Drone Detection Using YOLOv5 With GhostNet Backbone and Adaptive Fourier Neural Operator (AFNO 2D)

The growing presence of unmanned aerial vehicles (UAVs) in civilian, commercial, and military sectors has raised urgent and widespread security concerns. Real-time and reliable UAV detection remains a formidable challenge owing to their small size, rapid movement, and visual similarity to birds and other airborne objects in dynamic environments. This study proposes a lightweight yet high-accuracy drone detection framework based on the YOLOv5 object detection model, enhanced with a GhostNet backbone and Adaptive Fourier Neural Operator (AFNO 2D). The proposed architecture integrates Ghost Bottleneck modules, Spatial Pyramid Pooling Fast (SPPF), and frequency-domain learning via AFNO 2D to significantly improve global feature extraction while maintaining a compact model footprint. Additionally, Depthwise Convolution (DWConv) and C3 Ghost modules further optimize multi-scale feature fusion, ensuring robust and precise detection performance across varying altitudes, lighting conditions, and complex backgrounds. Compared with existing lightweight variants of You Only Look Once (YOLO), the proposed model demonstrates superior accuracy, achieving mAP@50:80.4% and mAP@50:95.5% while utilizing fewer parameters (2.6M) and offering faster inference across two distinct datasets. These characteristics make the proposed framework promising for real-time aerial surveillance and anti-drone defense applications on resource-constrained embedded platforms.