基于扩张图卷积网络的SAR图像分类

SAR images classification based on dilated graph convolutional network

合成孔径雷达(Synthetic Aperture Radar, SAR)图像存在相干斑噪声,且样本数量少,导致特征提取困难。为了提取更有表征力的特征,本文提出了一种基于扩张图卷积网络,用于SAR图像分类。该网络构造了一个新的残差扩张图卷积(Residual Dilated Graph Convolutional, RDGC)模块。RDGC模块包含3个不同扩张率的图结构,通过卷积提取不同感受野的特征,并能够调整感受野的大小,以适应不同尺度的特征信息;在此基础上,叠加多个RDGC模块,通过将每个RDGC模块的输出进行特征融合,提取较多的细节信息;最后,将SAR图像经过粗提取后的特征附加在其上,形成全局的残差连接,在高层语义特征中融入低层空间特征,进一步补充细节信息,且避免了特征丢失和梯度消失。在两幅真实SAR图像上进行实验,结果表明:改进的图卷积网络模型优于现有网络的分类效果和性能。

Speckle noise exists in synthetic aperture radar(SAR) images and the number of available samples is small, which makes feature extraction difficultl. In order to extract more representational features, dilated graph convolutional network for SAR image classification is proposed. The network constructs a new residual dilated graph convolutional module(RDGC). RDGC module is a graph structure with three different expansion rates. It extracts the features with different receptive fields through convolution and adjusts the size of receptive fields to adapt to feature information of different scales. On this basis, multiple RDGC modules are superimposed, and the output of each RDGC module is fused to extract more details. Finally, the coarse features of SAR images are added to them to form global residual connection, and the low-level spatial features are integrated into the high-level semantic features to further supplement details and avoid feature loss and gradient disappearance. Experiments are carried out on two groups of real SAR images, and a variety of algorithms are used for comparison and quantitative analysis. The experimental results verify the effectiveness of the improved network in SAR image classification.