Pith. sign in

REVIEW 3 major objections 5 minor 97 references

Modeling Dual-Exposure Quad-Bayer Patterns for Joint Denoising and Deblurring

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single dual-exposure Quad-Bayer RAW frame contains complementary short- and long-exposure information that a convolutional network can exploit to jointly denoise and deblur, outperforming prior methods with a…

desk verdict Solid engineering paper on dual-exposure Quad-Bayer imaging; the synthetic-to-real gap is a real concern but the central claim holds up. read the letter →

arxiv 2412.07256 v1 pith:XTV56X2G submitted 2024-12-10 eess.IV cs.CV

Yuzhi Zhao , Lai-Man Po , Xin Ye , Yongzhe Xu , Qiong Yan This is my paper · ORCID
classification eess.IVcs.CV
keywords ImageDenoisingDeblurringQuad-BayerSensordualexposureRAWrestorationB2QBsynthesisQRNetcomputationalphotography
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

Single-image restoration faces a tradeoff: short exposures freeze motion but amplify noise, while long exposures gather light but blur. The paper tries to break that tradeoff with a sensor pattern, not just a better network: a Quad-Bayer mosaic that packs short- and long-exposure pixels into one RAW frame, both exposures starting at the same instant. It argues that this single frame carries complementary information—sharp edges from short pixels, clean color from long pixels—so a network can jointly deblur and denoise without the misalignment that plagues burst capture. To train such a network, the authors introduce B2QB, a sampling scheme that synthesizes dual-exposure Quad-Bayer data from paired Bayer captures, and QRNet, a hierarchical convolutional network that maps degraded Quad-Bayer RAW to clean RGB. On their QR dataset and on real IMX586 captures, QRNet reports higher PSNR/SSIM and perceptual quality than state-of-the-art deblurring and denoising methods while using far fewer MACs.

What carries the argument

The load-bearing object is the dual-exposure Quad-Bayer CFA: a 4×4 color filter array in which half the pixels integrate a short exposure and half a long exposure, with both starting at the same time. Its companion mechanism is B2QB, a sampling scheme that turns paired short- and long-exposure Bayer RAW images into synthetic Quad-Bayer mosaics by sampling pixels according to the Quad-Bayer layout, so training pairs can be produced without a physical Quad-Bayer camera. The network side is carried by Pixel Unshuffle (which separates the mosaic into 16 exposure/color channels), the input enhancement block (which merges three average-pooled branches with the unshuffled channels via channel attention to reduce downsampling artifacts), and the five-level hierarchical feature extraction with inter-level connections. The frequency loss adds high-frequency supervision. Together these pieces let a single degraded RAW frame be mapped directly to a clean RGB image.

What would settle it

A decisive check would be to capture true dual-exposure Quad-Bayer images of a static scene with a controllable sensor, measure the actual per-pixel noise distribution of the short-exposure pixels, and compare it with the synthetic sta/A approximation used in training; if QRNet trained on the measured statistics does not match or beat QRNet trained on B2QB data, or if its real-image output quality drops sharply as ISO increases, the synthetic-to-real assumption is the weak link.

Watch

Extended reading notes

Core claim

The paper's central claim is that a dual-exposure Quad-Bayer sensor—where each 4×4 CFA block contains both short- and long-exposure pixels that start exposing at the same time—makes joint denoising and deblurring possible from a single RAW image. The short-exposure pixels record sharp but noisy motion information, and the long-exposure pixels record low-noise, color-accurate but blurred information, so one mosaic already contains the complementary cues that burst methods gather from multiple frames. To exploit this, the authors build QRNet: Pixel Unshuffle reorganizes the mosaic into 16 exposure/color channels, an input enhancement block smooths discontinuities created by downsampling, and a five-level hierarchy with inter-level feature interactions restores the RGB image under an L1 plus frequency loss. Because long- and short-exposure pixels are captured in the same frame, the paper argues there is no readout delay or misalignment, and the blur scale of long pixels can be inferred from neighboring short pixels. The authors also contribute B2QB, a Bayer-to-Quad-Bayer sampling scheme, and the 731-pair QR dataset, and report that QRNet outperforms existing methods on both synthetic and real Quad-Bayer data while using about 5% of the MACs of the strongest deblurring baseline.

Load-bearing premise

The load-bearing premise is that the synthetic B2QB training data faithfully represents real dual-exposure Quad-Bayer captures: the paper creates a 'short exposure' by taking a one-second, noise-free static image and dividing its intensities by the exposure ratio A=4, whereas a genuine short exposure has different photon shot noise, read noise, and motion behavior.

Editorial extensions

If this is right

  • A camera equipped with a dual-exposure Quad-Bayer sensor can perform joint denoising and deblurring from a single captured frame, so the motion-alignment and ghosting problems of burst-based methods disappear.
  • Because the short and long exposures share one readout, there is no readout delay between them, and the network can use neighboring short-exposure pixels to estimate the blur of long-exposure pixels.
  • The B2QB pipeline lets researchers generate large aligned Quad-Bayer training sets from ordinary Bayer captures, removing the need for a controllable Quad-Bayer camera during data collection.
  • With 34.6G MACs on a 512×512 patch, QRNet is substantially cheaper than the compared deblurring networks, suggesting the approach can run on power-constrained devices at interactive speeds.
  • Restoring the Quad-Bayer RAW directly rather than RGB leaves white balance, tone mapping, and color adjustment to the downstream ISP, which the paper demonstrates with its RAW-restoration experiments.

