CoChat
Sign Up
Cited by 2,116Image and Signal Denoising MethodsAdvanced Image Processing TechniquesAdvanced Image Fusion Techniques

Learning Deep CNN Denoiser Prior for Image Restoration

Open access
Zhang, Kai, Zuo, Wangmeng, Gu, Shuhang, Zhang, Lei·2017·2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
DOI 10.1109/cvpr.2017.300
Open this paper in CoChat — free accountOpen PDF

Abstract

Model-based optimization methods and discriminative learning methods have been the two dominant strategies for solving various inverse problems in low-level vision. Typically, those two kinds of methods have their respective merits and drawbacks, e.g., model-based optimization methods are flexible for handling different inverse problems but are usually time-consuming with sophisticated priors for the purpose of good performance, in the meanwhile, discriminative learning methods have fast testing speed but their application range is greatly restricted by the specialized task. Recent works have revealed that, with the aid of variable splitting techniques, denoiser prior can be plugged in as a modular part of model-based optimization methods to solve other inverse problems (e.g., deblurring). Such an integration induces considerable advantage when the denoiser is obtained via discriminative learning. However, the study of integration with fast discriminative denoiser prior is still lacking. To this end, this paper aims to train a set of fast and effective CNN (convolutional neural network) denoisers and integrate them into model-based optimization method to solve other inverse problems. Experimental results demonstrate that the learned set of denoisers can not only achieve promising Gaussian denoising results but also can be used as prior to deliver good performance for various low-level vision applications.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper trains a set of fast CNN Gaussian denoisers and uses them as modular priors inside a half-quadratic splitting optimization framework to solve image deblurring and super-resolution. Experiments show competitive denoising accuracy and flexible restoration performance across tasks without task-specific retraining.

Key findings

  • The proposed CNN denoiser improves gray-image denoising on BSD68 over BM3D, WNNM, TNRD, and MLP: at σ=15 it reaches 31.63 dB, at σ=25 it reaches 29.15 dB, and at σ=50 it reaches 26.19 dB.
  • For color denoising on BSD68, the proposed denoiser consistently beats CBM3D, e.g. 40.36 vs. 40.24 dB at noise level 5, 31.16 vs. 30.71 dB at 25, and 27.86 vs. 27.38 dB at 50.
  • The learned denoisers can be inserted into a model-based optimization method to handle deblurring and single-image super-resolution; for example, on the Leaves deblurring case with Gaussian blur σ=1.6 and noise level 2, the proposed method reports 29.78 dB versus 26.95 dB for IDDBM3D and 27.50 dB for NCSR.
  • Runtime is competitive: for a 256×256 image with noise level 25, the 24-feature-map version takes 0.10 s on CPU and 0.006 s on GPU, while BM3D takes 0.66 s on CPU and TNRD takes 0.47 s on CPU / 0.010 s on GPU.

Methods

  • The authors design a seven-layer CNN denoiser using dilated convolutions with dilation factors 1, 2, 3, 4, 3, 2, 1; middle layers use 64 feature maps, and the effective receptive field is 33×33.
  • Training uses 400 BSD images, 400 ImageNet validation images, and 4,744 Waterloo Exploration Database images, cropped into 35×35 patches with N = 256×4,000 patches; Adam is used with residual learning and batch normalization.
  • A set of 25 denoisers is trained for gray images and 25 for color images over noise levels from 0 to 50 in steps of 2; these denoisers are integrated into a half-quadratic splitting scheme, where the regularization subproblem becomes a Gaussian denoising step.

Limitations

  • The paper states that future work includes reducing the number of CNN denoisers and iterations, extending the framework to inpainting and blind deblurring, and using multiple complementary priors.
  • The experimental evaluation is limited to denoising, non-blind deblurring, and super-resolution; inpainting and blind deblurring are not tested.
  • The authors note that comparisons with SRCNN and VDSR for non-bicubic super-resolution are unfair because those methods were trained for bicubic degradation, which is evident from the reported setup rather than presented as a limitation of the proposed method.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/cvpr.2017.300

Cites

75 of 76 references verified0 citation errors · 1 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

2 of 26 cited sources checked · 1 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    SRCNN and VDSR are trained with bicubic blur kernel.

