Enhancing Galaxy Classification with U-Net Variational Autoencoders for Image Denoising

Sergey Mirzoyan

REVIEW 3 major objections 7 minor 39 references

Enhancing Galaxy Classification with U-Net Variational Autoencoders for Image Denoising

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Adding a U-Net variational autoencoder as a denoising preprocessing step improves galaxy morphology classification, with a group-equivariant CNN reaching 97.45 percent accuracy on Galaxy10 DECaLS images.

desk verdict The pipeline is sensible and the writing is clear, but the paper never runs the one experiment that would support its central claim: a same-protocol comparison on the original noisy Galaxy10 images. read the letter →

arxiv 2506.19434 v1 pith:QO3Z52SL submitted 2025-06-24 astro-ph.IM

Sergey Mirzoyan This is my paper

classification astro-ph.IM

keywords galaxymorphologyclassificationimagedenoisingU-NetvariationalautoencoderGalaxy10DECaLSEFIGIdatasetgroupequivariantCNND16symmetryastronomicalpreprocessing

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

Galaxy images arrive contaminated with foreground stars, satellite trails, and diffraction patterns, and this paper claims that cleaning them before classification is what unlocks better morphological classification. The proposed cleaning step is a U-Net variational autoencoder trained on EFIGI galaxy images into which such artifacts were artificially injected, and the resulting denoiser is applied to Galaxy10 DECaLS images before four classifiers see them. The best classifier, a group-equivariant CNN, reaches 97.45 percent accuracy, and all four architectures trained on the denoised images beat the previously published accuracies on the same dataset. The paper therefore argues that deep generative denoising deserves a place as a standard preprocessing step for astronomy image analysis.

What carries the argument

The U-Net variational autoencoder: an encoder–decoder network with skip connections whose latent space is regularized by a Kullback–Leibler divergence term, trained with a binary cross-entropy reconstruction loss to map contaminated images back to clean ones. Its training data are 1,000 uncontaminated EFIGI galaxies, each repeatedly overlaid with one to three simulated stars or galaxies at random positions, orientations, and brightness profiles, producing 10,000 noisy–clean image pairs. Once trained, the same network is applied to every Galaxy10 DECaLS image; the classifier then sees only the denoised output.

What would settle it

Train the same four classifier configurations, with identical hyperparameters, optimization settings, and data splits, on the original Galaxy10 DECaLS images without any denoising; if any of these matched-training runs reaches or exceeds the 97.45 percent accuracy reported for the denoised images, the claimed improvement would not be attributable to the denoising step.

Watch

Extended reading notes

Core claim

The central claim is that the denoising preprocessing step itself, not any single classifier, drives the accuracy gain. After training a U-Net variational autoencoder on 10,000 simulated contaminated versions of 1,000 clean EFIGI galaxies, the denoiser is applied to the Galaxy10 DECaLS dataset, and DenseNet-201, ResNet50, VGG16, and a D16-symmetric GCNN are trained on the denoised images. The GCNN achieves 97.45 percent accuracy with a 96.4 F1-score, and every tested architecture outperforms the corresponding results reported in earlier studies that used the same dataset without this denoising step. On the paper's reading, that consistent across-architecture improvement is evidence for the denoising step's effectiveness.

Load-bearing premise

The load-bearing premise is that a denoiser trained on EFIGI images with simulated stars, satellite trails, and diffraction patterns produces images that are genuinely better for classifying real Galaxy10 DECaLS galaxies, a transfer the paper never directly validates on the target dataset.

Editorial extensions

If this is right

Galaxy morphology surveys that adopt this preprocessing step could expect cleaner inputs to their classifiers, with accuracy gains that appear across model families rather than tied to one architecture.
The reported 97.45 percent accuracy on Galaxy10 DECaLS becomes a new target for subsequent classification work on that dataset.
The same trained denoiser, or a re-trained variant, could be applied to other contaminated astronomical images, such as searches for gravitational lenses or cosmic web structures, as the paper itself suggests.
The approach positions the denoiser as a model-agnostic preprocessing layer, meaning the choice of classifier can be optimized independently of the denoising.

