摘要
遥感影像语义分割是遥感图像处理分析的重要任务,尤其是在多类别语义分割方面。目前方法主要围绕卷积神经网络展开,但卷积仅关注图像局部信息而忽视全局信息。因此,受高分辨率网络(HRNet)和关系图卷积网络(R-GCN)启发,提出一种高分辨率关系图卷积网络(HRGCN),用于多类别语义分割。首先对原始图像做简单线性迭代聚类(SLIC),利用该结果分割HRNet输出的特征图,获得同质性高且包含多分辨率信息的超像素块;然后基于超像素块构建图节点和边,使用R-GCN对图节点分类,从而学习到不同地物间长距离依赖关系,并完成遥感影像的提取分类。利用HRGCN模型在Potsdam和Vaihingen数据集上进行实验,将结果与已有方法对比,F_(1)值和MIoU值均有不同程度提升,证明该方法具有较好的先进性。
Semantic segmentation of remote sensing images is an important task in remote sensing image processing and analysis,especially in multi-category semantic segmentation.Current methods mainly revolve around convolutional neural networks,but convolution only focuses on the local information of the image while ignoring the global information.Therefore,inspired by high resolution network(HRNet)and relational graph convolutional network(R-GCN),this paper proposes a high-resolution relational graph convolutional network(HRGCN)for multi-category semantic segmentation.Firstly,simple linear iterative clustering(SLIC)is done on the original image,and the result is used to segment the feature map output from HRNet to obtain superpixel blocks with high homogeneity and containing multi-resolution information;then graph nodes and edges are constructed based on the superpixel blocks,and R-GCN is used to classify the graph nodes,so as to learn the long-distance dependency between different features and complete the extraction and classification of remote sensing images.The HRGCN model designed in this paper is experimented on Potsdam and Vaihingen datasets,and the results are compared with the existing methods,and the F_(1)values and MIoU values are improved to certain degrees,which proves that the method has good advancement.
作者
王寅达
陈嘉辉
彭玲
李兆博
杨丽娜
WANG Yinda;CHEN Jiahui;PENG Ling;LI Zhaobo;YANG Lina(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《中国科学院大学学报(中英文)》
北大核心
2025年第1期107-115,共9页
Journal of University of Chinese Academy of Sciences
基金
全球能源互联网集团有限公司科技项目(SGGEIG00JYJS2100032)资助。
