联合Fisher核编码和卷积神经网络的影像场景分类  

Combined Fisher Kernel Coding Framework with Convolutional Neural Network for Remote Sensing Scene Classification

作　　者：刘异[1,2] 庄姊琪 闫利 廖明[2] LIU Yi;ZHUANG Ziqi;YAN Li;LIAO Ming(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China;Key Laboratory of Watershed Ecology and Geographical Environment Monitoring,NASG,Nanchang 330209,China)

机构地区：[1]武汉大学测绘学院,武汉430079 [2]流域生态与地理环境监测国家测绘地理信息局重点实验室,南昌330209

出　　处：《遥感信息》2018年第4期8-15,共8页Remote Sensing Information

基　　金：流域生态与地理环境监测国家测绘地理信息局重点实验室资助课题(WE2016016);国家重点研发计划项目(2016YFB0501403)

摘　　要：针对高分辨率遥感影像场景分类中使用中、低层特征不能有效表达高分影像的语义信息,造成分类精度不高的问题,提出了一种联合Fisher核编码和卷积神经网络的高分影像场景分类方法。首先利用Fisher核编码框架提取影像的中层语义特征,然后利用深度卷积神经网络提取影像高层语义特征,最后融合中、高层特征利用支持向量机进行分类。实验采用迁移学习方法来克服深度卷积神经网络对训练数据量的需求。实验数据采用UC-Merced 21类和WHURS 19类2个高分影像数据集。实验结果表明,中、高层融合特征包含更丰富的场景信息,增加了目标的可区分性,相比已有方法,该方法能够有效提高分类精度;迁移学习方法能够克服深度卷积神经网络对训练数据量的依赖性。Scene classification methods using middle and low-level artificial features cannot represent the scene information effectively and essentially and prevent them from achieving better performance.Aiming at this problem,a combined Fisher kernel coding framework with convolutional neural network method is proposed in this paper.Firstly,we obtain middle-level features under Fisher kernel coding framework.Secondly,we use deep convolutional neural network(DCNN)automatically to learn the high-level features.Finally,we embed the middle and high-level features into support vector machine(SVM)for image classification.We use transfer learning strategy to overcome the dependence of DCNN for a large amount of data.Experimental results using UC-Merced land-use dataset and WHU-RS land-use dataset show that combined middle and high-level features contain more scene information,which makes different scenes easier to distinguish.The proposed method can provide more accurate classification results than the state-of-the-art methods.And transfer learning strategy can overcome the dependence of DCNN for a large amount of data.

关 键 词：Fisher核 深度卷积神经网络 迁移学习 高分辨率遥感影像 场景分类 

分 类 号：P237[天文地球—摄影测量与遥感]