Conventional object detectors exhibit limited accuracy when tasked with detecting small objects in UAV imagery, this paper introduces a high-precision small object detection algorithm for UAV Imagery (HSU-YOLO). Firstly, the Cross Stage Partial-Grouping Multi-Scale Feature Fusion (CSP-GMSF2) module is designed through grouped convolution and cross-scale fusion strategies to achieve multi-scale feature fusion, enabling the algorithum to better extract features of small objects. Furthermore, a lightweight detection head specifically designed for tiny object detection is proposed. To improve small object feature extraction and reduce the detection head's computational complexity, detail-enhanced convolution and parameter sharing are integrated. Moreover, the DySample operator is adopted to optimize the up sampling process by generating content-aware sampling points, significantly enhancing the feature reconstruction fidelity of small objects in complex environments. Evaluated on the VisDrone2019 dataset, the proposed algorithm demonstrates superior performance to YOLO11, achieving gains of 5.0% (mAP@0.5), 4.1% (mAP@0.5:0.95), 6.6% (precision), and 4.0% (recall), while simultaneously reducing parameter counts. Meanwhile, compared with the newly proposed small object detection algorithums, the proposed algorithum achieves superior recognition precision. These improvements indicate effective balancing of detection accuracy and inference efficiency for UAV small object detection, with substantial practical utility.
 

UAV,Multi-Scale Feature Fusion,Small Object Detection,Parameter Sharing,DySample