Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Deep learning image burst stacking to reconstruct high-resolution ground-based solar observations

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A deep-learning model reconstructs bursts of 100 short-exposure solar images into a single high-quality frame in about half a second, matching speckle reconstructions and staying cleaner when speckle artifacts appear.

desk verdict Useful engineering paper with a circular evaluation; the fast speckle mimicry is real, but the 'high-resolution' claim needs an independent ground-truth test. read the letter →

arxiv 2506.04781 v1 pith:GTORFZSW submitted 2025-06-05 astro-ph.SR astro-ph.IMcs.CVphysics.comp-ph

Christoph Schirninger , Robert Jarolim , Astrid M. Veronig , Christoph Kuckein This is my paper · ORCID
classification astro-ph.SRastro-ph.IMcs.CVphysics.comp-ph
keywords imagereconstructionsolarobservationsburststackinggenerativeadversarialnetworksunpairedimage-to-imagetranslationspecklereal-timeprocessingGREGORtelescope
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Ground-based solar telescopes see through turbulent air, so observers capture bursts of short exposures and computationally fuse them into one sharp image; standard fusion methods are slow and can fail under bad seeing. This paper argues that an unpaired image-to-image translation network, applied to the entire 100-frame burst at once, reconstructs a single high-resolution image in real time (about 0.5 seconds on an A100 GPU) with quality comparable to speckle reconstruction. On an independent test set, the network beats the single-frame baseline on PSNR, SSIM, MAE, and FID for both G-band and Blue continuum data, and it produces perceptually cleaner outputs when the speckle reference itself contains artifacts. The authors read the result as evidence that deep burst stacking can serve as a fast, robust alternative to classical reconstruction, with reconstruction quality improving as more frames from the burst are used.

What carries the argument

The engine is unpaired image-to-image translation in the ITI framework: two U-Net generators (A-to-B and B-to-A) together with multi-scale discriminators, trained in three cycles that include degrading a speckle image into a synthetic burst, restoring it, imposing cycle consistency, and separating burst noise with a dedicated noise-estimator network. The central object is generator AB, which ingests a 100-frame burst as input channels and outputs a single high-resolution image, while a random multi-scale discriminator keeps the generator from memorizing frame order. The B-to-A generator and the cycle-consistency losses are what allow training without paired burst-to-speckle examples.

What would settle it

On an overlapping observation where both a speckle reconstruction and a diffraction-limited space-based image exist, compare the ITI output against the space-based image: if ITI scores well against speckle but misses small-scale structure present in the space-based image, or reproduces a known speckle artifact, the high-resolution claim reduces to imitation of the reference domain.

Watch

Extended reading notes

Core claim

The paper's central claim is that instrument-to-instrument translation (ITI), an unpaired GAN-based image-to-image method, can map a burst of 100 short-exposure, seeing-degraded solar images directly to a single speckle-like high-resolution image. Trained with speckle reconstructions as the target domain, the model learns a domain mapping rather than fitting individual pairs, which lets it exploit bursts that were never speckle-reconstructed (1078 versus 745 G-band bursts, 960 versus 392 Blue continuum) and gives it robustness when reference reconstructions show artifacts. Quantitatively, ITI improves PSNR from 16.23 to 20.30 (G-band, 100 frames) and from 14.40 to 18.20 (Blue continuum), raises SSIM from 0.36 to 0.66 and 0.33 to 0.65, lowers MAE, and drops FID from 37.69 to 6.99 and from 49.72 to 5.22. On a dedicated failed-speckle test set, ITI reaches FID 13.99 (G-band) and 11.42 (Blue continuum), lower than both the raw-burst baseline and the artifact-laden speckle reconstructions. Reconstruction runs at 0.48 seconds per burst on an NVIDIA A100, and quality rises with the number of input frames, so the paper concludes that the full 100-frame burst carries information the model genuinely learns to combine.

Load-bearing premise

The whole method is trained and evaluated against speckle reconstructions as the reference high-quality domain; if those references are biased, miss real small-scale structure, or contain their own artifacts, the network learns to imitate them and the metrics reward looking like speckle rather than being physically correct.

