Aerial image object detection algorithm based on the improved YOLOv11n

To address the challenges of small target size, low resolution, and complex backgrounds in unmanned aerial vehicle (UAV) aerial imagery for ground object detection, the article conducted an advanced YOLOv11n-based small target detection (TD) network. The key enhancements are: 1) Introducing a module, the original C3K2 structure is replaced with the novel C3k2-DFF1 feature extraction module, strengthening the network’s ability to feature tiny-pixel objects; 2) To enhance small object detection in aerial imagery, a P2 detection layer was specifically incorporated. 3)The original detection head (DH) was replaced with a dynamic head module to improve performance on densely clustered small targets, along with optimization of the loss function. The traditional NMS-CIOU algorithm was superseded by Soft-NMS-SIOU, and the NWD loss function was integrated into the bounding box regression module. This specifically addresses the issue of overlapping detection boxes in dense scenarios. Evaluated on the Visdrone2019 dataset, the enhanced YOLOv11n model demonstrates superior performance over YOLOv8s, achieving tested mAP@0.5 and mAP@0.5:0.95 metrics of $45.65 \%$ and $30.29 \%$, respectively. Compared to its original configuration prior to these improvements, this represents significant increases of $17.14 \%$ and $14.11 \%$ for each metric. Crucially, these gains are achieved while maintaining a manageable computational complexity of 20.7 GFLOPs. This effectively improves target detection (TD) accuracy in unmanned aerial vehicle (UAV) aerial imaging scenarios while meeting real-time processing requirements. Ablation studies confirm the individual contributions of the four key enhancements: the C3K2-DFF1 structural module, the P2 detection layer (PredictionLayer2), the Dynamic Head, and the optimized loss function suite. Each component was validated as effectively boosting the network’s overall detection performance.