Computer Engineering and Applications ›› 2020, Vol. 56 ›› Issue (24): 201-206.DOI: 10.3778/j.issn.1002-8331.1911-0102

Previous Articles     Next Articles

Hyperspectral Image Classification Based on MCFFN-Attention

CHENG Wenjuan, CHEN Wenqiang   

  1. School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230000, China
  • Online:2020-12-15 Published:2020-12-15



  1. 合肥工业大学 计算机与信息学院,合肥 230000


In view of the high dimensional characteristics of hyperspectral images and the limitation of the small number of samples, a multi-scale cross-layer feature fusion network attention mechanism method(MCFFN-Attention) is proposed. PCA reduction is performed on hyperspectral images. Based on 3D CNN, the center pixel and its adjacent pixels are input into the network as a whole, and the features obtained by different convolutional layers are fused. At the same time, spatial attention mechanism processing is carried out for the fused low-level features, and channel attention mechanism processing is carried out for the fused high-level features, and different weights are assigned to them to optimize the feature maps. Experiments on Indian Pines and University of Pavia data sets show that the classification accuracy of this method is improved compared with CNN, 3D CNN and M3D CNN.

Key words: hyperspectral image classification, multi-scale, feature fusion, attention mechanism


针对高光谱图像高维度的特性和样本数量少的局限性,提出了一个多尺度跨层特征融合注意力机制(MCFFN-Attention)的方法。对高光谱图像进行PCA降维,然后以3D CNN为基础,将中心像素和其相邻像素作为整体输入到网络中,对不同卷积层得到的特征进行融合。同时对融合的低层特征进行空间注意力机制处理,对融合的高层特征进行通道注意力机制处理,分配给它们不同的权重来优化特征图。在印第安松树和帕维亚大学数据集上进行实验,结果表明此方法相对于CNN、3D CNN和M3D CNN方法,分类精度得到了提升。

关键词: 高光谱图像分类, 多尺度, 特征融合, 注意力机制