Back

Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring

Seungjun NahTae Hyun KimKyoung Mu Lee

article2017CVPR2,553 citations

Presents an end-to-end multi-scale deep network paired with a realistic high-speed camera dataset, shifting dynamic scene deblurring from restrictive blur kernel estimations toward direct coarse-to-fine image restoration.

Listen

Motion blur from camera shake and fast-moving objects in dynamic scenes remains a persistent challenge in photography, as it produces complex, spatially varying artifacts that degrade image quality. Conventional deblurring methods struggle because they depend on simplified assumptions about blur kernels, such as uniform or locally linear motion, which fail at object boundaries, occlusions, and depth changes, often introducing ringing artifacts.

The article set out to develop an end-to-end method that restores sharp images directly from blurry inputs without estimating explicit blur kernels, while also creating a more realistic training dataset to support such learning. Researchers built a multi-scale convolutional neural network that processes images at multiple resolutions in a coarse-to-fine manner, trained it using a combination of multi-scale content loss and adversarial loss on pairs of blurry and sharp images, and generated a new dataset of 3,214 image pairs by averaging sequences of frames captured at 240 frames per second with a high-speed camera.

On the authors' GOPRO test set the approach achieved roughly 4–5 dB higher PSNR and substantially better SSIM scores than prior leading methods while running in a few seconds rather than minutes or hours; similar gains appeared on the Köhler and Lai datasets, with visibly cleaner object boundaries and fewer artifacts on both synthetic and real dynamic scenes. These results indicate that kernel-free, data-driven restoration can handle the full range of real-world motion blur sources more reliably than optimization-based techniques that rely on kernel estimation.

The work demonstrates that high-quality dynamic scene deblurring is now feasible in practical time frames, which could improve downstream tasks such as object recognition, surveillance, and consumer photo editing. Because the model learns directly from realistic blur examples, it avoids the artifacts that arise when kernels are misspecified.

Further gains may come from expanding the training data to additional camera types and lighting conditions, or from integrating the network into video pipelines. The main limitations are that performance still depends on the distribution of the training set and that quantitative metrics such as PSNR do not always match human perception of sharpness; readers should verify results on their own imagery before deployment at scale.

Automation and Systems Research InstituteSeoul National Universitycomputer visionconvolutional neural networksenergy optimizationdynamic scene deblurringdeep multi-scale convolutional neural networkblind deblurring
arXiv: 1612.02177SeungjunNah/DeepDeblur_release

No sufficiently relevant recommendations were found.

Cover for Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring

Abstract

Non-uniform blind deblurring for general dynamic scenes is a challenging computer vision problem as blurs arise not only from multiple object motions but also from camera shake, scene depth variation. To remove these complicated motion blurs, conventional energy optimization based methods rely on simple assumptions such that blur kernel is partially uniform or locally linear. Moreover, recent machine learning based methods also depend on synthetic blur datasets generated under these assumptions. This makes conventional deblurring methods fail to remove blurs where blur kernel is difficult to approximate or parameterize (e.g. object motion boundaries). In this work, we propose a multi-scale convolutional neural network that restores sharp images in an end-to-end manner where blur is caused by various sources. Together, we present multi-scale loss function that mimics conventional coarse-to-fine approaches. Furthermore, we propose a new large-scale dataset that provides pairs of realistic blurry image and the corresponding ground truth sharp image that are obtained by a high-speed camera. With the proposed model trained on this dataset, we demonstrate empirically that our method achieves the state-of-the-art performance in dynamic scene deblurring not only qualitatively, but also quantitatively.

