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

arXiv:2510.04718 · astro-ph.IM, hep-ex · Submitted 2025-10-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "BGRem: A background noise remover for astronomical images based on a diffusion model".

Jocelyn: Here's the gist of the paper. BGRem introduces BGRem, an ML-based tool to remove background noise from astronomical images using a diffusion model with an attention U-Net as its backbone This diffusion-based model with attention U-Net as its backbone can remove the background noise from astronomical images effectively with negligible artefacts Pruning redundant reasoning tokens The paper presents BGRem, a machine learning tool designed to remove background noise from astronomical images, aiming to improve image quality and enhance the performance of subsequent analysis pipelines The authors demonstrate that BGRem can improve the accuracy in source detection of traditional pixel-based methods by removing complex background noise Furthermore, BGRem shows generalizability across different data domains, including optical images from various telescopes and simulated gamma-ray data representative of the Fermi-LAT telescope How the training signal is built BGRem employs a supervised ML scheme trained on simulated images to learn to denoise astronomical images over several diffusion steps The training data consisted of 5 simulated full-field images from the MeerLICHT telescope and three more test data images, generated using the GalSim simulation software with corresponding background components including spatially varying sky background, detector-level effects like Gaussian read noise and dark current, and Poisson noise fluctuations Input images undergo pre-processing steps such as subtracting the median pixel value to estimate the background and then normalization to ensure pixel values fall within a similar range to the training data Where the gains come from The performance of BGRem was compared against SourceExtractor (SExtractor), showing 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 The model's architecture utilizes an attention U-Net as its backbone, which combines the concepts of a UNet from Ronneberger et al.

Vera: (2015) and the attention mechanism from Vaswani et al.

Jocelyn: (2017) The diffusion process is additive, where the clean source image remains constant while scaled background noise is superimposed according to a sine-based scaling factor Limitations and future scope Despite its effectiveness, BGRem exhibits a morphological bias where the Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component Consequently, BGRem may attenuate scientific signals from extended sources and is currently unsuitable for tasks requiring surface photometry or morphological analysis of such objects Additionally, the model tends to underestimate the flux of faint objects because it may misidentify low-level signals as stochastic background and remove them The current iteration fails to model and remove Poisson noise in gamma-ray data, and future work will focus on optimizing hyperparameters for better performance under the stochastic nature of poisson noise BGRem can serve as an effective pre-processing step for existing methods for building a pipeline to create an astronomical catalog 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 The computation time does not depend on the contents of the image, like the number of stars/galaxies in an image The first time a diffusion-based denoising model was shown to be effective in removing background from optical and gamma-ray images and help SExtractor to detect more sources The potential application of BGRem at a different wavelength, in γrays, with different background contaminations than optical images is highlighted References Abadi, M., Agarwal, A., Barham, P., et al.

Vera: 2015, TensorFlow: Large-scale machine learning on heterogeneous systems Abdollahi, S.

Title and authors: Vera: So we're looking at this paper titled BGRem: A background noise remover for astronomical images based on a diffusion model. It seems like they've put together a machine learning tool to take noisy astronomical pictures and clean them up using something called a diffusion model with an attention U-Net as its core structure.

Jocelyn: Right, it's about using this diffusion process over several steps to effectively remove background noise from those images without creating much of any unwanted artifacts. It’s basically trying to make the signal clearer for whatever comes next in the analysis pipeline.

Subrahmanyan: It sounds like they are tackling a fundamental problem: how to reliably separate faint astronomical sources from the complex, noisy background that is always present in the data we observe. I wonder what kind of noise characteristics this model can handle across different wavelengths or data types.

Vera: The paper shows that this tool improves source detection accuracy when compared to traditional pixel-based methods, specifically mentioning a seven percent increase in true positive sources using SExtractor for MeerLICHT data when BGRem was used as a pre-processing step.

Jocelyn: That's the practical result—showing real gains in what astronomers can actually find in the sky, which is pretty compelling when you’re looking at survey photometry or catalogs.

Subrahmanyan: A seven percent increase sounds modest for a pre-processing step, but if this holds up across different telescope data, it could significantly boost the reliability of large-scale surveys. What kind of generalizability are they talking about here?

