An improved IoU loss for small object detection via dynamic focusing and adaptive auxiliary-box mechanism

Abstract

Low-altitude unmanned aerial vehicle (UAV) detection is critical in applications such as airport surveillance, disaster response, and defense operations. However, diverse UAV types, extremely small target sizes, and large-scale variations present severe challenges for fine-grained recognition in complex environments. To address these limitations, this paper proposes an efficient detection framework, named Unmanned Aerial Vehicle Detection YOLOv11 (UAVD-YOLOv11). Specifically, to address feature loss and scale variance for small targets, a high-resolution P2 detection layer is integrated with a bidirectional feature pyramid network, enabling adaptive multi-scale feature fusion and the retention of rich spatial detail. Furthermore, to improve bounding box regression accuracy, an Inner-WIoUv3 loss function is proposed by coupling the Wise-IoU v3 (WIoUv3) dynamic focusing mechanism with Inner-IoU auxiliary bounding boxes, thereby speeding up convergence and effectively handling low-quality samples. Experimental results on the public complex background dataset demonstrate that UAVD-YOLOv11 achieves 94.8% mAP@0.5 and 57.7% mAP@0.5:0.95, representing significant gains of 4.2% and 4.4% over the baseline YOLOv11 model, respectively.