Editorial extensions

If this is right

  • Real-time reconstruction at roughly half a second per burst makes automatic processing pipelines and on-site quality checks feasible at large-aperture solar facilities.
  • Because reconstruction quality improves monotonically with input frame count and is highest for the full 100-frame burst, feeding the model all 500 originally recorded frames instead of only the best 100 could improve quality further without changing the method.
  • The unpaired setup grows the usable training data by about 44 percent in G-band and 145 percent in Blue continuum compared with paired training, easing the data scarcity that limits burst-based deep learning.
  • The robustness on failed-speckle cases suggests the method can rescue observations that classical reconstruction discards, enlarging the pool of scientifically usable data under poor seeing.
  • The architecture transfers to other wavelength channels of the same instrument once a suitable high-quality reference domain is available, as only the input channel count changes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported 0.48 seconds per burst is the GPU inference pass; an end-to-end pipeline including normalization, file I/O, and image handling will take longer and depend on hardware, so 'real time' should be read as the network pass only.
  • Because the target domain is speckle reconstruction, the method's physical fidelity inherits systematic errors of speckle: features speckle preserves will be sharp, but structure speckle loses will not be recovered. A stronger check would compare ITI outputs against diffraction-limited images from a space-based solar telescope on overlapping observations.
  • The monotone frame-count scaling suggests a testable extension: train with bursts of 200 to 500 frames and check whether the quality curve keeps rising or saturates, and whether the model's stacking behavior resembles averaging in Fourier space.
  • FID rewards global distribution similarity, so the low FID values on artifact cases show perceptual closeness to the high-quality domain, not per-pixel fidelity; pixel-level agreement on those cases would need a separate metric.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces ITI, a deep learning image burst stacking method for ground-based solar observations, which translates a burst of 100 short-exposure frames into a single high-resolution image using unpaired image-to-image translation. The model is trained with seeing-degraded bursts as input and speckle reconstructions as the target domain, and it is evaluated on an independent test set with PSNR, SSIM, MAE, and FID. The authors report real-time reconstruction (0.48 s per burst on an A100 GPU), improved metrics over a single-frame baseline, and claim increased robustness when speckle reconstructions show artifacts. A frame-ablation study shows that reconstruction quality improves with the number of input frames, supporting the claim that the model makes use of the full burst information.

Significance. If the central claim were independently established, the method would be practically significant: real-time burst reconstruction could enable automatic processing pipelines for large-aperture solar telescopes and would complement existing speckle and MOMFBD approaches. Strengths of the work include the public code repository, the use of full-resolution test images, the frame-ablation analysis, and the acknowledgment of model limitations in Section 4.6 and the Discussion. However, the quantitative evaluation is circular with respect to the training target: all metrics measure agreement with the same speckle-reconstruction domain that defines the target distribution. Consequently, the paper does not yet establish that ITI recovers true solar structure rather than merely imitating speckle-style outputs. This limits the significance of the reported quality improvements and the 'high-resolution' claim in the abstract and title.