Topics

computer visionconvolutional neural networksenergy optimizationmotion bluroptimizationblurred imagedeep learningtotal variation regularization

Keywords

dynamic scene deblurringdeep multi-scale convolutional neural networkblind deblurringmotion blurblind deconvolutionmulti-scale content losslatent image estimationblur kernelscamera shake

Table of Contents

  • 1 Introduction
  • 1.1 Related Works
  • 1.2 Kernel-Free Learning for Dynamic Scene Deblurring
  • 2 Blur Dataset
  • 3 Proposed Method
  • 3.1 Model Architecture
  • 3.2 Training
  • 4 Experimental Results
  • 4.1 GOPRO Dataset
  • 4.2 Köhler Dataset
  • 4.3 Dataset of Lai et al.
  • 5 Conclusion
  • References
  • A Appendix
  • A.1 Comparison of loss function
  • A.2 Comparison on GOPRO dataset
  • A.3 Comparison on Lai et al. [20] dataset
  • A.4 Comparison on real dynamic scenes

Knowls

  1. Knowl 1 — Multi-Scale Coarse-to-Fine Convolutional Architecture for Kernel-Free Deblurring

    model/method

    The multi-scale blind deblurring neural network restores sharp images from dynamic scenes in an end-to-end manner without estimating intermediate blur kernels. The architecture takes a Gaussian pyramid of blurry images as input, denoted by {Bk}k=1K\{B_k\}_{k=1}^K, where k=1k=1 is the finest scale (full resolution) and k=Kk=K is the coarsest scale (downsampled by a scale ratio of 0.5k−10.5^{k-1}), and outputs an estimated latent sharp image pyramid {Lk}k=1K\{L_k\}_{k=1}^K.

    In the standard configuration with K=3K = 3 scales, the network operates as follows:

    • Coarsest Level (k=3k = 3): A 5×55 \times 5 convolution layer maps the 3-channel coarsest patch (e.g., 64×6464 \times 64) to 64 feature channels. This is followed by a stack of 19 modified residual blocks (ResBlocks) and a final 5×55 \times 5 convolution projecting back to 3 channels to produce L3L_3. All convolutions employ zero-padding to preserve spatial resolution, totaling 40 convolution layers at this scale.
    • Finer Levels (k=2,1k = 2, 1): Latent feature representations from scale k+1k+1 pass through a learnable upconvolution layer to match the spatial dimensions of scale kk and are concatenated with the blurry image BkB_k. The concatenated tensor is processed by a 5×55 \times 5 convolution, 19 ResBlocks (64 feature channels), and an output convolution layer to produce latent image LkL_k.
    • Across all 3 scale stages, the complete network contains 120 convolution layers with uniform 5×55 \times 5 kernel sizes. At inference, the output at the finest level L1L_1 is taken as the final restored sharp image.

  2. Knowl 2 — High-Speed Video Integration and Gamma-Corrected Blurry Image Synthesis

    model/method

    Rather than convolving sharp images with artificial or uniform blur kernels, realistic dynamic scene blur is synthesized by modeling the physical accumulation of light over a camera exposure duration TT. The observed blurry image BB is represented by:

    B=g(1T∫0TS(t)dt)≃g(1M∑i=0M−1S[i])B = g\left(\frac{1}{T}\int_{0}^T S(t) dt\right) \simeq g\left(\frac{1}{M}\sum_{i=0}^{M-1} S[i]\right)

    where S(t)S(t) is the continuous sharp sensor signal, MM is the number of discrete high-speed frames sampled during exposure, S[i]S[i] is the ii-th sharp latent signal, and gg denotes the camera response function (CRF) mapping latent sensor signal to observed pixel values S^[i]=g(S[i])\hat{S}[i] = g(S[i]).

    To account for nonlinear camera response without ground-truth calibration, the CRF is modeled using a standard gamma curve:

    g(x)=x1/γ,γ=2.2g(x) = x^{1/\gamma}, \quad \gamma = 2.2

    Observed video frames are linearized into latent frame signals via inverse gamma correction S[i]=g−1(S^[i])=(S^[i])γS[i] = g^{-1}(\hat{S}[i]) = (\hat{S}[i])^{\gamma}, averaged in linear radiance space over MM successive frames, and mapped back to the observed blurry image B=(1M∑i=0M−1(S^[i])2.2)1/2.2B = \left(\frac{1}{M}\sum_{i=0}^{M-1} (\hat{S}[i])^{2.2}\right)^{1/2.2}.

