Pith. sign in

REVIEW 3 major objections 5 minor 46 references

DiMoSR: Feature Modulation via Multi-Branch Dilated Convolutions for Efficient Image Super-Resolution

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces DiMoSR, a lightweight image super-resolution network that uses multi-branch dilated convolutions to modulate feature maps, and reports that it outperforms prior lightweight methods on standard benchmarks, most…

desk verdict Efficient SR paper with a genuinely new block and top x4 results, but FFT-loss alone explains most of the margin over baselines, so the architecture's contribution is not yet established. read the letter →

arxiv 2505.21262 v1 pith:HDMCSLG5 submitted 2025-05-27 cs.CV eess.IV

M. Akin Yilmaz , Ahmet Bilican , A. Murat Tekalp This is my paper
classification cs.CVeess.IV
keywords singleimagesuper-resolutionlightweightdilatedconvolutionfeaturemodulationattentionmechanismFFTlossefficientnetworkdesignrestoration
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

DiMoSR asks whether lightweight image super-resolution can be improved by modulating feature maps with multi-branch dilated convolutions rather than by relying on attention alone. The paper introduces a compact network built from Dilated Modulation Blocks, each of which gathers information at four dilation rates and uses the fused result to predict per-channel scaling, shifting, and a gating weight. On the $\times4$ task the paper reports the best PSNR among the lightweight methods it compares on all four benchmark datasets (Set5, Set14, B100, Urban100), with 338K parameters and 20G FLOPs, and its smaller variant DiMoSR-S reaches nearly the same quality at about 250K parameters and 14G FLOPs. Ablations show that the modulation path and the attention gate each contribute, that their combination helps most on complex urban textures, and that a small FFT-based frequency loss adds consistent gains. The implied message is that feature modulation should sit alongside attention in the lightweight super-resolution toolkit.

What carries the argument

The central object is the Dilated Modulation Block (DMB). Its Feature Enhancement Block (FEB) applies LayerNorm, runs the features through four parallel branches (each a $1\times1$ convolution plus SiLU followed by a $3\times3$ dilated convolution with dilation rate 4, 8, 12, or 16), concatenates the branch outputs, and maps them with a $1\times1$ convolution into three tensors $\alpha$, $\beta$, $\gamma$. The block then applies affine modulation $\alpha X_{\mathrm{norm}}+\beta$ and gated attention $\sigma(\gamma)X_{\mathrm{norm}}$, concatenates the two results, fuses them with a $1\times1$ convolution, and adds a residual connection. An Efficient Residual Block (ERB) with LayerNorm, a channel-reducing $1\times1$, a $3\times3$ convolution, and a channel-expanding $1\times1$ completes the DMB. The four dilated branches widen the receptive field at low cost, and the fusion of modulation with attention is what the paper credits for the accuracy gains.

What would settle it

Retrain SAFMN and SPAN under DiMoSR's exact training protocol (128$\times$128 LR patches, 500,000 iterations, batch size 24, Adam with $\beta_2=0.99$, MAE plus 0.05 FFT loss) and evaluate $\times4$ on Set5, Set14, B100, and Urban100; if those baselines then meet or beat DiMoSR's reported numbers (32.31, 28.74, 27.65, and 26.25 dB), the gain is not attributable to the dilated-modulation architecture.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that feature modulation driven by multi-branch dilated convolutions is a stronger and cheaper building block for lightweight super-resolution than attention alone. The DiMoSR network stacks Dilated Modulation Blocks, each combining a Feature Enhancement Block -- four parallel $3\times3$ dilated convolutions with dilation rates 4, 8, 12, and 16, whose concatenated output predicts affine modulation coefficients $\alpha$, $\beta$ and a sigmoid attention gate $\sigma(\gamma)$ -- with an Efficient Residual Block. For $\times4$ upscaling the paper reports the best PSNR among the compared lightweight methods on Set5 (32.31 dB), Set14 (28.74 dB), B100 (27.65 dB), and Urban100 (26.25 dB), using 338K parameters and 20G FLOPs; DiMoSR-S, at 250K parameters and 14G FLOPs, nearly matches it. Ablation studies attribute the gain to both the modulation and attention paths, with the largest effect on Urban100, and to the FFT frequency loss.

Load-bearing premise

The comparison assumes that PSNR and SSIM numbers reported by earlier papers are directly comparable to DiMoSR's numbers, even though DiMoSR uses 128$\times$128 patches, 500,000 iterations, batch size 24, Adam with $\beta_2=0.99$, and an FFT loss, and no baseline is re-trained under that protocol.

