pith. sign in

arxiv: 2510.04718 · v3 · submitted 2025-10-06 · 🌌 astro-ph.IM

BGRem: A background noise remover for astronomical images based on a diffusion model

Rodney Nicolaas , Sascha Caron , Fiorenzo Stoppa , Saptashwa Bhattacharyya , Roberto Ruiz de Austri , Paul J. Groot , Andrew J. Levan This is my paper

Pith reviewed 2026-05-18 09:20 UTC · model grok-4.3

classification 🌌 astro-ph.IM
keywords background noise removaldiffusion modelsastronomical imagingsource detectionimage denoisingMeerLICHTFermi-LATmachine learning
0
0 comments X

The pith

BGRem uses a diffusion model trained on simulations to remove background noise from astronomical images, raising true positive source detections by about 7 percent with SExtractor on MeerLICHT data.

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

The paper presents BGRem as a machine learning tool that cleans background noise from astronomical images as a pre-processing step for source cataloging. It trains a diffusion model with an attention U-Net on simulated optical and gamma-ray images to learn denoising across multiple steps. Tests show this step lifts the number of true positive sources found by the standard SExtractor tool by roughly 7 percent on real MeerLICHT optical data. The same model also raises detection efficiency on optical images from other telescopes and on simulated gamma-ray data, indicating it can handle different noise patterns without retraining.

Core claim

BGRem is a diffusion-based model with an attention U-Net backbone trained in a supervised manner on simulated images to denoise astronomical data over several diffusion steps. When used as pre-processing, it increases the count of true positive sources detected by SExtractor by about 7 percent for MeerLICHT observations. The model further improves source detection efficiency on optical images from additional telescopes and on simulated gamma-ray images representative of Fermi-LAT, demonstrating adaptability across noise types and wavelengths.

What carries the argument

A diffusion model with attention U-Net backbone that performs supervised denoising of simulated astronomical images over multiple diffusion steps.

If this is right

  • BGRem pre-processing raises the yield of reliable sources extracted by pixel-based tools such as SExtractor.
  • The model applies directly to optical images from telescopes other than MeerLICHT without additional training.
  • BGRem maintains its denoising benefit when applied to simulated gamma-ray images with noise statistics different from optical data.
  • The cross-wavelength results position BGRem as a candidate general-purpose background removal step for multi-wavelength surveys.

Where Pith is reading between the lines

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

  • If the performance gain holds on real gamma-ray observations, the same model could assist source searches in high-energy surveys where background modeling is especially difficult.
  • Routine use of BGRem before cataloging could lower the detection threshold for faint sources in wide-field optical surveys.
  • The zero-shot transfer observed on other optical telescopes suggests the learned noise statistics may be broadly representative rather than instrument-specific.

Load-bearing premise

The simulated training images accurately capture the statistical properties of real background noise and source distributions so that results on held-out simulations transfer to actual telescope data.

What would settle it

A side-by-side run of SExtractor on a large collection of real MeerLICHT images processed with and without BGRem, counting whether true positive detections rise by 7 percent while false positives stay flat or drop.

Figures

Figures reproduced from arXiv: 2510.04718 by Andrew J. Levan, Fiorenzo Stoppa, Paul J. Groot, Roberto Ruiz de Austri, Rodney Nicolaas, Saptashwa Bhattacharyya, Sascha Caron.

Figure 1
Figure 1. Figure 1: Example of the working of a diffusion model with five diffusion steps. The shown input is already pre-processed, while between step 4 and output there is a model prediction and post￾processing, excluding denormalisation. that the total variance is constant (α 2 t + σ 2 t = 1) across all time steps and ensures numerically stable training. To inform the network of the current noise level, the squared noise s… view at source ↗
Figure 2
Figure 2. Figure 2: The fraction of noise and ground truth per di [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic of the diffusion model used in BGRem for 3 diffusion steps. The forward noise addition is only used during training. When making predictions, the model goes from right to left, iteratively removing the noise. starts with the simulated ground truth image and iteratively adds the background noise. When making predictions, BGRem starts with the noisy image and iteratively removes the noise [PITH_FU… view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the backbone architecture used in BGRem, modified from the original attention U-Net ( [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: MeerLICHT image (left) and the image with the back [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: A histogram of the pixel values for the original simulated image (red), the ground truth no background image (green), [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sources found by SExtractor in the ground truth image [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: An example of zero-shot inference of BGRem is shown [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Left: Part of the cutout from the Legacy Survey used for source localisation results shown in Figure [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Same as in figure 8, but now the application is on real data obtained from the Legacy Survey. was shown to be effective in removing background from optical and γ-ray images and help SExtractor to detect more sources. Our results demonstrate that this approach can serve as an ef￾fective pre-processing step for existing methods for building a pipeline to create an astronomical catalog. We also highlight the… view at source ↗
Figure 12
Figure 12. Figure 12: Examples of BGRem for simulated gamma-ray sky patches are shown here for two di [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: We show an example of mean pixel values of the de [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Same as Figures 8, 11, but here we show the results obtained with simulated γ-ray sky images. The work of S.B was supported by the Slovenian Research Agency under grants P1-0031, N1-0344, and J1-60014. PJG is partly supported by NRF SARChI grant 111692. The work of R. RdA was supported by PID2020-113644GB￾I00 from the Spanish Ministerio de Ciencia e Innovación and by the PROMETEO/2022/69 from the Spanish … view at source ↗
read the original abstract

