Photometric source detection and classification algorithm based on YOLOv11 and HybridViT embedded with Astrostem

Abstract

In large-scale astronomical data classification, the accurate classification of galaxies, stars, and quasars typically relies on spectral labels. However, spectral data account for only a small fraction of all astronomical observations, making the automatic fine classification of massive photometric data a persistent challenge. To tackle this challenge, we propose a photometric celestial source detection and classification algorithm based on YOLOv11 and a Hybrid Vision Transformer (HybridViT) embedded with the Astrostem convolutional module. We first utilize YOLOv11 to detect and localize photometric celestial sources, then feed the detected sources into the HybridViT network for fine-grained classification. By introducing a task-adaptive convolutional stem between the raw input and the Patch Embedding layer, HybridViT enhances the extraction of local morphological features for both point sources and extended sources. Experimental results demonstrate that YOLOv11 achieves an mAP@0.5 of 0.979 and an mAP@0.5:0.95 of 0.897. On the SDSS DR16 dataset, HybridViT achieves three-class classification of stars, quasars, and galaxies with an accuracy of 0.892. On the Galaxy Zoo 2 (GZ2) dataset, it further achieves fine-grained morphological classification of six galaxy types—barred spiral galaxies, cigar galaxies, elliptical galaxies, intermediate galaxies, spiral galaxies, and irregular galaxies—with an accuracy of 0.940. The proposed model outperforms mainstream CNN baselines across all metrics. The HybridViT model incorporating Astrostem can accurately extract morphological features from photometric sources, effectively improving the recognition accuracy for various celestial types, and is well suited for intelligent classification of large-scale sky survey photometric data.