Editorial extensions

If this is right

  • At $\times4$ upscaling, DiMoSR reports the top PSNR among the compared lightweight models on all four benchmarks, for 338K parameters and 20G FLOPs.
  • The smaller DiMoSR-S (250K parameters, 14G FLOPs) matches or nearly matches baselines with roughly twice its cost, so the design scales down without collapsing.
  • Removing either the attention gate or the modulation path lowers PSNR on every dataset, with the largest drop on Urban100, indicating the two paths carry complementary information.
  • Adding the FFT loss improves all four datasets (for example, $+0.10$ dB on Urban100 at $\times4$) and visibly sharpens edges, especially in repetitive urban structures.
  • Because the receptive field is widened by four dilation rates rather than by more layers or channels, the accuracy gain does not require a proportional increase in FLOPs.

Reading between the lines

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

  • The paper does not test whether the FEB-style modulation generalizes to other restoration tasks; the same affine-modulation-plus-gating construction could plausibly serve as a cheap context module for denoising or deblurring.
  • The dilation rates are fixed; an adaptive or searched assignment of per-branch rates might yield further gains, since the paper only shows that this particular set of four rates helps.
  • FLOPs measured at 1280$\times$720 output are not the same as deployment cost; on mobile hardware, dilated convolutions can have poor memory locality, so a latency-based comparison could rank the baselines differently.
  • The evaluations cover only $\times2$ and $\times4$ upscaling, so nothing in this paper establishes that the advantage persists at $\times8$ or under non-bicubic degradations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This manuscript proposes DiMoSR, a lightweight single-image super-resolution network built from Dilated Modulation Blocks (DMBs), each combining a Feature Enhancement Block (FEB) that uses multi-branch dilated convolutions to predict channel-wise modulation coefficients and a sigmoid attention weight, with an Efficient Residual Block (ERB). The network is trained on DF2K with a combined MAE and FFT frequency loss. The paper reports x2 and x4 results on Set5, Set14, B100, and Urban100, compares against a range of lightweight baselines, and presents ablations of the attention/modulation components and of the FFT loss.

Significance. If the empirical claims were fully controlled, DiMoSR would be a modest but useful contribution to the lightweight SISR literature: it demonstrates that feature modulation obtained from multi-branch dilated convolutions can be competitive with recent attention-based efficient networks, and the ablations in Tables 2 and 3 give quantitative information about each component. The authors also provide code and model weights, which is a genuine reproducibility asset. However, the central state-of-the-art claim is not yet supported by the evidence as presented, because the comparisons pool published numbers obtained under heterogeneous training protocols and because the paper's own FFT-loss ablation is large enough to account for most of the reported x4 margin over the best baselines. The x2 results also do not uniformly support the abstract's unqualified outperformance claim.

