V10Drone: Improved YOLOv10 with KAN for Small Object Detection in Aerial Images

Abstract

Object detection in aerial robots like small UAVs and drones are extensively applied in both military and civilian applications.However, accurate object detection remains challenging due to the high resolution and small object scales in drone-captured images. In this paper, an efficient and lightweight network V10Drone is proposed, and it is specifically designed for small object detection in aerial robots. Three key improvements are introduced to the network in detail. Firstly, convolutional Kolmogorov-Arnold Networks (KAN) layers are integrated into the backbone to improve feature extraction. Secondly, a bridge module is incorporated between the backbone and the neck to minimize loss in transmission of features. Thirdly, a new lightweight four-layer neck is redesigned by two different upsample and downsample modules to improve two abilities of multiscale feature fusion and small object detection. Experiments on the VisDrone dataset demonstrate that V10Drone achieves mAP50 of 43.4% and mAP of 36.3%, and these results exceed the baseline of the YOLOv10s model by 4.9% and 3.3%, respectively. In addition, V10Drone exhibits higher accuracy than the M-scale model.