pith. sign in

arxiv: 2604.09313 · v1 · submitted 2026-04-10 · 📡 eess.IV · cs.CV

Compositional-Degradation UAV Image Restoration: Conditional Decoupled MoE Network and A Benchmark

Jinquan Yan , Zhicheng Zhao , Zhengzheng Tu , Chenglong Li , Jin Tang , Bin Luo This is my paper

Pith reviewed 2026-05-10 17:02 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords UAV image restorationcompositional degradationmixture of expertsdegradation perceptionmulti-degradation benchmarkfactor-wise cuesconditional routing
0
0 comments X

The pith

DAME-Net improves UAV image restoration by decoupling explicit per-factor degradation perception from selective reconstruction.

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

The paper targets the problem of UAV images degraded by multiple simultaneous factors such as rain, haze, and noise, which hurt applications like mapping and inspection. It introduces DAME-Net to handle these compositional cases by first explicitly identifying each degradation factor separately instead of relying on a single entangled condition that mixes corrections and causes interference. A Factor-wise Degradation Perception module supplies interpretable per-factor cues through multi-label prediction, while a Conditioned Decoupled MoE module uses those cues for targeted, mask-constrained expert routing and hybrid spatial-frequency processing. The authors also release the MDUR benchmark covering 43 single-to-four-factor configurations with seen and unseen splits. Experiments show consistent gains over unified restoration networks, with larger benefits on unseen and higher-order composites, plus improved downstream object detection.

Core claim

DAME-Net decouples explicit degradation perception from degradation-conditioned reconstruction for compositional UAV image restoration. The Factor-wise Degradation Perception module provides per-factor cues via multi-label prediction with label-similarity-guided soft alignment, while the Conditioned Decoupled MoE module performs stage-wise conditioning, spatial-frequency hybrid processing, and mask-constrained decoupled expert routing to enable selective factor-specific correction and suppress irrelevant interference, yielding consistent improvements over unified methods on the MDUR benchmark especially for unseen and higher-order composites.

What carries the argument

DAME-Net's Factor-wise Degradation Perception module (FDPM) for multi-label per-factor cue prediction and Conditioned Decoupled MoE module (CDMM) for cue-guided selective expert routing and factor-specific correction.

If this is right

  • Restoration quality rises for images containing two or more simultaneous degradations.
  • Performance gains widen on unseen degradation combinations and higher-order composites.
  • Downstream UAV tasks such as object detection receive higher accuracy from the restored images.
  • Degradation handling becomes more interpretable because each factor receives its own explicit cue.
  • The MDUR benchmark supplies a standardized testbed for future compositional restoration methods.

Where Pith is reading between the lines

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

  • The same explicit-factor decoupling approach could transfer to other multi-degradation domains such as underwater or medical imaging.
  • Real captured UAV data rather than synthetic composites would be needed to confirm whether the gains hold outside the benchmark.
  • The benchmark's seen/unseen splits offer a template for evaluating generalization in other restoration tasks.
  • Selective expert routing may reduce compute compared with fully entangled models when only a subset of degradations is present.

Load-bearing premise

The Factor-wise Degradation Perception module produces accurate and generalizable per-factor cues that genuinely reduce interference during reconstruction.

What would settle it

Restoration metrics on a held-out set of real UAV images with independently verified multi-factor degradations showing equal or worse performance than a strong unified baseline would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.09313 by Bin Luo, Chenglong Li, Jinquan Yan, Jin Tang, Zhengzheng Tu, Zhicheng Zhao.

Figure 1
Figure 1. Figure 1: Real UAV examples illustrating the impact of adverse degradations on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Motivation for FDPM. Conventional hard alignment forces a composite [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of DAME-Net. FDPM predicts a degradation mask [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Architecture of the DC-MoE. The predicted degradation mask first [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on MDUR under single and composite degradations. From left to right: Input, Restormer, PromptIR, Ours, and Ground Truth (GT). Our method better preserves scene structure and color consistency while leaving fewer visible residual degradations. references, and restored outputs from all methods across all 43 configurations. We report Precision, Recall, mAP50, and mAP50-95, and define ∆m… view at source ↗
read the original abstract

