三维Gabor和多尺度残差网络的高光谱影像分类  

Hyperspectral Image Classification Based on Three-dimensional Gabor and Multi-scale Residual Network

作　　者：吕欢欢[1,2] 胡杨 张辉 LV Huanhuan;HU Yang;ZHANG Hui(College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China;School of Information Engineering,Huzhou University,Huzhou,Zhejiang 313000,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105 [2]湖州师范学院信息工程学院,浙江湖州313000

出　　处：《遥感信息》2023年第4期33-41,共9页Remote Sensing Information

基　　金：浙江省教育厅一般科研项目(Y202248546)。

摘　　要：为了减轻高光谱影像分类网络模型对训练样本的依赖性并解决网络层数加深产生的性能退化问题,文章研究了三维Gabor滤波和多尺度残差网络的高光谱影像分类方法。利用三维Gabor滤波器提取出有助于分类的光谱-纹理特征,引入扩张卷积和残差学习构建多尺度残差网络模型进行深层次特征提取,实现不同尺度下局部和全局特征融合和分类。在两幅高光谱影像上对该方法和其他方法进行实验比较。结果表明,该方法获得了最优的分类结果,能够在训练样本有限的情况下提高分类精度。In order to reduce the dependence of hyperspectral image classification network model on training samples and solve the performance degradation caused by deepening of network layers,a hyperspectral image classification method based on three-dimensional Gabor filtering and multi-scale residual network is studied.Spectral-texture features helpful for classification are extracted by using 3D Gabor filter,and a multi-scale residual network model is constructed by introducing dilated convolution and residual learning for deep feature extraction,thus achieving local and global feature fusion and classification at different scales.The experimental comparison between the proposed method and other methods on two hyperspectral images shows that the proposed method obtains the best classification results and can improve the classification accuracy under the condition of limited training samples.

关 键 词：卷积神经网络 高光谱影像分类 三维Gabor滤波 扩张卷积 残差学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]