Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections

Xiao-Jiao MaoChunhua ShenYubin Yang

article2016NeurIPS1,708 citations

Proposes a deep convolutional encoder-decoder architecture with symmetric skip connections that resolves vanishing gradients, preserves fine image details, and achieves state-of-the-art restoration quality across multiple tasks such as denoising and super-resolution using a single unified model.

Listen

Recovering high-quality visual content from degraded inputs is a fundamental challenge in digital imaging. Traditional restoration systems often struggle to preserve fine textural details while removing severe noise or increasing image resolution. Additionally, conventional machine learning approaches frequently require maintaining separate specialized models for every anticipated level of image degradation, creating significant operational and deployment overhead.

The article demonstrates that a very deep artificial neural network combining encoding and decoding layers with symmetric direct connections can systematically overcome these limitations. The primary objective is to demonstrate that this single deep learning framework can restore degraded images with superior accuracy while handling multiple levels and types of corruption within a single unified model.

To evaluate this approach, the researchers designed and tested network architectures of up to 30 layers trained on approximately 500,000 image patches derived from standard benchmark datasets. The model uses early convolutional layers to filter out noise and extract core structural features, while mirrored deconvolutional layers decode and reconstruct lost detail. Crucially, symmetric connections pass fine details directly across the network while facilitating gradient flow during training, preventing the performance degradation typical of very deep models. The framework was benchmarked against established algorithms across standard test sets for both random noise reduction and multi-scale image enhancement.

The analysis reveals several major findings. First, the 30-layer model consistently outperformed existing state-of-the-art methods across all test datasets in standard image quality and structural similarity metrics. Second, direct symmetric connections solved the vanishing gradient problem, allowing deeper 30-layer networks to achieve lower training loss and noticeably sharper reconstructions than shallower 10-layer configurations. Third, a single 30-layer model trained across varying degradation levels successfully handled multiple noise intensities and scaling factors with only negligible performance drops compared to dedicated models. Fourth, processing images through multi-orientation filtering during inference provided additional gains in smoothness and overall output fidelity.

These results provide important operational and technical implications. Organizations deploying image restoration can simplify their operational pipelines by replacing complex suites of narrow algorithms with a single, highly capable deep network. This substantially reduces model maintenance overhead while lowering the risk of choosing incorrect degradation parameters at runtime. Because the system learns end-to-end directly from raw training data, it eliminates manual feature engineering and avoids fragile assumptions regarding noise distributions.

Based on these findings, decision-makers should consider adopting deep symmetric encoder-decoder frameworks for demanding automated image restoration workflows. Implementation teams should leverage unified multi-degradation models to simplify software infrastructure and apply multi-orientation testing passes when maximum output quality is required. Further practical deployment should focus on evaluating hardware inference latencies and testing performance across wider real-world artifacts, such as optical blur and mixed compression artifacts.

Decision-makers can maintain high confidence in these findings across standard benchmark conditions, as the model was rigorously validated against multiple competitive baselines. However, stakeholders should note that initial evaluations focused primarily on single-channel luminance data and synthetic Gaussian degradations. A measured approach involves conducting targeted pilot validations on domain-specific, multi-channel color imagery before full-scale deployment.

arXiv: 1603.09056
Cover for Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections

Abstract

In this paper, we propose a very deep fully convolutional encoding-decoding framework for image restoration such as denoising and super-resolution. The network is composed of multiple layers of convolution and de-convolution operators, learning end-to-end mappings from corrupted images to the original ones. The convolutional layers act as the feature extractor, which capture the abstraction of image contents while eliminating noises/corruptions. De-convolutional layers are then used to recover the image details. We propose to symmetrically link convolutional and de-convolutional layers with skip-layer connections, with which the training converges much faster and attains a higher-quality local optimum. First, The skip connections allow the signal to be back-propagated to bottom layers directly, and thus tackles the problem of gradient vanishing, making training deep networks easier and achieving restoration performance gains consequently. Second, these skip connections pass image details from convolutional layers to de-convolutional layers, which is beneficial in recovering the original image. Significantly, with the large capacity, we can handle different levels of noises using a single model. Experimental results show that our network achieves better performance than all previously reported state-of-the-art methods.