major comments (4)
  1. [Sec. 2 and 4.1, Tables 3-5] The central claim of high-resolution reconstruction is not supported by the quantitative evaluation because the training target and the evaluation reference are the same speckle-reconstruction domain. The model is trained to map bursts into domain B (speckle reconstructions, Section 2), and the reported PSNR, SSIM, MAE, and FID values all compare ITI outputs against those same speckle reconstructions. The baseline is a single raw frame, so improvements over that baseline are expected for a network explicitly trained to produce speckle-like outputs. The artifact-robustness result in Table 5 is likewise computed with FID relative to high-quality speckle images, leaving the evaluation within the same circular framework. To substantiate the 'high-resolution' claim, the authors need an independent reference: for example, synthetic bursts with a known object, space-based observations of the same solar features, or a power-spectral-density analysis relative to an independent diffraction-limited estimate. The paper acknowledges this constraint in Section 4.6 and the Discussion, but the abstract and title nevertheless assert high-resolution reconstruction without the required independent verification.
  2. [Tables 4 and 5] The FID values are reported as single numbers without uncertainty estimates, even though the test sets are small (100 bursts in the main test set, and 150/107 bursts in the artifact test set). FID is known to have high variance on small samples, and the differences between ITI and the baselines (e.g., 6.99 vs. 37.69 in the G-band) may still be subject to considerable sampling noise. The authors should provide confidence intervals (e.g., bootstrap over the test set) or multiple evaluation runs to demonstrate that the reported perceptual-quality improvements are robust.
  3. [Sec. 4.5] The 'failed speckle reconstruction' test set is described as visually selected, which introduces a potential selection bias into the artifact-robustness claim. Because the selection criterion is the presence of artifacts in the speckle reconstruction, the FID comparison against the high-quality speckle domain in Table 5 may partly reflect the way the test set was chosen. Please specify the visual selection protocol (e.g., number of independent observers, explicit artifact criteria) and consider supplementing the FID evaluation with a quantitative artifact measure that does not depend on the same reference domain, such as local contrast or edge-preservation statistics.
  4. [Conclusion and Sec. 6] The conclusion states that ITI 'outperforms state-of-the-art methods' and can be considered an alternative to speckle reconstruction and MOMFBD. This claim is not supported by the experiments, which only compare ITI against a single-frame baseline in Tables 3 and 4 and against low-quality speckle in Table 5. No direct quantitative comparison is made between ITI outputs and speckle reconstructions (or MOMFBD) on the same test bursts using a common reference metric. If the authors wish to claim superiority over existing methods, they need such a direct comparison, or they should temper the wording to 'comparable to' rather than 'outperforms.'
minor comments (6)
  1. [Sec. 3] In the description of the loss functions, 'MEA loss' appears to be a typo for 'MAE loss' (mean absolute error). Please correct this.
  2. [Sec. 2] The normalization description states that speckle reconstructions use 'quantile normalization between 0.001 of the minimum value and 0.999 of the maximum value,' which is ambiguous; it should read 'between the 0.001 quantile and the 0.999 quantile.'
  3. [References] The reference for Jarolim et al. (2024) is cited as 'Nat. Commun. (in press)' with an arXiv number; please update to the published version and volume/pages if available.
  4. [Fig. 2 and Sec. 4.1] The histogram comparison in Fig. 2 is anecdotal; a quantitative distributional distance (e.g., Kolmogorov-Smirnov statistic) on the pixel-value distributions would strengthen the claim derived from that figure.
  5. [Sec. 4.1] The text mentions evaluation with MSE in the Discussion, but Table 3 reports MAE; please make the metric list consistent throughout the paper.
  6. [Sec. 1] There are occasional spelling inconsistencies, e.g., 'Fréchet' with and without accents, and 'von der Luehe' vs. 'von der Lühe.' Please unify these.

Circularity Check

3 steps flagged · score 6.0 of 10

All quantitative quality metrics compare ITI outputs to the speckle reconstructions used as the training target; the 'high-resolution' claim therefore needs an independent ground-truth test to avoid circularity.

  1. self definitional [Sect. 1 (domain definition)]
    "The high-quality domain refers to the speckle reconstruction and serves as a reference for our reconstruction, while the low-quality domain (the level 1 image bursts) serves as the low-quality input."

    This defines the training target 'high quality' as the speckle-reconstruction domain. Therefore any later claim to produce high-quality/high-resolution images is a claim to produce images in this domain. The definition makes the evaluation against speckle a check of self-consistency rather than independent physical fidelity.

  2. fitted input called prediction [Sect. 4.1 (quantitative evaluation)]
    "To compare the performance of ITI with this baseline, we also calculated the difference between the ITI reconstruction and the speckle reconstruction and then derived the performance metrics."

    The quantitative 'prediction' (PSNR/SSIM/MAE improvement over a raw frame) is computed against the same speckle reconstructions used as the training reference. Since the network is trained to translate bursts into the speckle domain, these metrics are forced to improve as the model learns its training target; they do not independently confirm that true solar structure is recovered.

1 more flagged steps
  1. fitted input called prediction [Sect. 4.5 (failed speckle reconstruction, Table 5)]
    "We evaluated the reconstruction performance on the high-quality test set (see Sect. 2). We made a comparison between the baseline, the speckle reconstruction showing artifacts, and ITI. Since pixel-based metrics cannot be used for the comparison, as the speckle reconstructions show artifacts, we only used the perceptual quality metric with the FID."

