Abstract
BACKGROUND: UAV-based power line inspections offer a safer, more efficient alternative to traditional methods, but insulator detection presents key challenges: multiscale object detection and intra-class variance. Insulators vary in size due to UAV altitude and perspective changes, while their visual similarities across types (e.g., glass, porcelain, composite) complicate classification. METHODS: To address these issues, we introduce APF-YOLO, an enhanced YOLOv8-based model integrating the Adaptive Path Fusion (APF) neck and the Adaptive Feature Alignment Module (AFAM). AFAM balances fine-grained detail extraction for small objects with semantic context for larger ones through local and global pathways by integrating advanced attention mechanisms. This work also introduces the Merged Public Insulator Dataset (MPID), a comprehensive dataset designed for insulator detection, representing diverse real-world conditions such as occlusions, varying scales, and environmental challenges. RESULTS: Evaluations on MPID demonstrate that APF-YOLO surpasses state-of-the-art models with different neck configurations, achieving at least a +2.71% improvement in mAP@0.5:0.9 and a +1.24% increase in recall, while maintaining real-time performance in server-grade environments. Although APF-YOLO adds computational requirements, these remain within acceptable limits for real-world applications. Future work will optimize APF-YOLO for edge devices through techniques such as model pruning and lightweight feature extractors, enhancing its adaptability and efficiency. CONCLUSION: Combined with MPID, APF-YOLO establishes a strong foundation for advancing UAV-based insulator detection, contributing to safer and more effective power line monitoring.