基于DenseNet-Attention模型的高光谱图像分类  

DenseNet-attention for hyperspectral remote sensing image classification

作　　者：张永鹏 张春梅[1] 白静[1] ZHANG Yong-peng;ZHANG Chun-mei;BAI Jing(School of Computer Science and Engineering,North Minzu University,Yinchuan Ningxia 750021,China)

机构地区：[1]北方民族大学计算机科学与工程学院,宁夏银川750021

出　　处：《图学学报》2020年第6期897-904,共8页Journal of Graphics

基　　金：国家自然科学基金项目(61461002,61762003);宁夏回族自治区重点研发计划项目(2019BDE03011);宁夏高等学校一流学科建设项目(电子科学与技术学科)(NXYLXK2017A07)。

摘　　要：针对高光谱图像标记样本量少,提取特征不充分以及提取到的特征不区分贡献度的问题,提出一个新型的DenseNet-Attention网络模型(DANet)。首先,该模型利用三维卷积核同步提取联合光谱空间特征,同时密集连接网络(DenseNet)的稠密连接块除了能够充分提取更加鲁棒的特征外,还减少了大量参数;其次,自注意力(self-attention)机制作为一个模块加入到稠密连接块中,可以使上层提取到的特征在进入下一层网络之前,经过该模块对其进行权重分配,使具有丰富的物类别信息的特征得到加强,进而区分特征的贡献度。网络模型以原始高光谱图像邻域块作为输入,无需任何预处理,是一个端对端学习的深度神经网络。在印第安松树林和帕维亚大学数据集上进行对比试验,网络模型的分类精度分别能够达到99.43%和99.99%,有效提高了高光谱图像分类精度。A new neural network,called DenseNet-Attention(DANet),was proposed in this paper for hyperspectral images classification to solve the problems of small sample quantity,insufficient features extraction,and indiscriminating contribution of the extracted features.First,it employed the three-dimensional convolution kernel to simultaneously extract both spectral and spatial features.Meanwhile,due to its dense blocks,DenseNet can not only fully extract more robust features,but reduce a large number of parameters.Second,the self-attention mechanism was added to the dense block as a module.Before the extracted feature was passed into the next layer of network,the weight was assigned to the feature according to its contribution through this model,thus strengthening the representation of the feature with ground object information.DANet was an end-to-end deep learning framework,which took the neighborhood block of the original hyperspectral image as an input without any preprocessing.Comparative experiments on Indian Pines and Pavia University datasets show that the classification accuracy of the network model proposed in this paper can reach 99.43%and 99.99%respectively,effectively enhancing the classification accuracy of hyperspectral images.

关 键 词：三维卷积 高光谱图像分类 稠密网络 自注意力机制 残差连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]