Distilling with Residual Network for Single Image Super Resolution

Furui Bai; Rui Wang; Wen Lu; Xiaopeng Sun

arxiv: 1907.02843 · v1 · pith:UF5U6LVWnew · submitted 2019-07-05 · 💻 cs.CV · eess.IV

Distilling with Residual Network for Single Image Super Resolution

Xiaopeng Sun , Wen Lu , Rui Wang , Furui Bai This is my paper

Pith reviewed 2026-05-25 02:25 UTC · model grok-4.3

classification 💻 cs.CV eess.IV

keywords single image super-resolutionresidual networkdistilling blockconvolutional neural networkimage reconstructionmodel efficiencydeep learningfeature distillation

0 comments

The pith

A residual distilling network for single image super-resolution delivers better performance with smaller model size than prior approaches.

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

The paper presents the distilling with residual network (DRN) to handle single image super resolution without the bloat and training issues that arise from standard residual or dense feature extractors. It defines residual distilling blocks that split work between a residual branch and a separate distillation branch to pull useful information from low-resolution inputs. These blocks are stacked into residual distilling groups that add long skip connections to merge local and global features before reconstruction. Experiments on benchmark datasets show the resulting model exceeds state-of-the-art methods specifically in the trade-off between accuracy and parameter count.

Core claim

The authors claim that residual distilling blocks, each containing one residual operation branch and one distillation branch, can be grouped with long skip connections into residual distilling groups to extract and fuse effective local and global features from low-resolution images, yielding networks that outperform existing methods on standard benchmarks while maintaining a superior performance-to-model-size ratio.

What carries the argument

The residual distilling block (RDB), a module with two parallel branches where one executes residual operations and the other distills effective information from the low-resolution input.

If this is right

Separating residual and distillation paths in each block permits more targeted feature handling than monolithic residual or dense designs.
Stacking the blocks into groups with long skips produces networks that combine local detail extraction with global context fusion.
The resulting architecture improves reconstruction metrics on common test sets while constraining total model size.
This block-and-group pattern directly counters the parameter growth and optimization problems noted for naive residual and dense networks.

Where Pith is reading between the lines

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

The block design could transfer to related low-level tasks such as image denoising by reusing the same split-branch structure.
Lower parameter counts achieved this way may support faster inference on devices with limited memory or compute.
The distillation branch could be inspected post-training to determine which input features it preferentially retains.

Load-bearing premise

The residual distilling block can effectively distill information from low-resolution images without causing the overall network to bloat or become difficult to train.

What would settle it

If direct tests on standard benchmarks such as Set5 or DIV2K show that the DRN fails to exceed the PSNR or SSIM of comparable state-of-the-art models at equal or smaller parameter counts, the superiority claim would be disproved.

read the original abstract

Recently, the deep convolutional neural network (CNN) has made remarkable progress in single image super resolution(SISR). However, blindly using the residual structure and dense structure to extract features from LR images, can cause the network to be bloated and difficult to train. To address these problems, we propose a simple and efficient distilling with residual network(DRN) for SISR. In detail, we propose residual distilling block(RDB) containing two branches, while one branch performs a residual operation and the other branch distills effective information. To further improve efficiency, we design residual distilling group(RDG) by stacking some RDBs and one long skip connection, which can effectively extract local features and fuse them with global features. These efficient features beneficially contribute to image reconstruction. Experiments on benchmark datasets demonstrate that our DRN is superior to the state-of-the-art methods, specifically has a better trade-off between performance and model size.

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.

Desk Editor's Note private letter to a colleague

DRN adds a two-branch RDB and RDG structure to keep residual networks for SISR from bloating, but the superiority claim rests on unreported benchmark numbers.

