AstroYOLO: A hybrid CNN–Transformer deep-learning object-detection model for blue horizontal-branch stars.

This article focuses on the development and evaluation of AstroYOLO, a novel deep-learning object-detection model designed to identify blue horizontal-branch stars (BHBs) in photometric sky survey images, particularly from the Sloan Digital Sky Survey (SDSS). AstroYOLO integrates convolutional neural networks (CNNs) for local feature extraction with Transformer blocks to capture global feature dependencies, addressing limitations of traditional CNN-based models like YOLOv3 and YOLOv4. Tested on a dataset of 4,799 BHB objects with cutout image sizes of 352×352 and 512×512 pixels, AstroYOLO outperformed existing YOLO models in average precision metrics (AP@50, AP@75, AP@95), demonstrating improved accuracy in detecting small-scale BHB stars. Ablation studies further identified optimal model configurations, including the use of CSPDarknet53 as backbone and three Transformer blocks in the multi-scale feature fusion module, highlighting the model's suitability for efficient and accurate BHB detection in large-scale astronomical surveys.