TOE-YOLO: Accurate and Efficient Detection of Tiny Objects in UAV Imagery

Abstract This paper proposes a redesigned YOLO-based model tailored for small object detection in UAV applications. To address the rotational characteristics of multi-scale objects in aerial imagery, this paper introduces a novel C3k2-ARC module designed to enhance the model’s capability in extracting features from rotated targets, thereby improving detection accuracy. Furthermore, we propose a novel feature fusion module named CL-Concat, which combines the traditional concatenation operation with channel attention and spatial attention mechanisms. This design significantly enhances the quality of multi-scale feature fusion. Experimental results demonstrate that the proposed approach achieves superior performance across multiple public benchmarks, with mAP improvements of 1.6% on the VISDRONE dataset, 1.4% on UAVDT, 0.2% on CARPK, and 0.1% on UAVROD.