Reading between the lines

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

A matched internal control—training the same classifiers on noisy versus denoised Galaxy10 DECaLS images under identical hyperparameters—would isolate the denoising contribution from the effect of simply retraining with different settings, something the paper does not provide.
If the accuracy gain is partly a smoothing or regularization effect rather than removal of the specific simulated artifacts, simpler preprocessing such as Gaussian smoothing might reproduce part of the gain; testing that would clarify the mechanism.
The denoiser was trained on EFIGI and applied to DECaLS without checking that its output is closer to the true DECaLS scene; measuring PSNR/SSIM against real DECaLS images with known injected artifacts would test whether the domain transfer actually holds.
Cross-survey generalization could be tested by applying the same pipeline to a third dataset, such as SDSS or JWST imaging, to see whether the preprocessing step still helps when the contamination statistics change.

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.

Desk Editor's Note

read the letter

Short version: I agree with the stress-test note. The paper's central claim—that U-Net VAE denoising improves galaxy classification—is untested, because there is no matched noisy baseline anywhere in the manuscript. Table 3 reports only denoised results; the 'noisy' numbers are borrowed from Yumeng, Wang, and Pandya, which use different splits, optimizers, and architectures. So the accuracy gains cannot be attributed to the denoiser.

What is genuinely new: the specific application of a U-Net VAE as a preprocessing step for Galaxy10 DECaLS classification, with a GCNN reaching 97.45% accuracy if it holds up. That is a reasonable engineering idea. The paper is also transparent about its pipeline: it documents the VAE architecture, the contamination simulation (1–3 objects per image via PyAutoLens), the training details (batch size, epochs, optimizer, loss), and the classifier hyperparameters. Public datasets are used throughout, which makes the work reproducible in principle. The paper is clearly written and the confusion matrix and per-class results are useful.

The soft spots are proportional to the claim. First, the missing baseline. The abstract asserts 'models trained on denoised images consistently outperform those trained on noisy images,' but the paper never presents that experiment. It compares denoised results to published numbers that were obtained with different training protocols. This is the central evidential gap. Second, the denoising quality metrics (PSNR 32.9 dB, SSIM 0.78) are computed on the EFIGI simulated test set, not on Galaxy10 DECaLS. The VAE is trained on EFIGI with simulated stars, trails, and diffraction patterns, and then applied to DECaLS without any validation of domain transfer. There may be a regularization/smoothing effect at play, but the paper does not demonstrate that actual artifacts are removed from DECaLS images. Third, the hyperparameter search for classifiers ('rigorous experimentation') is described only in prose; there is no ablations on data augmentation, training schedule, or seeds, so we cannot rule out that the gains come from the training configuration rather than the denoising.

None of this makes the paper incoherent. The architecture is sound and the writing is honest in its omissions. But as written, the evidence does not support the headline claim. The fix is straightforward: rerun the same classifiers on the same split of the original (non-denoised) Galaxy10 images, ideally with the same hyperparameters, and report the paired comparison. If the denoised version wins, the paper becomes a modest but useful contribution. If not, the conclusion changes completely. I would send this to a referee with a strong request to obtain that baseline before publication. It is a paper that can be repaired rather than one that should be discarded.

Who is the audience? Practitioners working on galaxy morphology classification pipelines. Not a methods paper; no new theory. I would not cite it yet, but I'd read a revised version.

Recommendation: engage, but require the matched noisy baseline before the claim is accepted.

Referee Report

3 major / 7 minor

Summary. The manuscript proposes a preprocessing pipeline in which a U-Net variational autoencoder, trained on EFIGI galaxy images with simulated contamination (stars, satellite trails, diffraction patterns), is applied to denoise images from the Galaxy10 DECaLS dataset. Four classifiers (DenseNet-201, ResNet50, VGG16, and a D16-equivariant GCNN) are then trained on the denoised images, with the GCNN reportedly reaching 97.45% test accuracy. The authors claim that this denoising step substantially improves classification relative to training on noisy images and to previously published results.