    In the GOPRO dataset, videos are captured at 240 frames per second using a GoPro Hero 4 Black camera. Successive sequences of M∈[7,13]M \in [7, 13] frames are averaged to simulate varying shutter speeds (e.g., averaging 15 frames simulates a 1/16 s exposure from 1/240 s individual frames). The sharp ground-truth image corresponding to BB is defined as the central sharp frame S^[⌊M/2⌋]\hat{S}[\lfloor M/2 \rfloor]. The resulting benchmark contains 3,214 pairs of blurry and sharp images at 1280×7201280 \times 720 resolution (2,103 training pairs and 1,111 testing pairs).

  3. Knowl 3 — Modified Residual Block Design for Image Restoration

    model/method

    The basic building block (ResBlock) of the multi-scale deblurring network adapts standard residual learning with two structural modifications tailored for low-level image restoration:

    1. Omission of Batch Normalization: Batch normalization layers are removed from the residual unit because deblurring networks are trained with small mini-batch sizes (mini-batch size of 2), where batch normalization statistics become noisy and degrade restoration accuracy.
    2. Removal of Post-Addition Activation: The rectified linear unit (ReLU) placed after the residual addition in standard ResNet blocks is eliminated. Each ResBlock consists of: Conv (5x5, 64 channels) -> ReLU -> Conv (5x5, 64 channels) + Identity Shortcut -> Output.

    Removing the final activation enables the network layers to directly learn unrestricted signed residual differences between blurry and sharp inputs across consecutive blocks, accelerating training convergence and improving deblurring quality.

  4. Knowl 4 — Multi-Scale Content and Adversarial Joint Loss Formulation

    equation

    The multi-scale deblurring network is trained using a composite loss function combining a normalized multi-scale mean squared error (MSE) content loss Lcont\mathcal{L}_{\text{cont}} and an adversarial loss Ladv\mathcal{L}_{\text{adv}}:

    Ltotal=Lcont+λLadv\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{cont}} + \lambda \mathcal{L}_{\text{adv}}

    where λ=1×10−4\lambda = 1 \times 10^{-4} is a weighting constant.

    The multi-scale content loss supervises each level of the generated Gaussian pyramid against the corresponding ground-truth sharp scale:

    Lcont=12K∑k=1K1ckwkhk∥Lk−Sk∥22\mathcal{L}_{\text{cont}} = \frac{1}{2K} \sum_{k=1}^K \frac{1}{c_k w_k h_k} \|L_k - S_k\|_2^2

    where KK is the total number of pyramid levels (K=3K=3), LkL_k and SkS_k denote the network's predicted output and the ground-truth sharp image at scale level kk, and ck,wk,hkc_k, w_k, h_k denote the channel count (ck=3c_k=3), width, and height of the patch at scale kk, normalizing the error by the total number of elements per scale.

    The adversarial loss acts on the finest-scale output (L1=G(B)L_1 = G(B)):

    Ladv=ES∼psharp(S)[log⁡D(S)]+EB∼pblurry(B)[log⁡(1−D(G(B)))]\mathcal{L}_{\text{adv}} = \mathbb{E}_{S \sim p_{\text{sharp}}(S)}[\log D(S)] + \mathbb{E}_{B \sim p_{\text{blurry}}(B)}[\log(1 - D(G(B)))]

    where GG is the multi-scale generator producing L1L_1 and DD is a discriminator trained to distinguish ground-truth sharp images from restored images.

  5. Knowl 5 — Discriminator Network Architecture for Adversarial Deblurring

    model/method

    To train the generator with adversarial feedback, a discriminator network DD takes either a full-resolution deblurred image L1L_1 or a ground-truth sharp image S1S_1 of dimension 3×H×W3 \times H \times W and outputs a scalar probability score. All convolution layers are followed by LeakyReLU activations.