major comments (3)
  1. [§4.2, Table 1 vs. §4.3.2, Table 3] The central claim that the DiMoSR architecture achieves state-of-the-art x4 performance is confounded by the training protocol. Table 3 shows that adding the FFT loss to MAE improves x4 PSNR by 0.05/0.09/0.03/0.10 dB on Set5/Set14/B100/Urban100, while the x4 PSNR advantage of DiMoSR over the best published baseline rows in Table 1 is approximately 0.07/0.08/0.03/0.07 dB. On Set14 and Urban100 the FFT-loss contribution alone essentially equals the entire reported margin, and under MAE-only training DiMoSR's x4 scores (32.26/28.65/27.62/26.15) are no longer consistently above the best published baselines (32.24/28.66/27.62/26.18). Since none of the baselines in Table 1 is retrained under DiMoSR's protocol (128x128 LR patches, 500k iterations, batch size 24, Adam beta2=0.99, MAE+FFT loss), the abstract's claim that the architecture outperforms state-of-the-art lightweight methods is not yet cleanly attributable to the dilated-modulation design. The authors should retrain the closest baselines (at least SAFMN, ShuffleMixer, SPAN, and RLFN) under an identical protocol, or otherwise directly compare MAE-only DiMoSR against baselines trained without FFT loss, and report those numbers in Table 1 or a separate table.
  2. [Abstract and §4.2, Table 1 (x2)] The abstract and conclusion state that DiMoSR outperforms state-of-the-art lightweight methods across diverse benchmark datasets, but the x2 rows of Table 1 do not support that unqualified claim: DiMoSR's Set5 PSNR (38.06) is below RLFN (38.07) and SPAN (38.08), and on Urban100 DiMoSR (32.30) is below RLFN (32.33), while the remaining margins are only 0.01-0.03 dB. The text in §4.2 is more careful, saying DiMoSR 'provides better PSNR values than RLFN, ShuffleMixer, and SPAN on several benchmarks,' but the abstract and conclusion should be scaled back to the x4 setting, or the x2 comparison should be made under controlled conditions before claiming general superiority.
  3. [§4.2, Table 1] All reported PSNR/SSIM values appear to be single runs, and many of the claimed advantages over baselines are 0.01-0.03 dB for x2 and 0.03-0.08 dB for x4, which is within the typical run-to-run variation for SISR training with fixed hyperparameters. For a claim of state-of-the-art performance, the authors should report the mean and standard deviation over at least three independent training runs for DiMoSR (and for any baseline they retrain), and should state explicitly which numbers in Table 1 are taken from original papers and which are produced by the authors under a common protocol.
minor comments (5)
  1. [§4.3.1, Table 2] The text says 'disabling attention causes a 0.10 dB decrease' on Urban100, but Table 2 shows a decrease from 26.25 dB to 26.14 dB, i.e., 0.11 dB; please correct this value.
  2. [§4.2] The sentence 'These empirical findings conclusively demonstrate' is too strong given the protocol confounds discussed in the major comments; 'suggest' or 'indicate' would be more appropriate.
  3. [Fig. 2 caption] The caption uses 'Shufflemixer' while the rest of the paper uses 'ShuffleMixer'; please standardize the capitalization.
  4. [Eq. (5)] The frequency loss L_freq = |F(I_SR) - F(I_HR)| should specify whether F denotes the magnitude of the two-dimensional FFT and whether the loss is averaged spatially; this detail matters for exact reproduction.
  5. [§3.1] The description of the DMB grouping is ambiguous: after each group of 6 DMBs a residual connection is added, and later 'the features are concatenated.' It should be clarified whether all three group outputs are concatenated, or only the final output, and how the 1x1 convolution operates on them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DiMoSR's claims rest on external benchmark comparisons and ablations, not on a fitted input renamed as a prediction.

full rationale

This paper is an empirical architecture study. The central claim—that DiMoSR outperforms lightweight SISR baselines—is supported by PSNR/SSIM values in Table 1, where baseline numbers are taken from previously published work and DiMoSR's numbers come from its own trained model evaluated on the same public benchmark datasets (Set5, Set14, B100, Urban100). No derivation chain exists in which an input quantity is defined in terms of an output quantity, and no fitted parameter is subsequently reported as a prediction. The FFT-loss ablation in Table 3 is an honest decomposition: it shows the marginal contribution of adding the frequency-domain loss to MAE, and the paper explicitly reports both configurations rather than hiding the effect. The skeptical concern that baseline protocols differ from DiMoSR's training recipe (128x128 patches, 500k iterations, batch size 24, Adam beta2=0.99, MAE+FFT loss) is a potential external-validity or comparability limitation, but it is not circularity: the paper does not use the benchmark numbers as training inputs or fit anything to them. The only self-citation is reference [19] in the introduction's generic statement that deep learning has advanced SISR; that citation is not load-bearing for any architectural claim or for the benchmark results. There is no imported uniqueness theorem, no ansatz smuggled in by citation, and no renaming of a known result as a new organization. Thus the circularity burden is not met, and the honest finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical, so the ledger lists the hand-chosen hyperparameters and domain assumptions that the benchmark results depend on. No new physical entities are introduced, and no mathematical derivation is claimed.

