摘要
近年来,低空无人机航拍图像和高空无人机遥感影像在土地资源管理、环境监测、地形测绘、城市规划等方面都发挥着重要的作用,如何准确地分割出低空无人机航拍图像和高空无人机遥感影像中的多个类别成为了当下的一个研究热点之一。针对低空无人机航拍图像和无人机遥感影像分割任务中远距离空间依赖关系难以捕捉和不同批次输入类别间的动态联系难以提取的问题,本文提出了一种基于ResNet和改进图卷积(GCN)模块的语义分割模型GRNet。改进图卷积以ResNet提取的特征为输入,将像素点转化为图节点,以像素之间的亲和力关系为边,使用坐标图卷积和通道图卷积对特征进行全局信息建模,同时引入类别动态图卷积,以输入信息的动态亲和力为边,对输入特征的类别动态相关性进行建模,最终将提取的特征融合到分割特征图,提高模型的分割精度。本文使用UDD5,UDD6两个低空无人机航拍数据集和ISPRS Potsdam高空无人机遥感数据集进行训练和验证,并在实验部分分别证明本文改进图卷积的不同模块的作用。实验结果证明所提出的GRNet具有很好的泛化性能,在无人机航拍图像数据集与无人机遥感影像数据集中都表现出良好的分割性能,优于多个对比网络。
In recent years,low-altitude UAV aerial images and high-altitude UAV remote sensing images have played an important role in land resource management,environmental monitoring,terrain mapping,urban planning,etc.How to accurately segment multiple categories in low-altitude UAV aerial images and high-altitude UAV remote sensing images has become one of the current research hotspots.For the low-altitude UAV aerial imagery and UAV remote sensing image segmentation tasks,it is difficult to capture long-distance spatial dependencies and difficult to extract dynamic connections between different batches of input categories,this paper proposes a semantic segmentation model GRNet based on ResNet and a modified graph convolution(GCN)module.Improved graph convolution takes the features extracted by ResNet as input,converts pixels into graph nodes,uses the affinity relationship between pixels as edges,and uses coordinate graph convolution and channel graph convolution to model global information for features,and at the same time,the category dynamic graph convolution is introduced,and the dynamic affinity of the input information is used as the edge to model the category dynamic correlation of the input features,and finally the extracted features are fused into the segmentation feature map to improve the segmentation accuracy of the model.This paper uses UDD5,UDD6 two low-altitude UAV aerial photography datasets and ISPRS Potsdam high-altitude UAV remote sensing dataset for training and verification,and in the experimental part respectively proves the role of different modules of this paper to improve graph convolution.The experimental results show that the proposed GRNet has good generalization performance,showing good segmentation performance in both the UAV aerial image data set and the UAV remote sensing image data set,which is better than multiple comparison networks.
作者
曾军英
王迎波
骆伟斌
陈宇聪
秦传波
翟懿奎
甘俊英
顾亚谨
ZENG Junying;WANG Yingbo;LUO Weibin;CHEN Yucong;QIN Chuanbo;ZHAI Yikui;GAN Junying;GU Yajin(Department of Intelligent Manufacturing,Wuyi University,Jiangmen,Guangdong 529020,China)
出处
《信号处理》
CSCD
北大核心
2023年第5期938-950,共13页
Journal of Signal Processing
基金
国家自然科学基金(61771347)
广东普通高校人工智能重点领域专项(2019KZDZX1017)
广东省基础与应用基础研究基金(2021A1515011576)
2021年度江门市基础与理论科学研究类科技计划项目(江科〔2021〕87号,2021030101900004391)
2022年广东省教育厅研究生教育创新计划项目(粤教研函〔2022〕1号)。
关键词
遥感图像
无人机航拍
语义分割
图卷积神经网络
深度学习
remote sensing image
drone aerial photography
semantic segmentation
graph convolutional neural networks
deep learning