Pith. sign in

REVIEW 3 major objections 5 minor 79 references

Adaptive Blind Super-Resolution Network for Spatial-Specific and Spatial-Agnostic Degradations

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

Pith's one-line read This paper proposes that real degradations split into spatial-agnostic (noise, downsampling) and spatial-specific (blur) classes, and builds a dual dynamic filter network, GLDFN, with a separate branch for each, reporting gains over…

desk verdict A legitimate blind-SR architecture paper with modest gains; the two-class degradation taxonomy is a useful framing but the paper's own ablations undercut the claimed branch specialization. read the letter →

arxiv 2506.07705 v1 pith:6N3MAV4N submitted 2025-06-09 cs.CV eess.IV

classification cs.CVeess.IV
keywords blindsuper-resolutionspatial-agnosticdegradationspatial-specificglobaldynamicfilteringlocalfilternetworkimagerestorationmodeling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that blind image super-resolution is better served by treating different degradations separately rather than with one uniform network. It classifies downsampling and noise as spatial-agnostic dominant degradations, meaning they affect the image more or less uniformly, and blur as a spatial-specific dominant degradation, because its type and strength vary across the image. On that basis it builds GLDFN, a network whose global dynamic filtering branch generates per-sample convolution weights for the spatial-agnostic class, while its local dynamic filtering branch generates per-region filters for the spatial-specific class. The reported experiments on synthetic benchmarks, a new spatially varying test set, and real historical photos place GLDFN at or near the top of the compared blind SR methods, with the clearest margins on anisotropic-kernel data. If true, the result would show that a simple two-way division of degradations can replace explicit blur-kernel estimation in practical super-resolution.

What carries the argument

The load-bearing mechanism is the pair of dynamic filtering layers inside each Dual Dynamic Filter Module. The global dynamic filtering layer builds on dynamic convolution: $K$ standard convolution kernels are weighted by attention scores $\pi_k(x)$ that depend on the input, so the effective kernel is assembled differently for every sample; this is what the paper uses to perceive spatial-agnostic degradations. The local dynamic filtering layer builds on decoupled dynamic filters and computes, for feature pixel $i$, a spatial dynamic filter $D^{\mathrm{sp}}_i$ and a channel dynamic filter $D^{\mathrm{ch}}_r$, so that the convolution operator applied in Eq. (5) changes from region to region; this is what perceives spatial-specific degradations. These two branches are organized into Dual Dynamic Filter Groups with residual-in-residual structure, fused after each group, and finally reconstructed with a sub-pixel convolution, with only L1 loss as supervision. The work of this machinery is to make the network input-adaptive at two granularities, per sample and per location, without ever estimating an explicit blur kernel.

What would settle it

An independent reimplementation of GLDFN trained under the paper's protocol would falsify the core competitive claim if it failed to reach or exceed DCLS's PSNR on DIV2KRK at 2x and 4x. A more targeted test of the taxonomy would be to evaluate the single-branch ablations on a texture-poor isotropic-blur dataset such as BSD100; if the local branch still beats the global branch there, the claim that global filters handle spatially invariant degradations would lose support.

Watch

Extended reading notes

Core claim

The central discovery is a division of labor for blind super-resolution. The authors show that the standard degradation model $y = (x \circledast k)\downarrow_s + n$ mixes two operationally different effects: blur, whose kernel can vary locally, and noise and downsampling, which act globally and uniformly. Instead of estimating a single blur kernel, GLDFN runs two dynamic filter branches in parallel inside each Dual Dynamic Filter Module. The global branch uses an attention mechanism to combine several parallel convolution kernels into sample-dependent weights, letting the network adapt to image-wide degradation; the local branch uses spatial and channel attention to produce a different filter at each pixel region, letting it adapt to spatially varying blur. After repeated groups of these modules with residual-in-residual connections and a fusion step per group, a sub-pixel convolution reconstructs the high-resolution image, trained only with an L1 loss. In the paper's comparisons, this combination gives the best or second-best numbers on most settings, and on the anisotropic Gaussian blur test set DIV2KRK it reports PSNR improvements of roughly 0.1 dB over the previous best method DCLS at both scale factors.

Load-bearing premise

The argument rests on the premise that practical degradations split cleanly into image-wide types (noise, downsampling) and location-specific types (blur), with the global branch for the first and the local branch for the second; if that split does not hold, the stated rationale for the architecture weakens even though the network may still function.

