文章编号:1002-2082(2023)01-0093-11基于MaskR-CNN结合边缘分割的颗粒物图像检测李轩,杨舟,陶新宇,王晓杰,莫绪涛,黄仙山(安徽工业大学数理科学与工程学院,安徽马鞍山243002)摘要:对颗粒物的尺寸检测是生产中重要的环节,使用相机采集图像并处理是常用的非接触检测方法。围绕颗粒物的识别与尺寸检测需求,选用沙粒为检测对象,提出了一种改进颗粒物边界掩膜的MaskR-CNN模型。该模型结合经典的边缘检测技术,并利用深度学习模型预测掩膜,根据边缘分割的结果来得到更高精度的掩膜。使用DenseNet作为检测网络的主干网络,使得整体网络参数量更少,并利用通道注意力机制加强网络的特征提取能力。实验结果表明,改进的网络可以提高检测的精度,且结合图像处理的方式能够改善掩膜尺寸检测的准确度,为颗粒物的工业检测提供了一种有意义的方法。关键词:颗粒物检测;深度学习;图像分割;机器视觉;尺寸分布中图分类号:TN911.73;TP391.4文献标志码:ADOI:10.5768/JAO202344.0102005ParticlesimagedetectionbasedonMaskR-CNNcombinedwithedgesegmentationLIXuan,YANGZhou,TAOXinyu,WANGXiaojie,MOXutao,HUANGXianshan(SchoolofMathematicsandPhysics,AnhuiUniversityofTechnology,Ma'anshan243002,China)Abstract:Particlessizedetectionisanimportantlinkinproduction,andtheuseofcamerastocaptureandprocessimagesisacommonly-usednon-contactdetectionmethod.Tomeettherequirementsofidentificationandsizedetectionofparticles,thesandparticleswereselectedasthedetectionobject,andaMaskR-CNNmodelwiththeimprovedboundarymaskofparticleswasproposed.Combinedwiththeclassicaledgedetectiontechnology,thedeeplearningmodelwasusedtopredictthemask,andthemaskwithhigherprecisionwasobtainedaccordingtotheresultsofedgesegmentation.TheDenseNetwasusedasthebackbonenetworkofthenetworkdetectiontoreducethenumberofnetworkparameters,andthechannelattentionmechanismwasusedtostrengthenthefeatureextractionabilityofthenetwork.Theexperimentsshowthattheimprovednetworkcanimprovethedetectionaccuracy,andthecombinationofimageprocessingcanimprovetheaccuracyofmasksizedetection,whichprovidesameaningfulmethodforindustrialdetectionofparticles.Keywords:particlesdetection;deeplearning;imagesegmentation...