arxiv: 2605.02471 · v1 · submitted 2026-05-04 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

Multispectral Blind Image Super-Resolution for Standing Dead Tree Segmentation

Mete Ahishali , Anis Ur Rahman , Einari Heinaro , Aysen Degerli , Samuli Junttila

Authors on Pith no claims yet

Pith reviewed 2026-05-08 18:24 UTC · model grok-4.3

classification 💻 cs.CV

keywords multispectral super-resolutionblind image super-resolutiondomain adaptationstanding dead tree segmentationaerial imageryunpaired domain adaptationforest monitoringimage restoration

0 comments

The pith

Blind super-resolution using unpaired domain adaptation enables standing dead tree segmentation in low-resolution multispectral aerial images

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

This paper establishes that a blind super-resolution method can restore low-resolution multispectral aerial images sufficiently well to support accurate segmentation of standing dead trees. The method relies on attention-guided networks to adapt between low- and high-resolution domains using only unpaired samples. This matters because high-resolution annotated data is scarce and expensive sensors limit widespread monitoring of forests affected by climate change. The framework also corrects for common issues like noise and low contrast in images from low-end sensors. It achieves Dice scores of 54 percent for segmentation when trained without any high-resolution labels and 64 percent when such labels are available.

Core claim

The paper claims to introduce the first real-world and generic super-resolution framework for multispectral data applied to standing dead tree segmentation. Using Attention-Guided Domain Adaptation Networks on unpaired samples, it learns the mapping from low-resolution to high-resolution images under realistic conditions where low-resolution data is not a downsampled version of high-resolution data. This results in segmentation performances of 54 percent Dice score without high-resolution annotations and 64 percent with them, while also serving as a general restorer for various image degradations.

What carries the argument

Attention-Guided Domain Adaptation Networks (ADA-Nets) that perform unpaired domain adaptation to map low-resolution multispectral images to high-resolution ones while addressing multiple degradation types.

If this is right

Super-resolved images support training of segmentation networks that generalize to real high-resolution data.
The method works on real-world unpaired data rather than synthetically degraded images.
It handles degradations such as saturation, noise, and low contrast in addition to low resolution.
The approach demonstrates the feasibility of using low-cost sensors for large-scale dead tree mapping.

Where Pith is reading between the lines

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

This could allow monitoring of forest changes over larger areas using more affordable equipment.
The technique might be adapted for other multispectral remote sensing applications where data quality varies.
Releasing the dataset publicly invites further development of methods for this task.

Load-bearing premise

Unpaired low- and high-resolution multispectral images share enough domain similarity for the adaptation network to learn a mapping that improves downstream segmentation on real high-resolution test data.

What would settle it

If the Dice score for segmentation on high-resolution test data does not exceed that obtained by using the original low-resolution images or standard synthetic super-resolution techniques, the effectiveness of the learned mapping would be called into question.

Figures

Figures reproduced from arXiv: 2605.02471 by Anis Ur Rahman, Aysen Degerli, Einari Heinaro, Mete Ahishali, Samuli Junttila.

**Figure 1.** Figure 1: Blind super-resolution framework with Attention-Guided Domain Adaptation Network (ADA-Net). The generator learns low-resolution (LR) to high view at source ↗

**Figure 2.** Figure 2: Several example image patches of low-resolution and high-resolution view at source ↗

**Figure 3.** Figure 3: False-color representations of extracted test image patches (a-e) with view at source ↗

**Figure 4.** Figure 4: Examples of full-scene low-resolution images and their super-resolved versions are given for both RGB (a–c) and false-color representations (d–f), view at source ↗

read the original abstract

Mapping standing dead trees is crucial for acquiring information on the effects of climate change on forests and forest biodiversity. However, leveraging high-quality aerial imagery for dead tree segmentation poses challenges due to limitations in sensor availability and the scarcity of annotated data. In this study, we propose a generic blind super-resolution framework that incorporates Attention-Guided Domain Adaptation Networks (ADA-Nets) to learn the mapping from low-resolution to high-resolution multispectral image domains. Our approach operates solely on unpaired samples, mimicking real-world conditions, i.e., low-resolution images are not synthetically obtained by downsampling the high-resolution images. Moreover, the proposed method serves as a general-purpose restorer addressing several image degradation types, including saturation, noise, and low contrast that typically occur in low-resolution images acquired by low-end sensors. To the best of our knowledge, this is the first study to perform real-world and generic super-resolution for multispectral data in the scope of standing dead tree segmentation. Experimental evaluations demonstrate segmentation performances of 54% and 64% in Dice scores. Notably, the first result is obtained without using any high-resolution annotations; the segmentation network is trained on super-resolved low-resolution images, while evaluation is performed on the high-resolution data. We publicly share the aerial multispectral dataset with manually annotated labels at https://www.kaggle.com/datasets/meteahishali/aerial-imagery-for-dead-tree-segmentation-poland.

