pith. sign in

arxiv: 1907.04253 · v2 · pith:CEPO2DWGnew · submitted 2019-07-09 · 💻 cs.CV

Gated Multiple Feedback Network for Image Super-Resolution

Qilei Li , Zhen Li , Lu Lu , Gwanggil Jeon , Kai Liu , Xiaomin Yang This is my paper

Pith reviewed 2026-05-25 00:22 UTC · model grok-4.3

classification 💻 cs.CV
keywords image super-resolutionfeedback networkgated moduleresidual dense blocksrecurrent unfoldingdeep learningfeature refinement
0
0 comments X

The pith

A network reroutes multiple high-level features back through gated modules to refine low-level features for image super-resolution.

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

The paper proposes that existing super-resolution networks suffer because their information flows only forward, leaving high-level features underused for contextual refinement of low-level ones. By cascading residual dense blocks and unfolding them recurrently across time steps, multiple feedback connections carry high-level information back to earlier layers. A gated feedback module then selects and enhances the most useful parts of these rerouted features before they refine the low-level representation. If this mechanism works as described, the resulting network produces higher-quality upscaled images on standard benchmarks than prior feedforward designs. The authors support the claim through quantitative metrics and visual comparisons showing reduced artifacts and better detail preservation.

Core claim

Cascading multiple residual dense blocks and recurrently unfolding them across time steps, combined with multiple feedback connections between adjacent time steps and a gated feedback module, allows high-level features captured under large receptive fields to efficiently enrich low-level features that lack sufficient contextual information, leading to more accurate single-image super-resolution.

What carries the argument

The gated feedback module that selects and further enhances useful information from multiple rerouted high-level features before refining the low-level features.

If this is right

  • The network outperforms state-of-the-art super-resolution methods on standard benchmarks in both quantitative metrics and visual quality.
  • High-level features from large receptive fields can be reused to supply missing context to low-level features that would otherwise remain under-informed.
  • Gated selection prevents irrelevant or noisy high-level information from degrading the low-level representation during refinement.
  • Recurrent unfolding of the blocks enables the feedback paths to operate without requiring an entirely new architecture for each time step.

Where Pith is reading between the lines

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

  • The same feedback-and-gating pattern could be tested on related tasks such as image denoising or deblurring where contextual enrichment is also needed.
  • If the recurrent structure proves stable, it might allow variable numbers of unfoldings at inference time to trade compute for quality on different images.
  • The gated module's selection logic could be inspected post-training to determine which high-level features are most frequently chosen for different image regions.

Load-bearing premise

The multiple feedback connections and gated selection will consistently enrich low-level features with useful high-level context without introducing training instability or artifacts during recurrent unfolding of the residual dense blocks.

What would settle it

Training the network without the feedback connections or gated module on the same datasets and observing no drop in PSNR, SSIM, or visual quality compared to the full model.

Figures

Figures reproduced from arXiv: 1907.04253 by Gwanggil Jeon, Kai Liu, Lu Lu, Qilei Li, Xiaomin Yang, Zhen Li.

Figure 1
Figure 1. Figure 1: Qualitative results for ×4 image SR on ‘img_092’ from Urban100 dataset. The proposed GMFN accurately recovers more image details compared with other state-of-the￾art image SR methods. image can be degraded from multiple HR images. In recent years, the development of deep learning (DL) based high-level vision (skip connections [8, 10] and attention mechanism [9]) helps networks for image SR become much deep… view at source ↗
Figure 2
Figure 2. Figure 2: The framework of our proposed gated multiple feedback network (GMFN). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Study of multiple feedback connections. Single-to-multiple (SM) feedback manner [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of averaged feature maps. move all GFMs or all refinement units in the GFMs, the communication between the two time step would be disconnected. Thus, we only investigate the necessity of the gate unit in the GFM. For N = 4 and M = 1, equipped with the gate unit, our model achieves a PSNR value of 26.13. After removing the gate unit, multiple high-level features are directly concatenated with … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of our GMFN with other methods on [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Feedback networks (FB) vs. feed￾forward networks (FF) 表格 1 Dataset T=1 T=2 T=3 T=4 Urban100 26.00 26.13 26.07 26.14 表格 1-1 S->M 26.03 PSNR 25.90 25.98 26.05 26.13 26.20 T=1 T=2 T=3 26.00 26.13 26.07 26.14 1 T=4 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: (a) Accuracy and numbers of parameters trade-off. (b) Accuracy and average [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative results on ‘img_044’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative results on ‘img_062’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative results on ‘ParaisoRoad’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results on ‘ToutaMairimasu’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative results on ‘UchiNoNyansDiary’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative results on ‘253027’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative results on ‘210088’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p016_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative results on ‘butterfly’ with scale factor [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
read the original abstract