Table of Contents

  • 1 Introduction
  • 2 Very deep RED-Net for Image Restoration
  • 2.1 Architecture
  • 2.1.1 Deconvolution decoder
  • 2.1.2 Skip connections
  • 2.2 Discussions
  • 2.3 Training
  • 2.4 Testing
  • 3 Experiments
  • 3.1 Image Denoising
  • 3.2 Image super-resolution
  • 3.3 Evaluation with a single model
  • 4 Conclusions
  • References

Knowls

  1. Knowl 1 — RED-Net Architecture for Image Restoration

    model/method

    The Residual Encoder-Decoder Network (RED-Net) is a fully convolutional and deconvolutional neural network designed for low-level image restoration tasks such as denoising and super-resolution.

    The network comprises an encoder chain of convolutional layers followed by a symmetric decoder chain of deconvolutional (transposed convolutional) layers:

    1. Encoder (Convolutional Layers): Multiple convolutional layers act as feature extractors that capture abstract representations of image content while eliminating noise and corruptions.
    2. Decoder (Deconvolutional Layers): A mirrored chain of deconvolutional layers reconstructs fine image details from the abstracted representations to produce a clean output image.
    3. Filter and Kernel Configuration: All convolutional and deconvolutional layers employ 3×33 \times 3 filter kernels and 64 feature channels. Rectified Linear Unit (ReLU) activation functions follow every convolution and deconvolution step.
    4. No Pooling or Unpooling: To avoid losing fine spatial details essential for image restoration, neither pooling nor unpooling operations are used; spatial resolution is preserved across layers.
    5. Symmetric Skip Connections: Feature maps from convolutional layers are connected to their mirrored deconvolutional counterparts every two layers via element-wise addition before applying the ReLU activation.

    Three standard depth configurations are defined: RED10 (5 convolutional and 5 deconvolutional layers without skip connections), RED20 (10 convolutional and 10 deconvolutional layers with skip connections every 2 layers), and RED30 (15 convolutional and 15 deconvolutional layers with skip connections every 2 layers). The network accepts arbitrary input image dimensions w×h×cw \times h \times c and outputs restored images of the same spatial dimension.

  2. Knowl 2 — Symmetric Skip Connections in Residual Encoder-Decoder Networks

    model/method

    In RED-Net, skip connections are placed symmetrically between corresponding convolutional (encoder) and deconvolutional (decoder) layers. If the encoder has layers l=1,…,Ll = 1, \dots, L and the decoder has mirrored layers l′=L+1,…,2Ll' = L+1, \dots, 2L, skip connections bridge convolutional layer ll and deconvolutional layer 2L−l+12L - l + 1 (typically every two layers).

    The feature map output from convolutional layer ll is passed forward and summed element-wise with the feature map output of the corresponding mirrored deconvolutional layer before passing through a rectification (ReLU) layer.

    This symmetric skip-connection design provides two primary functional benefits:

    1. Forward Detail Propagation: Convolutional layers eliminate corruption but progressively discard high-frequency spatial details. Passing early convolutional feature maps directly to mirrored deconvolutional layers injects fine-grained spatial details back into the decoder, improving reconstruction fidelity.
    2. Backward Gradient Propagation: Skip connections provide direct identity paths that allow gradient signals to back-propagate directly from top decoder layers to bottom encoder layers during training, resolving the vanishing gradient problem and enabling deeper models (such as 20-layer and 30-layer configurations) to achieve lower training loss and higher restoration quality than networks without shortcuts.

    Unlike standard residual networks where skip connections span sequential blocks within identical feature representations, symmetric skip connections establish element-wise correspondence across the encoding-decoding symmetry.