Vera: They are demonstrating that BGRem works across different data domains, including optical images from various telescopes and simulated gamma-ray data that mimics what you’d expect from the Fermi LAT telescope observation.

Jocelyn: So it's not just a tool for one specific telescope or one type of image; it seems adaptable to different noise characteristics, which is important when dealing with multi-wavelength surveys.

Subrahmanyan: Adaptability across optical and gamma-ray data suggests the underlying principles are robust enough to handle very different types of noise contamination, like the interstellar emission you mentioned in simulations. How does this model actually learn this variety?

Vera: The training signal for BGRem is built using a supervised machine learning scheme trained on simulated images over several diffusion steps. They used five simulated full-field images from the MeerLICHT telescope for training, plus three more test data images generated with the GalSim software.

Jocelyn: So they're simulating the noise—like spatially varying sky background and detector-level effects such as Gaussian read noise and dark current—and teaching the AI to denoise over those steps. That’s a solid setup for supervised learning.

Subrahmanyan: They pre-process the input images by subtracting the median pixel value to estimate that background, then they normalize everything so the pixel values fall into a similar range as their training data, which helps standardize things before the model starts working.

Vera: And architecturally, they use an attention U-Net as its backbone, which combines the structure of a UNet from Ronneberger et al. two thousand fifteen with an attention mechanism from Vaswani et al <ref:2510.04718#pg1>. two thousand seventeen <ref:2510.04718#pg1>. This architecture is key to how it handles the denoising process.

Jocelyn: The diffusion process itself is additive; they keep the clean source image constant while scaling the background noise up according to a sine-based scaling factor, which sounds like a very controlled way to introduce and remove noise iteratively.

Subrahmanyan: I’m interested in that scaling factor idea. It means the model knows exactly how much background noise to add or subtract at each step based on where it is in the diffusion process. That kind of iterative control over noise magnitude is interesting for theoretical modeling of signal separation.

Vera: Now, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on. That highlights where the complexity of different noise regimes still leaves a gap in this first version.

Vera: So, looking at what this paper suggests for future work, it seems the focus will shift toward optimizing those hyperparameters to handle the stochastic nature of Poisson noise more effectively in gamma-ray data.

Jocelyn: It’s clear that while they achieved some great cross-wavelength performance, there’s still a lot left to do to make it a truly general-purpose framework for all astronomical surveys.

Subrahmanyan: Ultimately, the conclusion is that BGRem can serve as an effective pre-processing step for building catalogs and the success in both simulated gamma-ray and optical data demonstrates its adaptability across very different observational domains.

Vera: It’s a big deal because it shows that this diffusion-based denoising model can actually help improve the accuracy of source detection when applied to existing methods, which is a practical win for pipeline building.

Jocelyn: And the computation time doesn't depend on what's in the image, like how many stars or galaxies are present, which is another nice quality for a tool you want to use on large surveys.

Subrahmanyan: The fact that its computational time isn't dependent on the contents of the image is important because it means we can apply this model across enormous datasets without worrying about scaling issues related to the object density.

Vera: So, to wrap up, BGRem is presented as a promising pre-processing step for catalog building, showing adaptability and improving detection sensitivity when applied to simulated gamma-ray images alongside optical data.

Jocelyn: It’s a significant first time we’re seeing a diffusion-based denoising model work effectively for background removal in both optical and gamma-ray images, which opens up new possibilities for cross-wavelength analysis.

Subrahmanyan: The potential application of BGRem at different wavelengths, with different background contaminations than optical images, really highlights its promise as a general framework for multi-wavelength surveys.

Vera: It’s a lot of work they did here on the paper "BGRem: A background noise remover for astronomical images based on a diffusion model." We’ve seen how this AI tool can improve source detection by about seven percent and adapt to different data types.

Jocelyn: Yeah, it’s definitely something worth keeping an eye on as they move into optimizing it for those harder noise types like Poisson noise in gamma-ray data.

Subrahmanyan: I'm looking forward to seeing if this framework can actually be applied successfully to real observational data that has those kinds of complex background components.

