Learning Relation by Graph Neural Network for SAR Image Few-shot Learning

Due to the data-driven training strategies, the applicability of supervised deep learning models is severely limited to the newly emerging or categories that lack annotated images. Therefore, few-shot learning methods are more applicable to synthetic aperture radar (SAR) image interpretation where numerous labeled data may not existed. In this paper, we introduce a few-shot learning method based on relation network and graph neural networks (GNNs). Relation network extracts the feature distance between query samples and support samples through the convolutional neural network, and has achieved good performance in few-shot learning problems. GNNs have received increasing attention in recent year, and have shown superior performance in relation extraction. We replace the relation module in the relation network with attention GNN to model the relationship between input images, and learn the metric of feature similarity. Experiments on MSTAR dataset demonstrate that the proposed method can better extract the relationship between query samples and support samples, thereby improving the performance of few-shot image classification tasks.

Recommended citation:

• Yang. R, Xu. X, Li, X.R., Wang. L, and Pu. F.L, “Learning Relation by Graph Neural Network for SAR Image Few-shot Learning”, in Proc. IEEE Int. Geosci. Remote Sens. Symp.(IGARSS), September. 2020.