Reading between the lines

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

  • Beyond the paper: the same mosaic-level fusion idea generalizes to other dual- or multi-exposure CFA layouts, such as RGBW or tri-exposure Quad-Bayer patterns, where a Pixel-Unshuffle-style reorganization plus enhancement block could be reused with minimal change.
  • Inference: a direct sensor-calibration study comparing true short-exposure pixel statistics on a Quad-Bayer sensor with the synthetic sta/A approximation would show how much of the reported gain survives the simulation-to-reality gap; the paper's real-image results are qualitative only.
  • A testable extension would be to sweep the exposure ratio A during training and inference; if the network is robust to A, the method could adapt online to changing light, which the current fixed-ratio dataset does not explore.
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 paper proposes a dual-exposure Quad-Bayer sensor pattern in which short- and long-exposure pixels are interleaved in a single RAW mosaic. It introduces a Bayer-to-Quad-Bayer (B2QB) synthesis pipeline that creates a QR dataset from paired Bayer captures, and a hierarchical convolutional network QRNet that maps the Quad-Bayer RAW input to a clean, sharp RGB image (or RAW output). On the synthetic QR validation set, QRNet reports higher PSNR/SSIM than a range of deblurring and denoising baselines while using fewer MACs, and on four real IMX586 images it reports higher no-reference LIQE values. The code, model, and dataset are promised to be public.

Significance. The dual-exposure Quad-Bayer idea is timely and practically motivated: a single-frame CFA that contains both short and long exposures could avoid burst misalignment while supplying complementary noise-blur information. The B2QB dataset is a concrete resource for the community, and QRNet is a reasonably efficient architecture with a sensible multi-level design. The paper also makes its code, model, and dataset available, which increases reproducibility. However, the central claims rest on a synthetic degradation model whose short-exposure channel is not a physical short exposure, and the real-camera evidence is very thin. As presented, the experimental support is not yet strong enough to establish the abstract's claim of superior performance on real-world datasets.

major comments (3)
  1. [IV (Data capturing) and V-B (Noise modeling)] The synthetic short-exposure data do not represent a physical short exposure. In Section IV the authors set sta' = sta/A by dividing a 1 s static, noise-free capture by A=4, so the short-exposure pixels are perfectly sharp and merely scaled in intensity. A true 1/4 s exposure of the moving scene would contain motion blur (less than the long exposure but non-zero) and different shot-noise and read-noise statistics. Moreover, the noise model in Eq. (6) is applied to this pre-scaled static image and then amplified by A in the pre-processing of Section III-B, which does not match the sensor's photon-limited noise process. Because Tables II-IV and the central claim that dual-exposure Quad-Bayer carries complementary noise-blur information are evaluated on this synthetic distribution, the reported gains may substantially overstate real-world performance. Additionally, the long-exposure channel is captured with a static camera viewing a monitor, so the only motion is scene motion; camera shake, which is typically dominant in handheld photography, is absent from the training data.
  2. [V-C (Evaluation with Dual-exposure-based Methods)] The real-camera evaluation is insufficient to validate transfer from the synthetic QR dataset. Only 24 indoor IMX586 images were captured, four are shown, there is no ground truth, and the quality claim relies entirely on the no-reference LIQE measure. The noise parameters are said to be 'estimated on real Quad-Bayer images based on [4]' without reporting a fidelity check, confidence bounds, or a sensitivity analysis. This evidence cannot support the abstract's claim of state-of-the-art performance on real-world datasets; the synthetic-to-real gap remains open.
  3. [V-B (Experiment settings) and Tables II-VI] The paper never states how the baseline networks consume a dual-exposure Quad-Bayer input. The 'Input data structure' column only labels the input as 'dual-exposure Quad-Bayer', but DeblurGAN, MPRNet, SRN, and similar architectures are designed for single-channel RGB or for a specific pre-processing. If all baselines were given the same 16-channel Pixel-Unshuffled tensor, that is a nonstandard input for them; if they were given the raw single-channel mosaic, they are at a disadvantage. The statement that 'the data selection scheme, training epochs, and batch size are identical to QRNet' does not address this, so the relative performance in Tables II-VI is not fully interpretable.
minor comments (5)
  1. [Figure 7 and Section V-F] The method name 'DPMHN' is a typo for 'DMPHN' in the Figure 7 caption and again in the method list of Section V-F.
  2. [Section IV (Data capturing)] The phrase 'All other lighting sources are closed except the monitor' should be reworded to 'turned off' or 'switched off' for clarity.
  3. [Section III-B (Pre-processing)] The notation NM(x)=x+n is imprecise because the noise term in Eq. (6) depends on the signal x; writing NM(x)=x+n(x) would clarify the signal-dependent nature of the shot noise.
  4. [Table I and Eq. (6)] In Table I, 'qu' is defined as a noise parameter, but Eq. (6) uses 'qu' as the quantization step; these notations should be aligned.
  5. [Section V-E (Short-exposure comparison)] Table V contains two QRNet rows (short-exposure and dual-exposure); the text says 'when changing the input data structure to the short-exposure Quad-Bayer for the QRNet, it still outperforms other methods,' which is confusing because QRNet is proposed for dual-exposure input. Consider labeling this clearly as an ablation-style comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: QRNet's claims rest on external benchmarks and a physically motivated synthetic pipeline; self-citations are not load-bearing.