  3. Knowl 3 — Training Objective and Loss Function for RED-Net

    model/method

    RED-Net is trained end-to-end to learn the mapping from corrupted input images to clean target images by minimizing the Mean Squared Error (MSE) / Frobenius norm loss over network parameters Θ\Theta:

    L(Θ)=1N∑i=1N∥F(Xi;Θ)−Yi∥F2\mathcal{L}(\Theta) = \frac{1}{N} \sum_{i=1}^{N} \|\mathcal{F}(X_i; \Theta) - Y_i\|_F^2

    where:

    • Θ\Theta represents the learnable weights and biases of the convolutional and deconvolutional kernels.
    • NN is the number of training image patch pairs.
    • Xi∈Rw×h×cX_i \in \mathbb{R}^{w \times h \times c} is the ii-th corrupted input patch (e.g., noisy or bicubic upscaled low-resolution patch).
    • Yi∈Rw×h×cY_i \in \mathbb{R}^{w \times h \times c} is the corresponding clean ground-truth image patch.
    • F(Xi;Θ)\mathcal{F}(X_i; \Theta) is the output produced by the network for input XiX_i.
    • ∥⋅∥F\|\cdot\|_F denotes the Frobenius norm.

    Training Setup:

    • Optimization is performed using the Adam optimizer with an initial base learning rate of 10−410^{-4} applied uniformly across all layers.
    • Training samples are extracted from 300 images in the Berkeley Segmentation Dataset (BSD), generating approximately 0.5×1060.5 \times 10^6 patches of size 50×5050 \times 50 pixels.
    • For image denoising, training targets are single-channel grayscale patches corrupted with additive Gaussian noise. For super-resolution, training is performed on the luminance (YY) channel of downsampled and upscaled patches.
  4. Knowl 4 — Multi-Orientation Geometric Kernel Transform Testing

    algorithm

    To improve output smoothness and boost restoration accuracy during inference, RED-Net uses multi-orientation testing. Because restoration filter kernels aim to eliminate corruption independently of content orientation, predictions under geometric transformations can be ensembled.

    Input: Corrupted test image XX, trained network F(⋅;Θ)\mathcal{F}(\cdot; \Theta)
    Output: Restored smooth image Y^\hat{Y}
    Initialize accumulator image S←0S \leftarrow 0
    Define geometric transformations T={t1,t2,…,tK}\mathcal{T} = \{t_1, t_2, \dots, t_K\} consisting of rotations by 0∘,90∘,180∘,270∘0^\circ, 90^\circ, 180^\circ, 270^\circ and horizontal mirror flips
    for each transformation tk∈Tt_k \in \mathcal{T} do
        Xk←tk(X)X_k \leftarrow t_k(X)
        Yk←F(Xk;Θ)Y_k \leftarrow \mathcal{F}(X_k; \Theta)
        Y~k←tk−1(Yk)\tilde{Y}_k \leftarrow t_k^{-1}(Y_k)
        S←S+Y~kS \leftarrow S + \tilde{Y}_k
    end for
    Y^←1∣T∣S\hat{Y} \leftarrow \frac{1}{|\mathcal{T}|} S
    return Y^\hat{Y}

    Alternatively, applying rotations and flips directly to the convolutional and deconvolutional filter kernels yields an equivalent ensemble that produces higher PSNR and SSIM than a single forward pass.

  5. Knowl 5 — Single Model Training Across Multiple Corruption Levels

    model/method

    Traditional restoration models require separate specialized networks trained individually for each noise variance σ\sigma or super-resolution scale ss, requiring prior corruption estimation. Due to the high model capacity of deep residual encoder-decoder architectures (e.g., RED30 with 30 layers and 64 feature channels), a single network can be trained on a combined dataset encompassing multiple noise levels (e.g., σ∈{10,30,50,70}\sigma \in \{10, 30, 50, 70\}) or multiple upscaling factors (e.g., s∈{2,3,4}s \in \{2, 3, 4\}).

    When trained across diverse corruption severities simultaneously, the single multi-level model achieves restoration performance nearly identical to models trained exclusively on specific noise or scale levels, while outperforming existing dedicated baseline models.

  6. Knowl 6 — Benchmark Image Denoising Evaluation Results

    data/table

    The denoising performance of RED-Net models (RED10, RED20, RED30) was evaluated on 14 standard benchmark images against BM3D, EPLL, NCSR, PCLR, PGPD, and WNNM under additive Gaussian noise with standard deviations σ∈{10,30,50,70}\sigma \in \{10, 30, 50, 70\}.