Editorial extensions

If this is right

  • Blind super-resolution can bypass explicit blur-kernel estimation, since the dual dynamic filters absorb degradation information implicitly and still handle blur, noise, and downsampling.
  • On anisotropic Gaussian blur (DIV2KRK), the proposed method reports higher PSNR than DCLS, DANv2, and KOALAnet at both 2x and 4x scales.
  • On spatially varying blur with random noise (COCO Valid200), the method reports the best PSNR and SSIM among the compared methods, with larger SSIM gains at higher noise levels.
  • On real historical images with no ground truth, the method reports a better NIQE score than the compared blind SR methods, matching the qualitative result of sharper textures and fewer ringing artifacts.
  • The ablation shows that the local branch alone can beat the global branch on isotropic Gaussian blur on Urban100, so the two branches are complementary in practice even where the paper's simple taxonomy does not predict which branch dominates.

Reading between the lines

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

  • A natural extension the authors leave unexplored is making the branch assignment learned rather than fixed by the taxonomy: the Setting 1 ablation already shows that texture content can outweigh degradation type, so an adaptive router could decide which branch handles which feature map.
  • The same spatial-agnostic versus spatial-specific split could apply to other restoration tasks, such as video super-resolution or joint deblurring and denoising, where camera noise is sensor-wide while motion blur varies per object, though the paper does not test those settings.
  • Because GLDFN reports larger gains as noise increases on the spatially varying test set, the method may be especially suited to high-noise real imagery such as surveillance or mobile photos, and a field test on such images would be needed to confirm.
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. The manuscript proposes GLDFN, a blind super-resolution network with two branches: a global dynamic filter layer (attention-weighted combination of convolution kernels) intended for spatial-agnostic degradations such as downsampling and noise, and a local dynamic filter layer (decoupled dynamic filter) intended for spatial-specific degradations such as blur. The method is evaluated under three synthetic degradation settings (isotropic Gaussian blur, anisotropic Gaussian blur, and spatially varying blur plus noise) and on real historical images, with ablations and comparisons against several recent blind SR methods. The central claim is that the two-branch design outperforms state-of-the-art blind SR methods by exploiting the proposed degradation taxonomy.

Significance. The paper addresses a timely problem and contains extensive experiments across multiple degradation settings, including a new spatially varying test set that the authors commit to release. The breadth of comparison and the ablation study are useful contributions. However, the evidence does not currently establish the central design rationale: the ablation results on Setting 1 directly contradict the proposed taxonomy, and the reported gains over the closest competitor, DCLS, are small and inconsistent across benchmarks. The manuscript would be strengthened by toning down the SOTA claim and by providing direct evidence of branch specialization.

major comments (3)
  1. [Abstract and Conclusions; Section IV.D, Table II] The statement that GLDFN 'outperforms state-of-the-art blind super-resolution algorithms' is not supported by the quantitative results. In Setting 1 (Table II), GLDFN is below DCLS on all five datasets at scale 4 (e.g., Set5 31.90 vs 32.12 PSNR) and on four of five datasets at scale 3; only at scale 2 is it competitive. The text in Section IV.D itself says performance is 'on par' for this setting. The abstract and conclusion should be revised to match the actual results, or the claimed advantage should be narrowed to specific settings such as Setting 2.
  2. [Section III.C and Section IV.C, Table I] The central taxonomy—global branch for spatial-agnostic degradations, local branch for spatial-specific degradations—is directly contradicted by the paper's own ablation. In Setting 1, the degradation is an isotropic Gaussian blur, which is spatially invariant and therefore spatial-agnostic under the paper's classification; nevertheless, the local branch (Exp. C) outperforms the global branch (Exp. B) on Urban100 (23.17 vs 23.13 PSNR). The authors acknowledge the contradiction and offer a post hoc explanation about Urban100 textures. This means the taxonomy is not predictive, and the claimed functional specialization of the two branches is unsupported. The corresponding Setting 2 ablation also shows only a 0.03 dB difference between B and C (32.65 vs 32.62), which is too small to support the claim that anisotropic blur is handled by the global branch. The paper should either provide evidence that the branches actually specialize (e.g., controlled experiments isolating noise-only vs blur-only vs spatially varying blur, or analysis of the learned filters) or reframe the contribution as an empirical architecture without the taxonomy as its foundation.
  3. [Section III.C, Global Dynamic Filter Layer] The assertion that the global dynamic filter layer 'can perceive the spatial-agnostic dominant degradation in different images' is not demonstrated by any experiment. The ablation in Table I only shows aggregate PSNR/SSIM differences; it does not isolate whether improvements come from the proposed degradation-type specialization or simply from additional capacity and input-dependent parameters. Without a direct test of specialization, the attribution of the reported gains to the taxonomy remains speculative.
minor comments (5)
  1. [Section III.D, Eq. (7)] The text says 'we adopt the L1 loss' but Eq. (7) defines L_sr = sqrt(||I_SR - I_GT||^2), which is the L2 norm. Please clarify which loss is actually used; if L1 is used, the equation is wrong, and if L2 is used, the description is wrong.
  2. [Section IV.D and Table II] The method name 'AdaTarget' is spelled inconsistently as 'AdaTargett' at several places in the text and in the Table IV caption; please unify the spelling.
  3. [Section 'Performance on Real Degradation Images'] There is a duplicated phrase 'ringing artifactsringing artifacts' in the sentence describing the qualitative comparison; please fix the typo.
  4. [Section IV.A, Setting 3] The synthesis of the COCO Valid200 test set is described only as 'Other settings are the same as described in [43]'; please provide the exact blur kernel, noise, and downsampling parameters needed to reproduce the test set.
  5. [Table I and Figure 3] In Table I, the 'Params (M)' column is not clearly separated from the metric columns, making the table hard to read; also the PSNR/SSIM values shown under the GT column in Figure 3 are unexplained and do not seem to correspond to a comparison with the ground truth.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central contribution is an empirical architecture evaluated on external benchmarks; no derivation step reduces to its own inputs.

