基于改进YOLOv11s的无人机航拍图像小目标检测模型

A Small Object Detection Model for UAV Aerial Images Based on an Improved YOLOv11s Approach

摘要: 无人机在巡查、农业监测等工程应用中具有重要的工程价值，但是无人机航拍图像目标尺寸小，分布密集、数量多，实际检测工作难度大，因此提高无人机航拍图像的小目标检测性能是当前重要而迫切的技术问题。本文提出了一种基于改进YOLOv11s的小目标检测模型，命名为RB-YOLOv11s。首先，设计了重参数化幽灵跨阶段高效聚合网络(RepGhostCSPELAN Net，简称RGNet)，该网络能够增强模型的表征能力，并且整合多层级特征，降低模型参数量和计算量。此外，以BiFPN-GLSA网络替换原有的路径聚合网络(PANet)，使得主干网络和颈部网络产生的特征层相互融合，增强模型对全局及局部空间信息的感知能力。在VisDrone2019无人机航拍图像数据集上的实验表明，RB-YOLOv11s模型展现出卓越性能，与原模型YOLOv11s相比，RB-YOLOv11s模型的计算量虽然小幅增加，但参数量下降了25.5%，检测精度提高了1.9%。本文模型有效解决了无人机航拍图像中小目标检测精度低的问题。

Abstract: Unmanned aerial vehicles (UAVs) hold significant engineering value in applications such as patrols and agricultural monitoring. However, UAV aerial images feature small target dimensions, dense distributions, and large quantities, making actual detection challenging. Therefore, improving small object detection performance in UAV aerial images is a critical and urgent technical issue. This paper proposes a small object detection model based on an improved YOLOv11s architecture, named RB-YOLOv11s. First, we designed the Reparameterized Ghost Cross-Stage Efficient Polymerization Network (RepGhostCSPELAN Net, abbreviated as RGNet). This network enhances the model’s representational capacity while integrating multi-level features, thereby reducing both the number of model parameters and computational requirements. Additionally, the original Path Aggregation Network (PANet) is replaced with the BiFPN-GLSA network. This enables the fusion of feature layers generated by the backbone network and the neck network, enhancing the model’s perception of both global and local spatial information. Experiments on the VisDrone2019 UAV aerial image dataset demonstrate that the RB-YOLOv11s model exhibits outstanding performance. Compared to the original YOLOv11s model, RB-YOLOv11s exhibits a slight increase in computational complexity but achieves a 25.5% reduction in parameters while improving detection accuracy by 1.9%. This model effectively addresses the issue of low detection accuracy for small objects in aerial drone imagery.