pith. sign in

arxiv: 1907.10128 · v1 · pith:VUTPJQKGnew · submitted 2019-07-23 · 📡 eess.IV · cs.CV

Blind Deblurring using Deep Learning: A Survey

Siddhant Sahu , Manoj Kumar Lenka , Pankaj Kumar Sa This is my paper

Pith reviewed 2026-05-24 16:48 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords blind deblurringdeep learningimage restorationend-to-end networksblur kernel estimationGOPRO datasetKohler datasetPSNR SSIM
0
0 comments X

The pith

Recent deep learning methods for blind image deblurring estimate the sharp image directly rather than first recovering the blur kernel.

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

This survey examines deep learning approaches to blind deblurring, where the blur is unknown. Early methods used networks to estimate features of the blur kernel, later ones predicted the entire kernel before applying non-blind deblurring. The latest techniques use end-to-end networks to map a blurred image straight to a sharp version without intermediate kernel steps. The paper also reports PSNR and SSIM scores for various models on the GOPRO and Kohler datasets. A reader would care because this tracks how the field moved toward simpler, more direct solutions.

Core claim

The paper establishes that deep learning solutions for blind deblurring have progressed from partial kernel feature estimation to complete kernel prediction and finally to end-to-end direct estimation of the latent sharp image from the blurred input, with the most recent state-of-the-art methods following the end-to-end approach.

What carries the argument

End-to-end convolutional neural networks that directly predict the latent sharp image from the blurred input.

If this is right

  • These methods eliminate the need for separate kernel estimation and non-blind deblurring stages.
  • Performance is measured by PSNR and SSIM values on the GOPRO and Kohler datasets.
  • The field can focus on improving network designs for better direct mapping from blurred to sharp images.
  • Surveys help identify the shift away from explicit kernel modeling in recent architectures.

Where Pith is reading between the lines

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

  • This direct estimation approach may extend to other image restoration tasks where explicit modeling of degradation proves less effective.
  • Testing these end-to-end models on additional real-world blurry images could reveal limits not captured by the standard benchmarks.
  • Kernel estimation might retain value for specific controlled blur types even if end-to-end methods dominate general cases.

Load-bearing premise

The deep learning solutions reviewed form a complete and representative sample of all work in the area through 2019.

What would settle it

Discovery of a major deep learning paper on blind deblurring published before 2019 that is not discussed in this survey.

Figures

Figures reproduced from arXiv: 1907.10128 by Manoj Kumar Lenka, Pankaj Kumar Sa, Siddhant Sahu.

Figure 2
Figure 2. Figure 2: Architecture used by Chakrabati [1] for prediction of Fourier coefficients for the deconvolution filter. Here H is high pass, B2, B1 are band pass, while L is low pass frequency band. The letters in bold are Fourier transforms of the corresponding symbols. patch Bp = {B[n] : n ∈ p} they considered the surrounding pixels of the patch while finding its Fourier coefficients for better results, let the blurry … view at source ↗
Figure 4
Figure 4. Figure 4: This shows given a pixel p (in yellow) how MRF smoothen its value based on N(p) i.e its neighboring pixels After estimating the motion kernels for all the patches they are concatenated and a Markov Random Function (MRF) is used to merge them all together, smoothen the transition of motion kernels amongst neighboring pixels (Fig.4) and generates a dense motion field by minimizing energy function, X p∈Ω [−C(… view at source ↗
Figure 3
Figure 3. Figure 3: Network architecture for predicting the motion kern [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture used by Gong et. al. [2] to predict the mo [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scale Recurrent Network Architecture used by ( [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Architecture used by Noorozi et. al. [5]. Here the three CNNs starting from the left denotes N1, N2, N3 respectively. Firstly the blurred image is given as input to the first network N1 (pure convolution) without any downsampling and its output is added with the downsampled version of the same blurred image by a factor of four. After this the first loss L1 is calculated using 21 by calculating the differen… view at source ↗
Figure 10
Figure 10. Figure 10: Multiscale architecture used by Nah. et. al [4] for that scale, then for each scale MSE(Mean Squared Error) with the sharp image is calculated and back propagation is done. The MSE for all the scales are averaged together to give the content loss as follows: Lcontent = 1 2K X K k=1 1 ckhkwk kLk − Ikk 2 (23) Here K is the total number of scales, ck, hk, wk are the channels, height and width of the k th sca… view at source ↗
Figure 9
Figure 9. Figure 9: The basic structure of a GAN, where G denotes the Generator and D denotes the discriminator. 2) With Adversarial Loss: Blind Deblurring can also be solved end-to-end by generative models like Generative Ad￾versarial Networks [14] [15] [16]. The approach Generative Adversarial Networks take is to have two different agents play a game against each other. One of the agents is a generator network which tries t… view at source ↗
read the original abstract

