电子测量与仪器学报

牟欣颖,何赟泽,王洪金,邓堡元,杨渊,周可,杨瑞珍.基于一维卷积神经网络的联动扫描热成像缺陷自动识别与深度回归[J].电子测量与仪器学报,2021,35(4):211-217

基于一维卷积神经网络的联动扫描热成像缺陷自动识别与深度回归

Joint scanning thermography defect automatic classifier anddepth regression based on 1D CNN

DOI：

英文关键词:infrared thermography machine vision deep learning defect

基金项目:国防科技大学装备综合保障技术重点实验室基金（6142003200205）、国家自然科学基金青年科学基金（Z20190142984）、湖南省科技创新计划项目科技人才专项（2018RS3039）、长沙市杰出创新青年培养计划（kq1802023）、长沙市科技计划项目（CSKJ2020 19）资助

作者	单位
牟欣颖	1. 湖南大学电气与信息工程学院
何赟泽	1. 湖南大学电气与信息工程学院
王洪金	1. 湖南大学电气与信息工程学院，2. 国防科技大学装备综合保障技术重点实验室
邓堡元	1. 湖南大学电气与信息工程学院
杨渊	1. 湖南大学电气与信息工程学院
周可	1. 湖南大学电气与信息工程学院
杨瑞珍	1. 湖南大学电气与信息工程学院，3. 长沙学院土木工程学院

Author	Institution
Mu Xinying	1. College of Electrical and Information Engineering, Hunan University
He Yunze	1. College of Electrical and Information Engineering, Hunan University
Wang Hongjin	1. College of Electrical and Information Engineering, Hunan University，2. Laboratory of Science and Technology on Integrated Logistics Support, National University of Defense Technology
Deng Baoyuan	1. College of Electrical and Information Engineering, Hunan University
Yang Yuan	1. College of Electrical and Information Engineering, Hunan University
Zhou Ke	1. College of Electrical and Information Engineering, Hunan University
Yang Ruizhen	1. College of Electrical and Information Engineering, Hunan University，3. School of Civil Engineering, Changsha University

摘要点击次数: 816

全文下载次数: 4

中文摘要:

联动扫描热成像(joint scanning thermography, JST)可以用于检测大面积对象的缺陷,但原始热图像缺陷信息模糊且无法实现缺陷定量。针对联动扫描热成像重构后的图像序列,提出了一种基于一维卷积神经网络( one-dimensional convolutional neural network,1D-CNN)的缺陷识别和定量方法,以图像序列中像素点对应的一维温度时间序列为网络输入,将缺陷深度作为输出,实现了碳纤维复合材料板中缺陷的自动检测和深度定量。实验结果显示,基于 1D-CNN 的检测方法准确实现了对缺陷自动检测,其对训练集数据的预测准确率最高可达 98. 8%,测试集准确率在 70%左右,相比传统处理方法取得了更好的效果。

英文摘要:

Joint scanning thermography(JST) can detect defection of large-area materials. The defection of raw images is inaccurate and the quantitative analysis is hard to achieve. According to the characteristics of images from the reconstruction of joint scanning thermography, a method based on an one-dimensional convolutional neural network ( 1D-CNN) is proposed to detect and quantitate defects. The one-dimensional temperature time series corresponding to the pixels of the pulse image sequences is applied as inputs for the network. This method could achieve defect detection automatically and defect quantification for carbon fiber reinforced polymer. As the result indicated, the 1D-CNN based method could detect defection automatically and accurately. It has a 98. 8% accuracy in defect classifying of training set and an about 70% accuracy in defect classifying of training set. The result is better than traditional method.

查看全文查看/发表评论下载PDF阅读器