IET Image Processing, Год журнала: 2025, Номер 19(1)

Опубликована: Янв. 1, 2025

ABSTRACT Object detection in Unmanned Aerial Vehicle (UAV) imagery plays an important role many fields. However, UAV images usually exhibit characteristics different from those of natural images, such as complex scenes, dense small targets, and significant variations target scales, which pose considerable challenges for object tasks. To address these issues, this paper presents a novel algorithm based on YOLOv8 (referred to OATF‐YOLO). First, orthogonal channel attention mechanism is added the backbone network imporve algorithm's ability extract features clear up any confusion between foreground background. Second, triple feature encoder scale sequence fusion module are integrated into neck bolster multi‐scale capability, thereby mitigating impact substantial differences scales. Finally, inner factor introduced loss function further upgrade robustness accuracy algorithm. Experimental results VisDrone2019‐DET dataset indicate that proposed significantly outperforms baseline model. On validation set, OATF‐YOLO achieves precision 59.1%, recall 40.5%, mAP50 42.5%, mAP50:95 25.8%. These values represent improvements 3.8%, 3.0%, 4.1%, 3.3%, respectively. Similarly, test 52.3%, 34.7%, 33.4%, 19.1%, reflecting enhancements 4.0%, 2.6%, validate model's scalability, experiments conducted NWPU‐VHR10 dataset, also excellent performance. Furthermore, compared several classical algorithms, demonstrates superior performance both datasets indicates it better suited image scenarios.

Язык: Английский