The paper's summary: Vera: So, this paper introduces BGRem, which is this machine learning tool designed to clean up astronomical images by removing background noise using a diffusion model with an attention U-Net backbone.

Jocelyn: Right, so they're using this whole diffusion process over several steps to effectively denoise those noisy pictures without creating much of any unwanted artifacts. It’s all about making the signal clearer for whatever comes next in the analysis pipeline.

Subrahmanyan: It sounds like they’re tackling that fundamental problem of separating faint astronomical sources from the complex, noisy background we always see in our data observations. I wonder what kind of noise characteristics this model can actually handle across different wavelengths or data types.

Vera: The authors show that BGRem improves source detection accuracy when compared to traditional pixel-based methods, specifically mentioning a seven percent increase in true positive sources using SExtractor for MeerLICHT data when BGRem was used as a pre-processing step.

Jocelyn: That's the practical result—showing real gains in what astronomers can actually find in the sky, which is pretty compelling when you’re looking at survey photometry or catalogs.

Subrahmanyan: A seven percent increase sounds modest for a pre-processing step, but if this holds up across different telescope data, it could significantly boost the reliability of large-scale surveys. What kind of generalizability are they talking about here?

Vera: They're demonstrating that BGRem works across different data domains, including optical images from various telescopes and simulated gamma-ray data that mimics what you’d expect from the Fermi LAT telescope observation.

Jocelyn: So it's not just a tool for one specific telescope or one type of image; it seems adaptable to different noise characteristics, which is important when dealing with multi-wavelength surveys.

Subrahmanyan: Adaptability across optical and gamma-ray data suggests the underlying principles are robust enough to handle very different types of noise contamination, like the interstellar emission you mentioned in simulations. How does this model actually learn this variety?

Vera: The training signal for BGRem is built using a supervised machine learning scheme trained on simulated images over several diffusion steps. They used five simulated full-field images from the MeerLICHT telescope for training, plus three more test data images generated with the GalSim software.

Jocelyn: So they're simulating the noise—like spatially varying sky background and detector-level effects such as Gaussian read noise and dark current—and teaching the AI to denoise over those steps. That’s a solid setup for supervised learning.

Subrahmanyan: They pre-process the input images by subtracting the median pixel value to estimate that background, then they normalize everything so the pixel values fall into a similar range as their training data, which helps standardize things before the model starts working.

Vera: And architecturally, they use an attention U-Net as its backbone, which combines the structure of a UNet from Ronneberger et al. and an attention mechanism from Vaswani et al. This architecture is key to how it handles the denoising process.

Jocelyn: The diffusion process itself is additive; they keep the clean source image constant while scaling the background noise up according to a sine-based scaling factor, which sounds like a very controlled way to introduce and remove noise iteratively.

Subrahmanyan: I’m interested in that scaling factor idea. It means the model knows exactly how much background noise to add or subtract at each step based on where it is in the diffusion process. That kind of iterative control over noise magnitude is interesting for theoretical modeling of signal separation.

Vera: Now, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on. That highlights where the complexity of different noise regimes still leaves a gap in this first version.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on. That highlights where the complexity of different noise regimes still leaves a gap in this first version.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on. That highlights where the complexity of different noise regimes still leaves a gap in this first version.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

The paper's improvements: Tom: So, we just covered how BGRem is basically an AI tool that cleans up astronomical images using diffusion models to remove background noise from both optical and gamma-ray data.

Vera: Right, so the authors are suggesting ways they could improve this process further by looking at its existing limitations and seeing where the next steps should go.

Jocelyn: They point out that while it works well for optical data, there's a problem with extended sources, like large galaxies.

Subrahmanyan: That means BGRem might actually weaken the signal from those big structures, which is a real issue for structure formation studies if we’re trying to map them accurately.

Vera: And they also flagged that the model tends to miss faint things because it might mistake low-level signals as background noise and just get rid of them.

Jocelyn: That risk for finding those rare, faint objects is pretty high if the AI decides something is just noise before we can see it.

Subrahmanyan: I also saw they mentioned that this first version doesn't really tackle the Poisson noise specifically in gamma-ray data, even though they plan to adjust the model settings later to handle that stochastic nature.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on.