Significance. If validated, the approach would offer a practical denoising preprocessing step for galaxy morphology classification, and the use of public datasets and a detailed architecture table are strengths. However, the central causal claim is not supported by the experiments as written: there is no same-protocol comparison between denoised and noisy inputs, and the denoising quality metrics are computed only on the simulation domain. The paper also describes its architecture as a U-Net despite providing no skip connections in Table 1. These issues are fixable but currently leave the main conclusion unsubstantiated.

major comments (3)

[Section 4, Table 3; Section 5] The central claim that models trained on denoised images outperform those trained on noisy images is never tested. All results in Table 3 are for classifiers trained on denoised Galaxy10 DECaLS images; there is no corresponding experiment using the original noisy images with the same train/validation/test split, the same architectures, and the same hyperparameters (batch size 128, learning rate 0.0001, Adagrad, 200 epochs, as described in Section 3.2). The comparisons to Yumeng (2023), Wang (2023), and Pandya et al. (2023) are cross-paper comparisons with different experimental setups, so the superior accuracies cannot be attributed to the denoising step. The statement in Section 5 that 'models trained on denoised images consistently outperformed those trained on noisy images' is therefore unsupported. A matched noisy baseline is required.
[Section 3.1, Table 1] The model is called a 'U-Net Variational Autoencoder' throughout the paper, but Table 1 shows no skip connections between the encoder and decoder. The decoder is a plain sequence of transposed convolutions with no concatenation of encoder feature maps, which is the defining characteristic of a U-Net. The authors should either implement actual skip connections or rename the architecture as a convolutional VAE; as written, the central methodological novelty is misrepresented.
[Section 3.1 and Section 4] The denoising quality metrics (PSNR 32.9 dB, SSIM 0.78) are computed exclusively on the simulated EFIGI test set. The VAE is then applied to Galaxy10 DECaLS images without any ground-truth evaluation or discussion of domain shift between the SDSS-based EFIGI training data and the DECaLS data. Because the subsequent classification improvements are attributed to denoising, the paper should demonstrate that the denoiser actually improves the Galaxy10 images, for example by reporting metrics on that dataset or by analyzing the denoised outputs qualitatively and quantitatively.

minor comments (7)

[Section 4] The text states 'the PSNR demonstrated an average of 32 dB' while Section 3.1 reports an average of 32.9 dB; these values should be consistent.
[Section 4] The phrase 'the SSIM achieved a mean of 0.78 and an average of 0.12' should read 'a standard deviation of 0.12'.
[References] The reference to Kingma & Welling (2013) should cite the 2014 publication 'Auto-Encoding Variational Bayes'.
[Figure 5] It should be stated which model's training and validation accuracy curves are shown; presumably the GCNN, but this is not specified.
[Table 2] The abbreviated column headers are difficult to read; using the full class names or a legend would improve clarity.
[Introduction and Abstract] The phrase 'traditional classification approaches' is vague; please specify the comparison baselines.
[Throughout] There are a few typographical issues, e.g., 'the used denoising procedure' in the Abstract, and duplicated 'instrumental artifacts' in Section 3.1.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: all tested claims rest on held-out data and empirical comparisons, although the causal attribution to denoising lacks a matched noisy baseline.

full rationale

The paper's derivation chain is not circular. The U-Net VAE is trained on EFIGI images with simulated contamination against clean EFIGI targets (Section 3.1), and the reported PSNR and SSIM values (Section 4, Figure 4) are computed on a separate 400-image EFIGI test subset explicitly described as 'the testing subset consists of 400 images'; this is standard in-distribution evaluation, not self-definition. The classification results in Table 3 are measured on the held-out 2,500-image Galaxy10 DECaLS test partition after applying the trained VAE, and the classifier hyperparameters ('128 batch size, 0.0001 for learning rate, categorical cross entropy for the loss and Adagrad for an optimizer') were fixed before testing. No fitted constant from the test set enters the reported accuracies. The only self-citation (Mirzoyan et al. 2019) is background on gravitational-lens pipelines and is not load-bearing. The statement that models trained on denoised images 'consistently outperform those trained on noisy images' is weaker than ideal because Table 3 only reports denoised runs and the noisy comparison comes from other papers (Yumeng 2023, Wang 2023, Pandya et al. 2023) with different training protocols; however, this is a missing-control or external-validity limitation, not circular reasoning. No step reduces to its own input by construction, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a small number of hand-chosen hyperparameters (VAE beta, classification settings) and several unverified domain assumptions, most notably the transferability of the denoiser from EFIGI to DECaLS. There are no invented physical entities. The free parameters and assumptions are used to set up the empirical pipeline rather than to derive the result mathematically, but the lack of domain-transfer validation is the most significant unaccounted-for premise.