full rationale

The paper contains no derivation in which a predicted quantity is defined from the quantity it predicts. The B2QB dataset generation uses a standard noise model [4] and an exposure-ratio scaling A to synthesize Quad-Bayer inputs, but the subsequent network is trained and evaluated on held-out validation pairs produced by the same pipeline, plus 24 real IMX586 images without ground truth and LIQE comparisons. Training and testing on the same synthetic degradation model is a standard benchmark practice and is not a circular reduction; the real-image evaluation is an external check, albeit thin. Self-citations [13] (D2HNet) and [61] (DSWN) are used only as related work and baselines, not to justify the core claim. The claim that dual-exposure contains complementary information is embedded in the sensor pattern being modeled (short pixels are a dark sharp version, long pixels are bright blurred), but this is an assumption about the sensor, not a 'prediction' derived from the network; the paper's contribution is showing a network can exploit it. No uniqueness theorem, no fitted-parameter-as-prediction, no ansatz smuggled via citation. Concerns about synthetic-to-real transfer are validity risks, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the fidelity of B2QB synthetic data to real dual-exposure sensor behavior. The free parameters are simulation and device choices, not fitted target outputs. No new physical entities are introduced.

free parameters (3)
  • Exposure ratio A = 4
    Set in Section IV, Data capturing, 'according to the real Quad-Bayer device'. It scales the synthetic short-exposure pixels and is central to the dual-exposure simulation.
  • Noise model hyperparameters K and sigma = K in {0.25, 0.5, 0.75}, sigma=5 for synthetic; K=5, sigma=1 for real
    Chosen to simulate different degradation levels in Section V-B and estimated from real IMX586 images based on [4]. They define the training and test degradation distribution.
  • White balance gain ranges = wg=1, wr in [1.6,1.8], wb in [1.4,1.6]
    Used when synthesizing RGB targets from Bayer data in Section IV, B2QB Scheme. Sampled uniformly, not estimated from the real device, which could affect color fidelity.
assumptions (4)
  • domain assumption The physical noise formation model in [4] (Poisson shot noise, Gaussian read noise, uniform quantization) accurately describes real Quad-Bayer RAW noise.
    Invoked in Sections III-B and V-B to generate noisy training data; no validation of the model on IMX586 noise statistics is reported.
  • domain assumption A real short exposure can be approximated by a 1-second static, noise-free capture divided by exposure ratio A=4.
    Section IV, Data capturing: 'we apply an exposure ratio A to the static image sta'. This ignores true short-exposure photon statistics and read noise.
  • domain assumption Motion blur produced by filming moving video on a monitor with a 1-second shutter is representative of real camera motion blur.
    Section IV data capture; the synthetic blur source is screen content, not a moving camera or scene in 3D.
  • domain assumption Interleaving independently captured long- and short-exposure Bayer frames produces a valid dual-exposure Quad-Bayer image, and both exposures start at the same time.
    Section IV B2QB Scheme and Figure 1(j); the same-start-time property is asserted, not measured on hardware.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modeling Dual-Exposure Quad-Bayer Patterns for Joint Denoising and Deblurring." pith.science (2026). https://pith.science/paper/XTV56X2G

@misc{pith2026241207256,
  author       = {Pith},
  title        = {Pith review of: Modeling Dual-Exposure Quad-Bayer Patterns for Joint Denoising and Deblurring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTV56X2G}},
  note         = {Machine review of arXiv:2412.07256}
}
read the original abstract

Image degradation caused by noise and blur remains a persistent challenge in imaging systems, stemming from limitations in both hardware and methodology. Single-image solutions face an inherent tradeoff between noise reduction and motion blur. While short exposures can capture clear motion, they suffer from noise amplification. Long exposures reduce noise but introduce blur. Learning-based single-image enhancers tend to be over-smooth due to the limited information. Multi-image solutions using burst mode avoid this tradeoff by capturing more spatial-temporal information but often struggle with misalignment from camera/scene motion. To address these limitations, we propose a physical-model-based image restoration approach leveraging a novel dual-exposure Quad-Bayer pattern sensor. By capturing pairs of short and long exposures at the same starting point but with varying durations, this method integrates complementary noise-blur information within a single image. We further introduce a Quad-Bayer synthesis method (B2QB) to simulate sensor data from Bayer patterns to facilitate training. Based on this dual-exposure sensor model, we design a hierarchical convolutional neural network called QRNet to recover high-quality RGB images. The network incorporates input enhancement blocks and multi-level feature extraction to improve restoration quality. Experiments demonstrate superior performance over state-of-the-art deblurring and denoising methods on both synthetic and real-world datasets. The code, model, and datasets are publicly available at https://github.com/zhaoyuzhi/QRNet.

Figures

Figures reproduced from arXiv: 2412.07256 by the authors.