    The artifact-robustness claim is also scored by FID against the high-quality speckle test set, which is drawn from the same speckle-reconstruction domain used for training. A lower FID for ITI means the output distribution is closer to the training-reference distribution; it is an imitation score, not an independent measure of reconstruction quality.

full rationale

The circularity is confined to the validation logic, not the engineering. The real-time speed (0.48 s on an A100) and the frame-ablation trend are independent, because they do not require a particular reference to show that more input frames improve the learned mapping. However, every metric used to support 'high-resolution reconstruction' — PSNR, SSIM, MAE and FID in Tables 3–5 — is computed against speckle reconstructions, which are the same domain used as the training reference (Sect. 1 and 2). The model is trained adversarially and with cycle/content losses to produce images in that domain, so high scores relative to a single raw frame largely express successful imitation of the training target. The paper itself states the limitation: 'We can only compare the method with state-of-the-art reconstruction techniques' (Sect. 4.6) and 'the method is constrained by the quality of the reference images' (Conclusions). The ITI self-citation (Jarolim et al. 2024) is not load-bearing in a circular way, because the code is public and the framework is reused, not re-validated by citation alone. Since no independent ground truth (synthetic scenes, space-based data, or point-spread-function-based physical metrics) is provided, the central claim of physical high-resolution reconstruction is not established beyond resemblance to the training reference.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical; it rests on the validity of the speckle reference, the static-object burst assumption, and the unpaired translation method, all standard domain assumptions but not independently verified here. Hand-tuned loss weights are the main ad hoc choices.

free parameters (3)
  • lambda_content = 10
    Hand-tuned weight for content loss; increases perceptual quality but is a manual choice.
  • lambda_id = 0.1
    Hand-tuned weight for identity loss, set to 0.1 per ITI defaults.
  • speckle normalization quantiles = 0.001 to 0.999
    Choices for quantile normalization, affecting contrast and training behavior.
assumptions (4)
  • domain assumption The 100 stored frames are the best of 500 and the burst can be treated as observing a static object (Eq. 1 with a single object O(x)).
    Assumed in Section 2 and Eq. 1; if solar features evolve faster than the burst, each frame sees a different object.
  • domain assumption Speckle reconstructions from the Wöger et al. pipeline are a valid high-resolution reference for the target domain.
    Used throughout as training reference and evaluation reference; the paper itself notes they can contain artifacts (Sect. 4.5).
  • domain assumption Unpaired image-to-image translation can transfer between the burst domain and the speckle domain without paired alignment.
    Core methodological assumption from ITI (Jarolim et al. 2024); necessary for the training to work.
  • domain assumption The convolutional architecture generalizes from 512x512 patches to full-resolution images.
    Training uses random patches; evaluation on full-resolution images assumes translation invariance and similar patch statistics (Sect. 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep learning image burst stacking to reconstruct high-resolution ground-based solar observations." pith.science (2026). https://pith.science/paper/GTORFZSW

@misc{pith2026250604781,
  author       = {Pith},
  title        = {Pith review of: Deep learning image burst stacking to reconstruct high-resolution ground-based solar observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GTORFZSW}},
  note         = {Machine review of arXiv:2506.04781}
}
read the original abstract

Large aperture ground based solar telescopes allow the solar atmosphere to be resolved in unprecedented detail. However, observations are limited by Earths turbulent atmosphere, requiring post image corrections. Current reconstruction methods using short exposure bursts face challenges with strong turbulence and high computational costs. We introduce a deep learning approach that reconstructs 100 short exposure images into one high quality image in real time. Using unpaired image to image translation, our model is trained on degraded bursts with speckle reconstructions as references, improving robustness and generalization. Our method shows an improved robustness in terms of perceptual quality, especially when speckle reconstructions show artifacts. An evaluation with a varying number of images per burst demonstrates that our method makes efficient use of the combined image information and achieves the best reconstructions when provided with the full image burst.

Figures

Figures reproduced from arXiv: 2506.04781 by the authors.

Figure 1
Figure 1. This training approach encompasses two types of map [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the image distribution between a single frame of the original burst and the speckle reconstruction (left) and our ITI re￾construction with the speckle reconstruction (right) from an observation on 11 April 2016. To evaluate the model’s performance on observations with lower-quality speckle reconstructions, we created a separate test set containing the original burst with its associated speckle re￾const… view at source ↗
Figure 3
Figure 3. Results of our ITI reconstructions for different solar features. Left: Single frame of original burst in the G-Band (430.7 nm) channel. Right: Corresponding ITI reconstructions. to translate images from domain B to domain B and generator Article number, page 4 of 10 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Example to illustrate the performance of an image reconstruction with ITI (lower panels) and the corresponding baseline (dif￾ference between single frame of original burst and speckle reconstruction; upper panels). Each panel shows the pixel-wise difference of the re￾c…
Figure 7
Figure 7. Figure 7: Visualization of a single frame from the original image burst on 13 June 2019 in the first row. The second and third rows show the speckle and ITI reconstruction, respectively. The first column shows the full-resolution observation, and the second column presents a zoo…
Figure 9
Figure 9. Figure 9: Evaluation of the quality metrics MAE, MSE, SSIM, and PSNR with a varying number of input frames of the burst for both the G-band (red) and Blue continuum (blue). The best quality for all metrics is achieved by using the whole image burst (100 frames) for the recon￾str…
Figure 10
Figure 10. Figure 10: Evaluation of speckle reconstruc￾tion showing artifacts in the Blue continuum (450.6 nm) from the 19 October 2019. The first column shows a single frame of the original burst, the second column shows the speckle re￾construction, and the third column shows our ITI reco…
Figure 11
Figure 11. Figure 11: Example where ITI shows limitations in the reconstruction of an image burst with an increased noise level from 29 May 2019. We show a single frame of the original burst in the first row, the speckle reconstruction in the second row, and our ITI reconstruction in the t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

  1. [1]

    2018, A&A, 620, A73 Asensio Ramos, A., Esteban Pozuelo, S., & Kuckein, C

    Asensio Ramos, A., de la Cruz Rodríguez, J., & Pastor Yabar, A. 2018, A&A, 620, A73 Asensio Ramos, A., Esteban Pozuelo, S., & Kuckein, C. 2023, Solar Phys., 298, 91 Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167 Berkefeld , T., Schmidt, D., Soltau, D., von der Lühe, O., & Heidecke, F. 2012, Astron. Nachr, 333, 863

    work page 2018

  2. [2]

    2012, Astron

    Collados, M., López, R., Páez, E., et al. 2012, Astron. Nachr., 333, 872

    work page 2012

  3. [3]

    2023, JATIS, 9, 015001

    Denker, C., Verma, M., Wi´sniewska, A., et al. 2023, JATIS, 9, 015001

    work page 2023

  4. [4]

    A., Conforto, V

    Fardo, F. A., Conforto, V . H., de Oliveira, F. C., & Rodrigues, P. S. 2016, arXiv e-prints, arXiv:1605.07116

    arXiv 2016

  5. [5]

    2020, Commun

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. 2020, Commun. ACM, 63, 139–144

    work page 2020

  6. [6]

    2017, in Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17 (Red Hook, NY , USA: Curran Associates Inc.), 6629–6640

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., & Hochreiter, S. 2017, in Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17 (Red Hook, NY , USA: Curran Associates Inc.), 6629–6640

    work page 2017

  7. [7]

    2023, IEEE Transactions on Pattern Analysis and Machine Intelligence, 45, 10173

    Huang, L., Qin, J., Zhou, Y ., et al. 2023, IEEE Transactions on Pattern Analysis and Machine Intelligence, 45, 10173

    work page 2023

  8. [8]

    2017, in 2017 IEEE International Confer- ence on Computer Vision (ICCV), 2458–2467

    Huang, R., Zhang, S., Li, T., & He, R. 2017, in 2017 IEEE International Confer- ence on Computer Vision (ICCV), 2458–2467

    work page 2017

Show all 29 references
  1. [9]

    Isola, P., Zhu, J.-Y ., Zhou, T., & Efros, A. A. 2017, in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 5967–5976

    2017

  2. [10]

    M., Pötzi, W., & Podladchikova, T

    Jarolim, R., Veronig, A. M., Pötzi, W., & Podladchikova, T. 2024, Nat. Commun. (in press), arXiv:2401.08057

    2024 arXiv

  3. [11]

    2020, ApJ, 903, L25

    Jeong, H.-J., Moon, Y .-J., Park, E., & Lee, H. 2020, ApJ, 903, L25

    2020

  4. [12]

    2019, ApJ, 881, L30

    Jia, P., Huang, Y ., Cai, B., & Cai, D. 2019, ApJ, 881, L30

    2019

  5. [13]

    2019, Nat

    Kim, T., Park, E., Lee, H., et al. 2019, Nat. Astron., 3, 397

    2019

  6. [14]

