改进YOLOv5s的遥感图像目标检测  

A remote sensing image object detection algorithm with improved YOLOv5s

作　　者：赵文清[1,2] 康怿瑾 赵振兵 翟永杰[1] ZHAO Wenqing;KANG Yijin;ZHAO Zhenbing;ZHAI Yongjie(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,China;Engineering Research Center of the Ministry of Education for Intelligent Computing of Complex Energy System,Baoding 071003,China;School of Electrical and Electronic Engineering,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学控制与计算机工程学院,河北保定071003 [2]复杂能源系统智能计算教育部工程研究中心,河北保定071003 [3]华北电力大学电气与电子工程学院,河北保定071003

出　　处：《智能系统学报》2023年第1期86-95,共10页CAAI Transactions on Intelligent Systems

基　　金：河北省自然科学基金项目(F2021502013);中央高校基本科研业务费面上项目(2020MS153,2021PT018);国家自然科学基金项目(61773160,61871182)。

摘　　要：针对遥感图像中感兴趣目标特征不明显、背景信息复杂、小目标居多导致的目标检测精度较低的问题,本文提出了一种改进YOLOv5s的遥感图像目标检测算法(Swin-YOLOv5s)。首先,在骨干特征提取网络的卷积块中加入轻量级通道注意力结构,抑制无关信息的干扰;其次,在多尺度特征融合的基础上进行跨尺度连接和上下文信息加权操作来加强待检测目标的特征提取,将融合后的特征图组成新的特征金字塔;最后,在特征融合的过程中引入Swin Transformer网络结构和坐标注意力机制,进一步增强小目标的语义信息和全局感知能力。将本文提出的算法在DOTA数据集和RSOD数据集上进行消融实验,结果表明,本文提出的算法能够明显提高遥感图像目标检测的平均准确率。Aiming at the low average target detection accuracy in remote sensing images caused by obscure features in the objects of interest,complex background information,and multiple small targets,we propose a new remote sensing image object detection algorithm with improved YOLOv5s(Swin-YOLOv5s).First,an efficient channel attention structure is added to the convolutional block of the backbone feature extraction network to suppress the interference of irrelevant information;second,cross-scale connection and contextual information weighting operations are performed to enhance detection target feature extraction on the basis of multiscale feature fusion,and the fused feature maps are composed into a new feature pyramid;finally,the Swin Transformer structure and coordinate attention mechanism are used to further enhance the semantic information and global perception ability of small targets.The result of a feature fusion elimination experiment performed on the DOTA and RSOD datasets shows that the proposed algorithm can significantly improve the average accuracy of object detection in remote sensing images.

关 键 词：遥感图像 感兴趣目标 目标检测 特征提取 轻量级通道注意力结构 多尺度特征融合 上下文信息 Swin变换器 坐标注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP391[自动化与计算机技术—控制科学与工程]