Vera: So, looking at what they found, they noted a morphological bias where this Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component when trying to remove noise.

Jocelyn: That’s a significant limitation because it means BGRem might actually attenuate scientific signals from sources that are naturally extended, like large galaxies or certain types of structures.

Subrahmanyan: If it struggles with extended sources, then its current application might be limited for tasks requiring surface photometry or detailed morphological analysis of those objects, which is a big constraint in structure formation studies.

Vera: Furthermore, the paper points out that this model tends to underestimate the flux of faint objects because it might misidentify low-level signals as stochastic background and remove them entirely.

Jocelyn: That sounds like a risk for finding very faint sources; if the AI decides something is just noise, it gets wiped out before we can see it, which is a problem when we’re hunting for rare objects.

Subrahmanyan: I also noticed the paper mentions that the current iteration doesn't actually model and remove Poisson noise in gamma-ray data specifically, though they plan to optimize hyperparameters for that later on.

Conclusion: Vera: So to wrap up, BGRem is essentially an AI tool that uses a diffusion model to clean up astronomical images by removing background noise from both optical and gamma-ray data, showing improvements in source detection efficiency on MeerLICHT data.

Jocelyn: It’s a powerful pre-processing step for building those catalogs we talk about, especially since it shows adaptability across different observational domains like optical and simulated gamma-ray observations.

Subrahmanyan: And this capability to handle distinct noise characteristics really highlights its potential as a general framework for multi-wavelength surveys across the whole sky.

Vera: The main thing is that this diffusion-based denoising model can actually help improve the accuracy of source detection when applied to existing methods, which is a practical win for pipeline building.

Jocelyn: It’s a significant first time we’re seeing a diffusion-based denoising model work effectively for background removal in both optical and gamma-ray images, opening up new possibilities for cross-wavelength analysis.

Subrahmanyan: The potential application of BGRem at different wavelengths, with different background contaminations than optical images, really highlights its promise as a general framework for multi-wavelength surveys.

Vera: It’s a lot of work they did here on the paper "BGRem: A background noise remover for astronomical images based on a diffusion model," and it shows how this AI tool can improve source detection by about seven percent.

Jocelyn: Yeah, it’s definitely something worth keeping an eye on as they move into optimizing it for those harder noise types like Poisson noise in gamma-ray data.

Subrahmanyan: I'm looking forward to seeing if this framework can actually be applied successfully to real observational data that has those kinds of complex background components.

Institute for Mathematics, Astrophysics and Particle Physics (IMAPP) Radboud University Nijmegen Department of Physics University of Oxford Centre for Astrophysics and Cosmology University of Nova Gorica Instituto de Física Corpuscular CSIC-UV Department of Astronomy and Inter-University Institute for Data Intensive Astronomy University of Cape Town South African Astronomical Observatory

astro-ph.IM, hep-ex

Submitted: 2025-10-06

Updated: 2026-05-11

Comments: Accepted version in A&A before any editorial changes

Journal ref: A&A 710, A131 (2026)

DOI: 10.1051/0004-6361/202557715

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: BGRem introduces BGRem, an ML-based tool to remove background noise from astronomical images using a diffusion model with an attention U-Net as its backbone This diffusion-based model with attention…

Key concepts

Diffusion Model
This is a type of generative AI model used here to learn how to remove noise. It works by gradually adding noise to an image until it becomes pure static, and then learning the reverse process—denoising the noisy image back into a clean version. This iterative process allows BGRem to effectively separate faint astronomical signals from complex background noise.
Attention U-Net Backbone
The core architecture of BGRem is an attention U-Net, which combines the structure of a standard UNet with an attention mechanism. The U-Net structure is excellent for image processing tasks like denoising because it captures both local details and global context. The attention mechanism helps the model focus on the most relevant parts of the image when deciding what noise to remove.
Supervised ML Scheme
The model was trained using a supervised machine learning approach. This means it learned by being shown many examples of images that already had noise removed (the training data). The training data included simulated astronomical images with known background components, such as sky background and detector noise, allowing the model to learn the specific patterns associated with astronomical signals versus unwanted noise.
Morphological Bias
This is a limitation where the Attention U-Net backbone tends to incorrectly interpret extended objects, like diffuse galaxies, as part of the background. Because it focuses on removing background noise, it can mistakenly attenuate or remove real scientific signals from these extended sources. This means BGRem might not be suitable for detailed morphological studies of large structures.

