作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (12): 215-221. doi: 10.19678/j.issn.1000-3428.0056791

• 图形图像处理 • 上一篇    下一篇

一种改进多尺度三维残差网络的高光谱图像分类方法

郑姗姗1a, 刘文1a, 单锐1a, 赵静一1b, 江国乾1c, 张智2   

  1. 1. 燕山大学 a. 理学院;b. 机械工程学院;c. 电气工程学院, 河北 秦皇岛 066004;
    2. 北京航天研究所, 北京 100094
  • 收稿日期:2019-12-04 修回日期:2020-01-07 发布日期:2020-01-10
  • 作者简介:郑姗姗(1995-),女,硕士研究生,主研方向为遥感图像分类;刘文、单锐、赵静一(通信作者),教授、博士;江国乾,讲师、博士;张智,工程师、博士。
  • 基金资助:
    国家自然科学基金(51675461,61803329);秦皇岛科技局项目(201703A020)。

A Hyperspectral Image Classification Method Based on Improved Multi-Scale Three-Dimensional Residual Network

ZHENG Shanshan1a, LIU Wen1a, SHAN Rui1a, ZHAO Jingyi1b, JIANG Guoqian1c, ZHANG Zhi2   

  1. 1a. School of Science;1b. School of Mechanical Engineering;1c. School of Electrical Engineering, Yanshan University, Qinhuangdao, Hebei 066004, China;
    2. Beijing Aerospace Research Institute, Beijing 100094, China
  • Received:2019-12-04 Revised:2020-01-07 Published:2020-01-10

摘要: 针对高光谱图像训练样本较少、光谱维度高导致分类精度较低的问题,提出一种利用改进多尺度三维残差卷积神经网络的高光谱图像分类方法。选择合适的卷积步长对网络首层光谱降维并提取浅层特征,使用三维卷积滤波器组中最大池化层减少整体网络训练参数量,改进多尺度滤波器组和三维残差单元提取图像深层局部空间-光谱联合特征,并将其输入Softmax函数层预测类别标签样本。实验结果表明,该方法在Indian Pines和Pavia University高光谱数据集上的总体分类精度分别为99.33%和99.83%,与SVM、SAE等方法相比,分类判别特征提取更准确,具有更高的图像分类精度。

关键词: 三维卷积块, 卷积神经网络, 高光谱图像, 多尺度滤波器, 残差单元

Abstract: To solve the low classification accuracy caused by the small number of training samples and high spectral dimension,this paper proposes a Hyperspectral Image(HSI) classification method based on improved multi-scale Three-Dimensional(3D) residual Convolutional Neural Networks(CNN).An appropriate convolution step size is selected to reduce the dimension of the first-layer spectrum of the network and extract the shallow features.The maximum pooling layer in the 3D convolution filter bank is used to reduce the training parameters of the whole network.The multi-scale filter bank and the 3D residual unit are improved to extract the deep local spatial-spectral joint features of the image,which are input into the Softmax function layer to predict the class label samples.Experimental results show that the overall classification accuracy of this method is 99.33% and 99.83% respectively on Indian Pines and Pavia University hyperspectral datasets.Compared with SVM and SAE methods,the proposed method can extract more accurate classification features and has higher image classification accuracy.

Key words: Three-Dimensional(3D) convolutional block, Convolutional Neural Networks(CNN), Hyperspectral Image (HSI), multi-scale filter, residual unit

中图分类号: