Scholarlink

<p>Artificial features for radar emitter recognition have clear physical meaning but evaluating their relevance and redundancy remains challenging, while deep neural networks yield powerful yet opaque representations. We propose DFSNet-IM, which employs a two-stage training strategy to quantify feature importance through weight and gradient values in the pairwise layer, capturing both the direct contribution of features to model output and the model’s sensitivity to feature changes. The method progressively eliminates unimportant features to achieve sparsification. We extend distance correlation to measure redundancy and incorporate it with minimal performance loss into the evaluation function for optimal subset selection. Applied to radar emitter signals, DFSNet-IM achieves 89.3% recognition accuracy at -5 dB signal-to-noise ratio, demonstrating effective feature selection for radar emitter recognition.</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

A novel deep feature selection network based on importance measure (DFSNet-IM) for radar emitter recognition

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Kang Yan,
  • Weidong Jin,
  • Yingkun Huang

摘要

Artificial features for radar emitter recognition have clear physical meaning but evaluating their relevance and redundancy remains challenging, while deep neural networks yield powerful yet opaque representations. We propose DFSNet-IM, which employs a two-stage training strategy to quantify feature importance through weight and gradient values in the pairwise layer, capturing both the direct contribution of features to model output and the model’s sensitivity to feature changes. The method progressively eliminates unimportant features to achieve sparsification. We extend distance correlation to measure redundancy and incorporate it with minimal performance loss into the evaluation function for optimal subset selection. Applied to radar emitter signals, DFSNet-IM achieves 89.3% recognition accuracy at -5 dB signal-to-noise ratio, demonstrating effective feature selection for radar emitter recognition.