MF-YOLOv8-based fine-grained detection of UAVs and birds for low-altitude security

Abstract

Aiming at the problem of false positives and missed detections induced by variable scales and similar textures between unmanned aerial vehicle (UAV) and bird targets in low-altitude security defense, existing detection models can hardly balance accuracy and real-time performance requirements. This paper proposes an MF-YOLOv8 model based on the YOLOv8 architecture, which aims to enhance multi-target discrimination capability under complex airspace conditions while showing potential for lightweight deployment on edge devices: 1) A BiFPN_Concat2 feature fusion module is designed to boost feature transmission of small targets through bidirectional cross-scale weighting; 2) A Multi-Spectral Channel Attention (MSCAAttention) mechanism is embedded to strengthen discriminative features of UAV rotors and bird feathers in the frequency domain; 3) A MobileViT-C2f hybrid backbone network is constructed to combine the merits of local perception and global context modeling. On a mixed dataset consisting of seven categories of UAVs and birds, the model achieves a detection precision of 92.6%, a recall of 92.1% and an mAP0.5 of 94.5%, with respective improvements of 3.7%, 3.6% and 3.1% over the baseline YOLOv8. It also outperforms advanced improved methods including YOLOv9 and YOLOv10. The proposed approach remarkably reduces the false positive rate of bird targets, delivering high-precision and real-time detection support for airspace security defense.