Detection and Classification of Astronomical Sources with Astro-RetinaNet in Crowded Stellar Fields

Abstract Upcoming next-generation sky surveys will detect a large number of faint objects with magnitudes larger than 25. When objects are crowded within a limited field of view, blending becomes unavoidable. Blending leads to the omission of many sources during photometry in these fields, which causes an underestimates of tens of percent in crowded fields, and remains a major challenge for existing source-extraction techniques. Although artificial neural networks have shown promising results in the detection and classification in wide-field surveys, they often fail with severely blended stars. We developed a robust deep learning model, Astro-RetinaNet, based on the RetinaNet algorithm to detect and classify blended sources in single-band astronomical images. After training and evaluating the performance of our network on simulated images, we find precisions of 0.96, 0.89, 0.70, 0.50, and 0.75 for single star, two-star, three-star, four-star, and five-or-more star blending cases, respectively, with star number density ∼22,000 stars arcmin −2 . We compare our method’s detection capability and completeness both on CSST-simulated NGC 2298 images and Hubble Space Telescope–observed M31 images. In crowded and noncrowded stellar fields of simulated NGC 2298, our results show that the model can recover 82% and 95% of sources, respectively, at a magnitude ( i band) of 25, while for SExtractor and Photutils, the completeness reduces to 20%, 59% and 60%, 88%, respectively. In the M31 case, as faint as 27 mag ( F 814 W ) in a crowded field, Astro-RetinaNet detects 2224 sources, significantly outperforming Photutils and SExtractor by factors of 3.4 and 7.1, respectively.