UAV images are critical for applications such as large-area mapping, infrastructure inspection, and emergency response. However, in real-world flight environments, a single image is often affected by multiple degradation factors, including rain, haze, and noise, undermining downstream task performance. Current unified restoration approaches typically rely on implicit degradation representations that entangle multiple factors into a single condition, causing mutual interference among heterogeneous corrections. To this end, we propose DAME-Net, a Degradation-Aware Mixture-of-Experts Network that decouples explicit degradation perception from degradation-conditioned reconstruction for compositional UAV image restoration. Specifically, we design a Factor-wise Degradation Perception module(FDPM) to provide explicit per-factor degradation cues for the restoration stage through multi-label prediction with label-similarity-guided soft alignment, replacing implicit entangled conditions with interpretable and generalizable degradation descriptions. Moreover, we develop a Conditioned Decoupled MoE module(CDMM) that leverages these cues for stage-wise conditioning, spatial-frequency hybrid processing, and mask-constrained decoupled expert routing, enabling selective factor-specific correction while suppressing irrelevant interference. In addition, we construct the Multi-Degradation UAV Restoration benchmark (MDUR), the first large-scale UAV benchmark for compositional UAV image restoration, with 43 degradation configurations from single degradations to four-factor composites and standardized seen/unseen splits.Extensive experiments on MDUR demonstrate consistent improvements over representative unified restoration methods, with greater gains on unseen and higher-order composite degradations. Downstream experiments further validate benefits for UAV object detection.

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

2 major / 2 minor

Summary. The manuscript proposes DAME-Net, a Degradation-Aware Mixture-of-Experts Network for compositional UAV image restoration. It decouples explicit degradation perception using the Factor-wise Degradation Perception module (FDPM), which performs multi-label prediction with label-similarity-guided soft alignment, from degradation-conditioned reconstruction via the Conditioned Decoupled MoE module (CDMM) that uses spatial-frequency hybrid processing and mask-constrained expert routing. The authors also present the Multi-Degradation UAV Restoration (MDUR) benchmark with 43 degradation configurations ranging from single to four-factor composites and standardized seen/unseen splits. Experiments demonstrate consistent improvements over unified methods, with larger gains on unseen and higher-order composites, and benefits for downstream UAV object detection.

Significance. This work tackles a practical challenge in UAV imaging where images often suffer from multiple simultaneous degradations. By making degradation perception explicit and interpretable, and enabling selective correction through MoE, the method has the potential to improve restoration quality in complex real-world scenarios. The MDUR benchmark is a valuable addition to the field as the first dedicated large-scale UAV dataset for compositional degradations. Strengths include the focus on generalizability to unseen composites and the downstream task validation. The significance would be higher with stronger evidence on the synthetic-to-real transfer.

major comments (2)
  1. [FDPM module and unseen split experiments] The central decoupling claim depends on FDPM producing accurate per-factor cues that generalize to unseen higher-order (3- and 4-factor) composites. The paper should report detailed multi-label classification metrics (e.g., per-factor F1 scores or accuracy) on the unseen test splits to verify that the soft alignment reduces interference in the CDMM stage, as opposed to the performance gains being attributable to benchmark-specific biases.
  2. [MDUR benchmark and synthetic degradation models] The MDUR benchmark relies on synthetic models for rain, haze, and noise. To support the claim of applicability to real UAV environments, the authors need to provide evidence (perhaps in the benchmark or experiments section) that the learned cues and routing do not overfit to simulation artifacts. A comparison with real multi-degraded UAV images or an analysis of domain gap would strengthen the load-bearing assumption.
minor comments (2)
  1. [Abstract] The abstract states 'consistent improvements' and 'greater gains' without providing numerical values, error bars, or specific metrics; including key quantitative results would improve clarity for readers.
  2. [Overall presentation] Ensure all figures and tables are clearly labeled and referenced in the text, particularly those showing ablation studies on the decoupling components.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript on DAME-Net and the MDUR benchmark. We appreciate the recognition of the practical challenges in compositional UAV image restoration and the value of the proposed benchmark. Below we provide point-by-point responses to the major comments, indicating revisions made to address them where feasible.

read point-by-point responses

  1. Referee: [FDPM module and unseen split experiments] The central decoupling claim depends on FDPM producing accurate per-factor cues that generalize to unseen higher-order (3- and 4-factor) composites. The paper should report detailed multi-label classification metrics (e.g., per-factor F1 scores or accuracy) on the unseen test splits to verify that the soft alignment reduces interference in the CDMM stage, as opposed to the performance gains being attributable to benchmark-specific biases.

