AOT: Aggregation Optimal Transport for Few-Shot SAR Automatic Target Recognition

The lack of labeled data presents challenges for synthetic aperture radar (SAR) in automatic target recognition. To address this problem, few-shot learning (FSL) approaches have been developed to extract more knowledge from limited labeled data and prevent overfitting. However, existing SAR FSL methods treat SAR images as optical images, disregarding the image blurring caused by different imaging mechanisms. This introduces more uncertainty in the feature space and affects the classification results. Existing class-level classification methods ignore fine-grained information in SAR images, while sample-level methods are negatively impacted by the uncertainty in SAR. In this article, we propose substructure-level prototypes match (SSPM) for SAR FSL and provide an implementation named aggregation optimal transport (AOT) based on the optimal transport algorithm. The AOT contains a two-layer OT structure. In the first layer, the model learns multiple substructure-level prototypes (SLP) using information from unlabeled data, which can effectively remove the effects of imaging mechanisms on fine-grained information extraction. In the second layer, the model learns class-level prototypes (CLPs) together using transfer probabilities from both unlabeled data to SLPs and SLPs to labeled data. Finally, the unlabeled data is classified by two probabilistic transfer matrices. Experiments on two public databases named MSTAR and OpenSARShip verify the effectiveness of the proposed AOT method.