    Method PSNR (dB) SSIM
    σ=10\sigma=10 σ=30\sigma=30 σ=50\sigma=50 σ=70\sigma=70 σ=10\sigma=10 σ=30\sigma=30 σ=50\sigma=50 σ=70\sigma=70
    BM3D 34.18 28.49 26.08 24.65 0.9339 0.8204 0.7427 0.6882
    EPLL 33.98 28.35 25.97 24.47 0.9332 0.8200 0.7354 0.6712
    NCSR 34.27 28.44 25.93 24.36 0.9342 0.8203 0.7415 0.6871
    PCLR 34.48 28.68 26.29 24.79 0.9366 0.8263 0.7538 0.6997
    PGPD 34.22 28.55 26.19 24.71 0.9309 0.8199 0.7442 0.6913
    WNNM 34.49 28.74 26.32 24.80 0.9363 0.8273 0.7517 0.6975
    RED10 34.62 28.95 26.51 24.97 0.9374 0.8327 0.7571 0.7012
    RED20 34.74 29.10 26.72 25.23 0.9392 0.8396 0.7689 0.7177
    RED30 34.81 29.17 26.81 25.31 0.9402 0.8423 0.7733 0.7206

    RED10 (without skip connections) outperforms the previous best method (WNNM). Increasing network depth while adding symmetric skip connections in RED20 and RED30 yields additional gains across all noise levels, exceeding WNNM by 0.32 dB at σ=10\sigma = 10, 0.43 dB at σ=30\sigma = 30, 0.49 dB at σ=50\sigma = 50, and 0.51 dB at σ=70\sigma = 70.

  7. Knowl 7 — Super-Resolution Benchmark Results on Set5, Set14, and BSD100

    data/table

    Super-resolution performance (PSNR in dB / SSIM) was evaluated on the Set5, Set14, and BSD100 datasets for scaling factors s∈{2,3,4}s \in \{2, 3, 4\}, comparing RED-Net against SRCNN, NBSRF, CSCN, CSC, TSE, and ARFL+.

    Dataset Scale SRCNN NBSRF CSCN CSC TSE ARFL+ RED10 RED20 RED30
    Set5 s=2s=2 36.66 36.76 37.14 36.62 36.50 36.89 37.43 37.62 37.66
    s=3s=3 32.75 32.75 33.26 32.66 32.62 32.72 33.43 33.80 33.82
    s=4s=4 30.49 30.44 31.04 30.36 30.33 30.35 31.12 31.40 31.51
    Set14 s=2s=2 32.45 32.45 32.71 32.31 32.23 32.52 32.77 32.87 32.94
    s=3s=3 29.30 29.25 29.55 29.15 29.16 29.23 29.42 29.61 29.61
    s=4s=4 27.50 27.42 27.76 27.30 27.40 27.41 27.58 27.80 27.86
    BSD100 s=2s=2 31.36 31.30 31.54 31.27 31.18 31.35 31.85 31.95 31.99
    s=3s=3 28.41 28.36 28.58 28.31 28.30 28.36 28.79 28.90 28.93
    s=4s=4 26.90 26.88 27.11 26.83 26.85 26.86 27.25 27.35 27.40

    RED10 outperforms earlier convolutional methods (e.g., SRCNN). RED30 achieves the highest PSNR and SSIM across all datasets and scales, outperforming CSCN on Set5 by 0.52 dB (s=2s=2), 0.56 dB (s=3s=3), and 0.47 dB (s=4s=4), and on BSD100 by 0.45 dB (s=2s=2), 0.35 dB (s=3s=3), and 0.29 dB (s=4s=4).

  8. Knowl 8 — Restoration Performance of Unified Single 30-Layer RED-Net

    data/table

    Quantitative performance of a single unified 30-layer RED-Net trained jointly to handle all corruption levels without requiring corruption-specific parameter tuning.

