The PAU Survey: Self-supervised image denoising for flux-preserving photometry

arXiv:2611.05447 · astro-ph.IM, astro-ph.CO, astro-ph.GA · Submitted 2026-11-03 · 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: "The PAU Survey".

Jocelyn: Here's the gist of the paper. The paper evaluates and improves Noise2Void (N2V) for denoising real astronomical images, focusing on preserving fluxes, reducing photometric scatter, and retaining source main features across a wide magnitude range Aims.

Vera: We evaluate and improve Noise2Void (N2V) for denoising real astronomical images, focusing on preserving fluxes, reducing photometric scatter, and retaining the source main features across a wide magnitude range Methods.

Jocelyn: We use real observations from the Physics of the Accelerating Universe Survey (PAUS) together with PAUS-like simulated images We first assess the standard N2V method and show that it systematically underestimates galaxy fluxes, especially for faint galaxies To mitigate this bias, we modify the N2V architecture with a bounded-output layer that limits large negative outputs We then fine-tune the model with low-rank adaptation (LoRA) and two flux-aware loss terms designed to preserve the flux scale and improve repeated-exposure flux consistency Results.

Vera: Standard N2V improves the visual appearance of the images, but introduces strong flux biases for faint galaxies For PAUS galaxies with 22 ≤ iAB < 23 in the NB645 band, the standard N2V flux-ratio distribution peaks far below unity, while the boundedoutput model restores the peak close to the original PAUS flux scale The flux-aware LoRA fine-tuning further reduces the width of the repeated-exposure flux-difference distribution to σ68 = 0.529 in NB645, corresponding to an effective 47% reduction in random flux scatter relative to the original PAUS uncertainty scale, while keeping the median flux ratio close to unity Conclusions.

Paper summary: Vera: We're looking at the paper "The PAU Survey: Self-supervised image denoising for flux-preserving photometry" and the core idea is that we need to fix systematic flux underestimation in noisy images from deep surveys like PAUS Jocelyn The thesis is that standard denoising methods fail because they introduce strong biases, especially for faint galaxies, which messes up quantitative measurements.

Subrahmanyan: They are using real observations from the Physics of the Accelerating Universe Survey together with simulated images to test this Vera They first show that standard Noise2Void systematically underestimates galaxy fluxes, particularly for faint galaxies Jocelyn This happens because pixel-wise denoising objectives don't align well with integrated photometric measurements.

Vera: To fix this, they modify the N2V architecture with a bounded-output layer to limit large negative predictions Subrahmanyan Then they fine-tune it using LoRA and two flux-aware loss terms designed to preserve the flux scale and improve repeated-exposure flux consistency Jocelyn This entire approach is about adapting self-supervised denoising to low signal-to-noise survey images while keeping the photometric information intact.

Subrahmanyan: The results show that the constrained model restores the peak of the flux-ratio distribution close to unity, and LoRA substantially improves repeated-exposure flux precision, with sigma sixty-eight equal to zero point five two nine in NB645 Vera That means they found a way to preserve the PAUS photometric scale.

Jocelyn: They also quantified how much scatter is reduced; for instance, the flux-difference distribution narrows significantly after the LoRA fine-tuning Vera This reduction in scatter translates into improved measurement precision across different wavelength ranges.

Conclusion: Vera: So, looking at this paper "The PAU Survey: Self-supervised image denoising for flux-preserving photometry," the authors are arguing that standard N2V isn't enough for unbiased faint-source photometry Jocelyn The main implication is that we have to evaluate these methods with photometric diagnostics instead of just image quality metrics.

Subrahmanyan: They conclude that flux-aware modifications combined with LoRA fine-tuning provide a practical route toward denoising low signal-to-noise survey images while preserving the photometric qualities needed for quantitative analysis Vera This means self-supervised denoising can be adapted to these kinds of observations, which is important for things like cosmological parameter inference.

Jocelyn: It’s about taking a technique that's good at image restoration and making it useful for real, noisy astronomical data where the underlying physics is what matters most Vera They aren't just recovering a perfect picture; they are recovering the flux information needed for accurate science.

