摘要
遥感场景分类是近年来计算机视觉和表示学习领域的热门研究课题,其主要工作是基于学习到的特征信息自动分类图像场景.传统上场景分类方法忽略了场景中多个子概念的学习,进而影响到场景语义识别.为了解决上述问题,文中提出一种弱监督多示例子概念学习(Weakly Supervised Multi-Instance Sub-concept Learning)的遥感场景分类方法.首先,基于弱监督定位网络从逐类响应图中预测峰值坐标,以定位感兴趣的示例区域;其次,将峰值坐标信息回溯到卷积层,自动截取多个示例特征组成示例袋作为多示例聚合网络的输入.然后,在多示例聚合网络上嵌入一个子概念层,迭代学习子概念与示例之间的匹配分数,再将所有的示例进行聚合生成示例袋概率分数;最后,组合两个损失函数,联合训练整个网络,得到富于判别的分类模型.在AID、NWPU-RESISC45和CIFAR10/100数据集上进行了分类实验,结果表明,所提方法有效提高了遥感场景分类性能.
Remote scene classification(RSC)is a hot research topic in the fields of computer vision and representation learning and the main work of RSC is to automatically classify image scene based on learned feature information.Traditional scene classification methods ignore the learning of multiple sub-concepts in the scene, which affects the scene semantic recognition.To this end, this paper proposes a weakly supervised multi-instance sub-concept learning approach for remote sensing scene classification.First, the weakly supervised localization network(WSL)is exploited to predict the peak locations in the class-wise response map such that one can localize the instances of interest;Secondly, the peak locations is backtracked to the last convolutional layer of the backbone network, and multiple instance features are extracted from each image to consist of an instance bag that will be fed to the multi-instance aggregation network(MIN).Then, we embed a sub-concept layer on the top of MIN such that the matching scores between each instance and all sub-concepts is learned.In the following, all instances are aggregated to generate an instance bag with probability score.Finally, two loss functions are combined to jointly train the entire network and to obtain discriminative classification model.remote sensing scene classification.The classification experiments of our approach are conducted on AID,NWPU-RESISC45 and CIFAR10/100,and prove that the proposed method can effectively improve the performance of remote sensing scene classification.
作者
陈春芳
边小勇
费雄君
杨博
张晓龙
CHEN Chun-fang;BIAN Xiao-yong;FEI Xiong-jun;YANG Bo;ZHANG Xiao-long(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Institute of Big Data Science and Engineering,Wuhan University of Science and Technology,Wuhan 430065,China;Key Laboratory of Hubei Province for Intelligent Information Processing and Real-time Industrial System,Wuhan 430065,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2022年第1期76-83,共8页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61972299,61806150)资助
校研究生创新基金项目(JCX201924,JCX201927)资助。
关键词
遥感场景分类
卷积神经网络
弱监督
多示例
子概念学习
remote sensing classification
convolutional neural network
weakly supervision
multiple instance
sub-concept learning