    Authors: We agree that explicit multi-label metrics on the unseen splits would provide stronger support for the generalization of the FDPM and the effectiveness of the soft alignment in reducing interference. In the revised manuscript we have added Table 3, which reports per-factor precision, recall, and F1 scores for FDPM on both seen and unseen test splits, stratified by composite order (1- to 4-factor). These results show that the label-similarity-guided soft alignment maintains competitive per-factor F1 scores (average >0.82) on unseen higher-order composites, with only modest degradation relative to seen cases. We have also included an ablation (Table 4) that isolates the contribution of the soft alignment to both perception accuracy and downstream restoration PSNR/SSIM, confirming that the observed gains are not explained by benchmark-specific biases alone. revision: yes

  2. Referee: [MDUR benchmark and synthetic degradation models] The MDUR benchmark relies on synthetic models for rain, haze, and noise. To support the claim of applicability to real UAV environments, the authors need to provide evidence (perhaps in the benchmark or experiments section) that the learned cues and routing do not overfit to simulation artifacts. A comparison with real multi-degraded UAV images or an analysis of domain gap would strengthen the load-bearing assumption.

    Authors: We acknowledge the importance of addressing potential overfitting to synthetic artifacts for real-world UAV applicability. In the revised manuscript we have added a new subsection (Section 5.4) containing (i) a qualitative evaluation of DAME-Net on a collection of real UAV images exhibiting visible rain, haze, and noise, and (ii) a domain-gap analysis via t-SNE visualization of FDPM features extracted from synthetic versus real degraded samples. These additions show that the learned cues and routing produce plausible restorations on real data without obvious simulation-specific artifacts. However, because paired clean ground-truth references for real multi-degraded UAV imagery are not available, we cannot provide quantitative metrics on real data; we have therefore explicitly stated this limitation and its implications in the revised text. revision: partial

Circularity Check

0 steps flagged

No circularity in architectural claims or benchmark evaluation

full rationale

The paper introduces DAME-Net as a new architecture with FDPM (multi-label degradation perception via label-similarity-guided soft alignment) and CDMM (conditioned decoupled MoE with mask-constrained routing), plus the MDUR benchmark with 43 synthetic configurations and explicit seen/unseen splits. All performance claims rest on empirical comparisons to unified restoration baselines rather than any derivation that reduces to fitted inputs, self-definitions, or self-citation chains. No equations or modules are shown to be equivalent to their own training targets by construction; the decoupling mechanism is presented as an explicit design choice evaluated on held-out composites.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 2 invented entities

The central claim rests on the assumption that explicit per-factor cues can be learned and used to route corrections without interference; the network itself contains many learned parameters typical of deep models.

free parameters (2)
  • FDPM multi-label prediction parameters
    Learned weights for predicting degradation presence and strength
  • CDMM expert routing and conditioning parameters
    Learned parameters controlling mask-constrained expert selection and spatial-frequency processing
axioms (2)
  • domain assumption Multiple degradation factors can be perceived independently without entanglement
    Invoked in the design of FDPM to replace implicit conditions
  • domain assumption Selective expert routing can suppress irrelevant interference in composite degradations
    Core premise of the CDMM module
invented entities (2)
  • DAME-Net no independent evidence
    purpose: Overall restoration architecture
    Newly proposed network combining FDPM and CDMM
  • MDUR benchmark no independent evidence
    purpose: Evaluation dataset with 43 degradation configurations
    Newly constructed benchmark for compositional UAV restoration

pith-pipeline@v0.9.0 · 5586 in / 1470 out tokens · 63403 ms · 2026-05-10T17:02:24.405662+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

42 extracted references · 42 canonical work pages

  1. [1]

    Ultrahigh-resolution boreal forest canopy map- ping: Combining UA V imagery and photogrammetric point clouds in a deep-learning-based approach,

    L. Li, X. Mu, F. Chianucci, J. Qi, J. Jiang, J. Zhou, L. Chen, H. Huang, G. Yan, and S. Liu, “Ultrahigh-resolution boreal forest canopy map- ping: Combining UA V imagery and photogrammetric point clouds in a deep-learning-based approach,”International Journal of Applied Earth Observation and Geoinformation, vol. 107, p. 102686, 2022. YANet al.: COMPOSITIO...

    work page 2022

  2. [2]

