [4]

    Chen, X.; Wang, X.; Zhou, J.; Qiao, Y.; and Dong, C. 2023 a . Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22367--22377

  5. [5]

    Chen, Z.; Zhang, Y.; Gu, J.; Kong, L.; Yang, X.; and Yu, F. 2023 b . Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, 12312--12321

  6. [6]

    Cheng, X.; Fu, Z.; and Yang, J. 2019. Multi-scale Dynamic Feature Encoding Network for Image Demoireing. arXiv:1909.11947

  7. [7]

    Cheng, Y.; Liu, X.; and Yang, J. 2023. Recaptured raw screen image and video demoireing via channel and spatial modulations. Advances in Neural Information Processing Systems, 36: 40414--40425

  8. [8]

    Chu, X.; Chen, L.; Chen, C.; and Lu, X. 2022. Improving image restoration by revisiting global information aggregation. In European Conference on Computer Vision, 53--71. Springer

Show all 67 references
  1. [9]

    Cui, Y.; Ren, W.; and Knoll, A. 2024. Omni-Kernel Network for Image Restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 1426--1434

  2. [10]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

  3. [11]

    W.; Khan, S.; Khan, F

    Dudhane, A.; Zamir, S. W.; Khan, S.; Khan, F. S.; and Yang, M.-H. 2022. Burst image restoration and enhancement. In Proceedings of the ieee/cvf Conference on Computer Vision and Pattern Recognition, 5759--5768

  4. [12]

    Fan, C.-M.; Liu, T.-J.; and Liu, K.-H. 2022. SUNet: Swin transformer UNet for image denoising. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), 2333--2337. IEEE

  5. [13]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

  6. [14]

    V.; and Timofte, R

    Gu, S.; Li, Y.; Gool, L. V.; and Timofte, R. 2019. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2511--2520

  7. [15]

    Guo, M.-H.; Lu, C.-Z.; Hou, Q.; Liu, Z.; Cheng, M.-M.; and Hu, S.-M. 2022. Segnext: Rethinking convolutional attention design for semantic segmentation. Advances in neural information processing systems, 35: 1140--1156

  8. [16]

    Guo, M.-H.; Lu, C.-Z.; Liu, Z.-N.; Cheng, M.-M.; and Hu, S.-M. 2023. Visual attention network. Computational visual media, 9(4): 733--752

  9. [17]

    Guo, X.; Dong, Y.; Chen, X.; Chen, W.; Li, Z.; Zheng, F.; and Pun, C.-M. 2024. Underwater Image Restoration via Polymorphic Large Kernel CNNs. arXiv preprint arXiv:2412.18459

  10. [18]

    He, B.; Wang, C.; Shi, B.; and Duan, L.-Y. 2019. Mop Moire Patterns Using MopNet. 2424--2432

  11. [19]

    He, B.; Wang, C.; Shi, B.; and yu Duan, L. 2020. FHDe2Net: Full High Definition Demoireing Network. In European Conference on Computer Vision

  12. [20]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

  13. [21]

    Johnson, J.; Alahi, A.; and Fei-Fei, L. 2016. Perceptual Losses for Real-Time Style Transfer and Super-Resolution. arXiv:1603.08155

  14. [22]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2017. Adam: A Method for Stochastic Optimization. arXiv:1412.6980

  15. [23]

    Kong, L.; Dong, J.; Ge, J.; Li, M.; and Pan, J. 2023. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5886--5895

  16. [24]

    Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25

  17. [25]

    W.; Po, L.-M.; and Rehman, Y

    Lau, K. W.; Po, L.-M.; and Rehman, Y. A. U. 2024. Large separable kernel attention: Rethinking the large kernel attention design in cnn. Expert Systems with Applications, 236: 121352

  18. [26]

    Lecun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11): 2278--2324

  19. [27]

    Li, Y.; Fan, Y.; Xiang, X.; Demandolx, D.; Ranjan, R.; Timofte, R.; and Van Gool, L. 2023. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18278--18289

  20. [28]

    Liang, J.; Cao, J.; Sun, G.; Zhang, K.; Van Gool, L.; and Timofte, R. 2021. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 1833--1844

  21. [29]

    Liu, F.; Yang, J.; and Yue, H. 2015. Moir \'e pattern removal from texture images via low-rank and sparse matrix decomposition. In 2015 Visual Communications and Image Processing (VCIP), 1--4. IEEE

  22. [30]

    Liu, L.; Liu, J.; Yuan, S.; Slabaugh, G.; Leonardis, A.; Zhou, W.; and Tian, Q. 2020. Wavelet-Based Dual-Branch Network for Image Demoireing. arXiv:2007.07173

  23. [31]

    Liu, S.; Chen, T.; Chen, X.; Chen, X.; Xiao, Q.; Wu, B.; K \"a rkk \"a inen, T.; Pechenizkiy, M.; Mocanu, D.; and Wang, Z. 2022. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620

  24. [32]

    Loshchilov, I.; and Hutter, F. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983

  25. [33]

    Lowe, D. G. 2004. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60: 91--110

  26. [34]

    Lu, Z.; Li, J.; Liu, H.; Huang, C.; Zhang, L.; and Zeng, T. 2022. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 457--466

  27. [35]

    Luo, P.; Xiao, G.; Gao, X.; and Wu, S. 2023. LKD-Net: Large kernel convolution network for single image dehazing. In 2023 IEEE International Conference on Multimedia and Expo (ICME), 1601--1606. IEEE

  28. [36]

    H.; Lee, S.-H.; and Lee, C

    Nguyen, D. H.; Lee, S.-H.; and Lee, C. 2023. Multiscale coarse-to-fine guided screenshot demoir \'e ing. IEEE Signal Processing Letters, 30: 898--902

  29. [37]

    Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part II...

