计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (5): 181-186.DOI: 10.3778/j.issn.1002-8331.1805-0134

• 图形图像处理 • 上一篇    下一篇

深度迁移学习在高光谱图像分类中的运用

王立伟1,李吉明2,周国民2,杨东勇1   

  1. 1.浙江工业大学 信息工程学院,杭州 310023
    2.浙江警察学院 计算机与信息技术系,杭州 310053
  • 出版日期:2019-03-01 发布日期:2019-03-06

Application of Deep Transfer Learning in Hyperspectral Image Classification

WANG Liwei1, LI Jiming2, ZHOU Guomin2, YANG Dongyong1   

  1. 1.College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023, China
    2.Department of Computer and Information Technology, Zhejiang Police College, Hangzhou 310053, China
  • Online:2019-03-01 Published:2019-03-06

摘要: 针对高光谱图像分类中,样本空间特征利用不足的问题。将深层残差网络作为特征提取器运用到高光谱图像分类中,利用深层残差网络更深的网络结构,挖掘样本邻域空间中的深层特征,实验证明此特征具有更好的可分性。同时,针对深层卷积网络有监督训练的过程中,由于有标签样本不足导致的过拟合现象,提出基于深度迁移学习方法的训练策略,通过迁移网络在另一相关数据集中训练得到的网络浅层卷积核参数,再使用目标数据集对深层卷积核参数进行微调,提高了残差网络在少量有标签样本情况下的分类效果。

关键词: 高光谱, 深层残差网络, 迁移学习

Abstract: In the field of hyperspectral image classification, the potential of spatial features is just taken into consideration in recent years and yet still not fully exploited. In this work, it generalizes the deep residual network to hyperspectral image classification as a feature extractor which is pre-trained on large-scale common image datasets, the discriminability of extracted features is verified on real data experiments and showed to be very promising. Moreover, under the supervised learning setting, aiming at the problem of overfitting due to insufficient label samples, a model-based transfer learning strategy is proposed. Through pre-training the deep residual network in another related hyperspectral data set, it then fixes the shallow convolution kernel parameters, and uses a small number of labeled samples of the target data set to fine-tune the network top-level convolution kernel parameters. The ability of generalization on new data set is also proved.

Key words: hyperspectral, deep residual network, transfer learning