    Denoising with a Single Model:

    Dataset 14 Benchmark Images BSD200
    Metric σ=10\sigma=10 σ=30\sigma=30 σ=50\sigma=50 σ=70\sigma=70 σ=10\sigma=10 σ=30\sigma=30 σ=50\sigma=50 σ=70\sigma=70
    PSNR (dB) 34.49 29.09 26.75 25.20 33.38 27.88 25.69 24.36
    SSIM 0.9368 0.8414 0.7716 0.7157 0.9280 0.7980 0.7119 0.6544

    Super-Resolution with a Single Model:

    Dataset Set5 Set14 BSD100
    Metric s=2s=2 s=3s=3 s=4s=4 s=2s=2 s=3s=3 s=4s=4 s=2s=2 s=3s=3 s=4s=4
    PSNR (dB) 37.56 33.70 31.33 32.81 29.50 27.72 31.96 28.88 27.35
    SSIM 0.9595 0.9222 0.8847 0.9135 0.8334 0.7698 0.8972 0.7993 0.7276

    Compared to dedicated models trained on individual noise levels or upscaling factors, the unified single model experiences only slight performance degradation (e.g., 34.49 dB vs 34.81 dB at σ=10\sigma=10 on the 14 images) and still outperforms prior state-of-the-art dedicated algorithms.

Coverage note — No substantial contributed material was omitted. Qualitative visual comparisons and preliminary toy ablations on 5-layer strided autoencoders were summarized within the architectural and skip connection knowls.