Figure 1
Figure 1. Illustration of (a) Normal Bayer pattern, (b-h) Other Bayer patterns [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of (a) Quad-Bayer pre-processing, (b) QRNet architecture (level 1 to 5 from top to bottom), and (c) input enhancement block. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of Pixel Unshuffle and Pixel Shuffle operations [16] for [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Illustration of the data capturing workflow for a moving image and a static image from an edited video. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Illustration of (a) Bayer to Quad-Bayer pattern sampling method (B2QB); (b) synthetic Quad-Bayer patterns, their corresponding RGB images, and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of restoration results of QRNet and other methods based on dual-exposure Quad-Bayer images. The noise parameter is [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: On one hand, since there is no ground truth for real [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 7
Figure 7. Figure 7: Illustration of 4 real input Quad-Bayer images and restoration results from different methods. For every full-resolution image, three patches are extracted [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Illustration of image restoration results of QRNet and other methods based on short-exposure images. The noise parameter is [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Illustration of image restoration results of QRNet and other baselines based on long-exposure images. The noise parameter is [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Illustration of image restoration results of QRNet and changing input Quad-Bayer data structure ablation study settings. The noise parameter is [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Illustration of image restoration results of QRNet and other ablation study settings. The noise parameter is [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

97 extracted references · 73 canonical work pages

  1. [4]

    A physics-based noise formation model for extreme low-light raw denoising,

    K. Wei, Y . Fu, J. Yang, and H. Huang, “A physics-based noise formation model for extreme low-light raw denoising,” in Proc. CVPR, 2020, pp. 2758–2767

    2020

  2. [1]

    Unprocessing images for learned raw denoising,

    T. Brooks, B. Mildenhall, T. Xue, J. Chen, D. Sharlet, and J. T. Barron, “Unprocessing images for learned raw denoising,” in Proc. CVPR, 2019, pp. 11 036–11 045

    2019

  3. [2]

    Cycleisp: Real image restoration via improved data synthesis,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Cycleisp: Real image restoration via improved data synthesis,” in Proc. CVPR, 2020, pp. 2696–2705

    2020

  4. [3]

    Noise flow: Noise modeling with conditional normalizing flows,

    A. Abdelhamed, M. A. Brubaker, and M. S. Brown, “Noise flow: Noise modeling with conditional normalizing flows,” in Proc. ICCV, 2019, pp. 3165–3173

    2019

  5. [5]

    Practical deep raw image denoising on mobile devices,

    Y . Wang, H. Huang, Q. Xu, J. Liu, Y . Liu, and J. Wang, “Practical deep raw image denoising on mobile devices,” in Proc. ECCV , 2020, pp. 1–16

    2020

  6. [6]

    Rethinking noise synthesis and modeling in raw denoising,

    Y . Zhang, H. Qin, X. Wang, and H. Li, “Rethinking noise synthesis and modeling in raw denoising,” in Proc. ICCV, 2021, pp. 4593–4601

    2021

  7. [7]

    Noise2noiseflow: Realistic camera noise modeling without clean im- ages,

    A. Maleky, S. Kousha, M. S. Brown, and M. A. Brubaker, “Noise2noiseflow: Realistic camera noise modeling without clean im- ages,” in Proc. CVPR, 2022, pp. 17 611–17 620

    2022

  8. [8]

    Deblurgan: Blind motion deblurring using conditional adversarial net- works,

    O. Kupyn, V . Budzan, M. Mykhailych, D. Mishkin, and J. Matas, “Deblurgan: Blind motion deblurring using conditional adversarial net- works,” in Proc. CVPR, 2018, pp. 8183–8192

    2018

Show all 97 references
  1. [9]

    Deblurgan-v2: Deblur- ring (orders-of-magnitude) faster and better,

    O. Kupyn, T. Martyniuk, J. Wu, and Z. Wang, “Deblurgan-v2: Deblur- ring (orders-of-magnitude) faster and better,” in Proc. ICCV, 2019, pp. 8878–8887

    2019

  2. [10]

    Rethinking coarse-to-fine approach in single image deblurring,

    S.-J. Cho, S.-W. Ji, J.-P. Hong, S.-W. Jung, and S.-J. Ko, “Rethinking coarse-to-fine approach in single image deblurring,” in Proc. ICCV , 2021, pp. 4641–4650

    2021

  3. [11]

    Lsd2 - joint denoising and deblurring of short and long exposure images with convolutional neural networks,

    J. Mustaniemi, J. Kannala, J. Matas, S. S ¨arkk¨a, and J. Heikkil ¨a, “Lsd2 - joint denoising and deblurring of short and long exposure images with convolutional neural networks,” in Proc. BMVC, 2020

    2020

  4. [12]

    Low-light image restoration with short-and long-exposure raw pairs,

    M. Chang, H. Feng, Z. Xu, and Q. Li, “Low-light image restoration with short-and long-exposure raw pairs,” IEEE Trans. Multimedia , vol. 24, pp. 702–714, 2021

    2021

  5. [13]

    D2hnet: Joint denoising and deblurring with hierarchical network for robust night image restoration,

    Y . Zhao, Y . Xu, Q. Yan, D. Yang, X. Wang, and L.-M. Po, “D2hnet: Joint denoising and deblurring with hierarchical network for robust night image restoration,” in Proc. ECCV, 2022, pp. 91–110

    2022

  6. [14]

