More diversity, less redundancy: Feature refinement network for few-shot SAR image classification

Abstract

Few-shot SAR image classification predicts new labels on SAR (synthetic aperture radar) images using only a few labeled samples. Due to the insufficient training data in few-shot learning, modern methods tend to extract as many features as possible, which ignore the redundancy caused by increasing features. In this paper, we observe that it is important to maintain a balance between the abundance and scarcity of features through feature refinement. Building on this insight, we proposed a Feature Refinement network for few-shot SAR image classification (FRSAR). FRSAR is a novel approach for feature balance through feature refinement, which consists of enrich feature extraction (EFE) block, redundant feature refinement (RFR) block, and key feature localization (KFL) block. The EFE block extracts complementary features to achieve a richer feature representation. The RFR block filters out redundant features through a two-stage reconstruction process. The KFL block further refines the classification features by adaptive computing thresholds. Extensive experiments on MSTAR dataset demonstrate that our FRSAR method can achieve better performance than other methods. For example, our method surpasses existing the state-of-the-art method by a large margin (+7.20/% on the 5-way 1-shot task and +7.22% on the 5-way 5-shot task). We believe the feature refinement framework can serve as a strong baseline for further research in wider communities.