full rationale

The paper is an empirical blind-super-resolution study. Its central claim, that integrating global and local dynamic filtering branches outperforms state-of-the-art blind SR methods, is supported by comparisons on external benchmarks (DIV2KRK, Urban100, Set5, Set14, BSD100, Manga109) and on a synthetic spatially varying test set built from the external DMBSR degradation model. No equation or fitted parameter is reused as a prediction, and no claimed result is defined in terms of the quantity it is supposed to predict. The spatial-agnostic versus spatial-specific degradation taxonomy is a conceptual categorization, not a self-referential derivation; the ablation section even admits a mismatch with the taxonomy ('This observation contradicts our initial assumption that global dynamic filters would be more effective for isotropic uniform blurring'), which is an empirical weakness, not a circular step. The only self-citation is reference [5] (a holistic attention network co-authored by W. Wen and W. Ren), which is cited in related work for introducing layer attention and channel-spatial attention and is not load-bearing for the paper's claims. There is no uniqueness theorem imported from the authors' prior work, no ansatz smuggled in via self-citation, and no renaming of a known result presented as a derivation. Accordingly, the derivation chain is self-contained with respect to circularity concerns, even though the stated design rationale is imperfectly supported by the ablation evidence.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The paper relies on the standard blind SR degradation model and on its own two-class taxonomy. The taxonomy is a design postulate rather than a proven fact, and the ablation partially undercuts it. There are several hand-picked hyperparameters and training choices, but no new physical entities or fitted theoretical constants.

free parameters (7)
  • Number of DDFGs = 5
    Chosen by hand; the depth of the refinement network. Performance likely depends on this.
  • Number of DDFMs per DDFG = 10
    Chosen by hand; determines capacity of each group.
  • Input patch size = 64x64
    Standard for SR training; manually selected.
  • Initial learning rate = 4e-4, halved every 2e5 iterations
    Manual schedule choice; affects convergence.
  • Training iterations = 5e5 for main models, 2e5 for ablations
    Manual stopping criterion, different for ablations.
  • Blur kernel width ranges = Isotropic: [0.2,2.0], [0.2,3.0], [0.2,4.0] for x2,x3,x4; anisotropic: (0.6,5) with angles in [-pi,pi]
    Training degradation distribution, manually chosen following prior work [11], [12].
  • Noise levels for Setting 3 test set = [0,5] and [5,10]
    Manually chosen for the synthesized COCO Valid200 test set.
assumptions (3)
  • domain assumption The degradation process follows y = (x ⊗ k) ↓_s + n (Eq. 1).
    Standard blind SR assumption adopted from prior work; the paper's entire training and evaluation pipeline depends on this model of real-world degradation.
  • ad hoc to paper Practical degradations can be divided into spatial-agnostic dominant (downsampling, noise) and spatial-specific dominant (blur) classes.
    This taxonomy is the paper's own contribution; it is not derived from first principles and is partially contradicted by the ablation in Setting 1 where local filters beat global filters.
  • standard math Standard CNN building blocks (convolution, residual learning, attention) behave as expected in this setting.
    Likely true; relies on established deep learning results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Blind Super-Resolution Network for Spatial-Specific and Spatial-Agnostic Degradations." pith.science (2026). https://pith.science/paper/6N3MAV4N

@misc{pith2026250607705,
  author       = {Pith},
  title        = {Pith review of: Adaptive Blind Super-Resolution Network for Spatial-Specific and Spatial-Agnostic Degradations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6N3MAV4N}},
  note         = {Machine review of arXiv:2506.07705}
}
read the original abstract

