基于CS-YOLOv5s的无人机航拍图像小目标检测  

Small target detection for UAV aerial images based on CS-YOLOv5s

作　　者：翁俊辉 成乐 黄曼莉 隋皓 朱宏娜[1] Weng Junhui;Cheng Le;Huang Manli;Sui Hao;Zhu Hongna(School of Physical Science and Technology,Southwest Jiaotong University,Chengdu 610031,China;School of Information Science and Technology,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学物理科学与技术学院,成都610031 [2]西南交通大学信息科学与技术学院,成都610031

出　　处：《电子测量技术》2024年第7期157-162,共6页Electronic Measurement Technology

摘　　要：无人机航拍图像存在小目标分布密集且目标尺度变化大等检测难点,本文提出一种面向无人机航拍图像小目标的跨尺度目标检测模型—CS-YOLOv5s。首先,在YOLOv5s基础上,引入小目标检测器,提高模型对小目标的捕捉能力;进一步,将最大池化分支嵌入上下文增强模块,提取并增强骨干网络尾部的深层特征,再注入PANet,实现深浅层特征有效融合和模型跨尺度检测能力的提升;同时采用SPDConv模块替换下采样卷积模块,实现无人机航拍图像中密集目标高效检测。实验表明,CS-YOLOv5s在数据集VisDrone2019达到42.0%mAP0.5,较基准模型提升9.8%,有效增强网络模型对无人机航拍图像小目标的识别能力,为无人机目标智能识别提供支撑。To address the challenges in detecting small targets with dense distribution and large-scale variations in UAV aerial images,a cross-scale target detection model for UAV aerial images,named CS-YOLOv5s,is proposed.Firstly,based on YOLOv5s,micro-object detector is utilized to improve the model ability for capturing small targets.Then,the max-pooling branch is embedded into the context augment model,extracting and enhancing deep feature maps at the tail of the backbone network.The PANet is injected to achieve effective fusion of deep and shallow features with enhancing the cross-scale detection capability.Furthermore,the down-sampling convolution module is replaced with the SPDConv module to achieve efficient detection of dense objects in UAV aerial images.Experiments demonstrate that CS-YOLOv5s achieves 42.0% mAP0.5 on the VisDrone2019 dataset,which is increased by 9.8%than that of the baseline model.Our model enhances the network ability to recognize small targets in UAV aerial images effectively,which provides a new way for intelligent targets recognition of UAV.

关 键 词：无人机航拍图像 YOLO 小目标检测器 上下文增强模块 SPDConv模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN919.8[自动化与计算机技术—计算机科学与技术]