摘要
关键部件缺陷图像自动检测对于复兴号动车组运营维护意义重大,但目前主要依靠专业人员对检测图像进行分析,耗费大量人力、物力,造成检测周期长,检测准确率无法保证。提出一种结合部件检测与缺陷分类流程的双通道缺陷检测框架MCDDF(Multi-channel Defect Detection Framework),部件检测通道基于目标检测算法实现动车组关键部件定位,定位后的关键部件经裁剪进行超分辨率提升,传入缺陷分类通道基于迁移学习方法实现缺陷类别的准确分类,结合两通道信息实现缺陷检测任务。实验分析两通道的性能提升方法,对比MCDDF与传统基于目标检测方法在铁路关键部件缺陷图像上的检测效果,验证了MCDDF方法的有效性。
Automatic detection of defect images of key components is of great significance for the operation and maintenance of the Fuxing Electric Multiple Units.However,the current reliance on manual analysis of detection images by professional personnel consumes manpower and resources,resulting in a long detection cycle and uncertain detection accuracy.In this paper,a framework which combines component detection and defect classification,called multi-channel defect detection framework (MCDDF ),was proposed.The component detection channel based on object detection algorithm can locate key components of EMU.The located components were clipped for super-resolution enhancement,then were sent to detect classification channel to realize accurate classification of defect categories based on transfer learning method.In the experiments,the performance improvement of two channels was analyzed respectively.The performances of MCDDF and traditional object detection methods on railway key components defect detection task were compared,where the effectiveness of the MCDDF method was verified.
作者
赵冰
代明睿
李平
马小宁
吴艳华
ZHAO Bing;DAI Mingrui;LI Ping;MA Xiaoning;WU Yanhua(Department of Postgraduates,China Academy of Railway Sciences,Beijing 100081,China;Railway Big Data Research and Application Innovation Center,China Academy of Railway Sciences,Beijing 100081,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2019年第8期67-73,共7页
Journal of the China Railway Society
基金
国家重点研发计划(2018YFB1201403)
中国铁路总公司科技研究开发计划(2017J003-D)
关键词
缺陷检测
卷积神经网络
目标检测
缺陷分类
图像超分辨率
defect detection
convolutional neural network
object detection
detect classification
image super resolution
