Journal Article

YOLOX-LS: Strong Gravitational Lenses Detection in the DECaLS with Deep Learning

Li, Jianxi, Tu, Liangping, Miao, Jiawei, Liu, Yangyang, Lin, Gengqi, Zhao, Chenying

2025

Overview

The identification of strong gravitational lenses is essential to facilitate many studies in astronomy. The search for strong gravitational lenses has become more challenging because of their scientific value and their rarity. In this paper, we construct a data set for strong gravitational lensing searches that consist of known lenses and lens candidates from the Dark Energy Spectroscopic Instrument Legacy Imaging Surveys (the DESI Legacy Imaging Surveys) and the Dark Energy Survey (DES) and non-lenses from the Legacy Surveys Data Release 9 (DR9). We apply the YOLOX algorithm as the basic framework and improve it by selecting the optimal optimizer, activation function, attention mechanism, and loss function. The improved YOLOX-LS algorithm achieves 97.87%, 97.51%, 0.97, 96.8%, and 53.1% in the evaluation metrics of precision, recall, F1 score, mean average precision (mAP)@0.5, and mAP@0.5:0.95, respectively. Compared with the YOLOX model, it improves by 0.63%, 0.26%, and 0.6% in the three metrics of precision, recall, and mAP@0.5, respectively. This paper presents the results of the trained YOLOX-LS algorithm applied to 4.75 million cutout images. These images are centered on the central source with magz ≤ 20 in the Dark Energy Camera Legacy Survey footprint from DESI DR9. Finally, we find 1697 lenses, including 303 known lenses or candidates, and 1394 new candidates, among which there are 102 high-quality candidates. This further verifies that the YOLOX-LS algorithm proposed in this paper can be effectively applied to the search for strong gravitational lenses. All visual results are displayed online at https://github.com/Young-mmm/YOLOX-LS.

• Publisher: IOP Publishing
• Subject: Algorithms / Astronomy / Dark energy / Deep learning / Gravitational lenses / Polls & surveys / Recall / Sky surveys (astronomy)