2024

Master’s Projects

@Physics/Astronomy

#diffusion models

#image processing

#deconvolution

 

🎉This project has been accepted at the ML4Physics workshop at NeurIPS!

Project Summary

This project tackles the problem of deconvolving astronomical images to uncover the intrinsic properties of celestial objects, especially in ground-based observations. It explores the use of diffusion models (DMs) and the Diffusion Posterior Sampling (DPS) algorithm to address this challenge. Score-based DMs, trained on high-resolution cosmological simulations, are used in a Bayesian framework to compute posterior distributions based on observations. By incorporating redshift and pixel scale, the method adapts to various datasets. Tests on Hyper Supreme Camera (HSC) data achieve resolutions comparable to Hubble Space Telescope (HST) images, while also quantifying uncertainties and identifying prior-driven features for scientific use.

 

Alexandre Boucaud
aboucaud@apc.in2p3.fr

  • Scientific Software Engineer at Laboratoire Astroparticule et Cosmologie (APC), CNRS

Alessio Spagnoletti
aspagnol@ens-paris-saclay.fr

Marc Huertas-Company

Projects in the same discipline