Work

A self-supervised learning approach for astronomical images

Year: 2021
Type: dissertation
Abstract: Modern astronomical sky surveys are providing us with a flood of images with unusual characteristics, such as numerous channels, saturated signals, faint signals, uncertainties, and varying signal-to-noise ratios. The complexity and diversity of these images make them an adequate use case for deep convolutional neural networks. Moreover, they yield millions of detected objects whose classes are mostly unknown. Given this context, the main objective of this work is to investigate deep representation learning approaches for multichannel astronomical images, focusing on finding reasonable representations that do not require labeled data and that make use of some domain knowledge. A reasonable representation may be thought of as one that contains enough discriminative information, that can be later used for higher-level tasks such as object classification, outlier detection and clustering. We propose a self-supervised learning approach that makes use of astronomical properties (more specifically, magnitudes) of the objects in order to pretrain deep neural networks with unlabeled data. We choose the task of classifying galaxies, stars and quasars as a baseline for quantifying the quality of the learned representations, and empirically demonstrate that our approach yields results that are better than -- or at least comparable to -- a benchmark RGB model pretrained on ImageNet.
Language: English
Open Access status: green
OpenAlex ID: W4212980283
Cites: 90
Cited by: 1
Related to: 9