Deep Learning-based Detection and Segmentation of Edge-on and Highly Inclined Galaxies

TL;DR

This paper presents a deep learning pipeline using YOLOv5 and SCSS-Net to detect and segment edge-on galaxies in SDSS images, resulting in a catalog of approximately 8,000 galaxies with analyzed properties.

Contribution

The authors developed a novel deep learning method combining detection and segmentation for identifying edge-on galaxies, enabling efficient catalog creation from large astronomical datasets.

Findings

Detected 8,000 edge-on galaxies from 16,000 candidates.

Most galaxies have redshifts between 0.02 and 0.10.

Galaxies are predominantly red with low b/a ratios.

Abstract

Edge-on galaxies have many important applications in galactic astrophysics, but they can be difficult to identify in vast amounts of astronomical data. To facilitate the search for them, we have developed a deep learning algorithm designed to identify and extract edge-on galaxies from astronomical images. We utilised a sample of edge-on spiral galaxies from the Galaxy Zoo database, retrieving the corresponding images from the Sloan Digital Sky Survey (SDSS). Our dataset comprises approximately $16,000$ galaxies, which we used to train the YOLOv5 algorithm for detection purposes. To isolate galaxies from their backgrounds, we trained the SCSS-Net neural network to generate segmentation masks. As a result, our algorithm detected $8,000$ edge-on galaxies, for which we compiled a catalogue including their parameters obtained from the SDSS database. We describe the basic properties of our sample, finding that most galaxies have redshifts $0.02<z<0.10$, have low values of $b/a$ and are mostly red, which is expected from edge-on galaxies and is consistent with our training sample, as well as other literature. The cutouts of the detected galaxies can be used for future studies and the algorithm can be applied to data from future surveys as well.