Prior methodologies have disregarded the diversities among distinct degradation types during image reconstruction, employing a uniform network model to handle multiple deteriorations. Nevertheless, we discover that prevalent degradation modalities, including sampling, blurring, and noise, can be roughly categorized into two classes. We classify the first class as spatial-agnostic dominant degradations, less affected by regional changes in image space, such as downsampling and noise degradation. The second class degradation type is intimately associated with the spatial position of the image, such as blurring, and we identify them as spatial-specific dominant degradations. We introduce a dynamic filter network integrating global and local branches to address these two degradation types. This network can greatly alleviate the practical degradation problem. Specifically, the global dynamic filtering layer can perceive the spatial-agnostic dominant degradation in different images by applying weights generated by the attention mechanism to multiple parallel standard convolution kernels, enhancing the network's representation ability. Meanwhile, the local dynamic filtering layer converts feature maps of the image into a spatially specific dynamic filtering operator, which performs spatially specific convolution operations on the image features to handle spatial-specific dominant degradations. By effectively integrating both global and local dynamic filtering operators, our proposed method outperforms state-of-the-art blind super-resolution algorithms in both synthetic and real image datasets.

Figures

Figures reproduced from arXiv: 2506.07705 by the authors.

Figure 1
Figure 1. This image is “img 87.png” from the DIV2KRK [8] dataset. The image is divided into three parts containing different degradation effects: blur, downsampling, and noise. On the left side of the image are four blur kernels, which are estimated from different positions in the blurred area. the simple down-sampling assumption. Although methods like [4]–[7] have shown promising results on synthetic bench￾mark datasets, th… view at source ↗
Figure 2
Figure 2. The architecture of the proposed global and local dynamic filters network (GLDFN). Given an LR image input, we obtain shallow blur features [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Ablation experiments for 4× SR with different network architectures on the Urban100 [57] dataset. Our algorithm GLDFN with all the proposed modules outperforms other configurations. TABLE I ABLATION STUDIES OF NETWORK ARCHITECTURES. RESULTS ARE REPORTED AS AVERAGE PSNR AND SSIM ON THREE DATASETS: URBAN100 [57] FOR SETTING 1, DIV2KRK [8] FOR SETTING 2, AND COCO V alid 200 [58] FOR SETTING 3. Exp. Global Local Setting… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visual results of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visual results of Img 14, Img 73 and Img 98 in DIV2KRK [8], for scale factor 4. Please zoom in for the best view. Setting 1: Isotropic Gaussian Kernel Setting. Firstly, we conduct blind SR experiments on isotropic Gaussian kernel setting. Following [11], [12], we unifo…
Figure 6
Figure 6. Figure 6: Visual results of Img 007 and Img 008 in Historical [69] dataset, for scale factor 2. Please zoom in for the best view. GT DANv2 DCLS KDSR RealDAN MRDA Ours Img 176 in COCO Img 015 in COCO [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Visual results of 015 and 176 in COCO V alid 200 dataset, for scale factor 4. Please zoom in for the best view. crucial components within the proposed approach GLDFN. To expedite the process, we conducted a mere 2 × 105 iterations for all ablation experiments (while th…
Figure 8
Figure 8. Figure 8: Visual results of Img 033, Img 044 and Img 052 in Urban100 [57] dataset, for scale factor 4. Please zoom in for the best view. TABLE V QUANTITATIVE COMPARISON ON COCO [58] IN THE PRESENCE OF SPATIALLY VARYING BLUR AND RANDOM NOISE. THE BEST AND SECOND BEST RESULTS ARE …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

79 extracted references · 76 canonical work pages

  1. [1]

    Learning a deep convolu- tional network for image super-resolution,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Learning a deep convolu- tional network for image super-resolution,” inEuropean Conference on Computer Vision, 2014, pp. 184–199. 1, 3

  2. [2]

    Deeply-recursive convolutional network for image super-resolution,

    J. Kim, J. Kwon Lee, and K. Mu Lee, “Deeply-recursive convolutional network for image super-resolution,” inIEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 1637–1645. 1

  3. [3]

    Enhanced deep resid- ual networks for single image super-resolution,

    B. Lim, S. Son, H. Kim, S. Nah, and K. Mu Lee, “Enhanced deep resid- ual networks for single image super-resolution,” inIEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 136–144. 1, 3, 12

  4. [4]

    Learning a single convolutional super- resolution network for multiple degradations,

    K. Zhang, W. Zuo, and L. Zhang, “Learning a single convolutional super- resolution network for multiple degradations,” inCVPR, 2018, pp. 3262–

  5. [5]

    Single image super-resolution via a holistic attention network,

    B. Niu, W. Wen, W. Ren, X. Zhang, L. Yang, S. Wang, K. Zhang, X. Cao, and H. Shen, “Single image super-resolution via a holistic attention network,” inECCV. Springer, 2020, pp. 191–207. 1, 3

  6. [6]

    Second-order atten- tion network for single image super-resolution,

    T. Dai, J. Cai, Y . Zhang, S.-T. Xia, and L. Zhang, “Second-order atten- tion network for single image super-resolution,” inIEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 11 065–11 074. 1, 3