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

Applies unpaired domain adaptation for blind multispectral SR to dead tree segmentation and releases a dataset, but provides no direct SR quality metrics to back the restoration claims.

read the letter

The paper takes attention-guided unpaired domain adaptation and uses it for blind super-resolution on real multispectral aerial images to help segment standing dead trees. They also release a new annotated dataset from Poland, which is the clearest concrete output here. Training the segmenter only on the super-resolved low-res images and testing on actual high-res data gets 54% Dice without any high-res labels for training, and 64% when high-res info is available. That setup matches real constraints where high-quality paired data and annotations are hard to get. The method targets actual degradations like saturation, noise, and low contrast without relying on synthetic downsampling, which fits low-end sensor scenarios in remote sensing. The novelty is in positioning this as the first generic real-world multispectral SR application for this specific task. The soft spots are in the missing validation for the super-resolution step itself. The abstract gives Dice numbers but no PSNR, SSIM, spectral angle, or vegetation index error on held-out real data, and no baselines or ablations are described. Without those, it is unclear whether the adaptation preserves multispectral statistics or simply produces images that happen to suit this segmenter. The stress-test point about possible task-specific hallucination rather than faithful restoration looks reasonable based on the reported evidence. This is for remote sensing and forest monitoring researchers who work with affordable multispectral imagery and need better tools for climate and biodiversity tracking. A reader in applied environmental computer vision would get value from the dataset and the practical experiment, even if the core claims need tightening. Send it for peer review. The public data and the real-world unpaired setup give it enough substance for referees to engage with, though revisions should add quantitative SR checks and comparisons.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a blind super-resolution framework for multispectral aerial images using Attention-Guided Domain Adaptation Networks (ADA-Nets) to improve standing dead tree segmentation. The method learns low-to-high resolution mappings from unpaired real-world samples and is presented as a generic restorer for degradations such as saturation, noise, and low contrast. It reports Dice scores of 54% (segmenter trained only on super-resolved LR images, no HR annotations used) and 64% on a publicly released dataset, claiming to be the first such real-world generic SR study in this application domain.

Significance. If the super-resolved outputs preserve spectral fidelity and the performance gains are robust, the work could meaningfully advance remote-sensing applications for forest monitoring under data scarcity, allowing low-end sensors to support climate-change studies. The public release of the annotated aerial multispectral dataset is a clear positive contribution to reproducibility.

major comments (2)

[Abstract and Experimental Evaluations] Abstract and Experimental Evaluations section: The headline 54% Dice result (segmenter trained exclusively on super-resolved LR images, tested on real HR data) is presented without any quantitative SR quality metrics (PSNR, SSIM, spectral-angle mapper, or vegetation-index error) on held-out real multispectral data. This directly undermines the claim that ADA-Nets produce a generic restorer rather than a task-specific mapping, as no evidence is given that multispectral statistics remain faithful.
[Experimental Evaluations] Experimental Evaluations section: No baselines, ablation studies, data-split details, or validation procedures are described for the reported Dice scores. Without these, it is impossible to isolate the contribution of the proposed ADA-Nets or to assess whether the 54%/64% figures are robust or sensitive to unstated choices.

minor comments (1)

[Abstract] Abstract: The two Dice scores are introduced without explicitly stating the precise experimental conditions (e.g., whether the 64% result uses HR annotations or a different training regime), which reduces clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below, indicating where revisions will be made to strengthen the presentation while respecting the constraints of our unpaired real-world data setting.

read point-by-point responses

Referee: [Abstract and Experimental Evaluations] Abstract and Experimental Evaluations section: The headline 54% Dice result (segmenter trained exclusively on super-resolved LR images, tested on real HR data) is presented without any quantitative SR quality metrics (PSNR, SSIM, spectral-angle mapper, or vegetation-index error) on held-out real multispectral data. This directly undermines the claim that ADA-Nets produce a generic restorer rather than a task-specific mapping, as no evidence is given that multispectral statistics remain faithful.