read the letter

The paper's core idea is a residual distilling block that splits into a standard residual branch and a second branch meant to pull out only the useful information from the low-resolution input. These blocks are then stacked into residual distilling groups that include a long skip connection to combine local and global features. The stated goal is to avoid the training difficulties that come from blindly stacking residuals and dense connections while still delivering good reconstruction quality at smaller model sizes. That motivation lines up with the design choices, and the long skip is a familiar move that often helps gradient flow. The architecture description itself is clear and internally consistent. What is actually new is the specific pairing of the distilling branch inside the RDB and the way the RDG wraps multiple RDBs; it is a variation rather than a wholesale departure from prior residual work in SISR. The abstract asserts that experiments on standard benchmarks show a better performance-to-size trade-off than existing methods. The soft spot is that no concrete PSNR, SSIM, or parameter counts appear in the abstract, so the size of the claimed advantage cannot be judged from the summary alone. If the full paper supplies fair comparisons on Set5, Set14, BSD100 and the usual ablation tables, the claim becomes testable; without them the result stays provisional. There are no load-bearing inconsistencies or circular definitions visible. This is the kind of paper that matters to people already working on lightweight SISR models who want another block-level option to try. It is not a field-shifting result, but the design is straightforward enough that a referee could evaluate it against the usual baselines without unusual effort. I would send it to peer review rather than desk-reject.

Referee Report

0 major / 3 minor

Summary. The manuscript proposes a Distilling with Residual Network (DRN) for single image super-resolution (SISR). It introduces the Residual Distilling Block (RDB) consisting of a residual branch and a distilling branch, and the Residual Distilling Group (RDG) formed by stacking multiple RDBs with a long skip connection. The central claim is that this design avoids network bloat while achieving a superior performance versus model-size trade-off compared to prior state-of-the-art SISR methods on standard benchmark datasets.

Significance. If the reported empirical results hold, the DRN architecture would supply a practical, parameter-efficient alternative for SISR that directly addresses the motivation of preventing bloated residual/dense networks. The explicit focus on the performance-size Pareto front is a useful contribution for deployment scenarios with limited compute.

minor comments (3)

Abstract: the claim of superiority would be strengthened by naming the specific benchmark datasets, the quantitative metrics (PSNR/SSIM), and at least two representative baselines (e.g., EDSR, RDN) together with the reported deltas.
§3.2 (RDG definition): the long-skip connection is described only qualitatively; a diagram or explicit equation showing how the fused feature is added to the final reconstruction head would improve reproducibility.
Table 1 (model-size comparison): ensure that parameter counts and FLOPs are measured under identical input resolutions and that the reported model sizes exclude any post-training quantization.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary and recommendation of minor revision. The recognition that DRN supplies a parameter-efficient alternative addressing network bloat is appreciated. No specific major comments appear in the provided report.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper proposes a new CNN architecture (DRN) for SISR by defining RDB (residual + distilling branches) and RDG (stacked RDBs + long skip) to address bloat in residual/dense nets. The central claim is an empirical result: DRN shows better performance vs. model-size trade-off on standard benchmarks. No derivation chain, equations, or first-principles predictions exist that reduce to self-definition, fitted inputs renamed as predictions, or self-citation load-bearing steps. The design is explicitly motivated and the validation is external experimental data, making the result self-contained against benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The central claim depends on the effectiveness of these new blocks, which are introduced without prior independent validation beyond the paper's experiments.

free parameters (1)

design parameters of RDB and RDG
The number of blocks, channels, and other hyperparameters are chosen to achieve the trade-off.

axioms (1)

domain assumption Residual connections facilitate training of deep networks by mitigating vanishing gradients.
Standard assumption in residual network designs for SISR.

invented entities (2)

Residual Distilling Block (RDB) no independent evidence
purpose: To extract features using a residual branch and a distilling branch.
New component proposed to address issues with standard residual and dense structures.
Residual Distilling Group (RDG) no independent evidence
purpose: To stack RDBs with long skip connection for local and global feature fusion.
New grouping structure for efficient feature extraction.

pith-pipeline@v0.9.0 · 5691 in / 1427 out tokens · 34359 ms · 2026-05-25T02:25:46.322061+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 2 canonical work pages · 2 internal anchors

[1]

The tasks of super-resolution are qu ite extensive, such as in the ﬁeld of video surveillance, medi- cal imaging, and target detection

INTRODUCTION The task of super-resolution(SR) is to reconstruct a high- resolution(HR) image consistent with it from a low- resolution(LR) image. The tasks of super-resolution are qu ite extensive, such as in the ﬁeld of video surveillance, medi- cal imaging, and target detection. However, SR is a reverse process of information loss. LR images have abunda...
[2]

Network Architecture As shown in Fig.3, the proposed DRN mainly consists three parts: low-level feature extraction(LFE), residual disti lling groups(RDGs), image reconstruction(IR)