Context: Astronomical imaging aims to maximize signal capture while minimizing noise. Enhancing the signal-to-noise ratio directly on detectors is difficult and expensive, leading to extensive research in advanced post-processing techniques. Aims: Removing background noise from images is a valuable pre-processing step catalog-building tasks. We introduce BGRem, a machine learning (ML) based tool to remove background noise from astronomical images. Methods: BGRem uses a diffusion-based model with an attention U-Net as backbone, trained on simulated images for optical and gamma ({\gamma})-ray data from the MeerLICHT and Fermi-LAT telescopes. In a supervised manner, BGRem learns to denoise astronomical images over several diffusion steps. Results: BGRem performance was compared with a widely used tool for cataloging astronomical sources, SourceExtractor (SExtractor). It was shown that the amount of true positive sources using SExtractor increased by about 7% for MeerLICHT data when BGRem was used as a pre-processing step. We also show the generalizability of BGRem by testing it with optical images from different telescopes and also on simulated {\gamma}-ray data representative of the Fermi-LAT telescope. We show that in both cases, BGRem improves the source detection efficiency. Conclusions: BGRem can improve the accuracy in source detection of traditional pixel-based methods by removing complex background noise. Using zero-shot approach, BGRem can generalize well to a wide range of optical images. The successful application of BGRem to simulated {\gamma}-ray images, alongside optical data, demonstrates its adaptability to distinct noise characteristics and observational domains. This cross-wavelength performance highlights its potential as a general-purpose background removal framework for multi-wavelength astronomical surveys.

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

2 major / 1 minor

Summary. The manuscript introduces BGRem, a diffusion-based model with an attention U-Net backbone for background noise removal in astronomical images. Trained in a supervised manner exclusively on simulated optical and gamma-ray data from MeerLICHT and Fermi-LAT, it is evaluated as a pre-processing step for SExtractor. The central empirical claim is an approximately 7% increase in true-positive sources detected on MeerLICHT data; additional tests demonstrate generalization to optical images from other telescopes and to simulated gamma-ray data.

Significance. If the simulation-to-real transfer is validated, BGRem would represent a useful addition to the toolkit for improving source detection efficiency in catalog-building pipelines, particularly for handling complex backgrounds that challenge traditional pixel-based methods. The reported cross-wavelength applicability and zero-shot generalization to varied optical data are potential strengths for multi-messenger or survey applications.

major comments (2)
  1. [Results (MeerLICHT evaluation)] Results section on MeerLICHT performance: the reported ~7% increase in true-positive sources detected by SExtractor is presented without quantitative validation that the simulated background noise and source distributions match the statistical properties of real MeerLICHT data (e.g., via power spectra, noise histograms, or source-density comparisons). This validation is load-bearing for the transfer of the performance gain from simulation to actual telescope observations.
  2. [Methods (training and evaluation)] Methods section on training and evaluation: no details are provided on the training/validation splits, the specific diffusion schedule and number of steps, or error bars/statistical significance tests for the 7% figure. These omissions prevent assessment of whether the improvement is robust or could be an artifact of the simulation setup.
minor comments (1)
  1. [Abstract] The abstract refers to a 'zero-shot approach' for generalization but does not define what this means operationally for the diffusion model or the held-out test sets.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our manuscript on BGRem. We have addressed each major comment below and will incorporate revisions to strengthen the presentation of our methods and results.

read point-by-point responses

  1. Referee: [Results (MeerLICHT evaluation)] Results section on MeerLICHT performance: the reported ~7% increase in true-positive sources detected by SExtractor is presented without quantitative validation that the simulated background noise and source distributions match the statistical properties of real MeerLICHT data (e.g., via power spectra, noise histograms, or source-density comparisons). This validation is load-bearing for the transfer of the performance gain from simulation to actual telescope observations.

