基于有监督对比学习的遥感图像场景分类  

Remote Sensing Image Scene Classification Based on Supervised Contrastive Learning

作　　者：郭东恩 夏英[1] 罗小波[1] 丰江帆[1] GUO Dongen;XIA Ying;LUO Xiaobo;FENG Jiangfan(Chongqing Engineering Research Center for Spatial Big Data Intelligent Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;School of Computer and Software,Nanyang Institute of Technology,Nanyang,Henan 473000,China)

机构地区：[1]重庆邮电大学重庆市空间大数据智能技术工程研究中心,重庆400065 [2]南阳理工学院计算机与软件学院,河南南阳473000

出　　处：《光子学报》2021年第7期79-90,共12页Acta Photonica Sinica

基　　金：国家自然科学基金(Nos.41971365,41871226,41571401);河南省科技攻关项目(No.212102210492)。

摘　　要：针对遥感场景图像中复杂背景以及类内多样性和类间相似性影响场景分类性能的问题,提出一种基于有监督对比学习的遥感场景分类方法。该方法包含判别性特征学习和线性分类两个阶段。在判别性特征学习阶段,引入有监督对比损失以拉近同类场景间的距离并增大不同类场景间的距离,提高类内多样性和类间相似性场景的判别能力;然后引入门控自注意模块对无用的背景信息进行过滤且聚焦关键场景区域,提高复杂背景的场景识别;最后引入一个预训练的Inception V3语义分支,把语义分支和原始模型提取的特征进行融合增强特征判别能力,以提高场景分类的整体性能。线性分类阶段通过对特征学习阶段训练的模型进行微调获得分类结果。在AID和NWPU-RESISC45数据集上的综合实验证明了所提方法的有效性。To solve the problem of scene classification performance caused by complex background,intraclass diversity and inter-class similarity in remote sensing scene images,a new remote sensing scene classification method based on supervised contrast learning is proposed.The method involves two stages:discriminative feature learning and linear classification.In the stage of discriminative feature learning,a supervised contrastive loss first is introduced to narrow the distance between similar scenes and increase the distance between different types of scenes,so as to improve the scene discriminative ability of intra-class diversity and inter-class similarity;secondly,a gated self-attention module is introduced to filter useless background information and focus on key scene parts for improving scene recognition capabilities with complex backgrounds;finally,the pre-trained Inception V3 branch is introduced,and the branch features are merged with the final features extracted by the original model to further enhance the feature discriminative ability for improving the overall performance of scene classification.In the linear classification stage,the classification results are obtained by fine-tuning the model trained in the first stage.Comprehensive experiments on AID and NWPU-RESISC45 datasets demonstrate the effectiveness of the proposed method.

关 键 词：有监督对比学习 特征融合 遥感场景分类 门控机制 自注意机制 遥感图像 预训练模型 

分 类 号：TP751.2[自动化与计算机技术—检测技术与自动化装置]