    “Note that SRCNN and VDSR are trained with bicubic blur kernel, thus it is unfair to use their models to super-resolve the low-resolution image with Gaussian kernel.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [21] 21. C. Dong, C. C. Loy, K. He, and X. Tang. Image super-resolution using deep convolutional networks. IEEE transactions on Pattern Analysis and Machine Intelligence, 38(2):295-307, 2016. 2, 7 source

      Section 3.3 (p. 5) says SRCNN training samples are blurred with a Gaussian kernel; bicubic interpolation is a later upscaling step, not the blur kernel. Section 4.4.1 (p. 9) also describes training a separate SRCNN network with a different Gaussian kernel for comparison. The paper does not address VDSR's training. (LLM, conf=0.87)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [35] 35. J. Kim, J. K. Lee, and K. M. Lee. Accurate image super- resolution using very deep convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1646-1654, 2016. 4, 7 source
  • Findingown data

    The learned set of CNN denoisers can be used as prior to deliver good performance for various low-level vision applications.

    “Experimental results demonstrate that the learned set of denoisers can not only achieve promising Gaussian denoising results but also can be used as prior to deliver good performance for various low-level vision applications.”
    • Also stated in: § 1. Introduction · boosted
      “Extensive experiments on classical IR problems, including deblurring and super-resolution, have demonstrated the merits of integrating flexible model-based optimization methods and fast CNN-based discriminative learning methods.”
    Quote in the paperSubstantial claim
  • Findingown data

    Batch normalization and residual learning enable fast transfer from one model to another with different noise level.

    “we empirically find it also can enable fast transfer from one model to another with different noise level.”
    Quote in the paperNot a substantial claim
  • Findingnone given

    If the size of the training patch is smaller than the receptive field, denoising performance decreases.

    “However, if the size of the training patch is smaller than the receptive field, the performance would decrease.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The proposed CNN denoiser has a PSNR gain of about 0.2 dB over WNNM, MLP and TNRD on gray image denoising on BSD68.

    “However, the proposed CNN denoiser can have a PSNR gain of about 0.2dB over those three methods.”
    Quote in the paperSubstantial claim
  • May combine two claims. 13
  • Classified as probably not a substantial claim. 10
  • The quote only partly states this claim. 1

  • May combine two claims.

    Premisecitation

    The combination of batch normalization and residual learning is helpful for Gaussian denoising, enabling fast and stable training and better denoising performance.

    “it has been pointed out that the combination of batch normalization and residual learning is particularly helpful for Gaussian denoising since they are beneficial to each other.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [65] 65. K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang. Be- yond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Process- ing, 2017. 4 source

  • May combine two claims.

    Findingown data

    The proposed denoiser with 24 feature maps per middle layer has comparable PSNR of 28.94 dB to TNRD but delivers faster speed.

    “It is worth emphasizing that the proposed denoiser with 24 feature maps of each layer has a comparable PSNR of 28.94dB to TNRD but delivers a faster speed.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Using a large training dataset does not improve PSNR results on BSD68 but can slightly improve performance on other testing images.

    “We empirically find using large dataset does not improve the PSNR results of BSD68 dataset [50] but can slightly improve the performance of other testing images.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [50] 50. S. Roth and M. J. Black. Fields of experts. International Journal of Computer Vision, 82(2):205-229, 2009. 1, 4, 5 source

  • May combine two claims.

    Findingown data

    The proposed CNN denoiser with dilated convolutions achieves average PSNR 29.15 dB on BSD68 at noise level 25, better than 28.94 dB of a 7-layer network with traditional 3×3 convolution and close to 29.20 dB of a 16-layer network.

    “It turns out that our designed model can have an average P-SNR of 29.15dB on BSD68 dataset [50], which is much better than 28.94dB of 7 layers network with traditional 3×3 convolution filter and very close to 29.20dB of 16 layers network.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [50] 50. S. Roth and M. J. Black. Fields of experts. International Journal of Computer Vision, 82(2):205-229, 2009. 1, 4, 5 source

  • May combine two claims.