    Dual convolutional neural networks for low-level vision,

    J. Pan, D. Sun, J. Zhang, J. Tang, J. Yang, Y.-W. Tai, and M.-H. Yang, “Dual convolutional neural networks for low-level vision,”International Journal of Computer Vision, vol. 130, no. 6, pp. 1440–1458, 2022

    work page 2022

  3. [3]

    Impact of adverse weather and image distortions on vision-based UA V detection: A performance evaluation of deep learning models,

    A. Munir, A. J. Siddiqui, S. Anwar, A. El-Maleh, A. H. Khan, and A. Rehman, “Impact of adverse weather and image distortions on vision-based UA V detection: A performance evaluation of deep learning models,”Drones, vol. 8, no. 11, p. 638, 2024

    work page 2024

  4. [4]

    Dc-mamba: A degradation-aware cross-modality framework for blind super-resolution of thermal UA V images,

    P. Si, M. Jia, H. Wang, J. Wang, L. Sun, and Z. Fu, “Dc-mamba: A degradation-aware cross-modality framework for blind super-resolution of thermal UA V images,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    work page 2025

  5. [5]

    arXiv preprint arXiv:2411.10708 , year=

    J. Mao, Y. Yang, X. Yin, L. Shao, and H. Tang, “Allrestorer: All-in-one transformer for image restoration under composite degradations,”arXiv preprint arXiv:2411.10708, 2024

    work page arXiv 2024

  6. [6]

    YOLOv8: A novel object detection algo- rithm with enhanced performance and robustness,

    R. Varghese and M. Sambath, “YOLOv8: A novel object detection algo- rithm with enhanced performance and robustness,” in2024 International conference on advances in data engineering and intelligent computing systems (ADICS). IEEE, 2024, pp. 1–6

    work page 2024

  7. [7]

    Dual-level prototype learning for composite degraded image restoration,

    Z. Wang, H. Zhao, L. Yao, J. Peng, and K. Zhao, “Dual-level prototype learning for composite degraded image restoration,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 14 006–14 016

    work page 2025

  8. [8]

    A comprehensive survey and taxonomy on single image dehazing based on deep learning,

    J. Gui, X. Cong, Y. Cao, W. Ren, J. Zhang, J. Zhang, J. Cao, and D. Tao, “A comprehensive survey and taxonomy on single image dehazing based on deep learning,”ACM Computing Surveys, vol. 55, no. 13s, pp. 1–37, 2023

    work page 2023

  9. [9]

    A survey of deep learning-based low-light image enhancement,

    Z. Tian, P. Qu, J. Li, Y. Sun, G. Li, Z. Liang, and W. Zhang, “A survey of deep learning-based low-light image enhancement,”Sensors, vol. 23, no. 18, p. 7763, 2023

    work page 2023

  10. [10]

    Phydae: Physics-guided degradation-adaptive experts for all-in-one remote sens- ing image restoration,

    Z. Dong, Z. Zhang, Y. Sun, H. Jiang, T. Liu, and Y. Gu, “Phydae: Physics-guided degradation-adaptive experts for all-in-one remote sens- ing image restoration,”IEEE Transactions on Geoscience and Remote Sensing, 2026

    work page 2026

  11. [11]

    Diffbir: Toward blind image restoration with generative diffusion prior,

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, Y. Qiao, W. Ouyang, and C. Dong, “Diffbir: Toward blind image restoration with generative diffusion prior,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 430–448

    work page 2024

  12. [12]

    Moe-diffir: Task-customized diffusion priors for universal compressed image restoration,

    Y. Ren, X. Li, B. Li, X. Wang, M. Guo, S. Zhao, L. Zhang, and Z. Chen, “Moe-diffir: Task-customized diffusion priors for universal compressed image restoration,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 116–134

    work page 2024

  13. [13]

    Unirestorer: Universal image restoration via adaptively estimating image degradation at proper granularity,

    J. Lin, Z. Zhang, W. Li, R. Pei, H. Xu, H. Zhang, and W. Zuo, “Unirestorer: Universal image restoration via adaptively estimating im- age degradation at proper granularity,”arXiv preprint arXiv:2412.20157, 2024

    work page arXiv 2024

  14. [14]

    Promptrestorer: A prompting image restoration method with degrada- tion perception,

