| 引用本文: | 孙延鹏，石增辉，屈乐乐. 基于多尺度特征匹配的小样本雷达无人机识别[J]. 雷达科学与技术, 2026, 24(1): 74-82.[点击复制] |
| SUN Yanpeng, SHI Zenghui, QU Lele. FewShot Radar UAV Recognition Based on MultiScale Feature Matching[J]. Radar Science and Technology, 2026, 24(1): 74-82.[点击复制] |
|
| 摘要: |
| 为了提高在数据不足的情况下对雷达无人机的分类效果，本文提出一种基于多尺度特征增强和度量学习的小样本分类方法——基于高效网络（EfficientNet）的多尺度学习网络（EfficientNet?based Multi?scale Learning Network,EMLNet）。该方法在轻量化EfficientNet网络中引入高效多尺度注意力机制（Efficient Multi?scale Attention，EMA）进行特征提取，通过多尺度并行子网络与跨空间依赖建模能力，有效增强了特征的稳定性与判别能力。分类阶段通过引入基于局部特征匹配的方法，实现支持集与查询集之间的细粒度相似性建模，为提升训练稳定性与泛化性能，本文进一步结合原型损失（Prototypical Loss）与交叉熵损失（Cross?Entropy Loss）构建复合损失函数（PCE Loss），引导模型在优化类间判别性的同时保持类内特征聚集性。最后在开源的多普勒雷达数据集上开展实验，实验结果表明，所提方法在小样本场景下表现出显著的性能优势。 |
| 关键词: 无人机分类 连续波雷达 小样本学习 EfficientNet网络 度量学习 时频图 |
| DOI：DOI:10.3969/j.issn.1672-2337.2026.01.008 |
| 分类号:TN957.5 |
| 基金项目:国家自然科学基金（61671310）；航空科学基金（2019ZC054004）；辽宁省高校基本科研业务费(LJ222410143071) |
|
| FewShot Radar UAV Recognition Based on MultiScale Feature Matching |
|
SUN Yanpeng, SHI Zenghui, QU Lele
|
|
College of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
|
| Abstract: |
| In order to improve the classification accuracy of radar?based UAV signals under limited data conditions, this paper proposes a few?shot classification framework EMLNet （EfficientNet?based Multi?scale Learning Network ） based on multi?scale feature enhancement and metric learning. The method integrates an efficient multi?scale attention mechanism（EMA） into the lightweight EfficientNet backbone for feature extraction, which effectively enhances the stability and discriminative ability of features through multi?scale parallel sub?network with cross?spatial dependence modeling. In the classification stage, a local feature matching strategy is adopted to achieve fine?grained similarity modeling between the support set and the query set. To further improve training stability and generalization performance, a composite loss function （PCE Loss） that combines prototypical loss and cross?entropy loss is introduced to optimize inter?class discrimination while maintaining intra?class feature aggregation. Experiments are carried out on the open?source Doppler radar dataset. The experimental results demonstrate that the proposed method achieves significant performance advantages in few?shot learning scenarios. |
| Key words: UAV classification continuous wave radar few‑shot learning EfficientNet metric learning time‑frequency map |