We inspect all the deep learning based solutions and provide holistic understanding of various architectures that have evolved over the past few years to solve blind deblurring. The introductory work used deep learning to estimate some features of the blur kernel and then moved onto predicting the blur kernel entirely, which converts the problem into non-blind deblurring. The recent state of the art techniques are end to end, i.e., they don't estimate the blur kernel rather try to estimate the latent sharp image directly from the blurred image. The benchmarking PSNR and SSIM values on standard datasets of GOPRO and Kohler using various architectures are also provided.

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

1 major / 1 minor

Summary. The manuscript is a survey of deep learning methods for blind image deblurring. It describes an evolutionary progression from early DL approaches that estimate blur-kernel features, through methods that predict the full blur kernel (reducing the problem to non-blind deblurring), to recent end-to-end architectures that directly regress the latent sharp image from the blurred input. Benchmark PSNR and SSIM values on the GOPRO and Kohler datasets are also compiled for the reviewed architectures.

Significance. If the reviewed set of papers is representative, the survey would usefully document the shift toward end-to-end DL solutions and aggregate quantitative results on two standard benchmarks. No machine-checked proofs or parameter-free derivations are present; credit is given for the explicit compilation of PSNR/SSIM numbers on GOPRO and Kohler.

major comments (1)
  1. [Abstract] Abstract: the assertion that the survey inspects 'all' deep learning solutions for blind deblurring lacks any description of literature-search criteria, date range, inclusion thresholds, or discussion of omitted high-impact works (2017-2019). This is load-bearing for the central claim that recent SOTA techniques are exclusively end-to-end rather than kernel-estimating, because an incomplete sample could alter the reported evolutionary narrative and the interpretation of the GOPRO/Kohler benchmarking table.
minor comments (1)
  1. The benchmarking table (mentioned in the abstract) should explicitly list the architectures compared, the exact PSNR/SSIM values, and any notes on training protocols to allow direct verification.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the abstract requires clarification on literature coverage and will revise the manuscript to address this.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that the survey inspects 'all' deep learning solutions for blind deblurring lacks any description of literature-search criteria, date range, inclusion thresholds, or discussion of omitted high-impact works (2017-2019). This is load-bearing for the central claim that recent SOTA techniques are exclusively end-to-end rather than kernel-estimating, because an incomplete sample could alter the reported evolutionary narrative and the interpretation of the GOPRO/Kohler benchmarking table.

    Authors: We agree that the claim to inspect 'all' solutions would be strengthened by explicit documentation of the search process. In the revised version we will insert a new subsection (likely in Section 1 or as an appendix) that states: searches were performed on Google Scholar and arXiv with keywords including 'blind image deblurring deep learning', 'CNN deblurring', 'end-to-end deblurring'; the temporal scope was 2015 through mid-2019; inclusion required a DL architecture for blind deblurring together with quantitative results on at least one standard benchmark. We will also add a short paragraph noting any high-impact 2017-2019 works that were omitted and the reasons (e.g., concurrent submission, focus on non-blind cases, or lack of benchmark numbers). The abstract will be rephrased to 'a broad set of' or 'the principal' deep-learning solutions, with a forward reference to the new methodology subsection. These changes will support rather than undermine the evolutionary narrative and the interpretation of the GOPRO/Kohler table. revision: yes

Circularity Check

0 steps flagged

Survey paper contains no derivations, predictions or self-referential steps

full rationale

This work is a literature review that summarizes external papers on blind deblurring. It states that recent SOTA methods are end-to-end but does so by citing outside literature rather than deriving or fitting anything internally. No equations, fitted inputs, uniqueness theorems, or ansatzes appear. The completeness of the sample is an external-validity issue, not a circularity issue within any derivation chain. Score 0 is the appropriate default for a self-contained review of external results.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

As a survey paper, there are no free parameters, axioms, or invented entities introduced by the authors.

pith-pipeline@v0.9.0 · 5630 in / 930 out tokens · 17733 ms · 2026-05-24T16:48:24.781447+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages · 4 internal anchors

  1. [1]

    A Neural Approach to Blind Motion Deblurring

    A. Chakrabarti, “A neural approach to blind motion deblurring,” CoRR, vol. abs/1603.04771, 2016. [Online]. Available: http://arxiv.org/abs/1603.04771

