Scholarlink

<p>Automotive radar needs to perceive various known and unknown targets in dynamic and complex environments. Moreover, the collection and annotation of radar data are usually costly, leading to a limited number of training samples. The existing methods suffer from two major limitations: inadequate handling of extreme samples and reliance on manually set thresholds for open-set recognition. To solve these two problems, a few-shot recognition method based on a filter system and prototype network (PN) is proposed. First, in the feature extraction stage, the mobile-ConViT network is proposed, which incorporates the MobileViTv2 Attention mechanism into the original PN structure, balancing model performance and real-time efficiency. Secondly, in the prototype computation module, a filtering system is employed to assign weights to samples based on their quality, thereby mitigating the impact of extreme samples. Furthermore, the metric module is extended for open-set recognition by introducing an adaptive threshold. Experimental results demonstrate that the proposed method improves recognition accuracy by approximately 3% compared to existing approaches and reduces the variance of accuracy across multiple runs by about 3%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Few-shot radar emitter signal recognition based on prototype network with filter system

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Yanping Liao,
  • Shengwen Lin,
  • Yihan He

摘要

Automotive radar needs to perceive various known and unknown targets in dynamic and complex environments. Moreover, the collection and annotation of radar data are usually costly, leading to a limited number of training samples. The existing methods suffer from two major limitations: inadequate handling of extreme samples and reliance on manually set thresholds for open-set recognition. To solve these two problems, a few-shot recognition method based on a filter system and prototype network (PN) is proposed. First, in the feature extraction stage, the mobile-ConViT network is proposed, which incorporates the MobileViTv2 Attention mechanism into the original PN structure, balancing model performance and real-time efficiency. Secondly, in the prototype computation module, a filtering system is employed to assign weights to samples based on their quality, thereby mitigating the impact of extreme samples. Furthermore, the metric module is extended for open-set recognition by introducing an adaptive threshold. Experimental results demonstrate that the proposed method improves recognition accuracy by approximately 3% compared to existing approaches and reduces the variance of accuracy across multiple runs by about 3%.