基于胶囊神经网络的合成孔径雷达图像分类方法

针对合成孔径雷达(synthetic aperture radar,SAR)图像的目标分类,传统的卷积神经网络(convolutional neural network,CNN)需要大量的数据样本进行训练,无法在小样本条件下进行,其应用受到限制。提出将胶囊神经网络(capsule network,Capsnet)算法用于SAR雷达图像的分类,针对小样本SAR数据集对Capsnet结构进行轻量化设计,并在MSTAR数据集上验证了该算法的有效性。结果表明,与CNN相比,基于Capsnet的SAR目标分类抗过拟合性强,泛化效果较好,具有更高的准确性,能够很好地实现SAR图像样本的分类。

SAR Image Classification Based on Capsule Network

For the target classification in Synthetic aperture radar (SAR) image ,the traditional convolution neural network (CNN) requires a great deal of data samples for training, which can not be used under the condition of small samples. This paper proposes a capsule neural network(Capsnet) based SAR image classification method, the lightweight design of the Capsnet structure was carried out for the SAR data set with small samples ,then uses the MSTAR dataset to verify the effectiveness of the algorithm. The results show that compared with CNN, Capsnet based method has stronger anti-overfitting, better generalization effect and higher accuracy, which can achieve the classification of SAR image samples well.