现代雷达2025，Vol.47Issue(11)：29-37,9.DOI:10.16592/j.cnki.1004-7859.2025062401
基于混合监督对比学习的小样本多视角SAR目标识别
Hybrid Supervised Contrastive Learning for Few-shot Multi-view SAR Target Recogniton
顾敏慧 1杨阳 2张仁李 1顾村锋2
作者信息
- 1. 南京理工大学 电子工程与光电技术学院,江苏 南京 210094
- 2. 上海机电工程研究所,上海 201109
- 折叠
摘要
Abstract
Insufficient labeled samples and diverse viewing angles are two major challenges that constrain the performance of syn-thetic aperture radar(SAR)target recognition.To address these issues,a hybrid supervised contrastive learning framework tailor-ed for few-shot multi-view SAR image recognition is proposed.The proposed method divides a training process into two stages:an upstream representation learning stage and a downstream classifier learning stage,incorporating a multi-view feature fusion strategy to exploit complementary information across different viewing angles.In the upstream stage,various random augmentation opera-tions are applied to SAR image patches to enlarge the effective batch size and construct diverse positive and negative sample pairs.These augmented samples are fed into a shared multi-view fusion network and a projection head,which maps the features into a contrastive space.A hybrid supervised contrastive module then jointly performs unsupervised instance discrimination and super-vised label discrimination,improving both the discriminability of the feature space and the intra-class compactness.In the down-stream stage,the projection head is removed,only a linear classifier is trained to perform the final recognition task.Experimental results on the few-shot moving and stationary target acquisition and recognition dataset demonstrate that the proposed method a-chieves significant improvements in recognition accuracy and generalization ability over conventional supervised and some unsuper-vised learning methods.These results validate the effectiveness and application potential of the proposed model in few-shot multi-view SAR recognition scenarios.关键词
合成孔径雷达/小样本目标识别/混合监督对比学习/图像注意力/多视角融合Key words
synthetic aperture radar(SAR)/few-shot target recognition/hybrid supervised contrastive learning/image attention/multi-view feature fusion分类
信息技术与安全科学引用本文复制引用
顾敏慧,杨阳,张仁李,顾村锋..基于混合监督对比学习的小样本多视角SAR目标识别[J].现代雷达,2025,47(11):29-37,9.
