测绘学报 ›› 2016, Vol. 45 ›› Issue (9): 1073-1080.doi: 10.11947/j.AGCS.2016.20150612

• 摄影测量学与遥感 • 上一篇    下一篇

联合显著性和多层卷积神经网络的高分影像场景分类

何小飞, 邹峥嵘, 陶超, 张佳兴   

  1. 中南大学地球科学与信息物理学院, 湖南 长沙 410083
  • 收稿日期:2015-12-04 修回日期:2016-07-01 出版日期:2016-09-20 发布日期:2016-09-29
  • 通讯作者: 陶超 E-mail:kingtaochao@126.com
  • 作者简介:何小飞(1991-),男,硕士,研究方向为高分遥感影像分类.E-mail:hxf0321@qq.com
  • 基金资助:
    国家自然科学基金(41301453;51479215);国家973计划(2012CB719903);教育部博士点基金(20130162120027)

Combined Saliency with Multi-Convolutional Neural Network for High Resolution Remote Sensing Scene Classification

HE Xiaofei, ZOU Zhengrong, TAO Chao, ZHANG Jiaxing   

  1. School of Geosciences and Info-Physics, Central South University, Changsha 410083, China
  • Received:2015-12-04 Revised:2016-07-01 Online:2016-09-20 Published:2016-09-29
  • Supported by:
    The National Natural Science Foundation of China (Nos.41301453;51479215);The National Basic Research Program of China (973 Program) (No.2012CB719903);Research Fund for the Doctoral Program of Higher Education (No.20130162120027)

摘要: 高分辨率遥感影像中的场景信息,对于影像解译和现实世界的理解具有重要意义。传统的场景分类方法多利用中、低层人工特征,但是高分辨率遥感影像的信息丰富,场景构成复杂,需要高层次的特征来表达。本文提出了一种联合显著性和多层卷积神经网络的方法,首先利用显著性采样获取包含影像主要信息的有意义的块,将这些块作为样本集输入卷积神经网络中进行训练,获得不同层次的特征表达,最后联合多层特征利用支持向量机进行分类。两组高分影像场景数据UC Merced 21类和Wuhan 7类试验表明,显著性采样能够有效地获取主要目标,减弱其他无关目标的影响,降低数据冗余;卷积神经网络能够自动学习高层次的特征,相比已有方法,本文方法能够有效提高分类精度。

关键词: 卷积神经网络, 显著性探测, 高分辨率遥感影像, 场景分类

Abstract: The scene information existing in high resolution remote sensing images is important for image interpretation and understanding of the real world. Traditional scene classification methods often use middle and low-level artificial features, but high resolution images have rich information and complex scene configuration, which need high-level feature to express. A joint saliency and multi-convolutional neural network method is proposed in this paper. Firstly, we obtain meaningful patches that include dominant image information by saliency sampling. Secondly, these patches will be set as a sample input to the convolutional neural network for training, obtain feature expression on different levels. Finally, we embed the multi-layer features into the support vector machine (SVM) for image classification. Experiments using two high resolution image scene data show that saliency sampling can effectively get the main target, weaken the impact of other unrelated targets, and reduce data redundancy; convolutional neural network can automatically learn the high-level feature, compared to existing methods, the proposed method can effectively improve the classification accuracy.

Key words: convolutional neural network, saliency detection, high resolution remote sensing image, scene classification

中图分类号: