电子学报 ›› 2020, Vol. 48 ›› Issue (9): 1824-1833.DOI: 10.3969/j.issn.0372-2112.2020.09.021

• 学术论文 • 上一篇    下一篇

联合多尺度多特征的高分遥感图像场景分类

黄鸿, 徐科杰, 石光耀   

  1. 重庆大学光电技术及系统教育部重点实验室, 重庆 400044
  • 收稿日期:2019-03-26 修回日期:2020-04-19 出版日期:2020-09-25
    • 通讯作者:
    • 黄鸿
    • 作者简介:
    • 徐科杰 男,1994年生于浙江舟山.2017年于浙江师范大学获得学士学位,现为重庆大学博士研究生.主要从事模式识别,图像处理,遥感影像分类等方面的研究.E-mail:xukejie@cqu.edu.cn
      石光耀 男,1988年生于河南项城.2015、2017年于重庆大学分别获得学士和硕士学位,现为重庆大学博士研究生.主要从事图像处理、遥感影像分类、机器视觉与目标追踪等方面的研究.E-mail:shiguangyao@cqu.edu.cn
    • 基金资助:
    • 重庆市基础与前沿研究计划 (No.cstc2018jcyjAX0093); 重庆市留学人员回国创业创新支持计划 (No.cx2019144); 重庆市研究生科研创新项目 (No.CYB19039); 重庆市教委科学技术研究计划 (No.KJZD-K201902501)

Scene Classification of High-Resolution Remote Sensing Image by Multi-scale and Multi-feature Fusion

HUANG Hong, XU Ke-jie, SHI Guang-yao   

  1. Key Laboratory of Optoelectronic Technique System of the Ministry of Education, Chongqing University, Chongqing 400044, China
  • Received:2019-03-26 Revised:2020-04-19 Online:2020-09-25 Published:2020-09-25
    • Corresponding author:
    • HUANG Hong
    • Supported by:
    • Chongqing Research Program of Basic and Frontier Technology (No.cstc2018jcyjAX0093); Venture & Innovation Support Program for Chongqing Overseas Returnees (No.cx2019144); Chongqing Postgraduate Research and Innovation Project (No.CYB19039); Science and Technology Research Project of Chongqing Municipal Education Commission (No.KJZD-K201902501)

摘要: 高分辨率遥感图像地物信息丰富,但场景构成复杂,目前基于手工设计的特征提取方法不能满足复杂场景分类的需求,而非监督特征学习方法尽管能够挖掘局部图像块的本征结构,但单一种类及尺度的特征难以有效表达实际应用中复杂遥感场景特性,导致分类性能受限.针对此问题,本文提出了一种基于多尺度多特征的遥感场景分类方法.该算法首先设计了一种改进的谱聚类非监督特征(iUFL-SC)以有效表征图像块的本征结构,然后通过密集采样提取每幅遥感场景的iUFL-SC、LBP、SIFT等三种多尺度局部图像块特征,并通过视觉词袋模型(BoVW)获得场景的中层特征表达,以实现更为准确详实的特征描述,最后基于直方图交叉核的支持向量机(HIKSVM)进行分类.在UC Merced数据集以及WHU-RS19数据集上的实验结果表明本文方法可对遥感场景进行鉴别特征提取,有效提高分类性能.

关键词: 遥感, 高分辨率影像, 场景分类, 非监督特征, 特征融合, 视觉词袋模型

Abstract: High resolution image possesses abundant information of ground objects. The hand-crafted features cannot meet the demand of complex scene classification due to complex scene distribution, while the unsupervised feature learning method can exploit the intrinsic structure of image patches to obtain effective discriminating features. However, single feature with a scale is difficult to represent the characteristics of complex scenes in practical applications, which restricts classification performance. To solve this problem, this paper proposed a new method based on multi-scale and multi-feature fusion (MMF) for remote sensing scene classification. At first, an improved unsupervised feature learning via spectral clustering (iUFL-SC) is designed to effectively reveal the intrinsic structure of image patches, and then the iUFL-SC, LBP, and SIFT features of image patches are extracted by dense sampling in each image. After that, the middle-level features of each scene are obtained through bag of visual words (BoVW) model for effective feature description. Finally, the fused features are classified by histogram intersection kernel SVM. Experimental results on two public data sets indicate that MMF can extract discriminant features of remote sensing image and subsequently improve the classification performance.

Key words: remote sensing, high resolution images, scene classification, unsupervised features, feature fusion, bag of visual words

中图分类号: