基于双通道空洞卷积神经网络的高光谱图像分类

Hyperspectral Image Classification Based on Dual-Channel Dilated Convolution Neural Network

受空洞卷积在图像信息方面保持优秀性能的启发,为进一步提高分类精度,提出一种基于双通道空洞卷积神经网络(DCD-CNN)的高光谱图像分类框架.空洞卷积可扩展滤波器的感受野,有效地避免图像信息丢失,从而提高分类精度.在该框架中,分别采用含有空洞卷积的一维卷积神经网络(1D-CNN)和二维卷积神经网络(2D-CNN)提取高光谱图像的光谱特征和空间特征.再采用加权融合方法对提取的空间特征和光谱特征进行融合.最后将融合后的特征输入支持向量机进行最终分类.对两个常用的高光谱图像数据集进行实验并与现有的4种分类方法进行比较,结果表明,所提框架具有更好的分类性能.

Based on the excellent hole convolution performance observed using the obtained image information,we propose a framework for performing hyperspectral image classification based on the dual-channel dilated convolution neural network(DCD-CNN)to improve the classification accuracy.The receptive field of the filters can be expanded via dilated convolution,which effectively avoides the loss of image information and improves the classification accuracy.In this proposed framework,one-dimensional CNN and two-dimensional CNN,exhibiting an empty convolution,are used to extract the spectral and spatial features of the hyperspectral images.Subsequently,these extracted features are combined using a weighted fusion method.Finally,the combined features are input into the support vector machine for performing final classification.The expreimental results on the two commonly used hyperspectral image datasets by the proposed framework are compared with that by the four existing classification methods,showing that the proposed framework exhibits improved classification performance.