Variational Inference for Deblending Crowded Starfields
Authors: Runjing Liu, Jon D. McAuliffe, Jeffrey Regier; Published in 2023, Volume 24(179):1−36
Abstract
In astronomical survey images, it is common for stars and galaxies to visually overlap. Deblending refers to the task of distinguishing and characterizing individual light sources in these images. This paper introduces StarNet, a Bayesian method designed to deblend sources in crowded star fields. StarNet utilizes recent advancements in variational inference, including amortized variational distributions and an optimization objective that targets an expectation of the forward KL divergence. Experimental results using SDSS images of the M2 globular cluster demonstrate that StarNet outperforms two competing methods, namely Probabilistic Cataloging (PCAT) and DAOPHOT. PCAT utilizes MCMC for inference, while DAOPHOT is a software pipeline employed by SDSS for deblending. Furthermore, the amortized inference approach employed by StarNet provides the necessary scalability to perform Bayesian inference on modern astronomical surveys.
[abs]