DISTILLING WITH RESIDUAL NETWORK 2.1. Network Architecture As shown in Fig.3, the proposed DRN mainly consists three parts: low-level feature extraction(LFE), residual disti lling groups(RDGs), image reconstruction(IR). Here, let’s deno te ILR and IHR as the input and output of DRN. As referred in [15, 7, 11], one convolutional layer is suitable to extrac...
[3]

EXPERIMENTAL RESUL TS 3.1. Implementation Details In the proposed networks, we set 3 ×3 as the size of all con- volutional layers with one padding and one striding except convolutional layers of local and global feature fusion. Th e ﬁlter size of local and global feature fusion is 1 ×1 with no padding and one striding. Low-level feature extraction lay ers...
[4]

Bas ed on resiudal distilling(RD), the DRN inherits the advantage s of the dense residue and connection paths, to achieve an ef- fective reuse and re-exploitation

CONCLUSION In this paper, we propose a simple and efﬁcient distilling wi th residual network(DRN) for SISR, which is better than most of the state-of-the-art methods and has fewer parameters. Bas ed on resiudal distilling(RD), the DRN inherits the advantage s of the dense residue and connection paths, to achieve an ef- fective reuse and re-exploitation. O...
[5]

Learning a deep convolutional network for image super-resolution,

Chao Dong, Chen Change Loy, Kaiming He, and Xi- aoou Tang, “Learning a deep convolutional network for image super-resolution,” in ECCV. Springer, 2014, pp. 184–199

2014
[6]

A+: Adjusted anchored neighborhood regression for fast super-resolution,

Radu Timofte, Vincent De Smet, and Luc V an Gool, “A+: Adjusted anchored neighborhood regression for fast super-resolution,” in ACCV. Springer, 2014, pp. 111–126

2014
[7]

Ac- curate image super-resolution using very deep convolu- tional networks,

Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee, “Ac- curate image super-resolution using very deep convolu- tional networks,” in CVPR, 2016, pp. 1646–1654

2016
[8]

Deeply-recursive convolutional network for image super-resolution,

Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee, “Deeply-recursive convolutional network for image super-resolution,” in CVPR, 2016, pp. 1637–1645

2016
[9]

Deep laplacian pyramid networks for fast and accurate super-resolution,

Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Y ang, “Deep laplacian pyramid networks for fast and accurate super-resolution,” in CVPR, 2017, pp. 5835–5843

2017
[10]

Deep residual learning for image recognition,

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

2016
[11]

Enhanced deep residual networks for single image super-resolution,

Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee, “Enhanced deep residual networks for single image super-resolution,” in CVPRW, 2017, pp. 1132–1140

2017
[12]

Memnet: A persistent memory network for image restoration,

Ying Tai, Jian Y ang, Xiaoming Liu, and Chunyan Xu, “Memnet: A persistent memory network for image restoration,” in CVPR, 2017, pp. 4539–4547

2017
[13]

Fast and accurate single image super-resolution via information distillation network,

Zheng Hui, Xiumei Wang, and Xinbo Gao, “Fast and accurate single image super-resolution via information distillation network,” in CVPR, 2018, pp. 723–731

2018
[14]

Residual dense network for image super- resolution,

Y ulun Zhang, Y apeng Tian, Y u Kong, Bineng Zhong, and Y un Fu, “Residual dense network for image super- resolution,” in CVPR, 2018

2018
[15]

Image Super-Resolution Using Very Deep Residual Channel Attention Networks

Y ulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bi- neng Zhong, and Y un Fu, “Image super-resolution using very deep residual channel attention networks,” arXiv preprint arXiv:1807.02758, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018
[16]

Learn- ing a single convolutional super-resolution network for multiple degradations,

Kai Zhang, Wangmeng Zuo, and Lei Zhang, “Learn- ing a single convolutional super-resolution network for multiple degradations,” in CVPR, 2018, vol. 6

2018
[17]

Multi-scale residual network for image super- resolution,

Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang, “Multi-scale residual network for image super- resolution,” in ECCV, 2018, pp. 517–532

2018
[18]

Real-time single image and video super-resolution using an efﬁcient sub-pixel convolu- tional neural network,

Wenzhe Shi, Jose Caballero, Ferenc Husz´ ar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang, “Real-time single image and video super-resolution using an efﬁcient sub-pixel convolu- tional neural network,”
[19]