    Findingown data

    Our method can super-resolve the Butterfly image in 0.5 second on GPU and 12 seconds on CPU, whereas NCSR spends 198 seconds on CPU.

    “As an example for the testing speed comparison, our method can super-resolve the Butterfly image in 0.5 second on GPU and 12 seconds on CPU, whereas NCSR spends 198 seconds on CPU.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The low-resolution image can be modeled by a blurring and subsequent down-sampling operation on a high-resolution one.

  • May combine two claims.

    Findingown data

    IDDBM3D, NCSR and MLP tend to smooth edges and generate color artifacts in image deblurring.

    “We can see that IDDBM3D, NCSR and MLP tend to smooth the edges and generate color artifacts.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The convolution is carried out with circular boundary conditions.

  • May combine two claims.

    Findingown data

    SRCNN and VDSR achieve promising results for bicubic kernel but their performance deteriorates seriously when the low-resolution image is not generated by bicubic kernel.

    “First, although S-RCNN and VDSR achieve promising results to tackle the case with bicubic kernel, their performance deteriorates seriously when the low-resolution image are not generated by bicubic kernel (see Figure 4).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The low-resolution image can be modeled by a blurring and subsequent down-sampling operation on a high-resolution one.

  • May combine two claims.

    Findingown data

    The proposed method can recover image sharpness and naturalness in image deblurring.

    “In contrast, the proposed method can recover image sharpness and naturalness.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The convolution is carried out with circular boundary conditions.

  • May combine two claims.

    Findingown data

    Both the gray and color CNN denoiser prior based optimization methods can produce promising results.

    “Third, both of the gray and color CNN denoiser prior based optimization methods can produce promising results.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The low-resolution image can be modeled by a blurring and subsequent down-sampling operation on a high-resolution one.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    CNN exploits external prior complementary to the internal prior of many existing denoisers such as BM3D.

    “Third, CNN exploits the external prior which is complementary to the internal prior of many existing denoisers such as BM3D.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Model-based optimization methods are flexible for handling different inverse problems but are usually time-consuming with sophisticated priors for good performance.

    “model-based optimization methods are flexible for handling different inverse problems but are usually time-consuming with sophisticated priors for the purpose of good performance”
    • Also stated in: § 1. Introduction · plain
      “model-based optimization methods are usually timeconsuming with sophisticated priors for the purpose of good performance [27].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [27] 27. Q. Gao and S. Roth. How well do filter-based MRFs model natural images? In Joint DAGM (German Association for Pattern Recognition) and OAGM Symposium, pages 62-72, 2012. 2 source

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    CNN exhibits powerful prior modeling capacity with deep architecture.

    “Second, CNN exhibits powerful prior modeling capacity with deep architecture.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    CNN inference is very efficient due to the parallel computation ability of GPU.

    “First, the inference of CNN is very efficient due to the parallel computation ability of GPU.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The number of denoisers (25) is much less than learning different models for different degradations.

    “Especially noteworthy is the number of the denoisers which is much less than that of learning different models for different degradations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Discriminative learning methods have fast testing speed but their application range is greatly restricted by the specialized task.

    “discriminative learning methods have fast testing speed but their application range is greatly restricted by the specialized task.”
    • Also stated in: § 1. Introduction · plain
      “discriminative learning methods are usually restricted by specialized tasks.”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    The good compromise between speed and performance over TNRD is attributed to three reasons: 3×3 convolution and ReLU are simple yet effective and efficient; the architecture encourages fluent information flow among layers, giving larger model capacity; batch normalization is adopted.

    “Such a good compromise between speed and performance over TNRD is properly attributed to the following three reasons.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    It is preferable to jointly handle RGB channels.