    The discriminator layer configuration is as follows:

    • Layer 1: Conv (32×3×5×532 \times 3 \times 5 \times 5, stride 2)
    • Layer 2: Conv (64×32×5×564 \times 32 \times 5 \times 5, stride 1)
    • Layer 3: Conv (64×64×5×564 \times 64 \times 5 \times 5, stride 2)
    • Layer 4: Conv (128×64×5×5128 \times 64 \times 5 \times 5, stride 1)
    • Layer 5: Conv (128×128×5×5128 \times 128 \times 5 \times 5, stride 4)
    • Layer 6: Conv (256×128×5×5256 \times 128 \times 5 \times 5, stride 1)
    • Layer 7: Conv (256×256×5×5256 \times 256 \times 5 \times 5, stride 4)
    • Layer 8: Conv (512×256×5×5512 \times 256 \times 5 \times 5, stride 1)
    • Layer 9: Conv (512×512×4×4512 \times 512 \times 4 \times 4, stride 4)
    • Layer 10: Fully Connected (512×1×1×1512 \times 1 \times 1 \times 1)
    • Layer 11: Sigmoid activation

  6. Knowl 6 — Inter-Scale Feature Propagation via Learnable Upconvolution

    model/method

    In the coarse-to-fine multi-scale deblurring network, transfer of information from a coarser stage k+1k+1 to the next finer stage kk is performed using a learnable upconvolution (transposed convolution) layer rather than non-parametric bilinear upsampling or spatial reshaping.

    Because sharp and blurry patches share common low-frequency structural content, passing the coarser scale latent sharp representation through a learnable upconvolution allows the network to adaptively transform coarser features into spatial dimensions matching the finer level while filtering redundant low-frequency components. The upconvolved feature map is concatenated along the channel dimension with the blurry input image patch BkB_k at scale level kk, forming the combined input for the finer-scale network stage.

  7. Knowl 7 — Deblurring Training Configuration and Data Augmentation Pipeline

    experimental setup

    The multi-scale deblurring model is trained on Gaussian pyramid patches of resolutions {256×256,128×128,64×64}\{256 \times 256, 128 \times 128, 64 \times 64\} extracted from the GOPRO dataset (2,103 training pairs). Pixel values are mapped to range [−0.5,0.5][-0.5, 0.5] by subtracting 0.5.

    The training pipeline employs several data augmentations:

    • Geometric: Random horizontal flips, random vertical flips, and random 90-degree rotations.
    • Color & Photometric: Random RGB channel permutations, random multiplication of HSV saturation by a scaling factor sampled uniformly from [0.5,1.5][0.5, 1.5].
    • Noise Robustness: Additive zero-mean Gaussian noise N(0,σ2)\mathcal{N}(0, \sigma^2) with σ=2/255\sigma = 2/255, followed by clipping to [0,1][0, 1].

    Optimization uses the ADAM optimizer with mini-batch size 2. The initial learning rate is set to 5×10−55 \times 10^{-5} for the first 3×1053 \times 10^5 iterations, then reduced to 5×10−65 \times 10^{-6} for the remaining training iterations, running for a total of 9×1059 \times 10^5 iterations to reach convergence on an NVIDIA GTX Titan X GPU.

  8. Knowl 8 — Deblurring Performance and Multi-Scale Analysis on the GOPRO Dataset

    data/table

    Quantitative deblurring performance was evaluated on the 1,111 test image pairs of the GOPRO dataset at 1280×7201280 \times 720 resolution. The proposed method was evaluated across different numbers of scale levels K∈{1,2,3}K \in \{1, 2, 3\} and compared against state-of-the-art dynamic scene deblurring methods:

    Measure Sun et al. [26] Kim and Lee [15] Ours
    K=1K = 1 K=2K = 2 K=3K = 3
    PSNR (dB) 24.64 23.64 28.93 29.23 29.08
    SSIM 0.8429 0.8239 0.9100 0.9162 0.9135
    Runtime 20 min 1 hr 7.21 s 4.33 s 3.09 s

    The proposed method outperforms prior approaches by over 4.4 dB PSNR and 0.07 SSIM while running orders of magnitude faster (seconds vs. tens of minutes or hours per image). Comparing scale configurations, K=2K = 2 yields the highest restoration accuracy (29.23 dB PSNR, 0.9162 SSIM), whereas K=3K = 3 provides the lowest computational latency (3.09 s runtime on an NVIDIA GTX Titan X GPU) with comparable quality.