  7. [7]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inIEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778. 1, 3 14 IEEE TRANSACTIONS ON IMAGE PROCESSING

  8. [8]

    Blind super-resolution kernel estimation using an internal-gan,

    S. Bell-Kligler, A. Shocher, and M. Irani, “Blind super-resolution kernel estimation using an internal-gan,”Advances in Neural Information Processing Systems, vol. 32, 2019. 1, 2, 4, 5, 8, 9, 11, 12

Show all 79 references
  1. [9]

    Deep constrained least squares for blind image super-resolution,

    Z. Luo, H. Huang, L. Yu, Y . Li, H. Fan, and S. Liu, “Deep constrained least squares for blind image super-resolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 17 642–17 652. 1, 7, 11, 12, 13

  2. [10]

    End-to-end alternating optimization for blind super resolution,

    Z. Luo, Y . Huang, S. Li, L. Wang, and T. Tan, “End-to-end alternating optimization for blind super resolution,” 2021. 1, 4, 7, 11, 12, 13

  3. [11]

    Unfolding the alter- nating optimization for blind super resolution,

    Z. Luo, Y . Huang, S. Li, L. Wang, and T. Tan, “Unfolding the alter- nating optimization for blind super resolution,” inAdvances in Neural Information Processing Systems, 2020. 1, 2, 4, 7, 9, 11, 12

  4. [12]

    Blind super-resolution with iterative kernel correction,

    J. Gu, H. Lu, W. Zuo, and C. Dong, “Blind super-resolution with iterative kernel correction,” inCVPR, 2019, pp. 1604–1613. 1, 2, 4, 7, 8, 9, 11, 12

  5. [13]

    Ddet: Dual-path dynamic enhancement network for real-world image super-resolution,

    Y . Shi, H. Zhong, Z. Yang, X. Yang, and L. Lin, “Ddet: Dual-path dynamic enhancement network for real-world image super-resolution,” IEEE Signal Processing Letters, vol. 27, pp. 481–485, 2020. 1

  6. [14]

    On efficient trans- former and image pre-training for low-level vision,

    W. Li, X. Lu, J. Lu, X. Zhang, and J. Jia, “On efficient trans- former and image pre-training for low-level vision,”arXiv preprint arXiv:2112.10175, 2021. 2, 3

  7. [15]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1833–1844. 2, 3

  8. [16]

    Activating more pixels in image super-resolution transformer,

    X. Chen, X. Wang, J. Zhou, Y . Qiao, and C. Dong, “Activating more pixels in image super-resolution transformer,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 367–22 377. 2, 3

  9. [17]

    Dynamic convolution: Attention over convolution kernels,

    Y . Chen, X. Dai, M. Liu, D. Chen, L. Yuan, and Z. Liu, “Dynamic convolution: Attention over convolution kernels,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 030–11 039. 2, 4, 6

  10. [18]

    Decoupled dynamic filter networks,

    J. Zhou, V . Jampani, Z. Pi, Q. Liu, and M.-H. Yang, “Decoupled dynamic filter networks,” inCVPR, 2021, pp. 6647–6656. 2, 4, 6

  11. [19]

    Accurate image super-resolution using very deep convolutional networks,

    J. Kim, J. Kwon Lee, and K. Mu Lee, “Accurate image super-resolution using very deep convolutional networks,” inIEEE Conference on Com- puter Vision and Pattern Recognition, 2016, pp. 1646–1654. 3

  12. [20]

    Real-esrgan: Training real- world blind super-resolution with pure synthetic data,

    X. Wang, L. Xie, C. Dong, and Y . Shan, “Real-esrgan: Training real- world blind super-resolution with pure synthetic data,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1905–1914. 3

  13. [21]

    Esrgan: Enhanced super-resolution generative adversar- ial networks,

    X. Wang, K. Yu, S. Wu, J. Gu, Y . Liu, C. Dong, Y . Qiao, and C. Change Loy, “Esrgan: Enhanced super-resolution generative adversar- ial networks,” inProceedings of the European conference on computer vision (ECCV) workshops, 2018, pp. 0–0. 3, 8

  14. [22]

    Ranksrgan: Generative adver- sarial networks with ranker for image super-resolution,

    W. Zhang, Y . Liu, C. Dong, and Y . Qiao, “Ranksrgan: Generative adver- sarial networks with ranker for image super-resolution,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 3096–3105. 3

  15. [23]

    Image super- resolution using very deep residual channel attention networks,

    Y . Zhang, K. Li, K. Li, L. Wang, B. Zhong, and Y . Fu, “Image super- resolution using very deep residual channel attention networks,” in European Conference on Computer Vision, 2018, pp. 286–301. 3, 6, 12

  16. [24]

