ET-YOLOv6: Tiny Object Detection Algorithm on UAV Targets Based on Star Operation and Cross-Level Fusion Mechanism

The proliferation of unauthorized Unmanned Aerial Vehicles (UAVs) has heightened security concerns, driving the need for advanced detection systems. YOLO-based algorithms show promise but struggle with detecting tiny UAVs (5x5 to 400x400 pixels) in complex backgrounds. To address these challenges, we introduce the Enhanced Tiny UAV-target YOLOv6n network (ET-YOLOv6n). Our contributions are threefold: Firstly, we enhance the backbone by introducing a C2f-Star Module, which replaces conventional blocks with Star Operation Blocks to improve feature extraction for tiny targets while reducing computational overhead. Secondly, in the neck, we propose a Cross-Two-Layer BiFPN that integrates features across both vertical and horizontal directions within the network graph, ensuring robust multi-scale feature fusion. Lastly, we augment the head with an additional P2 layer, leveraging detailed geometric information from the p2 feature map to predict extremely small objects more accurately. Empirical evaluations on the DUT Anti-UAV dataset demonstrate that our proposed ET-YOLOv6n achieves competitive and often superior performance in detecting tiny UAV targets compared to existing YOLO-based approaches, including both smaller and medium-sized models. Specifically, it outperforms the YOLOv6n baseline on mAP@0.5, increasing from 0.857 to 0.906, and on mAP@0.5:0.95, from 0.549 to 0.588. Notably, these improvements are achieved with over 30% fewer parameters and reduced computation flops, making ET-YOLOv6n not only more efficient but also highly effective for real-world applications.