J. Chen, M. Eriksen, D. Navarro-Gironés, P. Renard, E. Gaztanaga, J. Carretero, R. Casas, F. J. Castander, E Fernandez, J García-Bellido, R Miquel, C Padilla, E Sánchez, I Sevilla-Noarbe, P Tallada-Crespí

Institut de Física d’Altes Energies (IFAE) · Port d’Informació Científica (PIC) · Leiden University · Institute of Space Sciences (ICE, CSIC) · Institut d’Estudis Espacials de Catalunya (IEEC) · Institute of Cosmology and Gravitation, University of Portsmouth · Centro de Investigaciones Energéticas, Medioambientales y Tecnológicas (CIEMAT) · Instituto de Física Teórica CSIC/Universidad Autónoma de Madrid · Institució Catalana de Recerca i Estudis Avançats (ICREA)

astro-ph.IM, astro-ph.CO, astro-ph.GA

Submitted: 2026-11-03

Updated: 2026-11-03

Code: https://github.com/PAU-survey/nayo

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The paper evaluates and improves Noise2Void (N2V) for denoising real astronomical images, focusing on preserving fluxes, reducing photometric scatter…

Key concepts

Noise2Void (N2V)
N2V is a self-supervised method trained only on noisy observations. It works by masking parts of an image and training a network to predict the original noisy pixel values based on the surrounding context.
Bounded-Output Layer
This modification adds a layer after the main network output that limits how large negative predictions can be. This prevents the model from producing excessively low flux estimates, which helps mitigate systematic underestimation of faint galaxy fluxes.
LoRA-N2V Fine-tuning
This involves fine-tuning the model using two specific loss terms: a paired-difference loss (to improve repeated exposure consistency) and a photometric-bias loss (to prevent trivial flux rescaling). This specialized training ensures the denoising process preserves the actual flux information needed for quantitative analysis.

Terminology

Summary

The paper evaluates and improves Noise2Void (N2V) for denoising real astronomical images, focusing on preserving fluxes, reducing photometric scatter, and retaining source main features across a wide magnitude range Aims. p. 0 We evaluate and improve Noise2Void (N2V) for denoising real astronomical images, focusing on preserving fluxes, reducing photometric scatter, and retaining the source main features across a wide magnitude range Methods. p. 0 We use real observations from the Physics of the Accelerating Universe Survey (PAUS) together with PAUS-like simulated images We first assess the standard N2V method and show that it systematically underestimates galaxy fluxes, especially for faint galaxies To mitigate this bias, we modify the N2V architecture with a bounded-output layer that limits large negative outputs We then fine-tune the model with low-rank adaptation (LoRA) and two flux-aware loss terms designed to preserve the flux scale and improve repeated-exposure flux consistency Results. p. 0 Standard N2V improves the visual appearance of the images, but introduces strong flux biases for faint galaxies For PAUS galaxies with 22 ≤ iAB < 23 in the NB645 band, the standard N2V flux-ratio distribution peaks far below unity, while the boundedoutput model restores the peak close to the original PAUS flux scale The flux-aware LoRA fine-tuning further reduces the width of the repeated-exposure flux-difference distribution to σ68 = 0. p. 0529 in NB645, corresponding to an effective 47% reduction in random flux scatter relative to the original PAUS uncertainty scale, while keeping the median flux ratio close to unity Conclusions. p. 0 Our results demonstrate that denoising methods for astronomical surveys must be evaluated with photometric diagnostics, rather than image-quality metrics alone Flux-aware modifications to N2V reduce photometric biases and improve repeatedexposure consistency in PAUS, demonstrating that self-supervised denoising can be adapted to low-SNR survey images while preserving the flux information needed for quantitative analysis

Pruning redundant reasoning tokens

The context of the problem is that deep photometric surveys increasingly rely on low signal-to-noise imaging to detect and characterise faint galaxies, where noise degrades image quality and affects source detection, flux measurements, photometric-redshift estimation, and sample selection Image denoising methods can therefore improve the scientific exploitation of these data, provided that they preserve the photometric information encoded in the images Classical methods impose explicit image priors such as smoothness or sparsity together with assumptions about the noise properties In astronomy, however, high-precision measurements often avoids full image restoration because restoration priors can bias the quantities of interest Deep learning networks can learn image priors directly from training data (LeCun et al. 2015) Applying these methods to astronomical imaging is challenging for two main reasons: first, noiseless ground-truth images are generally unavailable for real astronomical observations, and second, many neural denoisers are trained or pre-trained on natural-image datasets whose image statistics differ markedly from those of astronomical images This motivates self-supervised denoising strategies that are trained directly on the target observations rather than on external clean reference images