  30. [38]

    Ruan, L.; Bemana, M.; Seidel, H.-p.; Myszkowski, K.; and Chen, B. 2023. Revisiting image deblurring with an efficient ConvNet. arXiv preprint arXiv:2302.02234

  31. [39]

    P.; Bishop, R.; Rueckert, D.; and Wang, Z

    Shi, W.; Caballero, J.; Husz \'a r, F.; Totz, J.; Aitken, A. P.; Bishop, R.; Rueckert, D.; and Wang, Z. 2016. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision a...

  32. [40]

    Siddiqui, H.; Boutin, M.; and Bouman, C. A. 2009. Hardware-friendly descreening. IEEE Transactions on Image Processing, 19(3): 746--757

  33. [41]

    Simonyan, K.; and Zisserman, A. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556

  34. [42]

    2018-2024

    Sovrasov, V. 2018-2024. ptflops: a flops counting tool for neural networks in pytorch framework

  35. [43]

    Sun, B.; Li, S.; and Sun, J. 2014. Scanned image descreening with image redundancy and adaptive filtering. IEEE transactions on image processing, 23(8): 3698--3710

  36. [44]

    Sun, L.; Dong, J.; Tang, J.; and Pan, J. 2023. Spatially-adaptive feature modulation for efficient image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, 13190--13199

  37. [45]

    Sun, Y.; Yu, Y.; and Wang, W. 2018. Moir \'e photo restoration using multiresolution convolutional neural networks. IEEE Transactions on Image Processing, 27(8): 4160--4172

  38. [46]

    Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; and Rabinovich, A. 2015. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1--9

  39. [47]

    Tsai, F.-J.; Peng, Y.-T.; Lin, Y.-Y.; Tsai, C.-C.; and Lin, C.-W. 2022. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, 146--162. Springer

  40. [48]

    Wang, C.; He, B.; Wu, S.; Wan, R.; Shi, B.; and Duan, L.-Y. 2023. Coarse-to-fine disentangling demoir \'e ing framework for recaptured screen images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8): 9439--9453

  41. [49]

    Wang, Y.; Li, Y.; Wang, G.; and Liu, X. 2024. Multi-scale attention network for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5950--5960

  42. [50]

    Wang, Z.; Bovik, A.; Sheikh, H.; and Simoncelli, E. 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600--612

  43. [51]

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

  44. [52]

    Xiao, Z.; Lu, Z.; and Wang, X. 2024. P-BiC: Ultra-High-Definition Image Demoireing via Patch Bilateral Compensation. In ACM Multimedia 2024

  45. [53]

    Xu, S.; Song, B.; Chen, X.; Liu, X.; and Zhou, J. 2024. Image demoireing in raw and srgb domains. In European Conference on Computer Vision, 108--124. Springer

  46. [54]

    Yang, J.; Liu, F.; Yue, H.; Fu, X.; Hou, C.; and Wu, F. 2017 a . Textured image demoir \'e ing via signal decomposition and guided filtering. IEEE Transactions on Image Processing, 26(7): 3528--3541

  47. [55]

    Yang, J.; Zhang, X.; Cai, C.; and Li, K. 2017 b . Demoir \'e ing for screen-shot images with multi-channel layer decomposition. In 2017 IEEE Visual Communications and Image Processing (VCIP), 1--4. IEEE

  48. [56]

    Yu, X.; Dai, P.; Li, W.; Ma, L.; Shen, J.; Li, J.; and Qi, X. 2022. Towards efficient and scale-robust ultra-high-definition image demoir \'e ing. In European Conference on Computer Vision, 646--662. Springer

  49. [57]

    Yue, H.; Cheng, Y.; Mao, Y.; Cao, C.; and Yang, J. 2022. Recaptured screen image demoir \'e ing in raw domain. IEEE Transactions on Multimedia, 25: 5589--5600

  50. [58]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; and Yang, M.-H. 2022. Restormer: Efficient Transformer for High-Resolution Image Restoration. arXiv:2111.09881

  51. [59]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; Yang, M.-H.; and Shao, L. 2020. Learning enriched features for real image restoration and enhancement. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XXV 16...

  52. [60]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; Yang, M.-H.; and Shao, L. 2021. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14821--14831

  53. [61]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018 a . The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. arXiv:1801.03924

  54. [62]

    Zhang, Y.; Li, K.; Li, K.; Wang, L.; Zhong, B.; and Fu, Y. 2018 b . Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), 286--301

  55. [63]

    Zhang, Y.; Tian, Y.; Kong, Y.; Zhong, B.; and Fu, Y. 2018 c . Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2472--2481

  56. [64]

    Zheng, B.; Yuan, S.; Slabaugh, G.; and Leonardis, A. 2020. Image Demoireing with Learnable Bandpass Filters. arXiv:2004.00406

  57. [65]

    Zheng, B.; Yuan, S.; Yan, C.; Tian, X.; Zhang, J.; Sun, Y.; Liu, L.; Leonardis, A.; and Slabaugh, G. 2021. Learning frequency domain priors for image demoireing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11): 7705--7717

  58. [66]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  59. [67]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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