Terminology

Summary

BGRem introduces BGRem, an ML-based tool to remove background noise from astronomical images using a diffusion model with an attention U-Net as its backbone This diffusion-based model with attention U-Net as its backbone can remove the background noise from astronomical images effectively with negligible artefacts

Pruning redundant reasoning tokens

The paper presents BGRem, a machine learning tool designed to remove background noise from astronomical images, aiming to improve image quality and enhance the performance of subsequent analysis pipelines The authors demonstrate that BGRem can improve the accuracy in source detection of traditional pixel-based methods by removing complex background noise Furthermore, BGRem shows generalizability across different data domains, including optical images from various telescopes and simulated gamma-ray data representative of the Fermi-LAT telescope

How the training signal is built

BGRem employs a supervised ML scheme trained on simulated images to learn to denoise astronomical images over several diffusion steps The training data consisted of 5 simulated full-field images from the MeerLICHT telescope and three more test data images, generated using the GalSim simulation software with corresponding background components including spatially varying sky background, detector-level effects like Gaussian read noise and dark current, and Poisson noise fluctuations Input images undergo pre-processing steps such as subtracting the median pixel value to estimate the background and then normalization to ensure pixel values fall within a similar range to the training data

Where the gains come from

The performance of BGRem was compared against SourceExtractor (SExtractor), showing 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 The model's architecture utilizes an attention U-Net as its backbone, which combines the concepts of a UNet from Ronneberger et al. (2015) and the attention mechanism from Vaswani et al. p. 3

Where the paper says it

This combines the successful concepts of a UNet from Ronneberger et al. (2015) and the attention mechanism from Vaswani et al.

(2017) The diffusion process is additive, where the clean source image remains constant while scaled background noise is superimposed according to a sine-based scaling factor

Limitations and future scope

Despite its effectiveness, BGRem exhibits a morphological bias where the Attention U-Net backbone tends to interpret diffuse, extended galaxies as part of the background component Consequently, BGRem may attenuate scientific signals from extended sources and is currently unsuitable for tasks requiring surface photometry or morphological analysis of such objects Additionally, the model tends to underestimate the flux of faint objects because it may misidentify low-level signals as stochastic background and remove them The current iteration fails to model and remove Poisson noise in gamma-ray data, and future work will focus on optimizing hyperparameters for better performance under the stochastic nature of poisson noise

Conclusion

BGRem can serve as an effective pre-processing step for existing methods for building a pipeline to create an astronomical catalog 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 The computation time does not depend on the contents of the image, like the number of stars/galaxies in an image The first time a diffusion-based denoising model was shown to be effective in removing background from optical and gamma-ray images and help SExtractor to detect more sources The potential application of BGRem at a different wavelength, in γrays, with different background contaminations than optical images is highlighted

References

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

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

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

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

Akhlaghi, M. & Ichikawa, T. p. 3

Where the paper says it

Several other recent alternatives to SExtractor, like ProFound (Robotham et al. 2018), NoiseChisel (Akhlaghi & Ichikawa 2015), were developed in recent years; however, it was shown through extensive analysis (Haigh et al.

2015, The Astrophysical Journal Supplement Series, 220, 1

Andrew L. Maas, Awni Y. p. 3

Where the paper says it

We used LeakyRelu (Andrew L. Maas 2013) as the activation function.

Hannun, A. Y. N., 2013 International Conference on Machine Learning, 30

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

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

Bertin, E. & Arnouts, S. p. 1

Where the paper says it

2018) as the backbone. To quantify and compare our results, we used the background subtraction method in SExtractor (Bertin & Arnouts 1996) for astronomical images.

