基于置信学习的低标注率辐射源个体识别算法
Low labeling rate specific emitter identification algorithm based on confidence learning
王艺卉 1闫文君 2凌青 2段可欣 3于楷泽3
作者信息
- 1. 海军航空大学,山东 烟台 264001;中国人民解放军31401部队,山东 烟台 264001
- 2. 海军航空大学,山东 烟台 264001
- 3. 海军航空大学,山东 烟台 264001;中国人民解放军91423部队,山东 烟台 264001
- 折叠
摘要
针对仅有少量标签数据的样本弱标注情况下辐射源个体识别难的问题,提出了一种基于置信学习的伪标签校正的辐射源个体识别方法.首先通过动态调整类内置信度,实现生成伪标签地及时校正;其次,分析样本价值,对影响模型性能的少量关键样本进行人工标注;然后利用联合交叉熵与中心损失函数,并叠加动态变化的伪标签置信度损失,同时关注类间、类内差异,最大化利用数据特征信息,实现的类内聚合和类间分离,最终实现了网络深度合理、精度与速度良好平衡的识别效果.实验结果表明:所提算法在有标记样本占比为 5%、10%、20%、50%、100%等多种条件下,均可实现辐射源个体有效识别,尤其在有标签数据低占比的情况下优势明显,识别准确率分别突破 70%与80%,有效减轻了有限标记样本的不足问题.
Abstract
In order to solve the problem of weak labeling of samples with only a small amount of labeled data,a false label correction method based on confidence learning is proposed to identify individual radiation sources.Firstly,by dynamically adjusting the in-class confidence,the generated false labels can be corrected in time.Secondly,the sample value is analyzed and a small number of key samples that affect the performance of the model are manually labeled.Then,by using the joint cross-entropy and center loss function,the dynamic pseudo-label confidence loss is superimposed,and the inter-class difference and intra-class difference are paid attention to,so as to maximize the intra-class aggregation and inter-class separation achieved by using the data feature information.Experimental results show that the proposed algorithm can achieve effective identification of individual radiation sources under various conditions,such as 5%,10%,20%,50%,100%,especially in the case of low proportion of labeled data,the advantages are obvious,and the recognition accuracy rate exceeds 70%and 80%respectively,effectively alleviating the shortage of limited labeled samples.The recognition effect of reasonable network depth,good balance between precision and speed is realized.
关键词
辐射源个体识别/伪标签校正/样本价值分析/伪标签损失函数Key words
specific emitter identification/pseudo label correction/sample value analysis/loss function of pseudo label引用本文复制引用
基金项目
国家自然科学基金面上项目(62271499)
电磁空间安全全国重点实验室开放基金()
出版年
2024
NETL
NSTL
万方数据