free parameters (4)

VAE KL weight beta = not reported
Appears in total loss Eq. 3; chosen by hand to balance reconstruction and KL terms, affecting the denoising behavior of the VAE.
Number of simulated contaminants per image = random integer 1-3
Section 3.1; chosen ad hoc to avoid overcrowding, shaping the noise distribution the VAE learns.
Classification hyperparameters (batch size, learning rate, optimizer, epochs) = 128, 1e-4, Adagrad, 200
Section 3.2; selected via 'rigorous experimentation' on the validation set, so the reported accuracies are tuned estimates rather than a single fixed configuration.
Simulated object physical parameters (Sersic index, ellipticity, effective radius) = varied within uncontrolled ranges
Section 3.1; these parameters are generated with PyAutoLens and affect the realism of the simulated contamination, but the exact ranges are not specified.

assumptions (5)

domain assumption EFIGI 'no contamination' flag correctly identifies clean galaxy images.
Section 3.1 uses this flag to select 1400 clean images; if the flag is imperfect, the VAE trains on mislabeled data.
ad hoc to paper Contamination simulated with PyAutoLens realistically represents the noise in real astronomical survey images.
The simulation is purpose-built for this study; no validation is provided that it matches the actual noise statistics of DECaLS or other survey images.
ad hoc to paper A VAE trained on EFIGI (SDSS-based) images transfers to Galaxy10 DECaLS images without fine-tuning.
Section 3.1 applies the trained VAE to Galaxy10 without any domain adaptation or validation on that dataset; this is a key unverified premise.
domain assumption Galaxy10 DECaLS labels are accurate.
The labels come from Galaxy Zoo and are used as ground truth; the paper does not account for label noise.
standard math Standard VAE reparameterization and BCE/KL losses behave as expected for image reconstruction.
Based on Kingma and Welling 2013; the paper relies on standard VAE machinery without re-deriving it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Galaxy Classification with U-Net Variational Autoencoders for Image Denoising." pith.science (2026). https://pith.science/paper/QO3Z52SL

@misc{pith2026250619434,
  author       = {Pith},
  title        = {Pith review of: Enhancing Galaxy Classification with U-Net Variational Autoencoders for Image Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QO3Z52SL}},
  note         = {Machine review of arXiv:2506.19434}
}

read the original abstract

AI-enhanced approaches are becoming common in astronomical data analysis, including in the galaxy morphological classification. In this study we develop an approach that enhances galaxy classification by incorporating an image denoising pre-processing step, utilizing the U-Net Variational Autoencoder (VAE) architecture and effectively mitigating noise in galaxy images and leading to improved classification performance. Our methodology involves training U-Net VAEs on the EFIGI dataset. To simulate realistic observational conditions, we introduce artifacts such as projected stars, satellite trails, and diffraction patterns into clean galaxy images. The denoised images generated are evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), to quantify the quality improvements. We utilize the denoised images for galaxy classification tasks using models such as DenseNet-201, ResNet50, VGG16 and GCNN. Simulations do reveal that, the models trained on denoised images consistently outperform those trained on noisy images, thus demonstrating the efficiency of the used denoising procedure. The developed approach can be used for other astronomical datasets, via refining the VAE architecture and integrating additional pre-processing strategies, e.g. in revealing of gravitational lenses, cosmic web structures.

Figures

Figures reproduced from arXiv: 2506.19434 by the authors.

