Mixed-Labeling Graph Prototype Neural Network for Few-Shot SAR Target Recognition

The graph neural networks (GNNs) have demonstrated unique advantages in few-shot synthetic aperture radar (SAR) target recognition in recent years. However, the existing GNN-based methods only explicitly update node features, with less attention paid to edge feature updates, which leads to insufficient exploration of intra-class similarity and inter-class difference in feature learning. To address this issue, a mixed-labeling graph prototype neural network (MGPNN) with both explicit node and edge feature updates is proposed. The MGPNN is trained to predict the edge-labels and node-labels simultaneously by iteratively updating the edges with explicit exploitation of the intra-class similarity and the inter-class difference. The prototype feature is used to solve the problem of inaccurate feature aggregation caused by intra-class difference. Based on the simultaneous prediction of the two types of labels, a mixed-labeling loss consisting of edge-labeling loss and node-labeling loss is designed to learn the model parameters of MGPNN, which can achieve better inter-class feature separability compared to a single-labeling loss. Experimental results on real SAR target recognition under few-shot settings show that the proposed method outperforms existing few-shot recognition methods.