Abstract: UAV aerial images suffer from blurred object boundaries, complex background interference, dense occlusion, and high localization sensitivity caused by center offsets. To address these problems, this paper proposes a lightweight UAV small object detection algorithm based on YOLOv8n, named FHG-YOLO. First, Feature Complementary Mapping (FCM) and Progressive Multi-kernel Depthwise Convolution (PzConv) are introduced into the backbone network to enhance the complementary interaction between spatial details and semantic features, thereby reducing the loss of small-object information during downsampling. Second, a Dynamic Selective Spatial Pyramid Pooling-Fast (DSSPF) module is designed to adaptively select effective contextual scales and enhance long-range contextual representation. The main innovations of this paper lie in the proposed Head Boundary-Context Mixer (HBCM) and WNW-Loss. Unlike conventional multi-branch feature fusion, HBCM establishes a collaborative mechanism among detail, context, and boundary information. Specifically, boundary responses are used to actively guide detail enhancement, encouraging the network to focus on discriminative texture regions around object contours. Meanwhile, contextual semantics are introduced to refine boundary responses and suppress misleading background edges. On this basis, a competitive branch fusion strategy is designed to adaptively allocate the contributions of detail, context, and boundary features according to the input feature distribution, enabling the model to dynamically select more effective information under different object scales, occlusion degrees, and background complexities. In addition, a WIoU-v2 guided NWD with Adaptive Weighting Loss, termed WNW-Loss, is proposed for bounding box regression. This loss function combines WIoU-v2 with Center-NWD and introduces scale and quality gates to strengthen the center constraint mainly for small-scale objects and low-quality samples, while avoiding unnecessary interference with high-quality regression samples. Experimental results on the VisDrone2019 dataset show that FHG-YOLO achieves 47.33% precision and 36.19% mAP50, which are 3.47 and 2.55 percentage points higher than those of YOLOv8n, respectively, while reducing the number of parameters by 65.8%. Experiments on the DOTA dataset further verify the generalization capability of the proposed method. The results demonstrate that FHG-YOLO improves small object detection accuracy and localization stability while maintaining a lightweight structure, making it suitable for resource-constrained UAV aerial image detection tasks.