    Under display camera quad bayer raw image restoration using deep learning,

    I. Kim, Y . Choi, H. Ko, D. Lim, Y . Seo, J. Lee, G. Lee, E. Heo, S. Song, and S. Lim, “Under display camera quad bayer raw image restoration using deep learning,” Electronic Imaging , vol. 2021, no. 7, pp. 67–1, 2021

    2021

  7. [15]

    Hdr video reconstruction with tri- exposure quad-bayer sensors,

    Y . Jiang, I. Choi, J. Jiang, and J. Gu, “Hdr video reconstruction with tri- exposure quad-bayer sensors,” arXiv preprint arXiv:2103.10982 , 2021

    2021 arXiv

  8. [16]

    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 Proc. CVPR, 2016, pp. 1874–1883

    2016

  9. [17]

    Color imaging array,

    B. E. Bayer, “Color imaging array,” U.S. Patent 3971065, Jul. 1976

    1976

  10. [18]

    Color filter arrays: Design and performance analysis,

    R. Lukac and K. N. Plataniotis, “Color filter arrays: Design and performance analysis,” IEEE Transactions on Consumer Electronics , vol. 51, no. 4, pp. 1260–1267, 2005

    2005

  11. [19]

    Solid state camera,

    S. Yamanaka, “Solid state camera,” U.S. Patent 4054906, Nov. 1977

    1977

  12. [20]

    A perceptually based design methodology for color filter arrays [image reconstruction],

    M. Parmar and S. J. Reeves, “A perceptually based design methodology for color filter arrays [image reconstruction],” in Proc. ICASSP, vol. 3, 2004, pp. iii–473

    2004

  13. [21]

    Bayesian-based iterative method of image restora- tion,

    W. H. Richardson, “Bayesian-based iterative method of image restora- tion,” JoSA, vol. 62, no. 1, pp. 55–59, 1972

    1972

  14. [22]

    Removing camera shake from a single photograph,

    R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis, and W. T. Freeman, “Removing camera shake from a single photograph,” in Proc. ACM SIGGRAPH, 2006, pp. 787–794

    2006

  15. [23]

    Edge-based blur kernel estimation using patch priors,

    L. Sun, S. Cho, J. Wang, and J. Hays, “Edge-based blur kernel estimation using patch priors,” in Proc. ICCP, 2013, pp. 1–8

    2013

  16. [24]

    Blind deconvolution using a normalized sparsity measure,

    D. Krishnan, T. Tay, and R. Fergus, “Blind deconvolution using a normalized sparsity measure,” in Proc. CVPR, 2011, pp. 233–240

    2011

  17. [25]

    Deblurring text images via l0-regularized intensity and gradient prior,

    J. Pan, Z. Hu, Z. Su, and M.-H. Yang, “Deblurring text images via l0-regularized intensity and gradient prior,” in Proc. CVPR , 2014, pp. 2901–2908

    2014

  18. [26]

    Learning a convolutional neural network for non-uniform motion blur removal,

    J. Sun, W. Cao, Z. Xu, and J. Ponce, “Learning a convolutional neural network for non-uniform motion blur removal,” in Proc. CVPR , 2015, pp. 769–777

    2015

  19. [27]

    From motion blur to motion flow: a deep learning solution for removing heterogeneous motion blur,

    D. Gong, J. Yang, L. Liu, Y . Zhang, I. Reid, C. Shen, A. Van Den Hengel, and Q. Shi, “From motion blur to motion flow: a deep learning solution for removing heterogeneous motion blur,” in Proc. CVPR, 2017, pp. 2319–2328

    2017

  20. [28]

    Deep multi-scale convolutional neural network for dynamic scene deblurring,

    S. Nah, T. Hyun Kim, and K. Mu Lee, “Deep multi-scale convolutional neural network for dynamic scene deblurring,” in Proc. CVPR , 2017, pp. 3883–3891

    2017

  21. [29]

    Scale-recurrent network for deep image deblurring,

    X. Tao, H. Gao, X. Shen, J. Wang, and J. Jia, “Scale-recurrent network for deep image deblurring,” in Proc. CVPR, 2018, pp. 8174–8182. IEEE TRANSACTIONS ON IMAGE PROCESSING 15

    2018

  22. [30]

    Deep stacked hierarchical multi-patch network for image deblurring,

    H. Zhang, Y . Dai, H. Li, and P. Koniusz, “Deep stacked hierarchical multi-patch network for image deblurring,” in Proc. CVPR , 2019, pp. 5978–5986

    2019

  23. [31]

    Multi-stage progressive image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Multi-stage progressive image restoration,” in Proc. CVPR , 2021, pp. 14 821–14 831

    2021

  24. [32]

    Dynamic scene deblurring using spatially variant recurrent neural networks,

    J. Zhang, J. Pan, J. Ren, Y . Song, L. Bao, R. W. Lau, and M.-H. Yang, “Dynamic scene deblurring using spatially variant recurrent neural networks,” in Proc. CVPR, 2018, pp. 2521–2529

    2018

  25. [33]

    Efficient dynamic scene deblurring using spatially variant deconvolution network with optical flow guided training,

    Y . Yuan, W. Su, and D. Ma, “Efficient dynamic scene deblurring using spatially variant deconvolution network with optical flow guided training,” in Proc. CVPR, 2020, pp. 3555–3564

    2020

  26. [34]

    Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training,

    D. Park, D. U. Kang, J. Kim, and S. Y . Chun, “Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training,” in Proc. ECCV , 2020, pp. 327– 343

    2020

  27. [35]

    Hierarchical integration diffusion model for realistic image deblurring,

    Z. Chen, Y . Zhang, D. Liu, B. Xia, J. Gu, L. Kong, and X. Yuan, “Hierarchical integration diffusion model for realistic image deblurring,” in Proc. NeurIPS, 2022, pp. 29 114–29 125

    2022

  28. [36]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in Proc. NeurIPS, 2014, pp. 2672–2680

    2014

  29. [37]

    Reblur2deblur: Deblurring videos via self-supervised learning,

    H. Chen, J. Gu, O. Gallo, M.-Y . Liu, A. Veeraraghavan, and J. Kautz, “Reblur2deblur: Deblurring videos via self-supervised learning,” in Proc. ICCP, 2018, pp. 1–9

    2018

  30. [38]

    Unsupervised class- specific deblurring,

    T. M. Nimisha, K. Sunil, and A. Rajagopalan, “Unsupervised class- specific deblurring,” in Proc. ECCV, 2018, pp. 353–369

    2018

  31. [39]

    Deblurring by realistic blurring,

    K. Zhang, W. Luo, Y . Zhong, L. Ma, B. Stenger, W. Liu, and H. Li, “Deblurring by realistic blurring,” in Proc. CVPR, 2020, pp. 2737–2746

    2020

  32. [40]

    Neural blind deconvolution using deep priors,

    D. Ren, K. Zhang, Q. Wang, Q. Hu, and W. Zuo, “Neural blind deconvolution using deep priors,” in Proc. CVPR, 2020, pp. 3341–3350

    2020

  33. [41]

    Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning,

    Z. Chi, Y . Wang, Y . Yu, and J. Tang, “Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning,” in Proc. CVPR, 2021, pp. 9137–9146

    2021

  34. [42]

    Scale-space and edge detection using anisotropic diffusion,

    P. Perona and J. Malik, “Scale-space and edge detection using anisotropic diffusion,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 12, no. 7, pp. 629–639, 1990

    1990

  35. [43]

    Nonlinear total variation based noise removal algorithms,

    L. I. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise removal algorithms,” Physica D: nonlinear phenomena , vol. 60, no. 1-4, pp. 259–268, 1992

    1992

  36. [44]

    Noise removal via bayesian wavelet coring,

    E. P. Simoncelli and E. H. Adelson, “Noise removal via bayesian wavelet coring,” in Proc. ICIP, vol. 1, 1996, pp. 379–382

    1996

  37. [45]

    A non-local algorithm for image denoising,

    A. Buades, B. Coll, and J.-M. Morel, “A non-local algorithm for image denoising,” in Proc. CVPR, vol. 2, 2005, pp. 60–65

    2005

  38. [46]

    New insights into the noise reduction wiener filter,

    J. Chen, J. Benesty, Y . Huang, and S. Doclo, “New insights into the noise reduction wiener filter,” IEEE/ACM Trans. Audio, Speech, Language Process., vol. 14, no. 4, pp. 1218–1234, 2006

    2006

  39. [47]

    K-svd: An algorithm for designing overcomplete dictionaries for sparse representation,

    M. Aharon, M. Elad, and A. Bruckstein, “K-svd: An algorithm for designing overcomplete dictionaries for sparse representation,” IEEE Trans. Signal Process. , vol. 54, no. 11, pp. 4311–4322, 2006

    2006

  40. [48]

    Image denoising via sparse and redundant representations over learned dictionaries,

    M. Elad and M. Aharon, “Image denoising via sparse and redundant representations over learned dictionaries,” IEEE Trans. Image Process. , vol. 15, no. 12, pp. 3736–3745, 2006

    2006

  41. [49]

    Non-local sparse models for image restoration,

    J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman, “Non-local sparse models for image restoration,” in Proc. ICCV , 2009, pp. 2272– 2279

    2009

  42. [50]

    Image denoising by sparse 3-d transform-domain collaborative filtering,

    K. Dabov, A. Foi, V . Katkovnik, and K. Egiazarian, “Image denoising by sparse 3-d transform-domain collaborative filtering,” IEEE Trans. Image Process., vol. 16, no. 8, pp. 2080–2095, 2007

    2007

  43. [51]

    Image denoising and inpainting with deep neural networks,

    J. Xie, L. Xu, and E. Chen, “Image denoising and inpainting with deep neural networks,” in Proc. NeurIPS, 2012, pp. 341–349

    2012

  44. [52]

    Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connec- tions,

    X. Mao, C. Shen, and Y .-B. Yang, “Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connec- tions,” in Proc. NeurIPS, 2016, pp. 2802–2810

    2016

  45. [53]

    Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration,

    Y . Chen and T. Pock, “Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 39, no. 6, pp. 1256–1272, 2016

    2016

  46. [54]

    Learning deep cnn denoiser prior for image restoration,

    K. Zhang, W. Zuo, S. Gu, and L. Zhang, “Learning deep cnn denoiser prior for image restoration,” in Proc. CVPR, 2017, pp. 3929–3938

    2017

  47. [55]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,” IEEE Trans. Image Process. , vol. 26, no. 7, pp. 3142–3155, 2017

