Source-linked AI summary
Generative Adversarial Networks recover features in astrophysical images of galaxies beyond the deconvolution limit
Kevin Schawinski, Ce Zhang, Hantian Zhang, Lucas Fowler, Gokula Krishnan Santhanam
TL;DR
Astronomical galaxy images are degraded by noise, blurring, and limits on conventional deconvolution. This paper trains a GAN with simulated degradation to recover galaxy features, achieving better recovery than classic deconvolution and supporting applications to existing and future survey data.
Problem
Noise, telescope and atmospheric blurring, and the Shannon-Nyquist limit constrain conventional recovery of features from galaxy images.
Method
The method uses a GAN trained on 4,550 nearby-galaxy images with automatically simulated telescope and atmospheric degradation.
Results
The GAN reliably recovers features from degraded galaxy images and outperforms classic deconvolution approaches.
Takeaways & Limitations
With a suitable training set, the method can be applied to existing and future astrophysical imaging data, including SDSS, Euclid, LSST, Hubble, and James Webb observations.
Takeaways & Limitations
Recovery is constrained by the training set and the method was applied to nearby galaxies observed with the same camera under similar conditions.
Abstract
from arXiv · showhide
Observations of astrophysical objects such as galaxies are limited by various sources of random and systematic noise from the sky background, the optical system of the telescope and the detector used to record the data. Conventional deconvolution techniques are limited in their ability to recover features in imaging data by the Shannon-Nyquist sampling theorem. Here we train a generative adversarial network (GAN) on a sample of $4,550$ images of nearby galaxies at $0.01<z<0.02$ from the Sloan Digital Sky Survey and conduct $10\times$ cross validation to evaluate the results. We present a method using a GAN trained on galaxy images that can recover features from artificially degraded images with worse seeing and higher noise than the original with a performance which far exceeds simple deconvolution. The ability to better recover detailed features such as galaxy morphology from low-signal-to-noise and low angular resolution imaging data significantly increases our ability to study existing data sets of astrophysical objects as well as future observations with observatories such as the Large Synoptic Sky Telescope (LSST) and the Hubble and James Webb space telescopes.
1 INTRODUCTION
Astrophysical images are degraded by noise, limited angular resolution, and atmospheric seeing, while deconvolution faces fundamental limits as an ill-posed inverse problem. The paper addresses these limits by using machine learning to introduce image priors learned from higher-quality galaxy data.
- Telescope images combine true light distributions with PSF-induced blurring and noise from the detector, sky background, optics, and atmosphere.Angular resolution depends on the telescope’s resolving power and, for ground observations, atmospheric seeing.
- Deconvolution is an ill-posed inverse problem because backwards modelling can yield no unique solution.
- Machine learning can introduce domain priors automatically by learning to recover information from poor-quality images using higher-quality training data.The learned priors help the system choose among possible solutions beyond the stated deconvolution limitation.
2 METHOD
The method trains a conditional GAN on paired degraded and non-degraded galaxy images, with degradation simulated from telescope and atmospheric effects. It evaluates recovery across galaxy samples, degradation settings, qualitative examples, and PSNR-based comparisons with deconvolution.
- 2 METHOD: The conditional GAN learns image-to-image translation from degraded galaxy images to corresponding non-degraded images.Training minimizes the difference between the recovered and non-degraded image in each pair.
- 2 METHOD: Degraded training pairs are generated automatically by simulating blurring and noise, enabling weak supervision without human annotation.The framework is designed to scale to large future sky surveys.
- 2 METHOD: 4,550 SDSS galaxies at 0.01 < z < 0.02 are evaluated with 10× cross validation, using 4,105 training and 455 testing images per fold.The images are processed into three-band RGB inputs using an asinh stretch.
- 2 METHOD: Figure 1 uses original images to generate degraded inputs for GAN training, while testing uses only the generator to recover images.
- 2 METHOD: The experiments vary Gaussian PSF FWHM from 1.4 to 2.5′′ and increase noise relative to the original images.The tested PSF range spans the approximate median SDSS seeing to substantially worse seeing.
- 2 METHOD: Representative results cover spiral galaxies, early-type galaxies, and mergers, with recovered details including star-forming regions, dust lanes, spiral-arm shapes, and merger structure.
3 RESULTS
The GAN is evaluated quantitatively and qualitatively against classical deconvolution on artificially degraded galaxy images. It achieves substantially higher PSNR in noisy cases, while examples show both successful recovery and failures on rare objects.
- Quantitative evaluation: 19.9 dB PSNR is achieved by blind deconvolution in the noisy-image comparison.
- Quantitative evaluation: 18.7 dB PSNR is achieved by Richardson deconvolution in the same comparison.
- Qualitative evaluation: Qualitative examples compare original, degraded, GAN-recovered, and deconvolved images across spiral, early-type, and merging galaxies.The examples use various degradation levels and contrast the GAN with deconvolution.
- Failure cases: The GAN generally fails on rare objects absent or underrepresented in its training set because it cannot reconstruct unlearned features.The paper describes these failures as rare and presents examples in Figure 4.
4 DISCUSSION
The method’s recovery ability is constrained by the representativeness of its training data and by intrinsically subtle distortions. With suitable training sets, the authors propose applications across current and future astronomical surveys, including high-redshift studies using simulations.
- Limitations: The training set limits recovery, especially when applying nearby-galaxy training to intrinsically different high-redshift morphologies.The authors suggest simulations of galaxy formation at the appropriate epoch as one possible remedy.
- Limitations: Weak lensing shear and other subtle distortions are impossible to recover through this route because they are truly irrecoverable.
- Limitations: Rare objects absent from the training set may cause the method to fail.
- Applications: With a suitable training set, the method can be applied to existing and future data from SDSS, Euclid, and LSST.
- Applications: Training sets from simulations could support high-redshift observations with Hubble and James Webb space telescopes to analyze early-universe galaxies.
- Reproducibility: The authors make the code available and provide instructions for accessing a reproducible virtual machine.
APPENDIX A: FULL GAN TEST OUTPUT
Appendix A provides the full GAN evaluation across the tested PSF FWHM and noise grid. It shows one object in the paper, while the remaining outputs and 10× cross-validation results are available online.
- Test grid: The full evaluation spans PSF FWHM values of 1.4, 1.8, 2.0, and 2.5′′ and noise levels of 1.0, 1.2, 2.0, 5.0, and 10.0σoriginal.
- Displayed outputs: Figure A1 displays the complete test grid for one object, while the remaining outputs are provided online-only.
- Cross validation: The online full set includes results from 10× cross validation.