    Advanced binary neural network for single image super resolution,

    J. Xin, N. Wang, X. Jiang, J. Li, and X. Gao, “Advanced binary neural network for single image super resolution,”International Journal of Computer Vision, vol. 131, no. 7, pp. 1808–1824, 2023. 3

  17. [25]

    Binarized neural network for single image super resolution,

    J. Xin, N. Wang, X. Jiang, J. Li, H. Huang, and X. Gao, “Binarized neural network for single image super resolution,” inComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16. Springer, 2020, pp. 91–107. 3

  18. [26]

    Wavelet-based dual recursive network for image super-resolution,

    J. Xin, J. Li, X. Jiang, N. Wang, H. Huang, and X. Gao, “Wavelet-based dual recursive network for image super-resolution,”IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 2, pp. 707–720,

  19. [27]

    Pre-trained image processing transformer,

    H. Chen, Y . Wang, T. Guo, C. Xu, Y . Deng, Z. Liu, S. Ma, C. Xu, C. Xu, and W. Gao, “Pre-trained image processing transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 12 299–12 310. 3

  20. [28]

    Eatformer: Improving vision transformer inspired by evolutionary al- gorithm,

    J. Zhang, X. Li, Y . Wang, C. Wang, Y . Yang, Y . Liu, and D. Tao, “Eatformer: Improving vision transformer inspired by evolutionary al- gorithm,”International Journal of Computer Vision, pp. 1–28, 2024. 3

  21. [29]

    Unified dynamic convolutional network for super-resolution with variational degradations,

    Y .-S. Xu, S.-Y . R. Tseng, Y . Tseng, H.-K. Kuo, and Y .-M. Tsai, “Unified dynamic convolutional network for super-resolution with variational degradations,” inCVPR, 2020, pp. 12 496–12 505. 3

  22. [30]

    Deep unfolding network for image super-resolution,

    K. Zhang, L. V . Gool, and R. Timofte, “Deep unfolding network for image super-resolution,” inCVPR, 2020, pp. 3217–3226. 3

  23. [31]

    Correction filter for single image super-resolution: Robustifying off-the-shelf deep super-resolvers,

    S. A. Hussein, T. Tirer, and R. Giryes, “Correction filter for single image super-resolution: Robustifying off-the-shelf deep super-resolvers,” inCVPR, 2020, pp. 1428–1437. 3, 12

  24. [32]

    Knowledge distillation based degradation estimation for blind super-resolution,

    B. Xia, Y . Zhang, Y . Wang, Y . Tian, W. Yang, R. Timofte, and L. Van Gool, “Knowledge distillation based degradation estimation for blind super-resolution,” inInternational Conference on Learning Representations, 2023. 3, 12, 13

  25. [33]

    Meta- learning based degradation representation for blind super-resolution,

    B. Xia, Y . Tian, Y . Zhang, Y . Hang, W. Yang, and Q. Liao, “Meta- learning based degradation representation for blind super-resolution,” IEEE Transactions on Image Processing, 2023. 3, 12, 13

  26. [34]

    “zero-shot

    A. Shocher, N. Cohen, and M. Irani, ““zero-shot” super-resolution using deep internal learning,” inCVPR, 2018, pp. 3118–3126. 3, 4, 7, 11, 12

  27. [35]

    Meta-transfer learning for zero-shot super-resolution,

    J. W. Soh, S. Cho, and N. I. Cho, “Meta-transfer learning for zero-shot super-resolution,” inCVPR, 2020, pp. 3516–3525. 3

  28. [36]

    Deep learning-based blind image super-resolution with iterative kernel reconstruction and noise estimation,

    H. F. Ates, S. Yildirim, and B. K. Gunturk, “Deep learning-based blind image super-resolution with iterative kernel reconstruction and noise estimation,”Computer Vision and Image Understanding, vol. 233, p. 103718, 2023. 4

  29. [37]

    End-to-end alternating optimization for real-world blind super resolution,

    Z. Luo, Y . Huang, S. Li, L. Wang, and T. Tan, “End-to-end alternating optimization for real-world blind super resolution,”International Journal of Computer Vision, vol. 131, no. 12, pp. 3152–3169, 2023. 4, 12, 13

  30. [38]

    Unsu- pervised degradation representation learning for blind super-resolution,

    L. Wang, Y . Wang, X. Dong, Q. Xu, J. Yang, W. An, and Y . Guo, “Unsu- pervised degradation representation learning for blind super-resolution,” inCVPR, 2021, pp. 10 581–10 590. 4

  31. [39]

    Degradation-aware self-attention based transformer for blind image super-resolution,

    Q. Liu, P. Gao, K. Han, N. Liu, and W. Xiang, “Degradation-aware self-attention based transformer for blind image super-resolution,”IEEE Transactions on Multimedia, 2024. 4

  32. [40]