**Figure 2.** Denoising VAE performance on real unseen images. The first row represents real noisy images, while [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Training and validation loss over 150 Epochs [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The figure illustrates the distribution and comparison of performance metrics for the dataset. The [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Training and validation accuracy over 200 Epochs [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages

[1]

Morphological Classification of Galaxies Through Structural and Star Formation Parameters Using Machine Learning

Aguilar-Arguello, G. et al,2025, Morphological Classification of Galaxies Through Structural and Star Formation Parameters Using Machine Learning, arXiv:2501.06340

work page Pith review arXiv 2025
[2]

et al, 2025, A benchmark analysis of saliency-based explainable deep learning methods for the morphological classification of radio galaxies, arXiv:2502.17207

Atemkeng M.T. et al, 2025, A benchmark analysis of saliency-based explainable deep learning methods for the morphological classification of radio galaxies, arXiv:2502.17207

arXiv 2025
[3]

et al, 2011 , The EFIGI catalogue of 4458 nearby galaxies with detailed morphology, Astronomy & Astrophysics, 532, A74

Baillard, A. et al, 2011 , The EFIGI catalogue of 4458 nearby galaxies with detailed morphology, Astronomy & Astrophysics, 532, A74

work page 2011
[4]

Binney J., Merrifield M., 1998, Galactic astronomy, Princeton, NJ : Princeton University Press, (Princeton series in astrophysics)

work page 1998
[5]

et al., 2022, A program to build E(N)-equivariant steerable CNNs

Cesa, G. et al., 2022, A program to build E(N)-equivariant steerable CNNs. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=WE4qe9xlnQw

work page 2022
[6]

Robustness of deep learning algorithms in astronomy -- galaxy morphology studies

\'Ciprijanovi\'c, A., Kafkes, D. et al., 2021, Robustness of deep learning algorithms in astronomy – galaxy morphology studies, arxiv.org/abs/2111.00961

work page Pith review arXiv 2021
[7]

et al, 2022, Deepadversaries: examining the robustness of deep learning models for galaxy morphology classification

\'Ciprijanovi\'c, A., Kafkes, D. et al, 2022, Deepadversaries: examining the robustness of deep learning models for galaxy morphology classification. Machine Learning: Science and Technology, 3(3):035007

work page 2022
[8]

and Welling, M., 2016, Group equivariant convolutional networks, arXiv:1602.07576

Cohen, T. and Welling, M., 2016, Group equivariant convolutional networks, arXiv:1602.07576

arXiv 2016

Show all 39 references

[9]

de la Calleja, J., Fuentes O., 2004, Machine learning and image analysis for morphological galaxy classification, MNRAS, Volume 349, Issue 1, pp. 87-93

2004
[10]

al., 2021, ROGER: Reconstructing orbits of galaxies in extreme regions using machine learning techniques, MNRAS, Volume 500, Issue 2, pp.1784-1794

de los Rios, M. al., 2021, ROGER: Reconstructing orbits of galaxies in extreme regions using machine learning techniques, MNRAS, Volume 500, Issue 2, pp.1784-1794

2021
[11]

G., et al

de Vaucouleurs, G., de Vaucouleurs, A., Corwin, Jr., H. G., et al. 1991, Third Reference Catalogue of Bright Galaxies Volume 1-3, XII, 2069 pp. 7 figs.. Springer-Verlag Berlin

1991
[12]

et al., 2025, Machine Learning Workflow for Morphological Classification of Galaxies, arXiv:2505.04676

Doser B. et al., 2025, Machine Learning Workflow for Morphological Classification of Galaxies, arXiv:2505.04676

2025 arXiv
[13]

et al., 2018, Robustness of rotation-equivariant networks to adversarial perturbations, arXiv:1802.06627

Dumont, B. et al., 2018, Robustness of rotation-equivariant networks to adversarial perturbations, arXiv:1802.06627

2018 arXiv
[14]

et al., 2025, DAWN JWST Archive: Morphology from profile fitting of over 340,000 galaxies in major JWST fields, Morphology evolution with redshift and galaxy type, arXiv:2505.21622

Genin, A. et al., 2025, DAWN JWST Archive: Morphology from profile fitting of over 340,000 galaxies in major JWST fields, Morphology evolution with redshift and galaxy type, arXiv:2505.21622

2025 arXiv
[15]

et al., 2025, Intrinsic galaxy alignments in the KiDS-1000 bright sample: dependence on colour, luminosity, morphology and galaxy scale, arXiv:2502.09452

Georgiou, C. et al., 2025, Intrinsic galaxy alignments in the KiDS-1000 bright sample: dependence on colour, luminosity, morphology and galaxy scale, arXiv:2502.09452

2025 arXiv
[16]

et al., 2025, Galaxy Morphological Classification with Zernike Moments and Machine Learning Approaches, arXiv:2501.09816

