pith. sign in

arxiv: 2605.31400 · v1 · pith:RRWREFMDnew · submitted 2026-05-29 · 💻 cs.CV

FSM-Net: An Efficient Frequency-Spatial Network for Real-World Deblurring

Vinh-Thuan Ly This is my paper

Pith reviewed 2026-06-28 22:33 UTC · model grok-4.3

classification 💻 cs.CV
keywords real-world image deblurringfrequency attentionefficient neural networkFFTE-Branchformerprogressive trainingRSBlur benchmarkNTIRE challenge
0
0 comments X

The pith

FSM-Net restores real-world blurred images to 33.144 dB PSNR using a dual frequency-spatial network with only 4.94 million parameters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Real-world image deblurring requires both high restoration quality and low computational demands, a trade-off that most existing methods do not satisfy well. FSM-Net tackles this through a dual-domain design that processes information in both frequency and spatial domains. A Frequency Attention module recovers high-frequency details explicitly via FFT, while a Cross-Gated Vision E-Branchformer models global dependencies at linear complexity in the bottleneck. Training proceeds progressively under a composite loss that includes multi-scale Charbonnier, structural edge, and frequency terms, yielding the reported performance on the RSBlur benchmark.

Core claim

FSM-Net is a Frequency-Spatial Multi-branch Network that applies a Frequency Attention module to recover high-frequency structural details via FFT operations and places a Cross-Gated Vision E-Branchformer at the bottleneck to capture global dependencies with linear complexity. A progressive curriculum training strategy guided by a composite loss function of Multi-Scale Charbonnier, Structural Edge, and Frequency terms enables robust convergence, resulting in 33.144 dB PSNR, 4.94M parameters, and 159.35 GMACs at 1920x1200 resolution on the RSBlur benchmark.

What carries the argument

Frequency Attention module that recovers high-frequency structural details via FFT, paired with Cross-Gated Vision E-Branchformer for global dependencies at linear complexity

If this is right

  • The dual-domain architecture achieves second place in the NTIRE 2026 Challenge on Efficient Real-World Deblurring.
  • The method improves the efficiency-quality Pareto frontier for real-world deblurring at 1920x1200 resolution.
  • The linear-complexity E-Branchformer combined with frequency attention supports high-fidelity output under tight parameter and GMAC budgets.
  • The composite loss and progressive training schedule enable reliable convergence on real captured blur without heavy overfitting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The frequency attention mechanism could be extended to video sequences by incorporating temporal frequency components.
  • Linear complexity in the bottleneck suggests the design may scale to higher resolutions without quadratic memory growth.
  • The same frequency-spatial split and composite loss might transfer to related tasks such as denoising or deraining.
  • Ablating the FFT-based attention versus the E-Branchformer on the same backbone would isolate which component drives the efficiency gain.

Load-bearing premise

The progressive curriculum training with the composite loss produces robust convergence on real-world data without overfitting to the RSBlur distribution.

What would settle it

Evaluating the same architecture on an independent real-world blur dataset with different capture conditions and measuring whether PSNR remains near 33 dB would test whether the training strategy generalizes.

Figures

Figures reproduced from arXiv: 2605.31400 by Vinh-Thuan Ly.

Figure 1
Figure 1. Figure 1 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall Architecture of FSM-Net. Our model follows an efficient U-shaped hierarchical structure. (a) The global pipeline with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual results of FSM-Net on the NTIRE 2026 test set. We present several representative examples to demonstrate the restoration capability of our proposed architecture. FSM-Net effectively recovers sharp structural details and suppresses complex motion artifacts, contributing to our 2nd place ranking on the official public leaderboard. These results highlight the model’s efficacy in resolving diverse real-… view at source ↗
read the original abstract