The rapid development of deep learning (DL) has driven single image super-resolution (SR) into a new era. However, in most existing DL based image SR networks, the information flows are solely feedforward, and the high-level features cannot be fully explored. In this paper, we propose the gated multiple feedback network (GMFN) for accurate image SR, in which the representation of low-level features are efficiently enriched by rerouting multiple high-level features. We cascade multiple residual dense blocks (RDBs) and recurrently unfolds them across time. The multiple feedback connections between two adjacent time steps in the proposed GMFN exploits multiple high-level features captured under large receptive fields to refine the low-level features lacking enough contextual information. The elaborately designed gated feedback module (GFM) efficiently selects and further enhances useful information from multiple rerouted high-level features, and then refine the low-level features with the enhanced high-level information. Extensive experiments demonstrate the superiority of our proposed GMFN against state-of-the-art SR methods in terms of both quantitative metrics and visual quality. Code is available at https://github.com/liqilei/GMFN.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript proposes the Gated Multiple Feedback Network (GMFN) for single-image super-resolution. It cascades multiple residual dense blocks (RDBs) and recurrently unfolds them, using multiple feedback connections between adjacent time steps together with a gated feedback module (GFM) that selects and enhances high-level features to refine low-level representations. The central empirical claim is that GMFN outperforms prior state-of-the-art SR methods on standard quantitative metrics and visual quality, with code released at the cited GitHub repository.

Significance. If the reported gains are reproducible, the work shows that gated multi-scale feedback can usefully enrich low-level features with high-level context in recurrent SR architectures. The public code release is a clear strength that supports direct verification and extension.

minor comments (3)
  1. [§3.2] §3.2 and Fig. 3: the gating equations inside the GFM are described only in prose; adding an explicit equation for the selection weights would improve precision and ease of re-implementation.
  2. [Table 2] Table 2: the PSNR/SSIM entries for competing methods on Urban100 and Manga109 are given without standard deviations or number of runs; reporting variability would strengthen the superiority claim.
  3. [§4.3] §4.3: the ablation study removes the GFM but does not isolate the effect of the number of feedback connections; a more granular ablation would clarify which component drives the reported gains.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive evaluation of our Gated Multiple Feedback Network (GMFN) and for recommending minor revision. We appreciate the recognition of the gated multi-scale feedback mechanism and the value placed on the public code release.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper proposes the GMFN architecture for single-image super-resolution by cascading and recurrently unfolding residual dense blocks with gated feedback modules. Its central claim is empirical superiority on quantitative metrics and visual quality versus prior SR methods, demonstrated via experiments on held-out test sets with released code. No derivation chain, mathematical prediction, or first-principles result is presented that reduces to its own inputs by construction; the design choices are architectural and validated externally rather than self-referential. No self-citation load-bearing steps, fitted parameters renamed as predictions, or ansatz smuggling appear in the performance claims.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on empirical validation of a new DL architecture rather than mathematical axioms or invented physical entities. No free parameters beyond standard network weights are introduced in the abstract.

pith-pipeline@v0.9.0 · 5741 in / 912 out tokens · 21565 ms · 2026-05-25T00:22:04.533765+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    Low-complexity single-image super-resolution based on nonnegative neighbor embed- ding

    Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embed- ding. In BMVC, 2012

    work page 2012

  2. [2]

    Human pose estimation with iterative error feedback

    Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In CVPR, 2016

    work page 2016

  3. [3]

    Learning a deep con- volutional network for image super-resolution

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep con- volutional network for image super-resolution. In ECCV, 2014

    work page 2014

  4. [4]