Authors: We acknowledge the referee's concern. Because the method is designed for blind super-resolution on unpaired real-world multispectral imagery (with no synthetic downsampling and thus no paired HR references), reference-based metrics such as PSNR, SSIM, or spectral-angle mapper cannot be computed on held-out real data. This limitation is inherent to the problem setting rather than an oversight. The primary evaluation is through the downstream segmentation task, which directly quantifies utility for dead-tree mapping under data scarcity. To better support the generic-restorer claim, we will add qualitative visualizations and vegetation-index preservation analysis in the revised Experimental Evaluations section, demonstrating that spectral characteristics are maintained across degradation types. revision: partial
Referee: [Experimental Evaluations] Experimental Evaluations section: No baselines, ablation studies, data-split details, or validation procedures are described for the reported Dice scores. Without these, it is impossible to isolate the contribution of the proposed ADA-Nets or to assess whether the 54%/64% figures are robust or sensitive to unstated choices.

Authors: We agree that these elements are necessary for reproducibility and for isolating the contribution of ADA-Nets. In the revised manuscript we will expand the Experimental Evaluations section to include: explicit data-split and cross-validation procedures, comparisons against relevant baseline blind SR methods, and ablation studies on the attention-guided domain adaptation components. These additions will clarify the robustness of the 54% and 64% Dice scores and allow readers to assess the specific impact of the proposed modules. revision: yes

Circularity Check

0 steps flagged

Empirical method with no self-referential derivations or load-bearing self-citations

full rationale

The paper describes an application of unpaired attention-guided domain adaptation (ADA-Nets) to perform blind multispectral super-resolution as a preprocessing step for dead-tree segmentation. All reported outcomes (54% Dice without HR annotations, 64% with) are obtained from direct experimental evaluation on a publicly released dataset. No equations, uniqueness theorems, or fitted parameters are presented that reduce by construction to the method's own inputs or prior self-citations. The unpaired real-world setting is explicitly stated as an operating assumption rather than derived, and the framework is positioned as a general-purpose restorer without renaming known results or smuggling ansatzes via self-reference. The derivation chain is therefore self-contained as an empirical pipeline.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the effectiveness of attention-guided domain adaptation for learning real-world degradations in multispectral imagery without paired data; no free parameters or invented entities are explicitly introduced beyond the network architecture itself.

axioms (1)

domain assumption Unpaired domain adaptation networks can learn a useful mapping from low-resolution to high-resolution multispectral domains that generalizes to downstream segmentation tasks.
Invoked in the description of ADA-Nets operating solely on unpaired samples to mimic real-world conditions.

pith-pipeline@v0.9.0 · 5568 in / 1350 out tokens · 55758 ms · 2026-05-08T18:24:33.639892+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith.Cost (J = ½(x+x⁻¹)−1, parameter-free) washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

hyperparameters are selected empirically: learning rate is 2×10^-6, λ=β=γ=ϑ=0.5, and τ=0.07 for the Ada-NET approach
IndisputableMonolith.Foundation.LogicAsFunctionalEquation derivedCost / J-uniqueness unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

L_ADA-Net = L_A + λ L_Spatial + β L_IDSpatial + γ L_Freq + ϑ L_IDFreq (adversarial + contrastive losses with tuned weights)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 4 canonical work pages · 1 internal anchor

[1]

Deep learning for multiple-image super-resolution,

M. Kawulok, P. Benecki, S. Piechaczek, K. Hrynczenko, D. Kostrzewa, and J. Nalepa, “Deep learning for multiple-image super-resolution,” IEEE Geoscience and Remote Sensing Letters, vol. 17, no. 6, pp. 1062– 1066, 2020

2020
[2]

A deep journey into super- resolution: A survey,

S. Anwar, S. Khan, and N. Barnes, “A deep journey into super- resolution: A survey,”ACM Computing Surveys (CSUR), vol. 53, no. 3, pp. 1–34, 2020

2020
[3]

Multi-image super-resolution for remote sensing using deep recurrent networks,

M. R. Arefin, V . Michalski, P.-L. St-Charles, A. Kalaitzis, S. Kim, S. E. Kahou, and Y . Bengio, “Multi-image super-resolution for remote sensing using deep recurrent networks,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 816–825

2020
[4]

Deep video super-resolution network using dynamic upsampling filters without explicit motion com- pensation,

Y . Jo, S. W. Oh, J. Kang, and S. J. Kim, “Deep video super-resolution network using dynamic upsampling filters without explicit motion com- pensation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 3224–3232

2018
[5]