References

  1. 1.H. C. Burger, C. J. Schuler, and S. Harmeling. Image denoising: Can plain neural networks compete with BM3D? In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 2392–2399, 2012.
  2. 2.P. Chatterjee and P. Milanfar. Clustering-based denoising with locally learned dictionaries. IEEE Trans. Image Process., 18(7):1438–1451, 2009.
  3. 3.F. Chen, L. Zhang, and H. Yu. External patch prior guided internal clustering for image denoising. In Proc. IEEE Int. Conf. Comp. Vis., pages 603–611, 2015.
  4. 4.Z. Cui, H. Chang, S. Shan, B. Zhong, and X. Chen. Deep network cascade for image super-resolution. In Proc. Eur. Conf. Comp. Vis., pages 49–64, 2014.
  5. 5.K. Dabov, A. Foi, V. Katkovnik, and K. O. Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Trans. Image Process., 16(8):2080–2095, 2007.
  6. 6.C. Dong, Y. Deng, C. C. Loy, and X. Tang. Compression artifacts reduction by a deep convolutional network. In Proc. IEEE Int. Conf. Comp. Vis., pages 576–584, 2015.
  7. 7.C. Dong, C. C. Loy, K. He, and X. Tang. Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 38(2):295–307, 2016.
  8. 8.W. Dong, L. Zhang, G. Shi, and X. Li. Nonlocally centralized sparse representation for image restoration. IEEE Trans. Image Process., 22(4):1620–1630, 2013.
  9. 9.S. Gu, L. Zhang, W. Zuo, and X. Feng. Weighted nuclear norm minimization with application to image denoising. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 2862–2869, 2014.
  10. 10.S. Gu, W. Zuo, Q. Xie, D. Meng, X. Feng, and L. Zhang. Convolutional sparse coding for image super-resolution. In Proc. IEEE Int. Conf. Comp. Vis., pages 1823–1831, 2015.
  11. 11.K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., volume abs/1512.03385, 2016.
  12. 12.S. Hong, H. Noh, and B. Han. Decoupled deep neural network for semi-supervised semantic segmentation. In Proc. Advances in Neural Inf. Process. Syst., 2015.
  13. 13.J. Huang, A. Singh, and N. Ahuja. Single image super-resolution from transformed self-exemplars. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 5197–5206, 2015.
  14. 14.Y. Huang, W. Wang, and L. Wang. Bidirectional recurrent convolutional networks for multi-frame super-resolution. In Proc. Advances in Neural Inf. Process. Syst., pages 235–243, 2015.
  15. 15.V. Jain and H. S. Seung. Natural image denoising with convolutional networks. In Proc. Advances in Neural Inf. Process. Syst., pages 769–776, 2008.
  16. 16.Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014.
  17. 17.D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proc. Int. Conf. Learn. Representations, 2015.
  18. 18.H. Liu, R. Xiong, J. Zhang, and W. Gao. Image denoising via adaptive soft-thresholding based on non-local samples. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 484–492, 2015.
  19. 19.J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 3431–3440, 2015.
  20. 20.D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. IEEE Int. Conf. Comp. Vis., volume 2, pages 416–423, July 2001.
  21. 21.P. Milanfar. A tour of modern image filtering: New insights and methods, both practical and theoretical. IEEE Signal Process. Mag., 30(1):106–128, 2013.
  22. 22.H. Noh, S. Hong, and B. Han. Learning deconvolution network for semantic segmentation. In Proc. IEEE Int. Conf. Comp. Vis., pages 1520–1528, 2015.
  23. 23.S. Osher, M. Burger, D. Goldfarb, J. Xu, and W. Yin. An iterative regularization method for total variation-based image restoration. Multiscale Modeling & Simulation, 4(2):460–489, 2005.
  24. 24.L. I. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based noise removal algorithms. Phys. D, 60(1-4):259–268, November 1992.
  25. 25.J. Salvador and E. Perez-Pellitero. Naive bayes super-resolution forest. In Proc. IEEE Int. Conf. Comp. Vis., pages 325–333, 2015.
  26. 26.S. Schulter, C. Leistner, and H. Bischof. Fast and accurate image upscaling with super-resolution forests. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 3791–3799, 2015.
  27. 27.K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. 2015.
  28. 28.R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. In Proc. Advances in Neural Inf. Process. Syst., pages 2377–2385, 2015.
  29. 29.P. Vincent, H. Larochelle, Y. Bengio, and P. Manzagol. Extracting and composing robust features with denoising autoencoders. In Proc. Int. Conf. Mach. Learn., pages 1096–1103, 2008.
  30. 30.Z. Wang, D. Liu, J. Yang, W. Han, and T. S. Huang. Deep networks for image super-resolution with sparse prior. In Proc. IEEE Int. Conf. Comp. Vis., pages 370–378, 2015.
  31. 31.Z. Wang, Y. Yang, Z. Wang, S. Chang, J. Yang, and T. S. Huang. Learning super-resolution jointly from external and internal examples. IEEE Trans. Image Process., 24(11):4359–4371, 2015.
  32. 32.J. Xie, L. Xu, and E. Chen. Image denoising and inpainting with deep neural networks. In Proc. Advances in Neural Inf. Process. Syst., pages 350–358, 2012.
  33. 33.J. Xu, L. Zhang, W. Zuo, D. Zhang, and X. Feng. Patch group based nonlocal self-similarity prior learning for image denoising. In Proc. IEEE Int. Conf. Comp. Vis., pages 244–252, 2015.
  34. 34.D. Zoran and Y. Weiss. From learning models of natural image patches to whole image restoration. In Proc. IEEE Int. Conf. Comp. Vis., pages 479–486, 2011.

Citation

MLA
Mao, X.-J., et al. “Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections”. arXiv, 2016, http://arxiv.org/abs/1603.09056v2.
APA
Mao, X.-J., Shen, C., & Yang, Y.-B. (2016). Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections. arXiv. http://arxiv.org/abs/1603.09056v2
Chicago
Mao, X.-J., C. Shen, and Y.-B. Yang. 2016. “Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections”. arXiv. http://arxiv.org/abs/1603.09056v2.
Harvard
Mao, X.-J., Shen, C. and Yang, Y.-B. (2016) “Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1603.09056v2.
Vancouver
1. Mao X-J, Shen C, Yang Y-B (2016) Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections. arXiv

BibTeX

@article{mao2016image,
  title = {Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections},
  author = {Mao, Xiao-Jiao and Shen, Chunhua and Yang, Yu-Bin},
  year = {2016},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1603.09056v2},
  eprint = {1603.09056}
}
Metadata:arXiv

Access the Paper

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

Open PDF
License: Authors