Real-world image deblurring demands both high-fidelity restoration and computational efficiency, a balance existing methods often struggle to achieve. In this paper, we propose FSM-Net (Frequency-Spatial Multi-branch Network), a highly efficient solution that secured 2nd place in the NTIRE 2026 Challenge on Efficient Real-World Deblurring. FSM-Net pioneers a dual-domain approach: a novel Frequency Attention module explicitly recovers high-frequency structural details via FFT, while a Cross-Gated Vision E-Branchformer at the bottleneck captures global dependencies with linear complexity. To ensure robust convergence, we employ a progressive curriculum training strategy guided by a composite loss function (Multi-Scale Charbonnier, Structural Edge, and Frequency). Evaluated on the RSBlur benchmark, FSM-Net achieves an outstanding 33.144 dB PSNR with only 4.94M parameters and 159.35 GMACs (at 1920x1200 resolution). By effectively pushing the Pareto frontier of efficiency and quality, FSM-Net establishes a strong baseline for resource-constrained image restoration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes FSM-Net, a dual-domain Frequency-Spatial Multi-branch Network for real-world image deblurring. It introduces a Frequency Attention module that uses FFT to recover high-frequency details, a Cross-Gated Vision E-Branchformer at the bottleneck for global dependencies with linear complexity, and a progressive curriculum training strategy driven by a composite loss (Multi-Scale Charbonnier + Structural Edge + Frequency). The central claim is that this architecture secured 2nd place in the NTIRE 2026 Challenge on Efficient Real-World Deblurring and delivers 33.144 dB PSNR on the RSBlur benchmark at 1920x1200 resolution using only 4.94M parameters and 159.35 GMACs, thereby advancing the efficiency-quality Pareto frontier.

Significance. If the reported metrics are supported by proper controls, the work would be significant for resource-constrained image restoration by showing that a compact dual-domain design can reach competitive real-world deblurring performance. The explicit frequency-domain recovery mechanism and linear-complexity global modeling are concrete contributions that could serve as a reproducible baseline for subsequent efficient restoration networks.

major comments (3)
  1. [§5] §5 (Experiments) and associated tables: No ablation studies isolate the contribution of the progressive curriculum stages or the individual terms in the composite loss. The headline 33.144 dB PSNR is therefore unattributed, leaving open whether the gain stems from the Frequency Attention module and E-Branchformer or from dataset-specific tuning of the free parameters (loss weights and curriculum schedule).
  2. [§5.1] §5.1 (Quantitative results): The manuscript states the RSBlur benchmark result and challenge ranking but supplies neither a comparison table against other NTIRE 2026 entries nor standard baselines (e.g., Restormer, NAFNet) with identical training protocols. This omission directly weakens the Pareto-frontier claim.
  3. [§4.3] §4.3 (Training strategy): The composite loss and curriculum are described at a high level without the exact weighting schedule or stage-transition criteria. Because these are the only free parameters identified, their omission prevents verification that convergence is robust rather than overfit to the RSBlur distribution.
minor comments (2)
  1. [Figure 2] Figure 2 (architecture diagram): The data-flow arrows between the Frequency Attention module and the E-Branchformer are not labeled with tensor dimensions, making it difficult to verify the claimed linear complexity.
  2. [Eq. (3)] Eq. (3) (Frequency loss): The notation for the FFT-based term is introduced without an explicit definition of the frequency weighting mask, which should be clarified for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [§5] §5 (Experiments) and associated tables: No ablation studies isolate the contribution of the progressive curriculum stages or the individual terms in the composite loss. The headline 33.144 dB PSNR is therefore unattributed, leaving open whether the gain stems from the Frequency Attention module and E-Branchformer or from dataset-specific tuning of the free parameters (loss weights and curriculum schedule).

    Authors: We agree that the current version lacks these ablations. In the revised manuscript we will add ablation studies in §5 isolating the progressive curriculum stages and each term of the composite loss to attribute the reported performance gains. revision: yes

  2. Referee: [§5.1] §5.1 (Quantitative results): The manuscript states the RSBlur benchmark result and challenge ranking but supplies neither a comparison table against other NTIRE 2026 entries nor standard baselines (e.g., Restormer, NAFNet) with identical training protocols. This omission directly weakens the Pareto-frontier claim.

    Authors: We acknowledge the omission. We will add a comparison table in the revised §5.1 that includes other NTIRE 2026 entries (where public) and standard baselines retrained under identical protocols on RSBlur to support the efficiency-quality claim. revision: yes

  3. Referee: [§4.3] §4.3 (Training strategy): The composite loss and curriculum are described at a high level without the exact weighting schedule or stage-transition criteria. Because these are the only free parameters identified, their omission prevents verification that convergence is robust rather than overfit to the RSBlur distribution.

