计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (5): 256-263.DOI: 10.3778/j.issn.1002-8331.2009-0144

• 图形图像处理 • 上一篇    下一篇

基于全卷积网络和自编码的高光谱图像分类

董朋欣,董安国,李楚婷,梁苗苗   

  1. 1.长安大学 理学院,西安 710064
    2.江西理工大学 信息工程学院,江西 赣州 341000
  • 出版日期:2022-03-01 发布日期:2022-03-01

Hyperspectral Image Classification Based on Fully Convolutional Network and Auto-Encoder

DONG Pengxin, DONG Anguo, LI Chuting, LIANG Miaomiao   

  1. 1.School of Science, Chang’an University, Xi’an 710064, China
    2.School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou, Jiangxi 341000, China
  • Online:2022-03-01 Published:2022-03-01

摘要: 针对高光谱图像空间信息利用不足、标记样本数量较少的问题,提出一种基于全卷积网络和堆栈稀疏自编码的高光谱图像分类算法。基于迁移学习的思想,利用预训练好的全卷积网络FCN-8s,挖掘图像潜在的多尺度几何结构特征;选取其特征的像素邻域信息,采用拼接融合的方法与原光谱信息进行融合;利用堆栈稀疏自编码网络完成最终的多尺度空谱特征提取,并通过Softmax分类器实现分类。对三组遥感图像进行实验,结果显示,所提算法极大改善了边界区域的分类效果。

关键词: 高光谱图像, 全卷积网络, 堆栈稀疏自编码, 迁移学习

Abstract: Aiming at the problems of insufficient utilization of hyperspectral image spatial information and the small number of labeled samples, a hyperspectral image classification algorithm based on fully convolutional network and stack sparse auto-encoder is proposed. First of all, based on the idea of transfer learning, the pre-trained fully convolutional network FCN-8s is used to mine the potential multi-scale geometric structure features of the image. Then the pixel neighborhood information of the features is selected, and the splicing fusion method is used to fuse with the original spectral information. Finally, the stacked sparse auto-encoder network is used to complete the final multi-scale spatial spectrum feature extraction, and the classification is realized by Softmax classifier.Experimental results on three groups of remote sensing images show that the proposed algorithm greatly improves the classification effect of boundary regions.

Key words: hyperspectral image, fully convolutional network, stacked sparse auto-encoder, transfer learning