    2017

  48. [56]

    Memnet: A persistent memory network for image restoration,

    Y . Tai, J. Yang, X. Liu, and C. Xu, “Memnet: A persistent memory network for image restoration,” in Proc. ICCV, 2017, pp. 4539–4547

    2017

  49. [57]

    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 Trans. Image Process. , vol. 27, no. 9, pp. 4608–4622, 2018

    2018

  50. [58]

    Multi-level wavelet- cnn for image restoration,

    P. Liu, H. Zhang, K. Zhang, L. Lin, and W. Zuo, “Multi-level wavelet- cnn for image restoration,” in Proc. CVPRW, 2018, pp. 773–782

    2018

  51. [59]

    Real image denoising with feature attention,

    S. Anwar and N. Barnes, “Real image denoising with feature attention,” in Proc. ICCV, 2019, pp. 3155–3164

    2019

  52. [60]

    Attention-guided cnn for image denoising,

    C. Tian, Y . Xu, Z. Li, W. Zuo, L. Fei, and H. Liu, “Attention-guided cnn for image denoising,” Neural Networks , vol. 124, pp. 117–129, 2020

    2020

  53. [61]

    Densely self-guided wavelet network for image denoising,

    W. Liu, Q. Yan, and Y . Zhao, “Densely self-guided wavelet network for image denoising,” in Proc. CVPRW, 2020, pp. 432–433

    2020

  54. [62]

    Learning enriched features for real image restoration and enhancement,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Learning enriched features for real image restoration and enhancement,” in Proc. ECCV, 2020, pp. 492–511

    2020

  55. [63]

    Nbnet: Noise basis learning for image denoising with subspace projection,

    S. Cheng, Y . Wang, H. Huang, D. Liu, H. Fan, and S. Liu, “Nbnet: Noise basis learning for image denoising with subspace projection,” in Proc. CVPR, 2021, pp. 4896–4906

    2021

  56. [64]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Proc. MICCAI, 2015, pp. 234– 241

    2015

  57. [65]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. CVPR, 2016, pp. 770–778

    2016

  58. [66]

    Toward robust image denoising via flow-based joint image and noise model,

    L. Guo, S. Huang, H. Liu, and B. Wen, “Toward robust image denoising via flow-based joint image and noise model,” IEEE Trans. Circuits Syst. Video Technol., vol. 34, pp. 6105–6115, 2024

    2024

  59. [67]

    Denoising diffusion models for plug-and-play image restoration,

    Y . Zhu, K. Zhang, J. Liang, J. Cao, B. Wen, R. Timofte, and L. V . Gool, “Denoising diffusion models for plug-and-play image restoration,” in Proc. CVPRW, 2023, pp. 1219–1229

    2023

  60. [68]

    Toward convolutional blind denoising of real photographs,

    S. Guo, Z. Yan, K. Zhang, W. Zuo, and L. Zhang, “Toward convolutional blind denoising of real photographs,” in Proc. CVPR, 2019, pp. 1712– 1722

    2019

  61. [69]

    Variational denoising network: Toward blind noise modeling and removal,

    Z. Yue, H. Yong, Q. Zhao, D. Meng, and L. Zhang, “Variational denoising network: Toward blind noise modeling and removal,” Proc. NeurIPS, vol. 32, pp. 1690–1701, 2019

    2019

  62. [70]

    Transfer learning from synthetic to real-noise denoising with adaptive instance normalization,

    Y . Kim, J. W. Soh, G. Y . Park, and N. I. Cho, “Transfer learning from synthetic to real-noise denoising with adaptive instance normalization,” in Proc. CVPR, 2020, pp. 3482–3492

    2020

  63. [71]

    Noise2noise: Learning image restoration without clean data,

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, and T. Aila, “Noise2noise: Learning image restoration without clean data,” in Proc. ICML, 2018, pp. 2965–2974

    2018

  64. [72]

    Noise2self: Blind denoising by self- supervision,

    J. Batson and L. Royer, “Noise2self: Blind denoising by self- supervision,” in Proc. ICML, 2019, pp. 524–533

    2019

  65. [73]

    Noise2void-learning denoising from single noisy images,

    A. Krull, T.-O. Buchholz, and F. Jug, “Noise2void-learning denoising from single noisy images,” in Proc. CVPR, 2019, pp. 2129–2137

    2019

  66. [74]

    Unpaired learning of deep image denoising,

    X. Wu, M. Liu, Y . Cao, D. Ren, and W. Zuo, “Unpaired learning of deep image denoising,” in Proc. ECCV, 2020, pp. 352–368

    2020

  67. [75]

    Learning invariant representation for unsupervised image restoration,

    W. Du, H. Chen, and H. Yang, “Learning invariant representation for unsupervised image restoration,” in Proc. CVPR , 2020, pp. 14 483– 14 492

    2020

  68. [76]

    Legacy photo editing with learned noise prior,

    Y . Zhao, L.-M. Po, T. Lin, X. Wang, K. Liu, Y . Zhang, W.-Y . Yu, P. Xian, and J. Xiong, “Legacy photo editing with learned noise prior,” in Proc. WACV, 2021, pp. 2103–2112

    2021

  69. [77]