Real-world single image super-resolution: A brief review,

H. Chen, X. He, L. Qing, Y . Wu, C. Ren, R. E. Sheriff, and C. Zhu, “Real-world single image super-resolution: A brief review,”Information Fusion, vol. 79, pp. 124–145, 2022

2022
[6]

Transformer for single image super-resolution,

Z. Lu, J. Li, H. Liu, C. Huang, L. Zhang, and T. Zeng, “Transformer for single image super-resolution,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops,, 2022, pp. 456–465

2022
[7]

Single-image super-resolution: A benchmark,

C.-Y . Yang, C. Ma, and M.-H. Yang, “Single-image super-resolution: A benchmark,” inEuropean Conference on Computer Vision (ECCV), 2014, pp. 372–386

2014
[8]

Frequency-assisted mamba for remote sensing image super-resolution,

Y . Xiao, Q. Yuan, K. Jiang, Y . Chen, Q. Zhang, and C.-W. Lin, “Frequency-assisted mamba for remote sensing image super-resolution,” IEEE Transactions on Multimedia, 2024

2024
[9]

Rethinking the upsampling layer in hyperspectral image super resolution,

H. Shi, F. Zhou, X. Sun, and J. Han, “Rethinking the upsampling layer in hyperspectral image super resolution,”IEEE Transactions on Multimedia, vol. 28, pp. 2824–2836, 2026

2026
[10]

Hyperspectral image super-resolution via boundary perception and topology inference,

H. Wang, C. Wang, and Y . Yuan, “Hyperspectral image super-resolution via boundary perception and topology inference,”IEEE Transactions on Multimedia, pp. 1–16, 2026

2026
[11]

Toward real-world single image super-resolution: A new benchmark and a new model,

J. Cai, H. Zeng, H. Yong, Z. Cao, and L. Zhang, “Toward real-world single image super-resolution: A new benchmark and a new model,” in IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 3086–3095

2019
[12]

An image is worth 16x16 words: Trans- formers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” inInternational Conference on Learning Representations (ICLR), 2021

2021
[13]

Individual tree detection and segmentation from unmanned aerial vehicle-lidar data based on a trunk point distribution indicator,

S. Deng, Q. Xu, Y . Yue, S. Jing, and Y . Wang, “Individual tree detection and segmentation from unmanned aerial vehicle-lidar data based on a trunk point distribution indicator,”Computers and Electronics in Agriculture, vol. 218, p. 108717, 2024

2024
[14]

Towards a global understanding of tree mortality,

I. T. M. Network, C. Senf, A. Esquivel-Muelbert, T. A. Pugh, W. R. Anderegg, K. J. Anderson-Teixeira, G. Arellano, M. Beloiu Schwenke, B. J. Bentz, H. J. Boehmeret al., “Towards a global understanding of tree mortality,”New Phytologist, vol. 245, no. 6, pp. 2377–2392, 2025

2025
[15]

Dual-task learning for dead tree detection and segmentation with hybrid self- attention u-nets in aerial imagery,

A. U. Rahman, E. Heinaro, M. Ahishali, and S. Junttila, “Dual-task learning for dead tree detection and segmentation with hybrid self- attention u-nets in aerial imagery,”International Journal of Applied Earth Observation and Geoinformation, vol. 144, p. 104851, 2025

2025
[16]

Deep learning-based automated forest health diagnosis from aerial images,

C.-Y . Chiang, C. Barnes, P. Angelov, and R. Jiang, “Deep learning-based automated forest health diagnosis from aerial images,”IEEE Access, vol. 8, pp. 144 064–144 076, 2020

2020
[17]

Instance segmentation of standing dead trees in dense forest from aerial imagery using deep learning,

A. Sani-Mohammed, W. Yao, and M. Heurich, “Instance segmentation of standing dead trees in dense forest from aerial imagery using deep learning,”ISPRS Open Journal of Photogrammetry and Remote Sensing, vol. 6, p. 100024, 2022

2022
[18]

Scattered tree death contributes to substantial forest loss in california,

Y . Cheng, S. Oehmcke, M. Brandt, L. Rosenthal, A. Das, A. Vrieling, S. Saatchi, F. Wagner, M. Mugabowindekwe, W. Verbruggenet al., “Scattered tree death contributes to substantial forest loss in california,” Nature Communications, vol. 15, no. 1, p. 641, 2024

2024
[19]

A machine learning approach to detect dead trees caused by longhorned borer in eucalyptus stands using uav imagery,

A. Duarte, N. Borralho, and M. Caetano, “A machine learning approach to detect dead trees caused by longhorned borer in eucalyptus stands using uav imagery,” inIEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2021, pp. 5818–5821

2021
[20]

Are transformers more robust than cnns?

