关键词: 高光谱图像, 深度可分离卷积, 残差网络, 压缩激活网络

Abstract: In response to the challenges posed by convolutional neural network (CNNs) commonly used for hyperspectral image classification, namely, their high parameter count, extended training times, and sensitivity to sample quantity, a classification network MDSR&SE-Net based on improved squeeze and excitation network and depth-separable residuals was proposed for limited training samples. First, the principal component analysis is employed in this model to reduce the dimension of the original HSI. Then, the multi-feature residual structure is connected by 3D convolutional neural network, and the spatial & spectral details of hyperspectral images are extracted by embedding the improved squeeze and excitation block. Finally, the extracted feature information is input into Softmax classifier to activate classification. To further lightweight the network, the number of parameters is reduced by using the depth separable convolution in the residual structure and introducing global average pooling. Experimental results show that overall accuracy of the three common hyperspectral data sets with the limited training samples are above 99%.

Key words: hyperspectral image, depth separable convolution, residual network, SE Net