free parameters (5)
  • Channel width C = 36 (DiMoSR), 32 (DiMoSR-S)
    Architecture capacity chosen by hand, no sensitivity analysis reported. Controls parameter count and FLOPs.
  • Number of DMBs = 18 (DiMoSR), 16 (DiMoSR-S)
    Depth chosen by hand; not ablated in the paper.
  • Dilation rates = 4, 8, 12, 16
    Chosen to cover multiple receptive fields; no ablation varying rates or branch count is provided.
  • FFT loss weight lambda = 0.05
    Borrowed from prior work [11, 31, 32]; ablation only toggles it off, no sweep is reported.
  • Training iterations / batch / patch size = 500k / 24 / 128x128
    Hand-chosen training budget; no analysis of whether longer training would change conclusions is given.
assumptions (4)
  • domain assumption PSNR and SSIM on the Y channel are accepted measures of super-resolution quality.
    Used for all comparisons in Table 1; the paper does not discuss perceptual metrics or human evaluation.
  • domain assumption Bicubic downscaling is the correct degradation model for evaluation.
    Standard in the SISR literature; LR inputs are generated with bicubic downscaling (Section 4.1).
  • domain assumption Cross-paper baseline numbers are comparable to numbers produced by DiMoSR's training protocol.
    Baselines are taken from original papers, not re-trained under identical conditions (Section 4.2, Table 1).
  • domain assumption FLOPs at 1280x720 output resolution are a meaningful proxy for efficiency.
    Stated in Table 1 caption; no runtime measurements are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiMoSR: Feature Modulation via Multi-Branch Dilated Convolutions for Efficient Image Super-Resolution." pith.science (2026). https://pith.science/paper/HDMCSLG5

@misc{pith2026250521262,
  author       = {Pith},
  title        = {Pith review of: DiMoSR: Feature Modulation via Multi-Branch Dilated Convolutions for Efficient Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDMCSLG5}},
  note         = {Machine review of arXiv:2505.21262}
}
read the original abstract

Balancing reconstruction quality versus model efficiency remains a critical challenge in lightweight single image super-resolution (SISR). Despite the prevalence of attention mechanisms in recent state-of-the-art SISR approaches that primarily emphasize or suppress feature maps, alternative architectural paradigms warrant further exploration. This paper introduces DiMoSR (Dilated Modulation Super-Resolution), a novel architecture that enhances feature representation through modulation to complement attention in lightweight SISR networks. The proposed approach leverages multi-branch dilated convolutions to capture rich contextual information over a wider receptive field while maintaining computational efficiency. Experimental results demonstrate that DiMoSR outperforms state-of-the-art lightweight methods across diverse benchmark datasets, achieving superior PSNR and SSIM metrics with comparable or reduced computational complexity. Through comprehensive ablation studies, this work not only validates the effectiveness of DiMoSR but also provides critical insights into the interplay between attention mechanisms and feature modulation to guide future research in efficient network design. The code and model weights to reproduce our results are available at: https://github.com/makinyilmaz/DiMoSR

Figures

Figures reproduced from arXiv: 2505.21262 by the authors.

Figure 1
Figure 1. Our proposed DiMoSR framework: a) Architecture consisting of Dilated Modula [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visual comparison on two Urban100 images. Left: full-size images with marked [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison on img062 from Urban100 ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 31 canonical work pages

  1. [36]

    Swift parameter-free attention network for efficient super- resolution

    Cheng Wan, Hongyuan Yu, Zhiqi Li, Yihang Chen, Yajun Zou, Yuqing Liu, Xuanwu Yin, and Kunlong Zuo. Swift parameter-free attention network for efficient super- resolution. InIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6246–6256, 2024

    work page 2024

  2. [1]

    Ntire 2017 challenge on single image super- resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super- resolution: Dataset and study. InIEEE/CVF Conf. on Computer Vision and Patt. Recog. Workshops (CVPRW), pages 1122–1131, 2017. doi: 10.1109/CVPRW.2017.150

    doi:10.1109/cvprw.2017.150 2017

  3. [2]

    Fast, accurate, and lightweight super-resolution with cascading residual network

    Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. InEuro. Conf. Computer Vision (ECCV), pages 256–272, 2018

    work page 2018

  4. [3]

    Low-Complexity Single-Image Super-Resolution based on Nonnegative Neighbor Em- bedding

    Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi Morel. Low-Complexity Single-Image Super-Resolution based on Nonnegative Neighbor Em- bedding. InBritish Machine Vision Conference (BMVC), Guildford, Surrey, UK, September 2012. URLhttps://inria.hal.science/hal-00747054

    work page 2012

  5. [4]

    Activating more pixels in image super-resolution transformer

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. InIEEE/CVF Conf. on Comp. Vis. Patt. Recog. (CVPR), pages 22367–22377, 2023. doi: 10.1109/CVPR52729.2023.02142

    arXiv 2023

  6. [5]

    Dual aggregation transformer for image super-resolution

    Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. InICCV, 2023

    work page 2023

  7. [6]

    Recursive generalization transformer for image super-resolution

    Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, and Xiaokang Yang. Recursive generalization transformer for image super-resolution. InICLR, 2024

    work page 2024

  8. [7]

    Image super-resolution using deep convolutional networks.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 38(2):295–307, 2016