    C. Wang, J. Pan, W. Wang, J. Dong, M. Wang, Y. Ju, and J. Chen, “Promptrestorer: A prompting image restoration method with degrada- tion perception,”Advances in Neural Information Processing Systems, vol. 36, pp. 8898–8912, 2023

    work page 2023

  15. [15]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProceedings of the 38th International Conference on Machine Learning. PMLR, 2021, pp. 8748–8763

    work page 2021

  16. [16]

    A comprehensive review of deep learning-based real-world image restoration,

    L. Zhai, Y. Wang, S. Cui, and Y. Zhou, “A comprehensive review of deep learning-based real-world image restoration,”IEEE Access, vol. 11, pp. 21 049–21 067, 2023

    work page 2023

  17. [17]

    Image dehaze algorithm based on improved atmospheric scattering models,

    W. Yan and L. Cui, “Image dehaze algorithm based on improved atmospheric scattering models,”IEEE Access, vol. 12, pp. 98 971– 98 976, 2024

    work page 2024

  18. [18]

    A lightweight degradation-aware framework for robust object detection in adverse weather,

    S. Park, J. Kuai, H. Kim, H. Ko, C. Jung, and Y. Son, “A lightweight degradation-aware framework for robust object detection in adverse weather,”Electronics, vol. 15, no. 1, p. 146, 2025

    work page 2025

  19. [19]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5728–5739

    work page 2022

  20. [20]

    Vision transformers for single image dehazing,

    Y. Song, Z. He, H. Qian, and X. Du, “Vision transformers for single image dehazing,”IEEE Transactions on Image Processing, vol. 32, pp. 1927–1941, 2023

    work page 1927

  21. [21]

    Multi-expert adaptive selection: Task-balancing for all-in-one image restoration,

    X. Yu, S. Zhou, H. Li, and L. Zhu, “Multi-expert adaptive selection: Task-balancing for all-in-one image restoration,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 5, pp. 4619– 4634, 2024

    work page 2024

  22. [22]

    Promptir: Prompting for all-in-one image restoration,

    V. Potlapalli, S. W. Zamir, S. H. Khan, and F. Shahbaz Khan, “Promptir: Prompting for all-in-one image restoration,”Advances in Neural Infor- mation Processing Systems, vol. 36, pp. 71 275–71 293, 2023

    work page 2023

  23. [23]

    Adair: Adaptive all-in-one image restoration via frequency mining and modulation,

    Y. Cui, S. W. Zamir, S. Khan, A. Knoll, M. Shah, and F. S. Khan, “Adair: Adaptive all-in-one image restoration via frequency mining and modulation,” inThe Thirteenth International Conference on Learning Representations, 2025, pp. 57 335–57 356

    work page 2025

  24. [24]

    All-in-one image restoration for unknown corruption,

    B. Li, X. Liu, P. Hu, Z. Wu, J. Lv, and X. Peng, “All-in-one image restoration for unknown corruption,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 17 452–17 462

    work page 2022

  25. [25]

    Transweather: Transformer-based restoration of images degraded by adverse weather conditions,

    J. M. J. Valanarasu, R. Yasarla, and V. M. Patel, “Transweather: Transformer-based restoration of images degraded by adverse weather conditions,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 2353–2363

    work page 2022

  26. [26]

    Complexity experts are task-discriminative learners for any image restoration,

    E. Zamfir, Z. Wu, N. Mehta, Y. Tan, D. P. Paudel, Y. Zhang, and R. Timofte, “Complexity experts are task-discriminative learners for any image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 12 753–12 763

    work page 2025

  27. [27]

    M2restore: Mixture-of-experts-based mamba-cnn fusion framework for all-in-one image restoration,

    Y. Wang, Y. Li, Z. Zheng, X.-P. Zhang, and M. Wei, “M2restore: Mixture-of-experts-based mamba-cnn fusion framework for all-in-one image restoration,”IEEE Transactions on Image Processing, vol. 34, pp. 8086–8100, 2025

    work page 2025

  28. [28]

    Restoring spatially-heterogeneous distortions using mixture of experts network,

    S. Kim, N. Ahn, and K.-A. Sohn, “Restoring spatially-heterogeneous distortions using mixture of experts network,” inProceedings of the Asian Conference on Computer Vision, 2020

    work page 2020

  29. [29]

