视点意识胶囊网络的遥感场景分类  

Viewpoint-Aware Capsule Networks for Remote Sensing Scene Classification

作　　者：孙学昊 边小勇[1,2] SUN Xuehao;BIAN Xiaoyong(College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan University of Science and Technology,Wuhan 430065)

机构地区：[1]武汉科技大学计算机科学与技术学院,武汉430065 [2]武汉科技大学智能信息处理与实时工业系统湖北省重点实验室,武汉430065

出　　处：《计算机与数字工程》2024年第10期2949-2954,共6页Computer & Digital Engineering

基　　金：国家自然科学基金面上项目(编号:62071456);湖北省自然科学基金项目(编号:2018CFB575)资助。

摘　　要：针对遥感图像的视点变化,提出了一种新的遥感场景分类方法,通过空间变换网络结合传感器视点估计,预测每个场景类别的主方向,并在网络内部计算二维仿射变换参数,对胶囊姿态向量进行编码,用以构建等变的、具有仿射变换鲁棒性的主胶囊层。对遥感图像的类内多样性问题进行了研究,提出了子概念动态路由,减少后续层中出现的噪声胶囊,生成场景的内部紧凑表示用于遥感场景分类。最后在两个公开的大型遥感场景数据集AID和NWPU-RESISC45进行测试,实验结果表明,该方法显著提高了遥感场景的分类准确率。A new remote sensing scene classification method is proposed for the viewpoint variation of remote sensing images,the main orientation of each scene class is predicted by a spatial transformation network combined with sensor viewpoint estimation,and the two-dimensional affine transformation parameters are computed inside the network to encode the capsule pose vectors for constructing an equivariant,affine transformation robust primary capsule layer.The intra-class diversity problem of remote sensing images is investigated,sub-concept dynamic routing is proposed to reduce the noisy capsules that appear in subsequent layers and generate an internal compact representation of the scene for remote sensing scene classification.Finally,tests are conducted on two publicly available large remote sensing scene datasets,AID and NWPU-RESISC45,the experimental results demonstrate that the method significantly improves the recognition accuracy of remote sensing scenes.

关 键 词：场景分类 胶囊网络 姿态估计 子概念 动态路由 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]