融合图卷积网络与节点相似度的遥感图像检索  

Remote sensing image retrieval based on fusion of graph convolutional networks and node similarity

作　　者：叶发茂 吴坤霖 王孟瑶 朱晓颖 张任高 YE Famao;WU Kunlin;WANG Mengyao;ZHU Xiaoying;ZHANG Rengao(Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake of Ministry of Natural Resources,East China University of Technology,Nanchang 330013,China;School of Surveying and Geoinformation Engineering,East China University of Technology,Nanchang 330013,China;School of Mathematics&Computer,Nanchang University,Nanchang 330013,China)

机构地区：[1]东华理工大学自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,南昌330013 [2]东华理工大学测绘与空间信息工程学院,南昌330013 [3]南昌大学数学与计算机学院,南昌330031

出　　处：《测绘科学》2023年第9期66-75,共10页Science of Surveying and Mapping

基　　金：江西省数字国土重点实验室开放基金项目(DLLJ201908);自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室基金项目(MEMI-2021-2022-22);江西省自然科学基金项目(20202BABL202030);国家自然科学基金项目(41801288)。

摘　　要：针对如何充分利用图像之间的语义关系提高检索特征表征能力的问题,该文提出了基于融合节点注意力机制的图卷积神经网络的遥感图像检索特征提取模型。该模型通过聚合节点和其邻域节点的特征构建图像的图聚合特征,获取更好的遥感图像检索特征。提出了综合图像特征的欧氏距离和图像集合的相似度的节点相似度作为图像相似度准则,提升度量图像距离的准确性;此外还将图卷积神经网络强大的节点分类能力和图像到类距离构建基于类别权重的节点相似度准则进行遥感图像检索,进一步提升遥感图像检索精度。在UCMD、SIRI_WHU和PatternNet 3个公开遥感图像数据集上进行了测试和验证,mAP较其他方法分别提升了0.91%、1.43%和0.08%。结果表明,该方法能够提升遥感图像检索精度。Aiming to address the issue of how to effectively utilize the semantic relationships between images to enhance the representation capability of retrieval features,this paper designs a feature extraction model of remote sensing image retrieval based on graph convolutional neural network with node attention mechanism.This model aggregates the features of nodes and their neighboring nodes to construct graph aggregation features for the image,aiming to obtain better retrieval features for remote sensing images.It introduces node similarity as a criterion for image similarity,which combines the Euclidean distance of comprehensive image features and the similarity of image collections,thereby enhancing the accuracy of measuring image distances.Furthermore,the powerful node classification capability of graph convolution neural network and image-to-class distance is used to construct node similarity criteria with class weight for remote sensing image retrieval,so as to further improve the accuracy of remote sensing image retrieval.The proposed method was tested and validated on three publicly available remote sensing image datasets,namely UCMD,SIRI_WHU,and PatternNet,resulting in improvements of 0.91%,1.43%,and 0.08%in mAP compared to other methods.These results demonstrate that the proposed method can enhance the precision of remote sensing image retrieval.

关 键 词：遥感图像检索 图卷积神经网络 节点注意力机制 类别权重 节点相似度准则 

分 类 号：P237[天文地球—摄影测量与遥感]