欢迎来到天天文库
浏览记录
ID:52176217
大小:1.92 MB
页数:7页
时间:2020-03-23
《基于深度学习特征匹配的铸件微小缺陷自动定位方法.pdf》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、第37卷第6期仪器仪表学报Vol37No62016年6月ChineseJournalofScientificInstrumentJun.2016基于深度学习特征匹配的铸件微小缺陷自动定位方法余永维,杜柳青,曾翠兰,张建恒(重庆理工大学机械工程学院重庆400054)摘要:针对射线实时成像检测中精密铸件微小缺陷自动定位的需要,提出一种基于深度学习特征匹配的铸件缺陷三维定位方法。模拟选择注意机制的中央周边差算法,提出以视觉显著度为尺度,从射线图像复杂背景中检测出微小缺陷及其区域,以定义的区域中央点为待匹配点;然后,提出构造深度卷积神经网络自动提取微小缺陷区域
2、的深度学习特征,通过深度学习特征矢量的相似度,实现在不同视角下投影图像中的同一微小缺陷点的自动匹配;最后,基于平移视差测距原理计算缺陷匹配点的三维空间坐标。实验表明,基于深度学习特征匹配的方法能够正确搜索平移前后投影图像中的同一缺陷点,以此为基础,利用视差测距原理实现了微小缺陷匹配点的自动准确定位,深度定位误差小于5.52%,能够满足对精密铸件微小缺陷智能评判的需要。关键词:射线图像;缺陷检测;深度学习;自动定位;神经网络+中图分类号:TP391.41TH878.1文献标识码:A国家标准学科分类代码:510.4050Automaticlocalizationm
3、ethodofsmallcastingdefectbasedondeeplearningfeatureYuYongwei,DuLiuqing,ZengCuilan,ZhangJianheng(CollegeofMechanicalEngineering,ChongqingUniversityofTechnology,Chongqing400054,China)Abstract:Aimingattherequirementoftheautomaticlocalizationofprecisecastingsmalldefectinradiographicreal
4、timeimagingdetection,athreedimensionallocalizationmethodforcastingdefectsbasedondeeplearningfeaturematchingisproposed.Simulatingthecentralperipheraldifferencealgorithmofselectiveattentionmechanism,takingthevisualsaliencyasthescale,thesmalldefectanditsregionaredetectedfromcomplexb
5、ackgroundoftherayimages.Thecentralpointdefinedinthedefectregionistakenasthepointtobematched.Thenthedeepconvolutionneuralnetworkisconstructedtoautomaticallyextractthedeeplearningfeaturesofthesmalldefectregion.Throughthesimilarityofthedeeplearningfeaturevector,theautomaticmatchingofth
6、esamesmalldefectpointintheprojectionimagesatdifferentviewinganglesisachieved.Finally,basedontheprincipleoftranslationparallaxdistancemeasurement,the3Dspatialcoordinatesofthedefectmatchingpointsarecalculated.Experimentresultsshowthatthemethodofdeeplearningfeaturematchingcancorrectlys
7、earchthesamedefectpointintheprojectionimagesbeforeandaftertranslation.Onthebasisofthisfeaturematching,theautomaticandaccuratelocalizationofthesmalldefectmatchingpointisrealizedusingtheprincipleofparallaxdistancemeasurement.Thedepthlocalizationerrorislessthan5.52%,whichcanmeettherequ
8、irementofthesmallde
此文档下载收益归作者所有