Semi-Supervised SAR Target Recognition with Graph Attention Network

Abstract:
This paper presents an new method for semi-supervised SAR target recognition using graph attention network. All images can be represented as graph-structured data by an improved graph modeling method. A symmetric autoencoder is used to extract node features and the adjacency matrix is initialized using a new similarity measurement method. Different attention mechanisms are used in the graph attentional layers to forecast the categories of unlabeled nodes. The adjacency matrix is updated during the training process. Results on the MSTAR data set indicate that the proposed approach can achieve 93.66% recognition accuracy for unlabeled nodes using 10% labeled data.