Htfd-yolo: Small target detection in drone aerial photography based on YOLOv8s

• Published in: The Journal of supercomputing Vol. 81; no. 4; p. 545
• Main Authors: Sun, Yuheng, Lan, Zhenping, Sun, Yanguo, Guo, Yuepeng, Li, Xinxin, Wang, Yuru, Meng, Yuwei
• Format: Journal Article
• Language: English
• Published: New York Springer Nature B.V 01.03.2025
• Subjects: Accuracy; Aerial photography; Algorithms; Artificial intelligence; Cameras; Drone aircraft; Drones; Feature selection; Localization; Modules; Multisensor fusion; Semantics; Spatial data; Target detection; Target recognition; Unmanned aerial vehicles
• ISSN: 1573-0484; 0920-8542; 1573-0484
• DOI: 10.1007/s11227-025-07067-3

The detection of small targets has significant value in the field of unmanned aerial vehicle (UAV) vision, yet it is also subject to certain challenges, including the use of images that are too small, difficulties in distinguishing the target from the background, and the presence of target-intensive. This paper presents a novel YOLO-based method for detecting small targets, specifically tailored to UAV photography. Firstly, a detection head is formulated for small targets to provide higher-resolution feature mapping. Secondly, a three-scale feature fusion module is proposed as a means of fusing the network features with the underlying features. This is intended to improve the deep semantic feature fusion and shallow texture feature fusion, provide rich spatial information for different detection heads and address the issue of feature loss. Furthermore, a module for Feature Selection Guidance Module is proposed, which enhances the ability to discriminate small targets by combining the CNN and the nonlinear learning operator. Finally, Soft_NMS is introduced and combined with DIOU, and the DIOU_Soft_NMS algorithm is proposed as a replacement for the original nonextremely large value suppression method. This new algorithm solves target crowding effectively and overlapping. Experimental results show that exhibits superior detection performance in UAV aerial photography scenarios, achieving remarkable outcomes on the VisDrone2019 dataset. In the test set, mAP0.5 reached 45%, representing a 12.1% improvement in comparison with YOLOv8, while mAP0.5 − 0.95 reached 34.1%, indicating an 11.4% improvement in comparison with YOLOv8. This suggests that the method will have potential for use in practical tasks in the field of UAVs. Furthermore, the results provide a solid foundation for future related research.