    Authors: We agree that explicit quantitative validation of the simulation-to-real match is important for supporting the claimed performance transfer. While our simulations were designed using known MeerLICHT instrument properties and noise characteristics from prior literature, the original manuscript did not include direct statistical comparisons. In the revised manuscript we will add these validations to the Results section, specifically including power spectrum comparisons, noise histogram matches, and source-density analyses between the simulated and real MeerLICHT data. revision: yes

  2. Referee: [Methods (training and evaluation)] Methods section on training and evaluation: no details are provided on the training/validation splits, the specific diffusion schedule and number of steps, or error bars/statistical significance tests for the 7% figure. These omissions prevent assessment of whether the improvement is robust or could be an artifact of the simulation setup.

    Authors: We acknowledge that the current Methods section lacks sufficient detail for full reproducibility and robustness evaluation. We will expand this section in the revision to report the training/validation split ratios and image counts, the precise diffusion schedule and number of steps used, and error bars on the 7% true-positive improvement together with statistical significance tests (such as paired t-tests across multiple evaluation runs). These additions will allow readers to assess whether the gain is robust. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper describes training a diffusion model with attention U-Net on simulated optical and gamma-ray images, then reports an empirical 7% increase in true-positive sources detected by SExtractor on real MeerLICHT data when BGRem is used as pre-processing. No mathematical derivation, first-principles prediction, or fitted parameter is presented that reduces by construction to its own inputs or to a self-citation chain. The performance number is an observed measurement on held-out real data rather than a quantity forced by the training procedure or by renaming a known result. The central claim therefore remains independent of the listed circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that simulated noise matches real telescope noise and that the supervised diffusion training objective produces a generalizable denoiser; no free parameters are explicitly named in the abstract, but the diffusion schedule and U-Net architecture choices function as implicit hyperparameters.

free parameters (1)
  • diffusion schedule and number of steps
    Standard diffusion hyperparameters that control the noise addition and removal process; their specific values are not stated in the abstract.
axioms (1)
  • domain assumption Simulated images faithfully reproduce the statistical properties of real background noise and source distributions for both optical and gamma-ray instruments.
    Invoked when the model trained on simulations is applied to real or other simulated data and performance gains are claimed.

pith-pipeline@v0.9.0 · 5870 in / 1299 out tokens · 25735 ms · 2026-05-18T09:20:39.720509+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 · 2 internal anchors

  1. [1]

    2015, TensorFlow: Large-scale ma- chine learning on heterogeneous systems

    Abadi, M., Agarwal, A., Barham, P., et al. 2015, TensorFlow: Large-scale ma- chine learning on heterogeneous systems

    work page 2015

  2. [2]

    Abdollahi, S. et al. 2020, The Astrophysical Journal Supplement Series, 247, 33

    work page 2020

  3. [3]

    2021, Frontiers in artificial intelligence, 4, 649917

    Acciarri, R., Adams, C., Andreopoulos, C., et al. 2021, Frontiers in artificial intelligence, 4, 649917

    work page 2021

  4. [4]

    Acero, F. et al. 2016, The Astrophysical Journal Supplement Series, 223, 26

    work page 2016

  5. [5]

    & Ichikawa, T

    Akhlaghi, M. & Ichikawa, T. 2015, The Astrophysical Journal Supplement Se- ries, 220, 1 Andrew L. Maas, Awni Y . Hannun, A. Y . N. 2013, International Conference on Machine Learning, 30

    work page 2015

  6. [6]

    Atwood, W. B. et al. 2009, The Astrophysical Journal, 697, 1071

    work page 2009

  7. [7]

    Bellm, E. C. et al. 2018, Publications of the Astronomical Society of the Pacific, 131, 018002

    work page 2018

  8. [8]

    & Arnouts, S

    Bertin, E. & Arnouts, S. 1996, Astronomy and astrophysics supplement series, 117, 393

    work page 1996

  9. [9]

    2023, Journal of Cosmology and As- troparticle Physics, 2023, 013

    Caron, S., Eckner, C., Hendriks, L., et al. 2023, Journal of Cosmology and As- troparticle Physics, 2023, 013

    work page 2023

  10. [10]