    Efficient and degradation-adaptive network for real-world image super-resolution,

    J. Liang, H. Zeng, and L. Zhang, “Efficient and degradation-adaptive network for real-world image super-resolution,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 574–591. 4

  33. [41]

    A dynamic kernel prior model for unsupervised blind image super- resolution,

    Z. Yang, J. Xia, S. Li, X. Huang, S. Zhang, Z. Liu, Y . Fu, and Y . Liu, “A dynamic kernel prior model for unsupervised blind image super- resolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 046–26 056. 4

  34. [42]

    Uncertainty learning in kernel estimation for multi-stage blind image super-resolution,

    Z. Fang, W. Dong, X. Li, J. Wu, L. Li, and G. Shi, “Uncertainty learning in kernel estimation for multi-stage blind image super-resolution,” in Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. Springer, 2022, ...

  35. [43]

    Deep model-based super- resolution with non-uniform blur,

    C. Laroche, A. Almansa, and M. Tassano, “Deep model-based super- resolution with non-uniform blur,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2023, pp. 1797–

  36. [44]

    Koalanet: Blind super-resolution using kernel-oriented adaptive local adjustment,

    S. Y . Kim, H. Sim, and M. Kim, “Koalanet: Blind super-resolution using kernel-oriented adaptive local adjustment,” inIEEE Conference on Computer Vision and Pattern Recognition, 2021, pp. 10 611–10 620. 4, 11, 12

  37. [45]

    Mutual affine network for spatially variant kernel estimation in blind image super- resolution,

    J. Liang, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Mutual affine network for spatially variant kernel estimation in blind image super- resolution,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 4096–4105. 4

  38. [46]

    Blind image super-resolution with spatially variant degra- dations,

    V . Cornillere, A. Djelouah, W. Yifan, O. Sorkine-Hornung, and C. Schroers, “Blind image super-resolution with spatially variant degra- dations,”ACM Transactions on Graphics (TOG), vol. 38, no. 6, pp. 1–13, 2019. 4

  39. [47]

    Edpn: Enhanced deep pyramid network for blurry image restoration,

    R. Xu, Z. Xiao, J. Huang, Y . Zhang, and Z. Xiong, “Edpn: Enhanced deep pyramid network for blurry image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 414–423. 4

  40. [48]

    Joint learning content and degradation aware feature for blind super- resolution,

    Y . Zhou, C. Lin, D. Luo, Y . Liu, Y . Tai, C. Wang, and M. Chen, “Joint learning content and degradation aware feature for blind super- resolution,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 2606–2616. 4

  41. [49]

    Dynamic filter networks,

    X. Jia, B. De Brabandere, T. Tuytelaars, and L. V . Gool, “Dynamic filter networks,”Advances in neural information processing systems, vol. 29,

  42. [50]

    Adaptive convolutional kernels,

    J. Zamora Esquivel, A. Cruz Vargas, P. Lopez Meyer, and O. Tickoo, “Adaptive convolutional kernels,” inICCV Workshops, 2019. 4

  43. [51]

    Solov2: Dynamic, faster and stronger,

    X. Wang, R. Zhang, T. Kong, L. Li, and C. Shen, “Solov2: Dynamic, faster and stronger,”arXiv preprint arXiv:2003.10152, 2020. 4

  44. [52]

    Conditional convolutions for instance segmentation,

    Z. Tian, C. Shen, and H. Chen, “Conditional convolutions for instance segmentation,” inECCV, 2020. 4

  45. [53]

    Dynamic high-pass filtering and multi-spectral attention for 15 image super-resolution,

    S. A. Magid, Y . Zhang, D. Wei, W.-D. Jang, Z. Lin, Y . Fu, and H. Pfister, “Dynamic high-pass filtering and multi-spectral attention for 15 image super-resolution,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 4288–4297. 4

  46. [54]

    Condconv: Conditionally parameterized convolutions for efficient inference,

    B. Yang, G. Bender, Q. V . Le, and J. Ngiam, “Condconv: Conditionally parameterized convolutions for efficient inference,” inNeurIPS, 2019. 4

  47. [55]

    Weightnet: Revisiting the design space of weight networks,

    N. Ma, X. Zhang, J. Huang, and J. Sun, “Weightnet: Revisiting the design space of weight networks,” inECCV, 2020. 4

  48. [56]

    Dynet: Dynamic con- volution for accelerating convolutional neural networks,

    Y . Zhang, J. Zhang, Q. Wang, and Z. Zhong, “Dynet: Dynamic con- volution for accelerating convolutional neural networks,”arXiv preprint arXiv:2004.10694, 2020. 4

  49. [57]

    Single image super-resolution from transformed self-exemplars,

    J.-B. Huang, A. Singh, and N. Ahuja, “Single image super-resolution from transformed self-exemplars,” inIEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 5197–5206. 5, 7, 9, 12

  50. [58]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2...