Ghaderi, H. et al., 2025, Galaxy Morphological Classification with Zernike Moments and Machine Learning Approaches, arXiv:2501.09816

2025 arXiv
[17]

et al., 2022, Galaxy Morphology Classification with DenseNet, Journal of Physics: Conference Series, Volume 2402, Issue 1, id.012009, 11 pp

Hui, W. et al., 2022, Galaxy Morphology Classification with DenseNet, Journal of Physics: Conference Series, Volume 2402, Issue 1, id.012009, 11 pp

2022
[18]

Martin, G., 2025, Morphological Feature Distances Among the Spectral Types of SDSS Galaxies, arXiv:2504.19484

Kim, D. ; Martin, G., 2025, Morphological Feature Distances Among the Spectral Types of SDSS Galaxies, arXiv:2504.19484

2025 arXiv
[19]

Kingma, D. P. and Welling, M., 2013, Auto-Encoding Variational Bayes, arXiv:1312.6114

2013 arXiv
[20]

Lintott, C. J. et al., 2008, Galaxy Zoo: morphologies derived from visual inspection of galaxies from the Sloan Digital Sky Survey, MNRAS, Volume 389, Issue 3, pp. 1179-1189

2008
[21]

Lintott, C. J. et al., 2011, Galaxy Zoo 1: data release of morphological classifications for nearly 900 000 galaxies, MNRAS, Volume 410, Issue 1, pp. 166-178

2011
[22]

et al., 2025, Evaluating the Accuracy of Non-parametric Galaxy Morphological Indicator Measurements in the CSST Imaging Survey, arXiv:2505.02127

Luo, Y. et al., 2025, Evaluating the Accuracy of Non-parametric Galaxy Morphological Indicator Measurements in the CSST Imaging Survey, arXiv:2505.02127

2025 arXiv
[23]

Masters K., 2025, Morphological Classification of Galaxies, arXiv:2502.09610

2025 arXiv
[24]

S.; Khachatryan, H.; Yegorian, G.; Gurzadyan, V

Mirzoyan, S. S.; Khachatryan, H.; Yegorian, G.; Gurzadyan, V. G., 2019, Machine learning and Kolmogorov analysis to reveal gravitational lenses, MNRAS (Letters), Volume 489, Issue 1, p.L32-L36

2019
[25]

et al., 2010, Galaxy Formation and Evolution, Cambridge University Press

Mo, H. et al., 2010, Galaxy Formation and Evolution, Cambridge University Press

2010
[26]

Nightingale, J. W. et al., 2021, PyAutoLens: Open-Source Strong Gravitational Lensing, Journal of Open Source Software, Volume 6, issue 58, id. 2825

2021
[27]

et al., 2023, E2 Equivariant Neural Networks for Robust Galaxy Morphology Classification, arXiv:2311.01500

Pandya S. et al., 2023, E2 Equivariant Neural Networks for Robust Galaxy Morphology Classification, arXiv:2311.01500

2023 arXiv
[28]

et al., 2003, HYPERLEDA

Paturel, G. et al., 2003, HYPERLEDA. I. Identification and designation of galaxies, A&A, Volume 412, pp. 45-55

2003
[29]

Petrillo, C. E. et al., 2017, Finding strong gravitational lenses in the Kilo Degree Survey with Convolutional Neural Networks, MNRAS, Volume 472, Issue 1, p.1129-1150

2017
[30]

et al., 2020, Data-Efficient Classification of Radio Galaxies, arXiv:2011.13311

Samudre, A. et al., 2020, Data-Efficient Classification of Radio Galaxies, arXiv:2011.13311

2020 arXiv
[31]

et al., 2025, The SRG/eROSITA all-sky survey: The morphologies of clusters of galaxies I: A catalogue of morphological parameters, arXiv:2502.02239

Sanders J.S. et al., 2025, The SRG/eROSITA all-sky survey: The morphologies of clusters of galaxies I: A catalogue of morphological parameters, arXiv:2502.02239

2025 arXiv
[32]

Sparke L., Gallagher J., 2007, Galaxies in the Universe: An Introduction, Cambridge University Press, New York