How the training signal is built

The standard Noise2Void (N2V) method is a self-supervised imagedenoising method trained only on noisy observations, assuming pixel-wise independent, zero-mean noise and a spatially correlated signal N2V uses a blind-spot strategy where selected pixels are masked in the input and the network is trained to predict their original noisy values from the surrounding context Formally, for a masked pixel i, the model prediction is sˆi = fθ xˆRF(i) The training target is the original noisy value xi, and training minimises a mean-squared error over the masked pixels Probabilistic Noise2Void (PN2V) extends N2V by predicting a distribution for the underlying signal at each pixel, combining it with an explicit noise model p(xi si) The model is trained by minimising the negative loglikelihood of the observed pixels LPN2V = X i − log (1/K X K k=1 p(xi s k i)) Constrained N2V adds an exponential linear unit (ELU) after the final linear output layer to bound negative output values, mitigating systematic flux biases The constrained N2V model suppresses large negative predictions and thereby reduces flux loss, but it does not exploit repeated-exposures or directly optimise flux consistency LoRA-N2V introduces a flux-aware fine-tuning stage based on two aperture-flux losses: a paired-difference loss and a photometric-bias loss The total fine-tuning objective is LLoRA−N2V = Lpb + Lpd, where the two loss terms are of comparable magnitude in our experiments

Where the gains come from

The gains come from modifying the N2V architecture with a bounded-output layer and by fine-tuning the model with flux-aware loss terms For standard N2V, negative stochastic predictions can shift the recovered profile downward and underestimate the flux, especially for weak galaxy signals The constrained model suppresses large negative outputs and reduces the downward shift in the average profile, supporting that negative stochastic predictions contribute to flux loss observed for faint simulated galaxies LoRA-N2V substantially improves repeated-exposure flux precision while preserving the median PAUS flux scale to within a few percent Specifically, LoRA-N2V narrows the repeated-exposure flux-difference distribution, with σ68 = 0.529 in NB645 and σ68 = 0. p. 0 p. 0415 for the faintest galaxies The photometric bias loss (Lpb) prevents the network from reducing the paired-difference loss through a trivial rescaling of the fluxes The paired-difference loss (Lpd) improves the repeatability of flux measurements from duplicate exposures

Evaluation and validation

We test standard N2V on simulations where noiseless reference images allow direct assessment of image quality and flux recovery In simulated images, N2V achieves the highest PSNR and preserves the main galaxy structure without pronounced artefacts, whereas classical methods leave more visible residual structure When applied to PAUS observations, standard N2V systematically underestimates the fluxes of faint galaxies, with the flux-ratio distribution peaking at approximately 0.36 for faint PAUS galaxies in NB645 The constrained N2V peaks at approximately 0. p. 698, showing that the output constraint strongly reduces the systematic flux bias LoRA-N2V substantially improves repeated-exposure flux precision, with σ68 = 0. p. 6529 in NB645 and σ68 = 0.415 for the faintest galaxies Furthermore, LoRA-N2V reduces the scatter of the D4000n measurements by 33–37% in the faint-dominated full sample This reduction in scatter is consistent with improved measurement precision and suggests that denoising lowers the effective photometric error entering the D4000 estimator

Downstream application

We test whether LoRA-N2V improves the repeatability of PAUS flux measurements across the full PAUS narrowband filter set The denoised fluxes show consistently lower σ68 values than the original measurements, indicating improved repeated-exposure flux precision across the wavelength range An illustrative coadded SED example shows that LoRA-N2V reduces band-to-band fluctuations while preserving physically interpretable spectral features, such as the redshifted 4000 Å break As a downstream application, we estimate the D4000 spectral break index from the denoised fluxes following the method of Renard et al. (2022) The measurement scatter decreases by 33–37% in the faint-dominated full sample, consistent with the improved flux precision The band-dependence of Rf across the D4000 continuum windows accounts for the observed shift, which should therefore be treated as a residual systematic rather than a physical effect