    Accelerating the super-resolution convolutional neural network

    Chao Dong, Chen Change Loy, and Xiaoou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016

    work page 2016

  5. [5]

    Image super-resolution via dual-state recurrent networks

    Wei Han, Shiyu Chang, Ding Liu, Mo Yu, Michael Witbrock, and Thomas S Huang. Image super-resolution via dual-state recurrent networks. In CVPR, 2018

    work page 2018

  6. [6]

    Deep back- projection networks for super-resolution

    Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back- projection networks for super-resolution. In CVPR, 2018

    work page 2018

  7. [7]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015

    work page 2015

  8. [8]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016

    work page 2016

  9. [9]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018

    work page 2018

  10. [10]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017

    work page 2017

  11. [11]

    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

    work page 2015

  12. [12]

    Multi-path feedback recurrent neural networks for scene parsing

    Xiaojie Jin, Yunpeng Chen, Zequn Jie, Jiashi Feng, and Shuicheng Yan. Multi-path feedback recurrent neural networks for scene parsing. In AAAI, 2017

    work page 2017

  13. [13]

    Accurate image super-resolution using very deep convolutional networks

    Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016

    work page 2016

  14. [14]

    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

    work page 2016

  15. [15]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014

    work page 2014

  16. [16]

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

    Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017

    work page 2017

  17. [17]

    Feed- back network for image super-resolution

    Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feed- back network for image super-resolution. In CVPR, 2019. 12 LI, LI, LU, JEON, LIU, Y ANG: GMFN FOR IMAGE SUPER-RESOLUTION

    work page 2019

  18. [18]

    Convolutional neural networks with intra- layer recurrent connections for scene labeling

    Ming Liang, Xiaolin Hu, and Bo Zhang. Convolutional neural networks with intra- layer recurrent connections for scene labeling. In NeurIPS, 2015

    work page 2015

  19. [19]

    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

    work page 2017

  20. [20]

    Non-local recurrent network for image restoration

    Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. In NeurIPS, 2018

    work page 2018

  21. [21]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics

    David Martin, Charless Fowlkes, Doron Tal, Jitendra Malik, et al. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001

    work page 2001

  22. [22]

    Sketch-based manga retrieval using manga109 dataset

    Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Ya- masaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 2017

    work page 2017

  23. [23]

    Recurrent convolutional neural networks for scene parsing

    Ronan Collobert Pedro HO Pinherio and H Pedro. Recurrent convolutional neural networks for scene parsing. In ICML, 2014

    work page 2014

  24. [24]

    Top-down feedback for crowd counting convolutional neural network

    Deepak Babu Sam and R Venkatesh Babu. Top-down feedback for crowd counting convolutional neural network. In AAAI, 2018

    work page 2018

  25. [25]

    Image super-resolution via deep recursive residual network

    Ying Tai, Jian Yang, and Xiaoming Liu. Image super-resolution via deep recursive residual network. In CVPR, 2017

    work page 2017

  26. [26]

    Image super-resolution using dense skip connections

    Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In ICCV, 2017

    work page 2017

  27. [27]

    Esrgan: Enhanced super-resolution generative adversarial networks

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In ECCV, 2018

    work page 2018

  28. [28]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, Eero P Simoncelli, et al. Image quality assessment: from error visibility to structural similarity. TIP, 2004

    work page 2004

  29. [29]

    Feedback networks

    Amir R Zamir, Te-Lin Wu, Lin Sun, William B Shen, Bertram E Shi, Jitendra Malik, and Silvio Savarese. Feedback networks. In CVPR, 2017

    work page 2017

  30. [30]

    On single image scale-up using sparse-representations

    Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces, 2010

    work page 2010

  31. [31]

    Progressive attention guided recurrent network for salient object detection

    Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, 2018

    work page 2018

  32. [32]

    Image super-resolution using very deep residual channel attention networks

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018

    work page 2018

  33. [33]

    Residual dense network for image super-resolution

    Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. LI, LI, LU, JEON, LIU, Y ANG: GMFN FOR IMAGE SUPER-RESOLUTION 13 Supplementary Material The following items are contained in the supplementary material: A. Feedback networks vs. feedforward networks. B. Study of time step. C. Mode...

    work page 2018