基于3D多尺度特征融合残差网络的高光谱图像分类

深度学习中用于训练的高光谱图像(HSI)数据十分有限,因此较深的网络不利于空谱特征的提取.为了缓解该问题,文中提出3D多尺度特征融合残差网络,利用深度学习和多尺度特征融合的方式对光谱-空间特征进行有序的学习.首先对3D-HSI数据进行自适应降维,将降维后的图像作为网络输入.然后,通过多尺度特征融合残差块依次提取光谱-空间特征,融合不同尺度的特征,通过特征共享增强信息流,获得更丰富的特征.最后以端到端的方式训练网络.在相关数据集上的测试表明,文中网络具有良好的分类性能.

Hyperspectral Image Classification Based on 3D Multi-scale Feature Fusion Residual Network

Hyperspectral image(HSI) data used for training in deep learning are insufficient, and therefore deeper network is unfavorable for spectral-spatial feature extraction. To solve this problem, a 3D multi-scale feature fusion residual network is proposed. Spectral-spatial features are learned by deep learning and multi-scale feature fusion. Firstly, the dimension of 3D-HSI data is adaptively reduced, and the images after dimensionality reduction are used as the input of the network. Secondly, spectral-spatial features are extracted successively through multi-scale feature fusion residual blocks and features of different scales are fused. The information flow is enhanced through sharing features and richer features are obtained. Finally, the network is trained end-to-end and tested on corresponding datasets. Experimental results show the satisfactory classification performance of the proposed network.