基于空谱分组卷积密集网络的高光谱图像分类  

Hyperspectral image classification based on spatial-spectral group convolution dense network

作　　者：欧阳宁[1,2] 李祖锋 林乐平[1,2] OUYANG Ning;LI Zu-feng;LIN Le-ping(Key Laboratory of Cognitive Radio and Information Processing of Ministry of Education,School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China;School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学信息与通信学院认知无线电与信息处理省部共建教育部重点实验室,广西桂林541004 [2]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《计算机工程与设计》2022年第7期2031-2039,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62001133、61661017、61362021);广西科技基地和人才专项基金项目(桂科AD19110060);广西自然科学基金项目(2017GXNSFBA198212);广西无线宽带通信与信号处理重点实验室基金项目(GXKL06200114)。

摘　　要：针对高光谱图像分类在特征提取过程中高分辨率信息丢失,导致分类精度下降的问题,提出一种基于空谱分组卷积密集网络的高光谱图像分类方法。设计光谱-空间三维分组卷积密集模块,对光谱与空间特征进行分步提取,利用分组卷积构造的密集网络能减少数据固有信息冗余,使高分辨率的特征进行重用,避免细节特征信息丢失;设计光谱残差注意力模块,该模块通过结合空-谱特征计算注意力权重,对提取到的光谱特征进行权重重分配,对光谱信息富有的区域进行增强。实验结果表明,相比于若干最优的深度网络方法,所提高光谱图像分类方法具有更好的分类性能。Aiming at the issues of high-resolution information loss in the process of feature extraction in hyperspectral image classification,which leads to the decline of classification accuracy,a spatial-spectral group-convolution dense network was proposed.The spatial-spectral 3D group-convolution densenet module was exploited to extract the spectral and spatial features step by step.The information redundancy of hyperspectral data was reduced in the feature extraction process,the high-resolution features were reused through dense connection when using the dense network constructed by group-convolution,avoiding the loss of detailed feature information.The spectral residual attention module was designed and employed to calculate the attention weight combined with the spatial-spectral information,and to redistribute the weight of the extracted spectral features to enhance the area with rich spectral information.Experimental results show that the proposed network performs better than the state-of-the-art neural network-based classification methods.

关 键 词：高光谱图像分类 三维分组卷积 密集网络 光谱残差注意力模块 空-谱特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]