FDA_YOLOv8: refined small object detection in unmanned aerial vehicle imagery.

Small object detection (SOD) is essential for security monitoring in unmanned aerial vehicle (UAV) imagery. However, the inherently low effective resolution, weak semantic representation, and cluttered background of small objects pose significant challenges. Although deep learning methods have been widely applied to extract multi-scale features from UAV images, their performance remains limited by the small size of objects and complex scene variations. In addition, the constrained computational resources of UAV platforms make achieving both accuracy and efficiency in SOD even more challenging. In this study, we propose an efficient small-object detection method, called FDA_YOLOv8, which is developed based on the YOLOv8s baseline and is designed to accurately detect small objects in UAV images under low computational cost. First, a four-head detection architecture is designed by introducing an additional lightweight detection head to enhance the sensitivity of smaller objects. Second, the dynamic head (Dyhead) framework is integrated to improve the representation capability of the detection head. Third, a FasterNet block is embedded into the C2f module to form the C2f_FA architecture, which improves spatial feature extraction while reducing model complexity and computational cost. Furthermore, an efficient multi-scale attention (EMA) mechanism is incorporated into the C2f_FA module, yielding the C2f_FE structure for better feature discrimination. Experiments on the VisDrone2019 dataset show that FDA_YOLOv8 achieves an mAP of 45.8%, outperforming YOLOv8s by 5.2%, confirming its effectiveness in refined small object detection for UAV imagery.