, 1996 Astronomy and astrophysics supplement series, 117, 393

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

Chollet, F. et al. 2015 Keras

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

Dey, A., Schlegel, D. p. 9

Where the paper says it

RdA was supported by PID2020-113644GBI00 from the Spanish Ministerio de Ciencia e Innovación and by the PROMETEO/2022/69 from the Spanish GVA. The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; Proposal ID #2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Prop.

J., Lang, D., et al. 2019a The Astronomical Journal, 157, 168

Dey, A. et al. 2019b The Astronomical Journal, 157, 168

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

Gheller, C. & Vazza, F. p. 1

Where the paper says it

However, generalizing these models to real data, which always includes noise, remains a significant challenge. For astronomical images, a deep convolutional autoencoder network was shown to be effective for denoising radio images and detecting faint sources (Gheller & Vazza 2021).

, 2021 Monthly Notices of the Royal Astronomical Society, 509, 990

Groot, P. J., Bloemen, S., Vreeswijk, P. p. 9

Where the paper says it

The authors are grateful to P. Vreeswijk for helpful comments to improve the draft.

p. 9
Where the paper says it

The authors are grateful to P. Vreeswijk for helpful comments to improve the draft.

p. 9
Where the paper says it

The authors are grateful to P. Vreeswijk for helpful comments to improve the draft.

, et al. 2024 Publications of the Astronomical Society of the Pacific, 136, 115003

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

Heymans, C. et al. 2012 Monthly Notices of the Royal Astronomical Society, 427, 146

Huber, P., 1964 The Annals of Mathematical Statistics, 35, 73

Ivezic, Ž., et al. 2019 The Astrophysical Journal, 873, 111

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

Kingma, D. P. & Ba, J., 2014 arXiv preprint arXiv:1412. p. 0

Where the paper says it

ML models are highly dependent on the quality and characteristics of their training data. Many models are trained on simulated data where the ground Article number, page 1 of 15 arXiv:2510.

6980

Loshchilov, I. & Hutter, F. p. 3

Where the paper says it

Maas 2013) as the activation function. The model was trained for ∼ 100 epochs using the AdamW optimizer (Kingma & Ba 2014; Loshchilov & Hutter 2017) with a starting learning rate of 10−3 and a steadily decreasing learning rate of 0.

, 2017 in International Conference on Learning Representations

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

Oktay, O., Schlemper, J., Folgoc, L. L., et al. 2018 arXiv preprint arXiv:1804. p. 0

Where the paper says it

ML models are highly dependent on the quality and characteristics of their training data. Many models are trained on simulated data where the ground Article number, page 1 of 15 arXiv:2510.

03999

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

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

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

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

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

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

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

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

Stoppa, F., Bhattacharyya, S., de Austri, R. R., et al. 2023a Astronomy & Astrophysics, 680, A109

Stoppa, F., de Austri, R. R., Vreeswijk, P., et al. 2023b Astronomy & Astrophysics, 680, A108

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

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

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

Vaswani, A., Shazeer, N., Parmar, N., et al. 2017 Advances in neural information processing systems, 30

Vojtekova, A. et al. 2020 Monthly Notices of the Royal Astronomical Society, 503, 3204

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

Zackay, B. & Ofek, E. O., 2017a The Astrophysical Journal, 836, 187

Zackay, B. & Ofek, E. O.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Background noise removal for astronomical images can be performed by BGRem, which uses a diffusion-based model with an attention U-Net backbone to learn to denoise astronomical images over several diffusion steps. This allows the AI system to improve image quality and enhance the performance of subsequent analysis pipelines.

  2. The AI system can achieve improved source detection efficiency when used as a pre-processing step, as shown by the result 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.

  3. The system can be generalized to perform background removal on different types of astronomical data, including optical images from various telescopes and simulated gamma-ray data representative of the Fermi-LAT telescope, demonstrating its adaptability to distinct noise characteristics and observational domains.

  4. By using a zero-shot approach, the BGRem model can generalize well to a wide range of optical images, enabling it to be applied effectively on new datasets without extensive retraining.

  5. The system can be used for catalog building tasks by generating a high-precision detection mask which is then applied to original data, improving the overall accuracy of source detection compared to traditional pixel-based methods.

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 (γ)-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 γ-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 γ-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.

Sources

Related papers