Few-Shot SAR Ship Recognition via Vision Mamba and Scattering Topology Fusion

Abstract

Synthetic aperture radar (SAR) ship recognition faces great challenges in few-shot scenarios, including insufficient global context modeling, underutilization of physical scattering topological characteristics, and poor generalization capability under limited labeled samples. To address these bottlenecks, this paper proposes a novel few-shot SAR ship recognition method integrating Vision Mamba and scattering topology fusion. A dual-branch complementary architecture is innovatively constructed to learn discriminative features from two heterogeneous perspectives. The visual semantic branch combines residual convolution and a state-space model in parallel, which enables the simultaneous capture of local fine-grained scattering traits and long-range global structural dependencies, breaking the inherent local receptive-field constraint of conventional convolutional neural network (CNN)-based schemes. The scattering topological branch innovatively introduces graph modeling of extracted strong-scattering points and leverages graph convolutional networks (GCNs) to extract implicit physical structural priors inherent in SAR ship targets, which are neglected by existing visual-only learning methods. A cross-branch feature fusion strategy is further developed to aggregate semantic and topological representations, yielding a robust feature embedding with strong intra-class compactness and inter-class separability under data scarcity. Experiments on the FUSARShip dataset validate that our method achieves superior performance compared with state-of-the-art competitors in both 1-shot and 5-shot tasks.