欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (5): 729-737.DOI: 10.11996/JG.j.2095-302X.2021050729

• 图像处理与计算机视觉 • 上一篇    下一篇

基于残差 3DCNN 和三维 Gabor 滤波器的 高光谱图像分类

  

  1. 1. 上海海洋大学信息学院，上海 201306；2. 上海电力大学，上海 200090；3. 自然资源部东海局，上海 200137
  • 出版日期:2021-10-31 发布日期:2021-11-03
  • 基金资助:
    国家自然科学基金项目(41906179)；上海市自然科学基金项目(18ZR1417300)；上海市科委部分地方高校能力建设项目(20050501900) 

Hyperspectral image classification based on residual 3DCNN and 3D Gabor filter

  1. 1. College of Information Technology, Shanghai Ocean University, Shanghai 201306, China; 2. Shanghai University of Electric Power, Shanghai 200090, China; 3. East China Sea Bureau, Ministry of Natural Resources, Shanghai 200137, China
  • Online:2021-10-31 Published:2021-11-03
  • Supported by:
    National Natural Science Foundation of China (41906179); Natural Science Foundation of Shanghai (18ZR1417300); The Capacity Development Project of Local Universities by Shanghai Science and Technology Commission (20050501900) 

PDF (PC)

130

摘要： 高光谱图像含有数百个波段，包含丰富的光谱信息，因此被广泛应用于地物分类中，但仍存在 着维数灾难的问题。高光谱图像中同时也含有丰富的纹理信息，有效利用纹理信息能够显著提高分类精度。三 维 Gabor 滤波器不仅能够保留图像丰富的光谱信息，还能提取到图像的纹理特征。为了充分利用高光谱图像的 特征，提出一种基于三维 Gabor 和残差三维卷积神经网络(Res-3DCNN)的分类方法。三维卷积神经网络(3DCNN) 能够直接对三维立方体数据进行处理，提取到深层纹理-光谱信息，然而随着网络层的加深会产生网络退化问 题，因此利用残差思想对 3DCNN 模型进行改进。在 PaviaU，Indian Pines 和 Salinas 3 个公共高光谱图像数据 集上进行实验，分别取得 99.17%，97.40%，98.56%的平均分类精度，结果表明该方法能有效提高高光谱图像 的地物分类精度。

关键词: 高光谱图像分类, 卷积神经网络, 三维 Gabor 滤波器, 三维卷积, 残差学习

Abstract: Hyperspectral remote sensing images contains hundreds of spectral bands and rich spectral information, resuling in wideapplications in the classification of ground objects, but there remains the problem of the curse of dimensionality. Hyperspectral images also contain rich texture information which can improve the classification precision significantly. 3D Gabor filter can not only keep rich spectral information of the images, but also extract the image texture features. In order to make full use of the features of hyperspectral images, this paper proposed a hyperspectral image classification model based on 3D Gabor and residual three-dimensional convolution neural network (Res-3DCNN). The 3DCNN can deal with three-dimensional cubic hyperspectral data and extract sufficient texture-spectral information. However, the deepening of convolutional neural network structure will lead to the problem of network degradation. Therefore, the idea of residual learning was applied to the improvement of the performance of 3DCNN. The proposed method was examined with three public hyperspectral data sets of PaviaU, Indian Pines and Salinas, reaching the average classification accuracy of 99.17%, 97.40% and 98.56%, respectively. Experimental results prove that the proposed method can effectively improve the ground targets classification accuracy of hyperspectral images. 

Key words: hyperspectral images classification, convolutional neural network, three-dimensional Gabor filter, three-dimensional convolution, residual learning 

中图分类号: