《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

人工智能46

1)提出了三种基于CNN的深度特征提取结构,用于提取HSI的光谱特征、空间特征和光谱-空间特征。所设计的三维CNN能有效提取光谱空间特征,具有较好的分类性能。

2)针对训练样本数量有限导致的过拟合问题,在训练过程中采用了L2正则化和dropout等正则化策略。

3)为了进一步提高训练性能,从成像过程的角度,提出了一种虚拟样本增强方法来创建训练样本。

4)首次可视化分析了HSI提取的不同深度的层次特征。

5)将所提方法应用于三个知名的高光谱数据集。在此背景下,我们从分类精度、复杂性分析和处理时间等不同角度将本文提出的方法与一些传统方法进行了比较。

Method

1.Spectral FE(feature extraction) Framework for HSI Classification:

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

网络结构比较简单:两层卷积+两层池化+逻辑回归分类。为了避免过拟合,使用 L2 正则化。下面为损失函数+正则化:

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

下面是对L1、L2正则化的解释,原文链接:https://blog.csdn.net/qq_19672707/article/details/88874629

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

2.Architecture of CNN with spatial features for HSI classification:

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

以下是3D卷积和2D卷积之间的区别。三位卷积可以同时提取空间特征和光谱特征。

[En]

The following is the difference between 3D convolution and 2D convolution. Three-bit convolution can extract spatial and spectral features at the same time.

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

3.Spectral–Spatial FE Framework:

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

为了解决三维CNN容易过拟合的问题,提出了一种基于稀疏约束的组合正则化策略,该策略包含了ReLU和dropout,并将dropout应用于全连接层。下图为网络参数:

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

针对训练样本有限的问题,提出了一种虚拟样本生成方法。

[En]

In this paper, a virtual sample generation method is proposed to solve the problem of limited training samples.

虚拟样本方法尝试从给定的训练样本中创建新的训练样本。由于大场景中照明情况复杂,同一类物体在不同位置表现出不同的特征。因此,我们可以通过将一个随机因子乘以一个训练样本并添加随机噪声来模拟一个虚拟样本。此外,我们可以从同一类的两个给定样本中以适当的比例生成一个虚拟样本。虚拟样本思想对CNN的训练很有帮助。

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

Experimental Result(Pavia)

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

添加虚拟样本后的训练结果:

[En]

Training results after adding virtual samples:

《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记

Conclusion

本文提出了三种网络模型用于高光谱图像分类:基于光谱特征的一维CNN,基于空间特征的2维CNN,基于光谱-空间特征的三维CNN。其中3D-CNN取得了最好的效果。在3D-CNN中加入创建的虚拟样本,效果进一步提升。

Original: https://www.cnblogs.com/AllFever/p/16691698.html
Author: AllFever
Title: 《Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks》论文笔记