To solve the problems of small object scale, dense object distribution, loss of texture details and unstable bounding-box regression in unmanned aerial vehicle aerial images, a lightweight small-object detection algorithm based on YOLO11n is proposed. First, an S2DResConv module is introduced into the shallow downsampling stage of the backbone network. Spatial rearrangement and residual convolution are used to preserve more edge and texture information of small objects. Second, an AFPN3Head progressive adaptive feature fusion detection head is constructed. The high-resolution P2 branch is added, P3 and P4 are retained, and the P5 branch with limited contribution to small objects is removed so that the limited computational budget can be transferred to the P2 branch. After this reconstruction, the stride tensor and grid-point generation in the TaskAlignedAssigner are synchronously updated, and P2, P3 and P4 participate in candidate generation with strides of 4,8 and 16, respectively. Finally, Shape-NWD regression loss is introduced, and Shape-IoU is used in positive sample matching to enhance localization stability. The results show that, on the VisDrone2019 dataset, the Precision, Recall, mAP@50 and mAP@50:95 of the improved model reach 48.3%, 38.3%, 38.6% and 23.1%, respectively, which are 4.2, 5.7, 5.8 and 4.0 percentage points higher than those of the baseline YOLO11n. The number of parameters is 2.45M. It is demonstrated that the proposed method improves the detection accuracy of small objects in UAV aerial images while maintaining low model complexity, which is suitable for UAV-side object detection scenarios sensitive to both accuracy and deployment cost.