  9. Knowl 9 — Quantitative Evaluation on the Köhler Deblurring Benchmark

    data/table

    The Köhler benchmark evaluates blind deconvolution across 4 latent images and 12 distinct 6D camera motion trajectories (48 test images total) under linear camera response conditions. The proposed model was evaluated by training with the camera response function gg set to the identity function:

    Measure Sun et al. [26] Kim and Lee [15] Ours
    K=1K = 1 K=2K = 2 K=3K = 3
    PSNR (dB) 25.22 24.68 25.74 26.02 26.48
    MSSIM 0.7735 0.7937 0.8042 0.8116 0.8079

    All multi-scale variants (K∈{1,2,3}K \in \{1, 2, 3\}) surpass the comparison methods in both metrics. The K=3K = 3 model achieves the highest PSNR (26.48 dB), exceeding Sun et al. by 1.26 dB and Kim and Lee by 1.80 dB, while the K=2K = 2 model achieves the highest Multi-Scale Structural Similarity (MSSIM = 0.8116).

  10. Knowl 10 — Quantitative Impact of Adversarial Loss on Image Restoration Metrics

    data/table

    An ablation study on the GOPRO test dataset (evaluated assuming linear camera response, with K=3K = 3 scale levels and adversarial weight λ=1×10−4\lambda = 1 \times 10^{-4}) demonstrates the trade-off introduced by incorporating adversarial loss into the objective:

    Loss Function Lcont (MSE)\mathcal{L}_{\text{cont}} \text{ (MSE)} Lcont+λLadv\mathcal{L}_{\text{cont}} + \lambda \mathcal{L}_{\text{adv}}
    PSNR (dB) 28.62 28.45
    SSIM 0.9094 0.9170

    Adding adversarial supervision results in a slight decrease in pixel-level reconstruction fidelity (-0.17 dB PSNR) while increasing structural preservation (+0.0076 SSIM). The adversarial loss regularizes output statistics against oversmoothing, yielding sharper structural boundaries.

Coverage note — No substantial contributed material was omitted; qualitative evaluations on the synthetic Lai et al. benchmark and real dynamic Sony RX100 captures were omitted as their numerical and architectural insights are captured by the benchmark data tables and model descriptions.

