林业工程学报,2019,4(3):112-117JournalofForestryEngineeringDOI:10.13360/j.issn.2096-1359.2019.03.017
收稿日期:2018-08-30㊀㊀㊀㊀修回日期:2018-10-10
舒肝和胃丸的成分基金项目:国家林业局 948 项目(2014-4-48);江苏省政策引导类计划(国际科技合作)项目(BZ2016028)㊂
作者简介:范佳楠,男,研究方向为木材无损检测㊁机电一体化㊂通信作者:刘英,女,教授㊂E⁃mail:lying_new@163.com基于FasterR⁃CNN的实木板材缺陷检测识别系统
范佳楠,刘英∗,胡忠康,赵乾,沈鹭翔,周晓林(南京林业大学机械电子工程学院,南京210037)好看的宫斗小说
儿童语言训练摘㊀要:我国木材资源有限,为了提高木材的利用率,采用机器视觉来实现木材缺陷快速而稳定的检测,不仅可以克服人工检测的低效率和木材缺陷识别的低准确率,而且对提高木材加工企业的智能化水平具有重要意义㊂为了高效㊁快速㊁准确地进行无损检测,采用深度学习方法,建立了一种基于快速深度神经网络的实木板材缺陷识别模型㊂首先采用ResnetV2结构对采集到的实木板材缺陷图像进
行特征提取,然后应用该模型对节子㊁孔洞等实木板材缺陷进行训练学习,最后构建了FasterR⁃CNN检测框架,并使用tensorflow开发平台对节子㊁孔洞等实木板材缺陷进行预测输出㊂具体选取了2000块杉木样本,通过旋转对原始的实木板材图像进行数据扩充,扩充后图像的80%作为训练集,20%作为验证集来进行仿真㊂仿真结果表明,该模型对实木板材节子缺陷检测正确率为98%,对实木板材孔洞缺陷检测正确率为95%,验证了将深度学习算法应用于实木板材缺陷检测中的有效性㊂
关键词:实木板材;板材缺陷识别;深度学习;FasterR⁃CNN;无损检测
中图分类号:TP391㊀㊀㊀文献标志码:A㊀㊀㊀文章编号:2096-1359(2019)03-0112-06
SolidwoodpaneldefectdetectionandrecognitionsystembasedonFasterR⁃CNN
管科FANJianan,LIUYing∗,HUZhongkang,ZHAOQian,SHENLuxiang,ZHOUXiaolin
(CollegeofMechanicalandElectronicEngineering,沉思默想
NanjingForestryUniversity,Nanjing210037,China)
Abstract:Chinahaslimitedtimberresources.Withthedevelopmentofthewoodprocessingindustry,thedemandsforthehighqualityofwoodproductshaveincreasedrapidly.Thetraditionalmethodsofwooddefectdetectioncouldnotmetthedemands.Itisveryimportanttoimprovethetechnologyofwooddefectdetectionandclassificationaccordingtothedifferentneeds.Inordertoimprovetheutilizationrateofwood,machinevisionwasusedfortherapidandsta⁃bledetectionofwooddefects.Themachinevisioncannotonlyovercomethelowefficiencyofthemanualdetectionandlowaccuracyofwooddefectidentification,butalsoplayanimportantroleinimprovingtheintelligencelevelofwoodprocessingenterprise
s.Inordertocarryoutnondestructivetestingefficiently,quicklyandaccurately,thisstudyuseddeeplearningmethodtoestablishasolidwoodplatedefectidentificationmodelbasedonfastdeepneuralnet⁃work.Firstly,theResnetV2structurewasusedtoextractthecollectedsolidwoodpaneldefectimagesforthefeatureextraction.Then,thismodelwasappliedtotrainandlearnthesolidwoodpaneldefects,suchasknotsandholes.Fi⁃nally,theFasterR⁃CNNdetectionframeworkwasconstructedandthedefectsoftheplanks,suchasknotsandholes,werepredictedandoutputusingthetensorflowdevelopmentplatform.Specifically,2000samplesofChinesefirwereselected,andtheoriginalsolidwoodplateimageswereexpandedbytherotation.80%oftheexpandedimageswereusedasthetrainingsetand20%oftheexpandedimageswereastheverificationsetforthesimul
ation.Thesimulationresultsshowedthattheaccuracyofthismodelwas98%forthedetectionofknotsdefectsinsolidwoodplates,and95%forthedetectionofholedefectsinsolidwoodplates,whichverifiedtheeffectivenessofapplyingthedeeplearn⁃ingalgorithmtothedetectionofdefectsofsolidwoodplates.Keywords:solidwoodpanel;paneldefectdetectionandrecognition;deeplearning;FasterR⁃CNN;non⁃destructivetesting㊀㊀目前,木材缺陷的分类识别尚未智能化,人工检测依然比较普遍,光靠人工检测不仅分类效率正月十五组合
心随风低,而且处理效果不够理想㊂因此需要计算机来参红巾军
与处理,很多学者在原有图像处理技术上不断研究㊁结合新的方法和理论,以实现木材缺陷的自动分类识别㊂支持向量机(supportvectormachine,SVM)是Vapnik提出的一种基于结构化风险最小化的统计学习方法[1],通过在特征空间构建具有