文献类型：期刊文献

中文题名：基于三维旋转卷积核的高光谱图像分类研究

英文题名：Research on Hyperspectral Image Classification Algorithm Based on Three Dimensional Rotating Convolution Kernel

作者：龙浩[1,2];徐聪[2];姚浩[2]

第一作者：龙浩

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100027

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2022

卷号：36

期号：4

起止页码：51-57

中文期刊名：北京联合大学学报

外文期刊名：Journal of Beijing Union University

基金：北京联合大学人才强校优选计划(BPHR2020CZ03);教育部高教司产学合作协同育人项目(201601011032)。

语种：中文

中文关键词：空谱分类;无监督学习;旋转卷积受限波尔兹曼机;高光谱图像

外文关键词：Spectral-spatial classification;Unsupervised learning;Rotating convolutional restricted Boltzmann machine;Hyperspectral image

摘要：针对在空谱信息特征提取过程中,由于降维造成部分高光谱信息丢失从而影响分类精度的问题,提出一种新型的三维旋转卷积核,并设计了无监督的旋转卷积受限波尔兹曼机。其从三维模型的原始表征中学习三维模型的高层局部特征,在原始数据上直接进行三维特征提取,获取表达力更强的局部表征,从而提高分类精度。将本文所提模型在Indian Pines和Pavia University公开数据集上进行验证,并同其他经典的分类方法进行实验对比。实验结果表明:该方法不仅能大幅度节省可学习的参数,降低模型的复杂度,而且表现出较好的分类性能。
In view of the problem that part of hyperspectral information is lost due to dimension reduction in the process of feature extraction of hyperspectral information,which affects the classification accuracy,a novel three-dimensional rotating convolution kernel is proposed,and an unsupervised rotating convolutional restricted Boltzmann machine is designed to learn the high-level local features from the original representation of the 3D model.We directly extract three-dimensional features from the original data to obtain more expressive local representations,thereby improving classification accuracy.The proposed methods are compared with state-of-the-art methods on the Indian Pines and Pavia University datasets.Experimental results show that the proposed unsupervised learning algorithms,which can extract more effective discriminant features,outperform the state-of-the-art supervised and semi-supervised learning classification methods,and achieve the best accuracy on all of the metrics.