2007
[33]

et al, 2025, Automatic Machine Learning Framework to Study Morphological Parameters of AGN Host Galaxies within z<1.4 in the Hyper Supreme-Cam Wide Survey, arXiv:2501.15739

Tian, C. et al, 2025, Automatic Machine Learning Framework to Study Morphological Parameters of AGN Host Galaxies within z<1.4 in the Hyper Supreme-Cam Wide Survey, arXiv:2501.15739

2025 arXiv
[34]

van den Bergh, S., 1998, Galaxy Morphology and classification, Cambridge University Press

1998
[35]

et al., 2022, Galaxy Zoo DECaLS: Detailed visual morphology measurements from volunteers and deep learning for 314000 galaxies, MNRAS, Volume 509, Issue 3, pp.3966-3988

Walmsley, M. et al., 2022, Galaxy Zoo DECaLS: Detailed visual morphology measurements from volunteers and deep learning for 314000 galaxies, MNRAS, Volume 509, Issue 3, pp.3966-3988

2022
[36]

Wang, G., 2023, Journal of Physics: Conference Series, Volume 2580, Issue 1, id.012064, 8 pp

2023
[37]

Conference on Neural Information Processing Systems (NeurIPS)

Weiler, M.e and Cesa, G., 2019, General E(2)-Equivariant Steerable CNNs. Conference on Neural Information Processing Systems (NeurIPS)

2019
[38]

Yumeng, Q., 2023, Performance comparison among VGG16, InceptionV3, and resnet on galaxy morphology classification, Journal of Physics: Conference Series, Volume 2580, Issue 1, id.012009, 7 pp

2023
[39]

et al., 2019, Galaxy morphology classification with deep convolutional neural networks, Astrophysics and Space Science, Volume 364, Issue 4, id

Zhu, X. et al., 2019, Galaxy morphology classification with deep convolutional neural networks, Astrophysics and Space Science, Volume 364, Issue 4, id. 55, 15 pp

2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Morphological Classification of Galaxies Through Structural and Star Formation Parameters Using Machine Learning

Aguilar-Arguello, G. et al,2025, Morphological Classification of Galaxies Through Structural and Star Formation Parameters Using Machine Learning, arXiv:2501.06340

work page Pith review arXiv 2025

[2] [2]

et al, 2025, A benchmark analysis of saliency-based explainable deep learning methods for the morphological classification of radio galaxies, arXiv:2502.17207

Atemkeng M.T. et al, 2025, A benchmark analysis of saliency-based explainable deep learning methods for the morphological classification of radio galaxies, arXiv:2502.17207

arXiv 2025

[3] [3]

et al, 2011 , The EFIGI catalogue of 4458 nearby galaxies with detailed morphology, Astronomy & Astrophysics, 532, A74

Baillard, A. et al, 2011 , The EFIGI catalogue of 4458 nearby galaxies with detailed morphology, Astronomy & Astrophysics, 532, A74

work page 2011

[4] [4]

Binney J., Merrifield M., 1998, Galactic astronomy, Princeton, NJ : Princeton University Press, (Princeton series in astrophysics)

work page 1998

[5] [5]

et al., 2022, A program to build E(N)-equivariant steerable CNNs

Cesa, G. et al., 2022, A program to build E(N)-equivariant steerable CNNs. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=WE4qe9xlnQw

work page 2022

[6] [6]

Robustness of deep learning algorithms in astronomy -- galaxy morphology studies

\'Ciprijanovi\'c, A., Kafkes, D. et al., 2021, Robustness of deep learning algorithms in astronomy – galaxy morphology studies, arxiv.org/abs/2111.00961

work page Pith review arXiv 2021

[7] [7]

et al, 2022, Deepadversaries: examining the robustness of deep learning models for galaxy morphology classification

\'Ciprijanovi\'c, A., Kafkes, D. et al, 2022, Deepadversaries: examining the robustness of deep learning models for galaxy morphology classification. Machine Learning: Science and Technology, 3(3):035007

work page 2022

[8] [8]

and Welling, M., 2016, Group equivariant convolutional networks, arXiv:1602.07576

Cohen, T. and Welling, M., 2016, Group equivariant convolutional networks, arXiv:1602.07576

arXiv 2016