Enhanced YOLOv11n for small object detection in UAV imagery: higher accuracy with fewer parameters

Object detection in UAV imagery faces several challenges due to high-altitude aerial capture: targets are densely distributed, small objects account for a large proportion, and onboard computing power is limited, leading to low detection accuracy and high rates of false and missed detections. To address these issues, this paper proposes an improved YOLOv11 model. First, we design a Multiscale Edge-Feature Adaptive Selection (MSEAF) module in the backbone to effectively cope with the predominance of small objects and weak edge information. Second, we use the ScalCat and Scal3DC modules to reconstruct the neck and add a P2 small-object detection head, alleviating feature degradation in multiscale processing and improving the utilization of high-resolution information. Finally, we design a shared, reparameterized lightweight detection head (SRepD) to resolve the computational redundancy and insufficient feature fusion of conventional heads. On the VisDrone2019 dataset, compared with the YOLOv11n baseline, our model increases mAP50 and Precision by 4.6% while reducing the number of parameters by approximately 8.5%. On the TinyPerson dataset, which contains only two extremely small categories, our model improves mAP50 and Precision by 5.5% and 5.6%, respectively, with a 7.7% reduction in parameters relative to YOLOv11n. Compared with the larger YOLOv11s, our model achieves gains of 3.8% in mAP50 and 3.2% in Precision while using only 25% of its parameters, demonstrating cross-scale performance superiority. On the HazyDet dataset, the first UAV-perspective hazy scene dataset, our model also demonstrates superior detection performance compared to the YOLOv11n baseline.