Summary and implications

Our results show that standard N2V is not sufficient for unbiased faint-source photometry in PAUS, as it introduces a strong flux-dependent bias with systematic flux underestimation for faint sources The limitation reflects a deeper mismatch between pixel-level denoising objectives and integrated photometric measurements, where the network can favour smoother or lowerflux reconstructions when the galaxy signal is comparable to the noise Flux-aware modifications to N2V, combined with efficient LoRA finetuning, provide a practical route toward denoising low-SNR survey images while preserving the photometric qualities needed for quantitative analysis The method preserves the PAUS photometric scale and improves repeated-exposure consistency, rather than demonstrating recovery of the true astrophysical flux of each source Future work will assess whether this strategy can be transferred to larger photometric surveys with different observing conditions, image characteristics, and photometric pipelines

Acknowledgements

We thank Shaolin Su, Kai Wang, Guangzai Ye, Hanyue Guo, Laura Cabayol Garcia, and Zijian Wang for useful discussions and their kind help We thank Gabriele Marchello for bringing Noise2Void to our attention in the context of liquid-phase electron microscopy J. Chen was supported by the China Scholarship Council (File No. p. 0 202306040042) ME acknowledges support from the Spanish Plan Nacional project PID2023-152069NA-I00 J. Carretero acknowledges support from the grant PID24-15942NB-C44, funded by MICIU/AEI/10. p. 013039/501100011 3 and by ERDF/EU CP acknowledges support from the Spanish Plan Nacional project PID22-141079NB-C32 JGB acknowledges the support of the Spanish Agencia Estatal de Investigación through the Research Project PID24-15942NB-C43 [MICINN-FEDER], and the grant IFT Centro de Excelencia Severo Ochoa CEX2020-001007 EG acknowledges support from PID24-156844NB-C21 and PID22-138896NB from MICINN/MICIU/AEI (/10.13039/501100011 3, Maria de Maeztu (CEX2020-M) grant, which include ERDF funds from the European Union, and the MaX-CSIC Excellence Award MaX4-SOMMAICE FJC acknowledges support from PID22-141079NB-C31 and PID25-173734NB-C31 from MICINN/MICIU/AEI (/10.13039/501 6>

Data availability

The PAUS catalogue data are publicly available through the PAUS public data release at https://pausurvey.org/public-data-release/ The PAUS Master Catalog (Navarro-Gironés et al. p. 0 2024), available through CosmoHub, is the recommended access point for PAUS photometry and photoz data Raw and reduced PAUS images, together with catalogue metadata, are also available through the PAUS public data-release infrastructure The intermediate files, trained model weights, and denoised flux products generated for this work are available on Zenodo at https://doi. p. 5org/10.5281/zenodo. 2611.05447v1 [astro-ph. p. 0IM] 3 Nov 2026 The code developed in this work is publicly available at https://github. p. 2com/KitfungChan/DenoisingPAUS The PAUS-like simulated images were generated with the simulation pipeline described in Appendix B, which is available at https://github. p. 2

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. Bounded-output layer modification to Noise2Void (N2V) reduces systematic flux biases by limiting large negative predictions, which is crucial for preserving faint galaxy fluxes during denoising.

  2. Flux-aware fine-tuning using low-rank adaptation (LoRA) with paired difference and photometric bias losses improves repeated exposure flux precision by reducing the width of the d fden distribution to sigma 68 = 0.529 in NB645, corresponding to an effective 47% reduction in random flux scatter.

  3. The LoRA-N2V model can reduce image quality metrics like PSNR and SSIM while simultaneously preserving the median PAUS flux scale to within a few percent, as shown by the median flux ratio Rf = 1.031.

  4. LoRA-N2V consistently reduces sigma 68 values of the d fden distribution across all PAUS narrow-band filters, indicating that improved repeated-exposure flux precision is not restricted to a single band but extends across the full filter set.

  5. The analysis demonstrates that for faint galaxies (21.5 <= iAB < 23) in NB645, the constrained N2V model peaks at Rf approximately 0.98, mitigating the main systematic flux loss where standard N2V peaks far below unity at approximately 0.36.

  6. LoRA-N2V reduces the scatter of the D4000n measurements by 33% to 37% in the faint-dominated full sample, which is consistent with improved flux precision and demonstrates that denoising lowers the effective photometric error entering the D4000 estimator.

  7. The system can be used to estimate galaxy flux from central 12x12 pixel regions of low-SNR images, allowing for aperture flux measurements that are directly comparable to PAUdm fluxes.

Sources

Related papers