An open-set recognition method for SAR images based on graph neural network

Abstract

Most synthetic aperture radar (SAR) image recognition systems operate under closed-set assumptions that all target categories are predefined. However, in real-world scenarios, unknown-class targets frequently appear, leading conventional closed-set models to misclassify them into known categories. To address this challenge, this paper proposes a novel open-set SAR image recognition approach based on graph neural network (GNN) and prototype learning. The method constructs a feature-to-prototype graph structure to dynamically model relationships between local SAR image features and class prototypes and incorporates a multi-level attention mechanism (including node-level and graph structural attention) to adaptively enhance correlations between critical target regions and prototypes. Specifically, node-level attention weights the saliency of local features, while graph structural attention captures global dependencies between features and prototypes, significantly improving the model's discriminative capability for known classes and rejection of unknown classes. Experimental results on the MSTAR and Truck datasets demonstrate the method significantly outperforms existing approaches in open-set scenarios.