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  2. [2]

    From motion blur to motion flow: A deep learning so lution for removing heterogeneous motion blur,

    D. Gong, J. Y ang, L. Liu, Y . Zhang, I. Reid, C. Shen, A. van d en Hengel, and Q. Shi, “From motion blur to motion flow: A deep learning so lution for removing heterogeneous motion blur,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    work page 2017

  3. [3]

    Deblurgan: Blind motion deblurring using conditional adv ersarial net- works,

    O. Kupyn, V . Budzan, M. Mykhailych, D. Mishkin, and J. Mat as, “Deblurgan: Blind motion deblurring using conditional adv ersarial net- works,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    work page 2018

  4. [4]

    Deep multi-scale convo lutional neural network for dynamic scene deblurring,

    S. Nah, T. Hyun Kim, and K. Mu Lee, “Deep multi-scale convo lutional neural network for dynamic scene deblurring,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    work page 2017

  5. [5]

    Motion Deblurring in the Wild

    M. Noroozi, P . Chandramouli, and P . Favaro, “Motion debl urring in the wild,” CoRR, vol. abs/1701.01486, 2017. [Online]. Available: http://arxiv.org/abs/1701.01486

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  6. [6]

    Deep Generative Filter for Motion Deblurring

    S. Ramakrishnan, S. Pachori, A. Gangopadhyay, and S. Ram an, “Deep generative filter for motion deblurring,” CoRR, vol. abs/1709.03481,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Available: http://arxiv.org/abs/1709.0 3481

    [Online]. Available: http://arxiv.org/abs/1709.0 3481

    work page

  8. [8]

    Learning to Deblur

    C. J. Schuler, M. Hirsch, S. Harmeling, and B. Sch¨ olkopf , “Learning to deblur,” CoRR, vol. abs/1406.7444, 2014. [Online]. Available: http://arxiv.org/abs/1406.7444

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  9. [9]

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

    J. Sun, W. Cao, Z. Xu, and J. Ponce, “Learning a convolutio nal neural network for non-uniform motion blur removal,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2015

    work page 2015

  10. [10]

    Scale-recurr ent network for deep image deblurring,

    X. Tao, H. Gao, X. Shen, J. Wang, and J. Jia, “Scale-recurr ent network for deep image deblurring,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    work page 2018

  11. [11]

    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. Y ang, “Dynamic scene deblurring using spatially variant recurrent neural networks,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    work page 2018

  12. [12]

    Deconvolution and checkerboard artifacts,

    A. Odena, V . Dumoulin, and C. Olah, “Deconvolution and checkerboard artifacts,” Distill, 2016. [Online]. Available: http://distill.pub/2016/deconv-checkerboard

    work page 2016

  13. [13]

    Discrim- inative non-blind deblurring,

    U. Schmidt, C. Rother, S. Nowozin, J. Jancsary, and S. Ro th, “Discrim- inative non-blind deblurring,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2013

    work page 2013

  14. [14]

    A machine learning approach for non-blind image deconvoluti on,

    C. J. Schuler, H. Christopher Burger, S. Harmeling, and B. Scholkopf, “A machine learning approach for non-blind image deconvoluti on,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2013

    work page 2013

  15. [15]

    Generative adversar ial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. War de-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversar ial nets,” in Advances in Neural Information Processing Systems 27 , Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, E ds. Curran Associates, Inc., 2014, pp. 2672–2680. [Online]. Av ailable: http://pa...

    work page 2014

  16. [16]

    Image-to-image translation with conditional adversarial networks,

    P . Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    work page 2017

  17. [17]

    Wasserstein g enerative adversarial networks,

    M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein g enerative adversarial networks,” in Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW , Australia, 6-11 August 2017 , 2017, pp. 214–223. [Online]. Available: http://proceedings.mlr.press/v70/arjovsky17a.html

    work page 2017

  18. [18]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in The IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) , July 2017

    work page 2017

  19. [19]

    Perceptual losse s for real- time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losse s for real- time style transfer and super-resolution,” in European Conference on Computer Vision, 2016

    work page 2016

  20. [20]

    Recording and playb ack of camera shake: Benchmarking blind deconvolution with a real -world database,

    M. B. S. B. H. S. Khler R., Hirsch M., “Recording and playb ack of camera shake: Benchmarking blind deconvolution with a real -world database,” in European Conference on Computer Vision (ECCV) , 2012

    work page 2012