Photo-realistic single image super- resolution using a generative adversarial network.,

Christian Ledig, Lucas Theis, Ferenc Husz´ ar, Jose Ca- ballero, Andrew Cunningham, Alejandro Acosta, An- drew P Aitken, Alykhan Tejani, Johannes Totz, Ze- han Wang, et al., “Photo-realistic single image super- resolution using a generative adversarial network.,” in CVPR, 2017, vol. 2, p. 4

2017
[20]

Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

Djork-Arn´ e Clevert, Thomas Unterthiner, and Sepp Hochreiter, “Fast and accurate deep network learn- ing by exponential linear units (elus),” arXiv preprint arXiv:1511.07289, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015
[21]

Dual path networks,

Y unpeng Chen, Jianan Li, Huaxin Xiao, Xiaojie Jin, Shuicheng Y an, and Jiashi Feng, “Dual path networks,” in NeurIPS, 2017, pp. 4467–4475

2017
[22]

Ac- celerating the super-resolution convolutional neural net- work,

Chao Dong, Chen Change Loy, and Xiaoou Tang, “Ac- celerating the super-resolution convolutional neural net- work,” in ECCV. Springer, 2016, pp. 391–407

2016
[23]

Image quality assessment: from error vis- ibility to structural similarity,

Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli, “Image quality assessment: from error vis- ibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

2004
[24]

A method for stochastic op- timization,

D Kinga and J Ba Adam, “A method for stochastic op- timization,” in ICLR, 2015, vol. 5

2015
[25]

Ntire 2017 challenge on single image super-resolution: Meth- ods and results,

Radu Timofte, Eirikur Agustsson, Luc V an Gool, Ming- Hsuan Y ang, Lei Zhang, Bee Lim, Sanghyun Son, Hee- won Kim, Seungjun Nah, Kyoung Mu Lee, et al., “Ntire 2017 challenge on single image super-resolution: Meth- ods and results,” in CVPRW. IEEE, 2017, pp. 1110– 1121

2017
[26]

Low-complexity single- image super-resolution based on nonnegative neighbor embedding,

Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel, “Low-complexity single- image super-resolution based on nonnegative neighbor embedding,” 2012

2012
[27]

On single image scale-up using sparse-representations,

Roman Zeyde, Michael Elad, and Matan Protter, “On single image scale-up using sparse-representations,” in International Conference on Curves and Surfaces . Springer, 2010, pp. 711–730

2010
[28]

Single image super-resolution from transformed self- exemplars,

Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja, “Single image super-resolution from transformed self- exemplars,” in CVPR, 2015, pp. 5197–5206

2015
[29]

A database of human segmented natural im- ages and its application to evaluating segmentation al- gorithms and measuring ecological statistics,

David Martin, Charless Fowlkes, Doron Tal, and Jiten- dra Malik, “A database of human segmented natural im- ages and its application to evaluating segmentation al- gorithms and measuring ecological statistics,” in ICCV. IEEE, 2001, vol. 2, pp. 416–423

2001
[30]

Sketch-based manga retrieval using manga109 dataset,

Y usuke Matsui, Kota Ito, Y uji Aramaki, Azuma Fu- jimoto, Toru Ogawa, Toshihiko Y amasaki, and Kiy- oharu Aizawa, “Sketch-based manga retrieval using manga109 dataset,” Multimedia T ools and Applications, vol. 76, no. 20, pp. 21811–21838, 2017

2017
[31]

Seven ways to improve example-based single image super resolution,

Radu Timofte, Rasmus Rothe, and Luc V an Gool, “Seven ways to improve example-based single image super resolution,” in CVPR, 2016, pp. 1865–1873. 0 20 40 60 80 100 epoch 31.0 31.5 32.0 32.5 33.0 33.5 34.0 34.5PSNR(dB) DRN with RDB DRN+ DRN without RDB

2016

[1] [1]

The tasks of super-resolution are qu ite extensive, such as in the ﬁeld of video surveillance, medi- cal imaging, and target detection

INTRODUCTION The task of super-resolution(SR) is to reconstruct a high- resolution(HR) image consistent with it from a low- resolution(LR) image. The tasks of super-resolution are qu ite extensive, such as in the ﬁeld of video surveillance, medi- cal imaging, and target detection. However, SR is a reverse process of information loss. LR images have abunda...