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional networks.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 38(2):295–307, 2016. doi: 10.1109/TPAMI.2015.2439281

    arXiv 2016

Show all 46 references
  1. [8]

    Accelerating the super-resolution convolutional neural network

    Chao Dong, Chen Change Loy, and Xiaoou Tang. Accelerating the super-resolution convolutional neural network. InEuropean Conference Computer Vision (ECCV), pages 391–407. Springer International Publishing, 2016. ISBN 978-3-319-46475-6

    2016

  2. [9]

    Fast and memory-efficient network towards efficient image super-resolution

    Zongcai Du, Ding Liu, Jie Liu, Jie Tang, Gangshan Wu, and Lean Fu. Fast and memory-efficient network towards efficient image super-resolution. InIEEE/CVF Conf. on Computer Vision and Patt. Recog. Workshops (CVPRW), pages 852–861,

  3. [10]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, 2018

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, 2018. ISSN 0893-6080. doi: https://doi.org/10.1016/j.neunet.2017.12

    2018 doi

  4. [11]

    Fourier space losses for efficient per- ceptual image super-resolution

    Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient per- ceptual image super-resolution. InIEEE/CVF Int. Conf. on Computer Vision (ICCV), pages 2340–2349, 2021. doi: 10.1109/ICCV48922.2021.00236

    2021

  5. [12]

    Special issue on deep reinforcement learning

    URLhttps://www.sciencedirect.com/science/article/pii/ S0893608017302976. Special issue on deep reinforcement learning

  6. [13]

    Drct: Saving image super- resolution away from information bottleneck

    Chih-Chung Hsu, Chia-Ming Lee, and Yi-Shiuan Chou. Drct: Saving image super- resolution away from information bottleneck. InIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6133–6142, June 2024

    2024

  7. [14]

    Feature distillation interaction weighting network for lightweight image super-resolution

    Guangwei Gao, Wenjie Li, Juncheng Li, Fei Wu, Huimin Lu, and Yi Yu. Feature distillation interaction weighting network for lightweight image super-resolution. In AAAI Conf. on Artificial Intelligence, volume 36, pages 661–669, 2022. YILMAZ, BILICAN, TEKALP: DIMOSR11

    2022

  8. [15]

    Lightweight image super- resolution with information multi-distillation network

    Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super- resolution with information multi-distillation network. InACM Int. Conf. on Multime- dia (ACM MM), pages 2024–2032, 2019

    2024

  9. [16]

    Single image super-resolution from transformed self-exemplars

    Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. InIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5197–5206, 2015. doi: 10.1109/CVPR.2015.7299156

    2015

  10. [17]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Int. Conf. Learning Representation (ICLR), 2015

    2015

  11. [18]

    Accurate image super-resolution using very deep convolutional networks

    Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. InIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1646–1654, 2016. doi: 10.1109/CVPR.2016.182

    2016 doi

  12. [19]

    Murat Tekalp, and Zafer Dogan

    Cansu Korkmaz, A. Murat Tekalp, and Zafer Dogan. Training transformer models by wavelet losses improves quantitative and visual performance in single image super- resolution. InIEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR), June 2024

    2024

  13. [21]

    La- par: Linearly-assembled pixel-adaptive regression network for single image super- resolution and beyond.Advances in Neural Information Processing Systems, 33, 2020

    Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. La- par: Linearly-assembled pixel-adaptive regression network for single image super- resolution and beyond.Advances in Neural Information Processing Systems, 33, 2020

    2020

  14. [22]

    Deep lapla- cian pyramid networks for fast and accurate super-resolution

    Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep lapla- cian pyramid networks for fast and accurate super-resolution. InIEEE Conf. on Com- puter Vision and Patt. Recog. (CVPR), pages 5835–5843, 2017. doi: 10.1109/CVPR. 2017.618

    2017 doi

  15. [23]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Tim- ofte. Swinir: Image restoration using swin transformer. InIEEE/CVF Int. Conf. on Computer Vision Workshops (ICCVW), pages 1833–1844, 2021. doi: 10.1109/ ICCVW54120.2021.00210

    2021

  16. [24]

    Blueprint separable residual network for efficient image super-resolution

    Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 832–842, 2022. doi: 10.1109/CV...

    2022

  17. [25]

    SGDR: Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. InInt. Conf. on Learning Representations (ICLR), 2017

    2017

  18. [26]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. InIEEE Conf. on Computer Vision and Patt. Recog. Workshops (CVPRW), pages 1132–1140, 2017. doi: 10.1109/ CVPRW.2017.151. 12YILMAZ, BILICAN, TE...

