JKAN-YOLO: A Method for Small Object Detection in UAV Aerial Images

Fast and accurate detection of small, dense, and background-complex objects in UAV aerial images has become a key challenge in the field of object detection. To address this issue, this paper proposes an improved model named JKAN-YOLO based on the YOLOv10s detection framework. The proposed method replaces part of the convolutional structures in YOLOv10s with JacobiKANConv modules to enhance the model's sensitivity to edge details and small-scale targets, while effectively reducing computational resource consumption. Experimental results on the VisDrone2019 dataset show that JKAN-YOLO improves [email protected], Precision, and Recall by 4.60%, 2.82%, and 2.61%, respectively, with [email protected] reaching 41.0%, and FLOPs reduced from 24.5 G to 22.1 G. On DOTA-v1. 0 and Tiny Person datasets, the model also demonstrates strong generalization ability, achieving [email protected] scores of 40.5% and 25.1%, respectively. The experimental results indicate that JKAN-YOLO significantly improves the detection accuracy and stability of small objects while maintaining detection speed, showing promising application potential in real-world scenarios.