DAID-YOLO: Small Object Detection Algorithm for Drone Aerial Images

作者：Han Ping, Luo Jie, H. Zuo · 发表于：IEEE International Conference on Signal and Image Processing · 年份：2024 · DOI：10.1109/ICSIP61881.2024.10671547 · 被引用次数：3

Since the height and angle are not fixed when the drone is flying, the taken images have the faults of diverse scales, complex backgrounds, severe occlusions, meanwhile the objects are relatively small and dense. Aiming at the problem of poor object detection in drone aerial photography, the DAID-YOLO(Drone Aerial Images Detection YOLO) model is proposed. Firstly, the SimAM attention mechanism is introduce on the basis of YOLOv8s, and an additional small object detection head is added. In addition, a cross-layer multi-scale feature fusion structure is designed to increase the model's ability of extracting spatial and semantic information. Finally, the Focal EIOU is used to replace the original model's bounding box regression loss function CIOU. On the VisDrone2019 test set, the average mean accuracy of DAID-YOLO reaches 38.2%, which is 3.9% higher than that of the baseline method YOLOv8s. It also achieves higher detection accuracy compared with other mainstream detection methods. The results show that the DAID-YOLO algorithm has a better performance for drone aerial photography detection tasks.