Y . Bai, J. Mei, A. L. Yuille, and C. Xie, “Are transformers more robust than cnns?”Advances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 26 831–26 843, 2021

2021
[21]

Ada- net: Attention-guided domain adaptation network with contrastive learning for standing dead tree segmentation using aerial imagery,

M. Ahishali, A. U. Rahman, E. Heinaro, and S. Junttila, “Ada- net: Attention-guided domain adaptation network with contrastive learning for standing dead tree segmentation using aerial imagery,” arXiv:2504.04271, 2025

work page arXiv 2025
[22]

Image-to-image translation with conditional adversarial networks,

P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” inProceedings of the IEEE 10 Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 5967–5976

2017
[23]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

2016
[24]

Attention is all you need,

A. Vaswani, “Attention is all you need,”Advances in Neural Information Processing Systems (NIPS), 2017

2017
[25]

Saf-net: Self-attention fusion network for myocardial infarction detec- tion using multi-view echocardiography,

I. Adalioglu, M. Ahishali, A. Degerli, S. Kiranyaz, and M. Gabbouj, “Saf-net: Self-attention fusion network for myocardial infarction detec- tion using multi-view echocardiography,” inComputing in Cardiology (CinC), vol. 50, 2023, pp. 1–4

2023
[26]

Crossvit: Cross-attention multi- scale vision transformer for image classification,

C.-F. R. Chen, Q. Fan, and R. Panda, “Crossvit: Cross-attention multi- scale vision transformer for image classification,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 347–356

2021
[27]

Generative adversarial networks,

I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

2020
[28]

R2c-gan: Restore-to-classify generative adversarial networks for blind x-ray restoration and covid-19 classification,

M. Ahishali, A. Degerli, S. Kiranyaz, T. Hamid, R. Mazhar, and M. Gab- bouj, “R2c-gan: Restore-to-classify generative adversarial networks for blind x-ray restoration and covid-19 classification,”Pattern Recognition, vol. 156, p. 110765, 2024

2024
[29]

Contrastive learning for unpaired image-to-image translation,

T. Park, A. A. Efros, R. Zhang, and J.-Y . Zhu, “Contrastive learning for unpaired image-to-image translation,” inEuropean Conference on Computer Vision (ECCV), 2020, pp. 319–345

2020
[30]

Representation Learning with Contrastive Predictive Coding

A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv:1807.03748, 2018

work page Pith review arXiv 2018
[31]

Focal frequency loss for image reconstruction and synthesis,

L. Jiang, B. Dai, W. Wu, and C. C. Loy, “Focal frequency loss for image reconstruction and synthesis,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 13 899– 13 909

2021
[32]

Orthoimage, false colour 2008- 2020, all images, 1:10 000, etrs-tm35fin,

National Land Survey of Finland, “Orthoimage, false colour 2008- 2020, all images, 1:10 000, etrs-tm35fin,” http://urn.fi/urn:nbn:fi: csc-kata00001000000000000199, 2015, CSC – IT Center for Science

2008
[33]

Orthoimage, rgb or grayscale 2004-2020, all images, 1:10 000, etrs-tm35fin,

——, “Orthoimage, rgb or grayscale 2004-2020, all images, 1:10 000, etrs-tm35fin,” http://urn.fi/urn:nbn:fi:csc-kata20171228102116763542, 2017, CSC – IT Center for Science

2004
[34]

Orthophotomap (orto),

“Orthophotomap (orto),” https://www.geoportal.gov.pl/en/data/ orthophotomap-orto, 2025

2025
[35]

Adam: A Method for Stochastic Optimization

D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv:1412.6980, 2014

work page internal anchor Pith review arXiv 2014
[36]

Unpaired Image- to-Image Translation using Cycle-Consistent Adversarial Networks,

J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image- to-image translation using cycle-consistent adversarial networks,” arXiv:1703.10593, 2020

work page arXiv 2020
[37]

Detecting and simulating artifacts in gan fake images,

X. Zhang, S. Karaman, and S.-F. Chang, “Detecting and simulating artifacts in gan fake images,” inIEEE International Workshop on Information Forensics and Security (WIFS), 2019, pp. 1–6

2019
[38]

On hallucinating context and background pixels from a face mask using multi-scale gans,

S. Banerjee, W. Scheirer, K. Bowyer, and P. Flynn, “On hallucinating context and background pixels from a face mask using multi-scale gans,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2020, pp. 300–309

2020