    “and it is preferable to jointly handle RGB channels.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [42] 42. T. Miyata. Inter-channel relation based vectorial total varia- tion for color image recovery. In IEEE International Confer- ence on Image Processing,, pages 2251-2255, 2015. 3

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    To compare with TNRD under similar PSNR performance, the run time of the proposed denoiser with 24 feature maps per middle layer was additionally provided.

    “To make a further comparison with TNRD under similar PSNR performance, we additionally provide the run time of the proposed denoiser where each middle layer has 24 feature maps.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The denoiser in Eqn (9) acts its own role regardless of the noise type and noise level of the image to be denoised.

    “the denoiser here just acts its own role regardless of the noise type and noise level of the image to be denoised.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    To show the advantage of the proposed design over a 7-layer traditional 3×3 network and a 16-layer network, three models were trained on noise level 25 with same training settings.

    “To show the advantage of our design over the above two cases, we have trained three different models on noise level 25 with same training settings.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    The iterative regularization step of methods such as NCSR and WNNM corresponds to solving Eqn.(6a).

    “It is worthy noting that the iterative regularization step of methods such as NCSR and WNNM actually corresponds to solving Eqn.(6a).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductown data

    The number of iterations set to 30 is large enough to obtain a satisfying performance.

    “The number of iterations is set to 30 as we find it is large enough to obtain a satisfying performance.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The convolution is carried out with circular boundary conditions.

  • May combine two claims.

    Interpretationreasoning

    The proposed deep CNN denoiser prior based optimization method is flexible in handling various tasks while producing very favorable results.

    “On the other hand, different from discriminative learning methods which are specialized for certain image restoration tasks, the proposed deep CNN denoiser prior based optimization method is flexible in handling various tasks while can produce very favorable results.”
    Quote in the paperSubstantial claim
23 of 23

  • The paper relies on this without support.

    The study of integration with fast discriminative denoiser prior is still lacking.

    “However, the study of integration with fast discriminative denoiser prior is still lacking.”
    “However, to the best of our knowledge, the study of integration with discriminative denoiser is still lacking.”

  • The authors state this as an assumption.

    The convolution is carried out with circular boundary conditions.

    “In addition, we assume the convolution is carried out with circular boundary conditions.”

  • The paper relies on this without support.

    The low-resolution image can be modeled by a blurring and subsequent down-sampling operation on a high-resolution one.

    “In general, the low-resolution (LR) image can be modeled by a blurring and subsequent down-sampling operation on a high-resolution one.”
3 of 3

Cites

Flagged first, then by citations

5 of the 81 references couldn’t be matched to a known paper.

Research works printed but not linked: 2

  1. foi, 2006, Pointwise shape adaptive DCT denoising with structure preservation in luminance-chrominance space, International Workshop on Video Processing and Quality Metrics for Consumer Electronics
  2. danielyan, 2010, Image deblurring by augmented lagrangian with BM3D frame prior, Workshop on Information Theoretic Methods in Science and Engineering, 16

Cited by

Sorted by citations

Related

Authors
Zhang, K., Zuo, W., Gu, S., & Zhang, L.
Affiliations
  • Hong Kong Polytechnic University (HK)
  • Harbin Institute of Technology (CN)
Venue
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
Publisher
IEEE
Published
2017-07
Pages
2808-2817
DOI
10.1109/cvpr.2017.300
OpenAlex
W2613155248
Cited by
2,116
References
81
Citations by year
2026: 80 · 2025: 191 · 2024: 248 · 2023: 293 · 2022: 272 · 2021: 338 · 2020: 297 · 2019: 240 · 2018: 135 · 2017: 10
FWCI
49.67
Type
conference paper
Language
en
Open access
Locations
  • IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings (publishedVersion)
  • PolyU Institutional Research Archive (Hong Kong Polytechnic University) (submittedVersion)
  • TUScholarShare (Temple University) (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Signal Denoising Methods
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image Processing Techniques
  • Physical Sciences › Engineering › Media Technology › Advanced Image Fusion Techniques
Keywords
image restoration, model-based optimization, discriminative learning, variable splitting, inverse problems, low-level vision, Gaussian denoising, image deblurring

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review