基于人工-真实样本度量学习的指针式仪表检测方法  

Pointer meter detection method based on artificial-real sample metric learning

作　　者：翟永杰[1] 赵振远 王乾铭 白康[1] Zhai Yongjie;Zhao Zhenyuan;Wang Qianming;Bai Kang(Department of Automation,North China Electric Power University,Baoding 071003,Hebei,China)

机构地区：[1]华北电力大学自动化系,河北保定071003

出　　处：《电测与仪表》2022年第10期174-183,共10页Electrical Measurement & Instrumentation

基　　金：国家自然科学基金资助项目(U21A20486,61871182);河北省自然科学基金资助项目(F2020502009,F2021502008,F2021502013)。

摘　　要：为缓解指针式仪表检测精度对样本数量的严重依赖,有效提升少样本情况下指针式仪表的检测精度,提出了基于人工-真实样本度量学习的指针式仪表检测方法。通过对指针式仪表结构进行统计分析,提取其显著特征进行建模,用以生成所需要的人工基准样本,弥补真实场景下指针式仪表数据缺乏的问题;结合度量学习的特性,以Faster R-CNN为基线模型,引入特征相似性度量模块,从低维特征向量空间降低或消除人工基准样本和真实样本之间的分布差异,并加强特征提取网络对指针式仪表显著特征的学习。实验证明,较基线模型,基于人工-真实样本度量学习的指针式仪表检测方法AP^(75)提升了22.14%,有效提高了少样本情况下指针式仪表检测的精度。In order to alleviate the severe dependence of the detection accuracy of pointer meter on the number of samples,and effectively improve the detection accuracy of pointer meter under the condition of few shot,a pointer meter detection method based on artificial-real sample metric learning is proposed.Firstly,the structure of pointer meter is statistically analyzed,and its significant features are extracted for modeling,so as to generate the required artificial benchmark samples to make up for the lack of pointer meter data in real scenario.Then,combined with the characteristics of metric learning,the Faster R-CNN is used as the baseline model to introduce the feature similarity metric module to reduce or eliminate the distribution difference between the artificial benchmark sample and the real sample in the low-dimensional feature vector space,and strengthen the feature extraction network to learn the significant features of the pointer meter.Experimental results show that,compared with the baseline model,AP^(75) of the pointer meter detection method based on artificial-real sample metric learning improves by 22.14%,which effectively improves the accuracy of pointer meter detection in the case of few shot.

关 键 词：指针式仪表 度量学习 目标检测 少样本 人工样本 

分 类 号：TM933[电气工程—电力电子与电力传动] TP216[自动化与计算机技术—检测技术与自动化装置] TP391.4[自动化与计算机技术—控制科学与工程]