    2017

  19. [27]

    Multi-attention based ultra lightweight image super-resolution

    Abdul Muqeet, Jiwon Hwang, Subin Yang, JungHeum Kang, Yongwoo Kim, and Sung- Ho Bae. Multi-attention based ultra lightweight image super-resolution. InEuro. Conf. Computer Vision (ECCV) Workshops, pages 103–118, 2020

    2020

  20. [28]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented nat- ural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InIEEE Int. Conf. on Computer Vision (ICCV), volume 2, pages 416–423, 2001. doi: 10.1109/I...

    2001

  21. [29]

    Courville

    Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron C. Courville. Film: Visual reasoning with a general conditioning layer. InAAAI, 2018

    2018

  22. [30]

    Single image super-resolution via a holistic attention network

    Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. InEuro. Conf. Comp. Vision (ECCV), pages 191–207, 2020

    2020

  23. [31]

    ShuffleMixer: An efficient convnet for image super-resolution

    Long Sun, Jinshan Pan, and Jinhui Tang. ShuffleMixer: An efficient convnet for image super-resolution. InAdvances in Neural Information Processing Systems, 2022

    2022

  24. [32]

    Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang

    Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network. InIEEE/CVF Conf. on Computer Vision and Patter...

    2016

  25. [33]

    Lightweight image super-resolution with enhanced cnn.Knowledge- Based Systems, page 106235, 2020

    Chunwei Tian, Ruibin Zhuge, Zhihao Wu, Yong Xu, Wangmeng Zuo, Chen Chen, and Chia-Wen Lin. Lightweight image super-resolution with enhanced cnn.Knowledge- Based Systems, page 106235, 2020

    2020

  26. [34]

    Spatially-adaptive feature modulation for efficient image super-resolution

    Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Spatially-adaptive feature modulation for efficient image super-resolution. InICCV, 2023

    2023

  27. [35]

    Exploring sparsity in image super-resolution for efficient inference

    Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In IEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR), pages 4915–4924, 2021

    2021

  28. [37]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4):600–612, 2004. doi: 10.1109/TIP.2003.819861. YILMAZ, BILICAN, TEKALP: DIMOSR13

    2004

  29. [38]

    The super-resolution recon- struction algorithm of multi-scale dilated convolution residual network.Fron- tiers in Neurorobotics, V olume 18, 2024

    Shanqin Wang, Miao Zhang, and Mengjun Miao. The super-resolution recon- struction algorithm of multi-scale dilated convolution residual network.Fron- tiers in Neurorobotics, V olume 18, 2024. doi: 10.3389/fnbot.2024.1436052. URLhttps://www.frontiersin.org/journals/neurorobotic...

    2024

  30. [39]

    On single image scale-up using sparse-representations

    Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. InCurves and Surfaces, pages 711–730, Berlin, Heidelberg,

  31. [40]

    See more details: Efficient image super-resolution by experts mining

    Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yulun Zhang, and Radu Timofte. See more details: Efficient image super-resolution by experts mining. InInt. Conf. on Machine Learning (ICML). PMLR, 2024

    2024

  32. [41]

    Edge-oriented convolution block for real- time super resolution on mobile devices

    Xindong Zhang, Hui Zeng, and Lei Zhang. Edge-oriented convolution block for real- time super resolution on mobile devices. InACM Int. Conf. on Multimedia, pages 4034–4043, 2021

    2021

  33. [42]

    Efficient long-range attention network for image super-resolution

    Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. InEuropean Conf. on Computer Vision, 2022

    2022

  34. [43]

    Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution.arXiv preprint arXiv:2208.11247, 2022

    Dafeng Zhang, Feiyu Huang, Shizhuo Liu, Xiaobing Wang, and Zhezhu Jin. Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution.arXiv preprint arXiv:2208.11247, 2022

    2022 arXiv

  35. [44]

    Residual dense network for image super-resolution

    Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. InCVPR, 2018

    2018

  36. [45]

    Efficient image super-resolution using pixel attention

    Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. InEuropean Conf. on Computer Vision, pages 56–72. Springer, 2020

    2020

  37. [46]

    Image super-resolution using very deep residual channel attention networks

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. InECCV, 2018

    2018

  38. [2012]

    ISBN 978-3-642-27413-8

    Springer Berlin Heidelberg. ISBN 978-3-642-27413-8

Pith tools

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