GFPN-YOLOv8m: A Deep Learning Network for Astronomical Source Detection

In astronomical image processing, issues such as varying target source sizes, partial overlap between sources, and confusion between some sources and the background often cause a decrease in detection accuracy. To cope with these issues, this paper proposed an improved GFPN-YOLOv8m network. Specifically, Coordinate Attention is introduced into the backbone network, embedding positional data into the attention map to achieve more precise target localization and recognition. The Global Feature Pyramid Network (GFPN) is integrated into the neck network to fuse multi-scale features more efficiently. Moreover, an additional layer dedicated to detecting small objects has been incorporated into the detection head, enhancing the ability to detect small objects. Comprehensive experimental results reveal that the proposed network surpasses existing alternatives, achieving an accuracy of 0.984, a recall of 0.971, a mAP50 of 0.993, and a mAP50-95 of 0.779. Therefore, the improved GFPN-YOLOv8m network exhibits higher accuracy and robustness, providing astronomers with more detailed and comprehensive data.