References

  1. 1.A. Chakrabarti. A neural approach to blind motion deblurring. In ECCV, 2016. 1, 2
  2. 2.F. Couzinie-Devy, J. Sun, K. Alahari, and J. Ponce. Learning to estimate and remove non-uniform image blur. In CVPR, 2013. 2
  3. 3.J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 2
  4. 4.E. L. Denton, S. Chintala, R. Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems, pages 1486–1494, 2015. 3, 5, 6
  5. 5.A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazirbas, V. Golkov, P. van der Smagt, D. Cremers, and T. Brox. Flownet: Learning optical flow with convolutional networks. In CVPR, pages 2758–2766, 2015. 3
  6. 6.D. Eigen and R. Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In ICCV, pages 2650–2658, 2015. 3, 5
  7. 7.D. Eigen, D. Krishnan, and R. Fergus. Restoring an image taken through a window covered with dirt or rain. In ICCV, pages 633–640, 2013. 2
  8. 8.D. Eigen, C. Puhrsch, and R. Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in Neural Information Ppocessing Ssytems, pages 2366–2374, 2014. 3, 5
  9. 9.I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 2, 6
  10. 10.A. Gupta, N. Joshi, C. L. Zitnick, M. Cohen, and B. Curless. Single image deblurring using motion density functions. In ECCV, pages 171–184. Springer, 2010. 1
  11. 11.S. Harmeling, H. Michael, and B. Schölkopf. Space-variant single-image blind deconvolution for removing camera shake. In Advances in Neural Information Processing Systems, pages 829–837, 2010. 1
  12. 12.K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 4
  13. 13.M. Hirsch, C. J. Schuler, S. Harmeling, and B. Schölkopf. Fast removal of non-uniform camera shake. In ICCV, 2011. 1, 3
  14. 14.T. H. Kim, B. Ahn, and K. M. Lee. Dynamic scene deblurring. In ICCV, 2013. 1
  15. 15.T. H. Kim and K. M. Lee. Segmentation-free dynamic scene deblurring. In CVPR, 2014. 1, 6, 7, 8, 13, 19
  16. 16.T. H. Kim and K. M. Lee. Generalized video deblurring for dynamic scenes. In CVPR, 2015. 2
  17. 17.T. H. Kim, S. Nah, and K. M. Lee. Dynamic scene deblurring using a locally adaptive linear blur model. arXiv preprint arXiv:1603.04265, 2016. 2
  18. 18.D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6
  19. 19.R. Köhler, M. Hirsch, B. Mohler, B. Schölkopf, and S. Harmeling. Recording and playback of camera shake: Benchmarking blind deconvolution with a real-world database. In ECCV, pages 27–40. Springer, 2012. 6
  20. 20.W.-S. Lai, J.-B. Huang, Z. Hu, N. Ahuja, and M.-H. Yang. A comparative study for single image blind deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1701–1709, 2016. 8, 16
  21. 21.Y. Li, S. B. Kang, N. Joshi, S. M. Seitz, and D. P. Huttenlocher. Generating sharp panoramas from motion-blurred videos. In CVPR, 2010. 2
  22. 22.J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431–3440, 2015. 5
  23. 23.M. Mathieu, C. Couprie, and Y. LeCun. Deep multi-scale video prediction beyond mean square error. arXiv preprint arXiv:1511.05440, 2015. 3, 5
  24. 24.A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 6
  25. 25.C. J. Schuler, M. Hirsch, S. Harmeling, and B. Schölkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439–1451, 2016. 1, 2, 3
  26. 26.J. Sun, W. Cao, Z. Xu, and J. Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In CVPR, pages 769–777. IEEE, 2015. 1, 2, 3, 6, 7, 8, 13, 19
  27. 27.Y.-W. Tai, X. Chen, S. Kim, S. J. Kim, F. Li, J. Yang, J. Yu, Y. Matsushita, and M. S. Brown. Nonlinear camera response functions and image deblurring: Theoretical analysis and practice. PAMI, 35(10):2498–2512, 2013. 3
  28. 28.O. Whyte, J. Sivic, A. Zisserman, and J. Ponce. Non-uniform deblurring for shaken images. 2010. 1
  29. 29.L. Xu, J. S. Ren, C. Liu, and J. Jia. Deep convolutional neural network for image deconvolution. In Advances in Neural Information Processing Systems, pages 1790–1798, 2014. 1, 2
  30. 30.D. Zoran and Y. Weiss. From learning models of natural image patches to whole image restoration. In ICCV, pages 479–486. IEEE, 2011. 3

Citation

MLA
Nah, S., et al. “Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring”. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 257–65, https://doi.org/10.1109/CVPR.2017.35.
APA
Nah, S., Kim, T. H., & Lee, K. M. (2017). Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 257–265. https://doi.org/10.1109/CVPR.2017.35
Chicago
Nah, S., T. H. Kim, and K. M. Lee. 2017. “Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring”. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 257–65. https://doi.org/10.1109/CVPR.2017.35.
Harvard
Nah, S., Kim, T.H. and Lee, K.M. (2017) “Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring”, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 257–265. Available at: https://doi.org/10.1109/CVPR.2017.35.
Vancouver
1. Nah S, Kim TH, Lee KM (2017) Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 257–265

BibTeX

@inproceedings{Nah_2017, title={Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring}, url={http://dx.doi.org/10.1109/CVPR.2017.35}, DOI={10.1109/cvpr.2017.35}, booktitle={2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Nah, Seungjun and Kim, Tae Hyun and Lee, Kyoung Mu}, year={2017}, month=July, pages={257–265} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: IEEE