基于多分支自编码器的高光谱图像聚类  

Hyperspectral Image Clustering Based on Multi-branch Autoencoder

作　　者：金桂芳 王丽娜 宋佳力 JIN Guifang;WANG Lina;SONG Jiali(CEPREI,Guangzhou 511370,China;Ningbo CEPREI IT Research Institute Co.,Ltd.,Ningbo 315040,China)

机构地区：[1]工业和信息化部电子第五研究所,广东广州511370 [2]宁波赛宝信息产业技术研究院有限公司,浙江宁波315040

出　　处：《电子质量》2023年第11期39-44,共6页Electronics Quality

摘　　要：利用卷积网络提取高光谱图像的空谱特征时,空间信息会对目标地物边缘的点的特征产生不良影响,即产生空间噪声。针对该问题,提出了基于多分支自编码网络的高光谱图像聚类方法。通过基于卷积网络的空间自编码器提取目标点的空谱特征,然后通过光谱自编码网络引入目标点的光谱信息,对空谱特征再次进行编码解码,提纯空谱特征中光谱信息的成分占比,从而降低空谱特征中空间噪声的影响。再通过改进K-means聚类算法,整体提高聚类效果。实验表明,新提出的聚类方法比其他传统的聚类方法取得了更优异的聚类效果。When using convolutional networks to extract spatial spectral features from hyperspectral images,the spatial information will have a bad effect on the features of the points on the edge of the target object,that is,the spatial noise will be generated.To address this problem,a hyperspectral image clustering method based on multi-branch self-coding networks is proposed.By extracting the spatial spectral features of the target point based on convolutional networks,and by introducing the spectral information of the target point through spectral self-encoding networks,the spatial spectral features are encoded and decoded again to purify the spectral information component ratio in the spatial spectral features,thereby reducing the influence of spatial noise in the spatial spectral features.Then,by improving the K-means clustering algorithm,the overall clustering effect is improved.Experiments show that the newly proposed clustering method achieves better clustering results than other traditional clustering methods.

关 键 词：高光谱图像 特征提取 自编码器 聚类算法 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]