2023年第8期仪表技术与传感器InstrumentTechniqueandSensor收稿日期:2022-12-11基于CascadeRCNN的热轧带钢表面缺陷检测陆尧,薛林,王云森,王豪(大连理工大学机械工程学院,辽宁大连116024)摘要:在工业生产过程中,带钢表面产生的缺陷影响其质量和使用性能,需要对表面缺陷进行检测,为此提出了基于深度学习的缺陷检测模型CascadeRCNN。首先,对骨干网络进行改进,将标准卷积替换为可切换空洞卷积,在不增加参数量的情况下,增大输出单元的感受野。其次,改变特征金字塔FPN,结构不变的情况下添加了自上而下的连接方式,同时使用特征上采样算子CARAFE替换最邻近上采样,提高了上采样精度和定位精度。最后,将损失函数换为FocalLoss,解决目标检测过程中正负样本不平衡问题。结果显示:通过以上方法的改进,检测精度有大幅提升,平均均值精度提高了7.61%,达到77.82%,各类缺陷的检测精度都得到了提高;与其他检测模型对比,模型的检测能力得到了提高,采用的改进方法有一定的应用价值。关键词:深度学习;缺陷检测;CascadeRCNN;热轧带钢;特征上采样算子;FocalLoss中图分类号:TP391文献标识码:A文章编号:1002-1841(2023)08-0101-06SurfaceDefectDetectionofHotRolledStripBasedonCascadeRCNNLUYao,XUELin,WANGYunsen,WANGHao(SchoolofMechanicalEngineering,DalianUniversityofTechnology,Dalian116024,China)Abstract:Defectsonthesurfaceofthestripwillaffectitsqualityandserviceabilityduringindustrialproduction.Detectionofsurfacedefectsisnecessary.Therefore,adefectdetectionmodelbasedondeeplearning,CascadeRCNN,wasproposed.Firstly,thebackbonenetworkwasimprovedbyreplacingthestandardconvolutionwithswitchablevoidconvolution,whichincreasedthereceptivefieldoftheoutputunitwithoutincreasingthenumberofparameters.Secondly,thefeaturepyramidFPNwaschanged,andthetop-downconnectionmodewasaddedwithoutchangingthestructure.Meanwhile,thefeatureupsamplingoperatorCA-RAFEwasusedtoreplacethenearestupsampling,whichimprovedtheupsamplingaccuracyandpositioningaccuracy.Finally,theFocalLossreplacedbylossfunctionwasusedtosolvetheproblemofunbalancedpositiveandnegativesamples.Theresultsshowthatthedetectionaccuracyisgreatlyimprovedbytheimprovementoftheabovemethods.Themeanaverageprecisionwasin-creasedby7.61%,whichcanreach77.82%,...
