YOLO-LiRa: lightweight detection algorithm for small aerial targets

Abstract With the rapid development of drone technology, aerial small target detection has become an important research direction in the field of computer vision. This paper proposes a lightweight object detection algorithm based on YOLOv11, YOLO-LiRa (Lightweight Intelligent Remote-sensing Algorithm), to address the issues of small size, complex background, and dense targets faced by small object detection in aerial images. Incorporating Monte Carlo attention mechanism and partial convolution into the C3K2 module enhances the extraction of small target features and reduces computational complexity. Referring to the lightweight design concept of MobileNetV3 to optimize the backbone network, combined with GSConv and VoVGSCSP modules, the multi-scale feature fusion capability of the neck network is enhanced. Moreover, the feature map resolution and detection performance were optimized using the DySample upsampling operator. Experiments on the publicly available AI-TOD datasets have shown that YOLO-LiRa achieves 0.27 on the mAP50-95 evaluation metric, reducing the parameter count by 24.4% and computational complexity by 12.5% compared to the Baseline model, while achieving a good balance between detection accuracy and speed. Compared with other mainstream object detection algorithms, YOLO-LiRa exhibits more competitive performance in small object detection tasks. The method proposed in this article is suitable for applications such as unmanned aerial vehicle monitoring, intelligent transportation, and agricultural monitoring that require high lightweight and real-time performance.