  51. [59]

    Rectified linear units improve restricted boltzmann machines

    V . Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines.” inICML, 2010. 6

  52. [60]

    What is the best multi-stage architecture for object recognition?

    K. Jarrett, K. Kavukcuoglu, M. Ranzato, and Y . LeCun, “What is the best multi-stage architecture for object recognition?” inThe IEEE International Conference on Computer Vision (ICCV), 2009. 6

  53. [61]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” inIEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141. 6

  54. [62]

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

    N. Ahn, B. Kang, and K.-A. Sohn, “Fast, accurate, and lightweight super-resolution with cascading residual network,” inEuropean Confer- ence on Computer Vision, 2018, pp. 252–268. 7, 11

  55. [63]

    Deblurring images via dark channel prior,

    J. Pan, D. Sun, H. Pfister, and M.-H. Yang, “Deblurring images via dark channel prior,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no. 10, pp. 2315–2328, 2017. 7, 11

  56. [64]

    Tackling the ill-posedness of super-resolution through adaptive target generation,

    Y . Jo, S. W. Oh, P. Vajda, and S. J. Kim, “Tackling the ill-posedness of super-resolution through adaptive target generation,” inIEEE Conference on Computer Vision and Pattern Recognition, 2021, pp. 16 236–16 245. 7, 11, 12

  57. [65]

    Low- complexity single-image super-resolution based on nonnegative neighbor embedding,

    M. Bevilacqua, A. Roumy, C. Guillemot, and M. L. Alberi-Morel, “Low- complexity single-image super-resolution based on nonnegative neighbor embedding,”British Machine Vision Conference, 2012. 7, 9

  58. [66]

    On single image scale-up using sparse-representations,

    R. Zeyde, M. Elad, and M. Protter, “On single image scale-up using sparse-representations,” inInternational conference on curves and sur- faces, 2010, pp. 711–730. 7, 9

  59. [67]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,

    D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,” inIEEE International Conference on Computer Vision, vol. 2, 2001, pp. 416–423. 7, 9

  60. [68]

    Sketch-based manga retrieval using manga109 dataset,

    Y . Matsui, K. Ito, Y . Aramaki, A. Fujimoto, T. Ogawa, T. Yamasaki, and K. Aizawa, “Sketch-based manga retrieval using manga109 dataset,” Multimedia Tools and Applications, vol. 76, no. 20, pp. 21 811–21 838,

  61. [69]

    Deep laplacian pyramid networks for fast and accurate super-resolution,

    W.-S. Lai, J.-B. Huang, N. Ahuja, and M.-H. Yang, “Deep laplacian pyramid networks for fast and accurate super-resolution,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 624–632. 10, 11

  62. [70]

    Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network,

    W. Shi, J. Caballero, F. Husz ´ar, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert, and Z. Wang, “Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network,” in CVPR, 2016, pp. 1874–1883. 8

  63. [71]

    Image super-resolution by neural texture transfer,

    Z. Zhang, Z. Wang, Z. Lin, and H. Qi, “Image super-resolution by neural texture transfer,” inIEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 7982–7991. 8

  64. [72]

    Perceptual losses for real- time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real- time style transfer and super-resolution,” inEuropean Conference on Computer Vision, 2016, pp. 694–711. 8

  65. [73]

    Learning parallax attention for stereo image super-resolution,

    L. Wang, Y . Wang, Z. Liang, Z. Lin, J. Yang, W. An, and Y . Guo, “Learning parallax attention for stereo image super-resolution,” inIEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 12 250–12 259. 8

  66. [74]

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

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” inIEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017, pp. 126–135. 8

  67. [75]

    Ntire 2017 challenge on single image super-resolution: Methods and results,

    R. Timofte, E. Agustsson, L. Van Gool, M.-H. Yang, and L. Zhang, “Ntire 2017 challenge on single image super-resolution: Methods and results,” inIEEE Conference on Computer Vision and Pattern Recogni- tion Workshops, 2017, pp. 114–125. 8

  68. [76]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Signal processing letters, vol. 20, no. 3, pp. 209–212, 2012. 8, 11, 13

  69. [77]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. 9

  70. [78]

    Deep back-projection networks for super-resolution,

    M. Haris, G. Shakhnarovich, and N. Ukita, “Deep back-projection networks for super-resolution,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1664–1673. 12 Weilei Wenreceived the master’s degree from the Xidian University, in 2021. ...

  71. [2017]

    Ming-Husan Yang as a joint-training Ph.D

    From 2015 to 2016, he was supported by the China Scholarship Council and worked with Prof. Ming-Husan Yang as a joint-training Ph.D. student with the Electrical Engineering and Com- puter Science Department, University of California at Merced. He is currently a Professor with ...

Pith tools

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