    Chollet, F. et al. 2015, Keras,https://keras.io

    work page 2015

  11. [11]

    T., & Shah, M

    Croitoru, F.-A., Hondru, V ., Ionescu, R. T., & Shah, M. 2023, IEEE Transactions on Pattern Analysis and Machine Intelligence, 45, 10850–10869

    work page 2023

  12. [12]

    2022, Monthly Notices of the Royal Astronomical Society, 515, 5185

    Ehlert, S., Chen, C.-T., Swartz, D., et al. 2022, Monthly Notices of the Royal Astronomical Society, 515, 5185

    work page 2022

  13. [13]

    & Vazza, F

    Gheller, C. & Vazza, F. 2021, Monthly Notices of the Royal Astronomical Soci- ety, 509, 990

    work page 2021

  14. [14]

    J., Bloemen, S., Vreeswijk, P., et al

    Groot, P. J., Bloemen, S., Vreeswijk, P., et al. 2024, Publications of the Astro- nomical Society of the Pacific, 136, 115003

    work page 2024

  15. [15]

    Haigh, C. et al. 2021, Astronomy & Astrophysics, 645, A107

    work page 2021

  16. [16]

    Heymans, C. et al. 2012, Monthly Notices of the Royal Astronomical Society, 427, 146 Ivezi´c, Ž. et al. 2019, The Astrophysical Journal, 873, 111

    work page 2012

  17. [17]

    Jia, P. et al. 2023, Monthly Notices of the Royal Astronomical Society, 527, 6581 Jonathan Ho, Ajay Jain, P. A. 2020, Advances in Neural Information Processing Systems, 33, 6840

    work page 2023

  18. [18]

    Kingma, D. P. & Ba, J. 2014, arXiv preprint arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  19. [19]

    Oetelaar, C. et al. 2021, PoS, ICRC, 663

    work page 2021

  20. [20]

    Attention U-Net: Learning Where to Look for the Pancreas

    Oktay, O., Schlemper, J., Folgoc, L. L., et al. 2018, arXiv preprint arXiv:1804.03999

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  21. [21]

    2021, Astronomy & astrophysics, 656, A62

    Panes, B., Eckner, C., Hendriks, L., et al. 2021, Astronomy & astrophysics, 656, A62

    work page 2021

  22. [22]

    Robotham, A. et al. 2018, Monthly Notices of the Royal Astronomical Society, 476, 3137

    work page 2018

  23. [23]

    2015, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (Springer International Publishing), 234–241

    Ronneberger, O., Fischer, P., & Brox, T. 2015, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (Springer International Publishing), 234–241

    work page 2015

  24. [24]

    Roscani, V . et al. 2020, Astronomy & Astrophysics, 643, A43

    work page 2020

  25. [25]

    T., Jarvis, M., Mandelbaum, R., et al

    Rowe, B. T., Jarvis, M., Mandelbaum, R., et al. 2015, Astronomy and Comput- ing, 10, 121

    work page 2015

  26. [26]

    2015, in In- ternational conference on machine learning, PMLR, 2256–2265

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., & Ganguli, S. 2015, in In- ternational conference on machine learning, PMLR, 2256–2265

    work page 2015

  27. [27]

    Song, Y . et al. 2020, International Conference on Learning Representations

    work page 2020

  28. [28]

    Stetson, P. B. 1987, Publications of the Astronomical Society of the Pacific, 99, 191

    work page 1987

  29. [29]

    2022, Astronomy & Astrophysics, 662, A109

    Stoppa, F., Vreeswijk, P., Bloemen, S., et al. 2022, Astronomy & Astrophysics, 662, A109

    work page 2022

  30. [30]

    Tian, C. et al. 2020, Neural Networks, 131, 251

    work page 2020

  31. [31]

    2024, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1059, 169010

    Tung, Y .-C., Li, J., Hsiung, Y ., et al. 2024, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1059, 169010

    work page 2024

  32. [32]

    2017, Advances in neural information processing systems, 30 V ojtekova, A

    Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, Advances in neural information processing systems, 30 V ojtekova, A. et al. 2020, Monthly Notices of the Royal Astronomical Society, 503, 3204

    work page 2017

  33. [33]

    N., Jones, B

    Wolf, T. N., Jones, B. A., & Bowler, B. P. 2024, The Astronomical Journal, 167, 92

    work page 2024