[2] [2]

Network Architecture As shown in Fig.3, the proposed DRN mainly consists three parts: low-level feature extraction(LFE), residual disti lling groups(RDGs), image reconstruction(IR)

DISTILLING WITH RESIDUAL NETWORK 2.1. Network Architecture As shown in Fig.3, the proposed DRN mainly consists three parts: low-level feature extraction(LFE), residual disti lling groups(RDGs), image reconstruction(IR). Here, let’s deno te ILR and IHR as the input and output of DRN. As referred in [15, 7, 11], one convolutional layer is suitable to extrac...

[3] [3]

EXPERIMENTAL RESUL TS 3.1. Implementation Details In the proposed networks, we set 3 ×3 as the size of all con- volutional layers with one padding and one striding except convolutional layers of local and global feature fusion. Th e ﬁlter size of local and global feature fusion is 1 ×1 with no padding and one striding. Low-level feature extraction lay ers...

[4] [4]

Bas ed on resiudal distilling(RD), the DRN inherits the advantage s of the dense residue and connection paths, to achieve an ef- fective reuse and re-exploitation

CONCLUSION In this paper, we propose a simple and efﬁcient distilling wi th residual network(DRN) for SISR, which is better than most of the state-of-the-art methods and has fewer parameters. Bas ed on resiudal distilling(RD), the DRN inherits the advantage s of the dense residue and connection paths, to achieve an ef- fective reuse and re-exploitation. O...

[5] [5]

Learning a deep convolutional network for image super-resolution,

Chao Dong, Chen Change Loy, Kaiming He, and Xi- aoou Tang, “Learning a deep convolutional network for image super-resolution,” in ECCV. Springer, 2014, pp. 184–199

2014

[6] [6]

A+: Adjusted anchored neighborhood regression for fast super-resolution,

Radu Timofte, Vincent De Smet, and Luc V an Gool, “A+: Adjusted anchored neighborhood regression for fast super-resolution,” in ACCV. Springer, 2014, pp. 111–126

2014

[7] [7]

Ac- curate image super-resolution using very deep convolu- tional networks,

Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee, “Ac- curate image super-resolution using very deep convolu- tional networks,” in CVPR, 2016, pp. 1646–1654

2016

[8] [8]

Deeply-recursive convolutional network for image super-resolution,

Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee, “Deeply-recursive convolutional network for image super-resolution,” in CVPR, 2016, pp. 1637–1645

2016

[9] [9]

Deep laplacian pyramid networks for fast and accurate super-resolution,

Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Y ang, “Deep laplacian pyramid networks for fast and accurate super-resolution,” in CVPR, 2017, pp. 5835–5843

2017

[10] [10]

Deep residual learning for image recognition,

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

2016

[11] [11]

Enhanced deep residual networks for single image super-resolution,

Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee, “Enhanced deep residual networks for single image super-resolution,” in CVPRW, 2017, pp. 1132–1140

2017

[12] [12]

Memnet: A persistent memory network for image restoration,

Ying Tai, Jian Y ang, Xiaoming Liu, and Chunyan Xu, “Memnet: A persistent memory network for image restoration,” in CVPR, 2017, pp. 4539–4547

2017

[13] [13]

Fast and accurate single image super-resolution via information distillation network,

Zheng Hui, Xiumei Wang, and Xinbo Gao, “Fast and accurate single image super-resolution via information distillation network,” in CVPR, 2018, pp. 723–731

2018

[14] [14]

Residual dense network for image super- resolution,

Y ulun Zhang, Y apeng Tian, Y u Kong, Bineng Zhong, and Y un Fu, “Residual dense network for image super- resolution,” in CVPR, 2018

2018

[15] [15]

Image Super-Resolution Using Very Deep Residual Channel Attention Networks

Y ulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bi- neng Zhong, and Y un Fu, “Image super-resolution using very deep residual channel attention networks,” arXiv preprint arXiv:1807.02758, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018

[16] [16]

Learn- ing a single convolutional super-resolution network for multiple degradations,

Kai Zhang, Wangmeng Zuo, and Lei Zhang, “Learn- ing a single convolutional super-resolution network for multiple degradations,” in CVPR, 2018, vol. 6

2018

[17] [17]

Multi-scale residual network for image super- resolution,

Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang, “Multi-scale residual network for image super- resolution,” in ECCV, 2018, pp. 517–532

