Abstract
In UAV-captured images, the high density of objects and the large proportion of small targets pose significant challenges to YOLO-based object detection algorithms. This study presents an enhanced object detection framework derived from the YOLOv10s architecture, aiming to achieve superior detection accuracy. First, an Adaptive Progressive Feature Unification (APFU) module is proposed to effectively integrate multi-level feature representations, ensuring a balanced fusion of high-level semantic information from low-resolution features and fine-grained spatial details from high-resolution features. Second, a Feature Enhancement and Attention (FEA) module is introduced to adaptively recalibrate feature responses, emphasizing informative features while suppressing irrelevant noise and interference. Finally, based on these modules, the APFU-YOLOv10 network is built to effectively improve the network's perception ability of objects at different scales. Experimental results on the VisDrone dataset demonstrate the superior performance of the proposed algorithm: mAP@0.5 increased from 42.6% to 43.5%, a relative improvement of approximately 2.11%; mAP@0.5:0.95 improved from 25.4% to 26.2%, a relative increase of about 3.15%; recall improved from 0.410 to 0.416, further reducing missed detections and enhancing object coverage. The method achieves significant improvements in detection accuracy under medium to high IoU thresholds, validating the effectiveness of multi-scale feature fusion and adaptive attention mechanisms in small object detection for UAV imagery.