    Towards unified image deblurring using a mixture-of-experts decoder,

    D. Feijoo, P. Garrido-Mellado, J. Rim, A. Garcia, and M. V. Conde, “Towards unified image deblurring using a mixture-of-experts decoder,” arXiv preprint arXiv:2508.06228, 2025

    work page arXiv 2025

  30. [30]

    Freprompter: Frequency self-prompt for all-in-one image restoration,

    Z. Wu, W. Liu, J. Wang, J. Li, and D. Huang, “Freprompter: Frequency self-prompt for all-in-one image restoration,”Pattern Recognition, vol. 161, p. 111223, 2025

    work page 2025

  31. [31]

    Selective frequency network for image restoration,

    Y. Cui, Y. Tao, Z. Bing, W. Ren, X. Gao, X. Cao, K. Huang, and A. Knoll, “Selective frequency network for image restoration,” inThe Eleventh International Conference on Learning Representations, 2023

    work page 2023

  32. [32]

    Degradation-aware feature perturbation for all-in-one image restoration,

    X. Tian, X. Liao, X. Liu, M. Li, and C. Ren, “Degradation-aware feature perturbation for all-in-one image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 28 165–28 175

    work page 2025

  33. [33]

    Mixed degrada- tion image restoration via local dynamic optimization and conditional embedding,

    Y. Gu, Y. Meng, X. Sun, J. Ji, W. Ruan, and R. Ji, “Mixed degrada- tion image restoration via local dynamic optimization and conditional embedding,”arXiv preprint arXiv:2411.16217, 2024

    work page arXiv 2024

  34. [34]

    Ffdnet: Toward a fast and flexible solution for CNN-based image denoising,

    K. Zhang, W. Zuo, and L. Zhang, “Ffdnet: Toward a fast and flexible solution for CNN-based image denoising,”IEEE Transactions on Image Processing, vol. 27, no. 9, pp. 4608–4622, 2018

    work page 2018

  35. [35]

    Deep unfolding network for image super-resolution,

    K. Zhang, L. V. Gool, and R. Timofte, “Deep unfolding network for image super-resolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 3217–3226

    work page 2020

  36. [36]

    Parame- ter efficient adaptation for image restoration with heterogeneous mixture- of-experts,

    H. Guo, T. Dai, Y. Bai, B. Chen, X. Ren, Z. Zhu, and S.-T. Xia, “Parame- ter efficient adaptation for image restoration with heterogeneous mixture- of-experts,”Advances in Neural Information Processing Systems, vol. 37, pp. 13 522–13 547, 2024

    work page 2024

  37. [37]

    Promptcir: Blind compressed image restoration with prompt learning,

    B. Li, X. Li, Y. Lu, R. Feng, M. Guo, S. Zhao, L. Zhang, and Z. Chen, “Promptcir: Blind compressed image restoration with prompt learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 6442–6452

    work page 2024

  38. [38]

    All-in-one multi- degradation image restoration network via hierarchical degradation rep- resentation,

    C. Zhang, Y. Zhu, Q. Yan, J. Sun, and Y. Zhang, “All-in-one multi- degradation image restoration network via hierarchical degradation rep- resentation,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 2285–2293

    work page 2023

  39. [39]

    Dpmambair: All-in-one image restoration via degradation-aware prompt state space model,

    Z. Liu, S. Zhou, Y. Dai, Y. Wang, Y. An, and X. Zhao, “Dpmambair: All-in-one image restoration via degradation-aware prompt state space model,”arXiv preprint arXiv:2504.17732, 2025

    work page arXiv 2025

  40. [40]

    Uni- versal image restoration pre-training via degradation classi- fication.arXiv preprint arXiv:2501.15510, 2025

    J. Hu, L. Jin, Z. Yao, and Y. Lu, “Universal image restoration pre-training via degradation classification,”arXiv preprint arXiv:2501.15510, 2025

    work page arXiv 2025

  41. [41]

    Controlling vision- language model for enhancing image restoration,

    M. Shao, W. Liu, Q. Li, L. Meng, and Y. Wan, “Controlling vision- language model for enhancing image restoration,”Image and Vision Computing, vol. 158, p. 105538, 2025

    work page 2025

  42. [42]

    Vision-language model guided image restoration,

    C. Yang, R. Dong, and K.-M. Lam, “Vision-language model guided image restoration,”arXiv preprint arXiv:2512.17292, 2025

    work page arXiv 2025