2018

[18] [18]

Real-time single image and video super-resolution using an efﬁcient sub-pixel convolu- tional neural network,

Wenzhe Shi, Jose Caballero, Ferenc Husz´ ar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang, “Real-time single image and video super-resolution using an efﬁcient sub-pixel convolu- tional neural network,”

[19] [19]

Photo-realistic single image super- resolution using a generative adversarial network.,

Christian Ledig, Lucas Theis, Ferenc Husz´ ar, Jose Ca- ballero, Andrew Cunningham, Alejandro Acosta, An- drew P Aitken, Alykhan Tejani, Johannes Totz, Ze- han Wang, et al., “Photo-realistic single image super- resolution using a generative adversarial network.,” in CVPR, 2017, vol. 2, p. 4

2017

[20] [20]

Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

Djork-Arn´ e Clevert, Thomas Unterthiner, and Sepp Hochreiter, “Fast and accurate deep network learn- ing by exponential linear units (elus),” arXiv preprint arXiv:1511.07289, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015

[21] [21]

Dual path networks,

Y unpeng Chen, Jianan Li, Huaxin Xiao, Xiaojie Jin, Shuicheng Y an, and Jiashi Feng, “Dual path networks,” in NeurIPS, 2017, pp. 4467–4475

2017

[22] [22]

Ac- celerating the super-resolution convolutional neural net- work,

Chao Dong, Chen Change Loy, and Xiaoou Tang, “Ac- celerating the super-resolution convolutional neural net- work,” in ECCV. Springer, 2016, pp. 391–407

2016

[23] [23]

Image quality assessment: from error vis- ibility to structural similarity,

Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli, “Image quality assessment: from error vis- ibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

2004

[24] [24]

A method for stochastic op- timization,

D Kinga and J Ba Adam, “A method for stochastic op- timization,” in ICLR, 2015, vol. 5

2015

[25] [25]

Ntire 2017 challenge on single image super-resolution: Meth- ods and results,

Radu Timofte, Eirikur Agustsson, Luc V an Gool, Ming- Hsuan Y ang, Lei Zhang, Bee Lim, Sanghyun Son, Hee- won Kim, Seungjun Nah, Kyoung Mu Lee, et al., “Ntire 2017 challenge on single image super-resolution: Meth- ods and results,” in CVPRW. IEEE, 2017, pp. 1110– 1121

2017

[26] [26]

Low-complexity single- image super-resolution based on nonnegative neighbor embedding,

Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel, “Low-complexity single- image super-resolution based on nonnegative neighbor embedding,” 2012

2012

[27] [27]

On single image scale-up using sparse-representations,

Roman Zeyde, Michael Elad, and Matan Protter, “On single image scale-up using sparse-representations,” in International Conference on Curves and Surfaces . Springer, 2010, pp. 711–730

2010

[28] [28]

Single image super-resolution from transformed self- exemplars,

Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja, “Single image super-resolution from transformed self- exemplars,” in CVPR, 2015, pp. 5197–5206

2015

[29] [29]

A database of human segmented natural im- ages and its application to evaluating segmentation al- gorithms and measuring ecological statistics,

David Martin, Charless Fowlkes, Doron Tal, and Jiten- dra Malik, “A database of human segmented natural im- ages and its application to evaluating segmentation al- gorithms and measuring ecological statistics,” in ICCV. IEEE, 2001, vol. 2, pp. 416–423

2001

[30] [30]

Sketch-based manga retrieval using manga109 dataset,

Y usuke Matsui, Kota Ito, Y uji Aramaki, Azuma Fu- jimoto, Toru Ogawa, Toshihiko Y amasaki, and Kiy- oharu Aizawa, “Sketch-based manga retrieval using manga109 dataset,” Multimedia T ools and Applications, vol. 76, no. 20, pp. 21811–21838, 2017

2017

[31] [31]

Seven ways to improve example-based single image super resolution,

Radu Timofte, Rasmus Rothe, and Luc V an Gool, “Seven ways to improve example-based single image super resolution,” in CVPR, 2016, pp. 1865–1873. 0 20 40 60 80 100 epoch 31.0 31.5 32.0 32.5 33.0 33.5 34.0 34.5PSNR(dB) DRN with RDB DRN+ DRN without RDB

2016