BMDnet: an efficient UAV small object detection model via multi-scale feature enhancement and dynamic attention

Small object detection in UAV aerial images is challenged by significant scale variation, weak edge information, and complex backgrounds. To address these issues, an efficient model termed BMDnet is proposed based on YOLOv11n. The proposed method introduces a collaborative optimization framework from three aspects: feature fusion, feature enhancement, and detection optimization. Specifically, a bidirectional multi-scale feature fusion network (BIMAFPN) is designed to enhance cross-layer feature interaction; a multi-scale edge information enhancement module (MEIE) is introduced to strengthen fine-grained structural representation; and a dynamic attention head (DAH) is employed to adaptively refine feature representations via scale-aware, spatial-aware, and task-aware mechanisms. Experimental results on the VisDrone2019 dataset demonstrate that BMDnet improves mAP50 and mAP50–95 by 3.5 and 2.2 percentage points, respectively, with only a moderate increase in computational cost. Additional experiments on the HIT-UAV dataset further verify the generalization capability of the proposed model. The results indicate that BMDnet achieves an effective balance between detection accuracy and computational efficiency.