    Zero-shot noise2noise: Efficient image denoising without any data,

    Y . Mansour and R. Heckel, “Zero-shot noise2noise: Efficient image denoising without any data,” in Proc. CVPR, 2023, pp. 14 018–14 027

    2023

  70. [78]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. ICLR, 2014

    2014

  71. [79]

    Image deblurring with blurred/noisy image pairs,

    L. Yuan, J. Sun, L. Quan, and H.-Y . Shum, “Image deblurring with blurred/noisy image pairs,” in Proc. ACM SIGGRAPH , 2007, pp. 1–es

    2007

  72. [80]

    Non-uniform deblurring for shaken images,

    O. Whyte, J. Sivic, A. Zisserman, and J. Ponce, “Non-uniform deblurring for shaken images,” Int. J. Comput. Vis. , vol. 98, no. 2, pp. 168–186, 2012

    2012

  73. [81]

    Rgb-nir imaging with exposure bracketing for joint denoising and deblurring of low-light color images,

    H. Yamashita, D. Sugimura, and T. Hamamoto, “Rgb-nir imaging with exposure bracketing for joint denoising and deblurring of low-light color images,” in Proc. ICASSP, 2017, pp. 6055–6059

    2017

  74. [82]

    Deformable convolutional networks,

    J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, and Y . Wei, “Deformable convolutional networks,” in Proc. ICCV, 2017, pp. 764–773

    2017

  75. [83]

    Spatial temporal video enhancement using alternating exposures,

    W. Shen, M. Cheng, G. Lu, G. Zhai, L. Chen, M. S. Asif, and Z. Gao, “Spatial temporal video enhancement using alternating exposures,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 8, pp. 4912–4926, 2021

    2021

  76. [84]

    Learning to see in the dark,

    C. Chen, Q. Chen, J. Xu, and V . Koltun, “Learning to see in the dark,” in Proc. CVPR, 2018, pp. 3291–3300

    2018

  77. [85]

    Raw image deblur- ring,

    C.-H. Liang, Y .-A. Chen, Y .-C. Liu, and W. Hsu, “Raw image deblur- ring,” IEEE Trans. Multimedia , vol. 24, pp. 61–72, 2022

    2022

  78. [86]

    Deep joint demo- saicking and denoising,

    M. Gharbi, G. Chaurasia, S. Paris, and F. Durand, “Deep joint demo- saicking and denoising,” ACM Trans. on Graphics , vol. 35, no. 6, pp. 1–12, 2016

    2016

  79. [87]

    Saliency map- aided generative adversarial network for raw to rgb mapping,

    Y . Zhao, L.-M. Po, T. Zhang, Z. Liao, X. Shi et al. , “Saliency map- aided generative adversarial network for raw to rgb mapping,” in Proc. ICCVW, 2019, pp. 3449–3457. IEEE TRANSACTIONS ON IMAGE PROCESSING 16

    2019

  80. [88]

    Joint demosaicing and denoising with self guidance,

    L. Liu, X. Jia, J. Liu, and Q. Tian, “Joint demosaicing and denoising with self guidance,” in Proc. CVPR, 2020, pp. 2240–2249

    2020

  81. [89]

    End-to-end learning for joint image demosaicing, denoising and super-resolution,

    W. Xing and K. Egiazarian, “End-to-end learning for joint image demosaicing, denoising and super-resolution,” in Proc. CVPR, 2021, pp. 3507–3516

    2021

  82. [90]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proc. CVPR, 2018, pp. 7132–7141

    2018

  83. [91]

    Self-guided network for fast image denoising,

    S. Gu, Y . Li, L. V . Gool, and R. Timofte, “Self-guided network for fast image denoising,” in Proc. ICCV, 2019, pp. 2511–2520

    2019

  84. [92]

    Demosaicing with directional filtering and a posteriori decision,

    D. Menon, S. Andriani, and G. Calvagno, “Demosaicing with directional filtering and a posteriori decision,” IEEE Trans. Image Process., vol. 16, no. 1, pp. 132–141, 2006

    2006

  85. [93]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

    W. Zhang, G. Zhai, Y . Wei, X. Yang, and K. Ma, “Blind image quality assessment via vision-language correspondence: A multitask learning perspective,” in Proc. CVPR, 2023, pp. 14 071–14 081

    2023

  86. [94]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Trans. Image Process. , vol. 13, no. 4, pp. 600–612, 2004

    2004

  87. [95]

    Pytorch-opcounter,

    “Pytorch-opcounter,” https://github.com/Lyken17/pytorch-OpCounter, accessed: Sep. 30, 2021. [Online]

    2021

  88. [96]

    Hinet: Half instance normalization network for image restoration,

    L. Chen, X. Lu, J. Zhang, X. Chu, and C. Chen, “Hinet: Half instance normalization network for image restoration,” in Proc. CVPRW , 2021, pp. 182–192. Yuzhi Zhao (S’19) received the B.Eng. degree in Electronic and Information Engineering from Huazhong University of Science and...

    2021

  89. [2013]

    Xin Ye received the B.S

    He also served on the Organizing Committee, of the IEEE International Conference on Acoustics, Speech and Signal Processing in 2003, and the IEEE International Conference on Image Processing in 2010. Xin Ye received the B.S. degree in physics from Zhejiang University in 2014 a...

    2003

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.