    Kingma, D. P. & Ba, J. 2014, CoRR, abs/1412.6980

    2014 arXiv

  7. [15]

    2020, A&A, 641, A27

    Kleint, L., Berkefeld, T., Esteves, M., et al. 2020, A&A, 641, A27

    2020

  8. [16]

    2017, in Fine Structure and Dynam- ics of the Solar Atmosphere, ed

    Kuckein, C., Denker, C., Verma, M., et al. 2017, in Fine Structure and Dynam- ics of the Solar Atmosphere, ed. S. Vargas Domínguez, A. G. Kosovichev, P. Antolin, & L. Harra, V ol. 327, 20–24

    2017

  9. [17]

    2017, in 2017 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 105–114 Löfdahl, M., Noort, M., & Denker, C

    Ledig, C., Theis, L., Huszár, F., et al. 2017, in 2017 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 105–114 Löfdahl, M., Noort, M., & Denker, C. 2007, Modern Solar Facilities – Advanced Solar Science, 119 Löfdahl, M. G., Bones, P. J., & Millane, R. P. 2...

    2017

  10. [18]

    W., Weigelt, G., & Wirnitzer, B

    Lohmann, A. W., Weigelt, G., & Wirnitzer, B. 1983, Appl. Opt., 22, 4028 M. Van Noort, e. a. 2005, Solar Phys., 228, 191

    1983

  11. [19]

    2017, in 2017 IEEE International Conference on Computer Vision (ICCV) (Los Alamitos, CA, USA: IEEE Computer Soci- ety), 2813–2821

    Mao, X., Li, Q., Xie, H., et al. 2017, in 2017 IEEE International Conference on Computer Vision (ICCV) (Los Alamitos, CA, USA: IEEE Computer Soci- ety), 2813–2821

    2017

  12. [20]

    2019, ApJ, 884, L23

    Park, E., Moon, Y .-J., Lee, J.-Y ., et al. 2019, ApJ, 884, L23

    2019

  13. [21]

    2017, in NIPS-W Quintero Noda, C., Schlichenmaier, R., Bellot Rubio, L

    Paszke, A., Gross, S., Chintala, S., et al. 2017, in NIPS-W Quintero Noda, C., Schlichenmaier, R., Bellot Rubio, L. R., et al. 2022, A&A, 666, A21

    2017

  14. [22]

    & Marino, J

    Rimmele, T. & Marino, J. 2011, Living Reviews in Solar Phys., 8, 2

    2011

  15. [23]

    R., Warner, M., Keil, S

    Rimmele, T. R., Warner, M., Keil, S. L., et al. 2020, Solar Phys., 295, 172

    2020

  16. [24]

    2012, Astronom

    Schmidt, W., von der Lühe, O., V olkmer, R., et al. 2012, Astronom. Nachr., 333, 796

    2012

  17. [25]

    2020, ApJ, 895, L16

    Shin, G., Moon, Y .-J., Park, E., et al. 2020, ApJ, 895, L16

    2020

  18. [26]

    2021, ApJ, 920, 101 The SunPy Community, Barnes, W

    Son, J., Cha, J., Moon, Y .-J., et al. 2021, ApJ, 920, 101 The SunPy Community, Barnes, W. T., Bobra, M. G., et al. 2020, ApJ, 890, 68 von der Luehe, O. 1993, A&A, 268, 374 von der Lühe, O., V olkmer, R., Kentischer, T. J., & Geißler, R. 2012, Astron. Nachr., 333, 894

    2021

  19. [27]

    2018, in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8798–8807

    Wang, T.-C., Liu, M.-Y ., Zhu, J.-Y ., et al. 2018, in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8798–8807

    2018

  20. [28]

    2004, Image Processing, IEEE Transactions on, 600 Wöger, F., von der Lühe, O., & Reardon, K

    Wang, Z., Bovik, A., Sheikh, H., & Simoncelli, E. 2004, Image Processing, IEEE Transactions on, 600 Wöger, F., von der Lühe, O., & Reardon, K. 2008, A&A, 488, 375

    2004

  21. [29]

    Zhu, J.-Y ., Park, T., Isola, P., & Efros, A. A. 2017, in 2017 IEEE International Conference on Computer Vision (ICCV), 2242–2251 Article number, page 10 of 10

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.