    Authors: We will expand §4.3 in the revision to include the precise loss weights and stage-transition criteria, improving reproducibility and allowing verification of robust convergence. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical performance claims are independent of inputs

full rationale

The paper describes a novel dual-domain architecture (Frequency Attention via FFT, Cross-Gated Vision E-Branchformer) plus a progressive curriculum with composite loss (Multi-Scale Charbonnier + Structural Edge + Frequency), then reports an empirical PSNR result on the RSBlur benchmark from the NTIRE 2026 challenge. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the provided text. The reported 33.144 dB PSNR at given parameter/GMAC counts is presented as an outcome of training and evaluation, not derived by construction from the architecture description itself. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

Abstract-only review limits visibility into design choices; the performance claim rests on unstated hyperparameters for the composite loss and curriculum schedule plus the assumption that the proposed modules function as described on the target data distribution.

free parameters (2)
  • Composite loss weights
    Relative weights among Multi-Scale Charbonnier, Structural Edge, and Frequency losses are not stated and must be selected to reach the reported PSNR.
  • Curriculum schedule parameters
    Details of the progressive difficulty ramp in training are omitted and likely tuned for convergence on RSBlur.
axioms (2)
  • domain assumption FFT-based Frequency Attention can explicitly recover high-frequency structural details from blurred images
    This is the stated mechanism of the frequency branch.
  • domain assumption Cross-Gated Vision E-Branchformer captures global dependencies at linear complexity
    This is presented as the property of the bottleneck module.

pith-pipeline@v0.9.1-grok · 5709 in / 1566 out tokens · 33391 ms · 2026-06-28T22:33:49.898220+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references

  1. [1]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization, 2016. 3

    2016

  2. [2]

    Charbonnier, L

    P. Charbonnier, L. Blanc-Feraud, G. Aubert, and M. Bar- laud. Two deterministic half-quadratic regularization algo- rithms for computed imaging. InProceedings of 1st Inter- national Conference on Image Processing, pages 168–172 vol.2, 1994. 5

    1994

  3. [3]

    Pre-trained image processing transformer

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yip- ing Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 12299–12310,

  4. [4]

    Hinet: Half instance normalization network for image restoration

    Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Cheng- peng Chen. Hinet: Half instance normalization network for image restoration. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182–192, 2021. 2

    2021

  5. [5]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII, page 17–33, Berlin, Heidelberg, 2022. Springer-Verlag. 1, 2, 3, 7, 8

    2022

  6. [6]

    Fast fourier convolu- tion

    Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolu- tion. InAdvances in Neural Information Processing Systems, pages 4479–4488. Curran Associates, Inc., 2020. 2

    2020

  7. [7]

    Rethinking coarse-to-fine approach in sin- gle image deblurring

    Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in sin- gle image deblurring. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 4641–4650, 2021. 2, 8

    2021

  8. [8]

    Selective frequency network for image restoration

    Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. InThe Eleventh In- ternational Conference on Learning Representations, 2023. 3

    2023

  9. [9]

    Dauphin, Angela Fan, Michael Auli, and David Grangier

    Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional net- works. InProceedings of the 34th International Confer- ence on Machine Learning - Volume 70, page 933–941. JMLR.org, 2017. 3

    2017

  10. [10]

    Repvgg: Making vgg-style convnets great again

    Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13733–13742, 2021. 2

    2021

  11. [11]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 1, 2

    2021

  12. [12]

    Ef- ficient real-world deblurring using single images: Aim 2025 challenge report

    Daniel Feijoo, Paula Garrido, Marcos V Conde, Jaesung Rim, Alvaro Garcia, Sunghyun Cho, and Radu Timofte. Ef- ficient real-world deblurring using single images: Aim 2025 challenge report. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 5713–5721,

    2025

  13. [13]

    Gaussian error linear units (gelus), 2023

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus), 2023. 4

    2023

  14. [14]

    Xydeblur: Divide and conquer for single image deblur- ring

    Seo-Won Ji, Jeongmin Lee, Seung-Wook Kim, Jun-Pyo Hong, Seung-Jin Baek, Seung-Won Jung, and Sung-Jea Ko. Xydeblur: Divide and conquer for single image deblur- ring. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 17421– 17430, 2022. 2

    2022

  15. [15]

    Focal frequency loss for image reconstruction and synthesis

    Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13919–13929, 2021. 2, 5

    2021

  16. [16]

    Percep- tual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Percep- tual losses for real-time style transfer and super-resolution. InComputer Vision – ECCV 2016, pages 694–711, Cham,

    2016

  17. [17]

    Springer International Publishing. 1

  18. [18]

    Han, and Shinji Watanabe

    Kwangyoun Kim, Felix Wu, Yifan Peng, Jing Pan, Prashant Sridhar, Kyu J. Han, and Shinji Watanabe. E-branchformer: Branchformer with enhanced merging for speech recogni- tion. In2022 IEEE Spoken Language Technology Workshop (SLT), pages 84–91, 2023. 2, 4

    2023

  19. [19]

    Efficient frequency domain-based trans- formers for high-quality image deblurring

    Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based trans- formers for high-quality image deblurring. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 5886–5895, 2023. 2

    2023

  20. [20]

    Deblurgan: Blind motion deblurring using conditional adversarial networks

    Orest Kupyn, V olodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Ji ˇr´ı Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. InPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1

    2018

  21. [21]

    Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better

    Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2019. 8

    2019

  22. [22]

    Ntire 2022 challenge on efficient super- resolution: Methods and results

    Yawei Li et al. Ntire 2022 challenge on efficient super- resolution: Methods and results. In2022 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition Work- shops (CVPRW), pages 1061–1101, 2022. 2

    2022

  23. [23]

    Ntire 2023 challenge on image denoising: Methods and results

    Yawei Li et al. Ntire 2023 challenge on image denoising: Methods and results. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1905–1921, 2023. 2

    2023

  24. [24]

    Swinir: Image restoration us- ing swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) Workshops, pages 1833–1844, 2021. 1

    2021

  25. [25]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012–10022, 2021. 4

    2021

  26. [26]

    Sgdr: Stochastic gradient descent with warm restarts, 2017

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts, 2017. 6

    2017

  27. [27]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6

    2019

  28. [28]

    Intriguing findings of frequency selection for image deblurring.Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):1905–1913, 2023

    Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring.Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):1905–1913, 2023. 2

    1905

  29. [29]

    Marr and E

    D. Marr and E. Hildreth. Theory of edge detection.Proceed- ings of the Royal Society of London. B. Biological Sciences, 207(1167):187–217, 1980. 5

    1980

  30. [30]

    Exponential moving average of weights in deep learning: Dynamics and benefits.Transactions on Machine Learning Research, 2024

    Daniel Morales-Brotons, Thijs V ogels, and Hadrien Hen- drikx. Exponential moving average of weights in deep learning: Dynamics and benefits.Transactions on Machine Learning Research, 2024. 5

    2024

  31. [31]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 3883– 3891, 2017. 1, 8

    2017

  32. [32]

    Ntire 2021 challenge on image deblur- ring

    Seungjun Nah et al. Ntire 2021 challenge on image deblur- ring. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 149– 165, 2021. 2

    2021

  33. [33]

    Conformer: Local fea- tures coupling global representations for visual recognition

    Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local fea- tures coupling global representations for visual recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 367–376, 2021. 2

    2021

  34. [34]

    Real-world blur dataset for learning and benchmarking de- blurring algorithms

    Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking de- blurring algorithms. InComputer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV, page 184–201, Berlin, Heidelberg,

    2020

  35. [35]

    Springer-Verlag. 1, 7

  36. [36]

    Realistic blur synthesis for learning image deblurring

    Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. InComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII, page 487–503, Berlin, Heidelberg, 2022. Springer-Verlag. 5

    2022

  37. [37]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2

    2018

  38. [38]

    Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang

    Wenzhe Shi, Jose Caballero, Ferenc Huszar, 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. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3

    2016

  39. [39]

    Maitreya Suin, Kuldeep Purohit, and A. N. Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  40. [40]

    Scale-recurrent network for deep image deblurring

    Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Ji- aya Jia. Scale-recurrent network for deep image deblurring. In2018 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 8174–8182, 2018. 8

    2018

  41. [41]

    Stripformer: Strip transformer for fast image deblurring

    Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for¬†fast image deblurring. InComputer Vision – ECCV 2022, pages 146–162, Cham, 2022. Springer Nature Switzerland. 2

    2022

  42. [42]

    Maxim: Multi-axis mlp for image processing

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5769–5780, 2022. 1

    2022

  43. [43]

    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. 6

    2004

  44. [44]

    Uformer: A general u-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 17683–17693, 2022. 2, 7, 8

    2022

  45. [45]

    Fda: Fourier domain adaptation for semantic segmentation

    Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2

    2020

  46. [46]

    Multi-stage progressive image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821–14831, 2021. 2, 7, 8

    2021

  47. [47]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728–5739, 2022. 1, 2, 4, 8

    2022

  48. [48]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018. 6

    2018

  49. [49]

    Ntire 2024 challenge on bracketing image restoration and enhancement: Datasets methods and results

    Zhilu Zhang et al. Ntire 2024 challenge on bracketing image restoration and enhancement: Datasets methods and results. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6153–6166, 2024. 2

    2024

  50. [50]

    Deep fourier up-sampling

    Man Zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. InAdvances in Neural Information Processing Systems, pages 22995–23008. Curran Associates, Inc., 2022. 3

    2022