文章编号：1671-7872(2023)02-0166-07基于轻量级卷积神经网络的注油孔检测算法孟瑞1a,1b，卢宗远1a，丛英浩2(1.安徽工业大学a.机械工程学院，b.工程研究院,安徽马鞍山243032；2.上海梅山钢铁股份有限公司,江苏南京210039)摘要：为解决当前注油机器人目标检测算法对现场适应性弱、识别率低等问题，通过改进YOLOv5算法，提出一种新的注油孔检测算法YOLOv5-S。将ShuffleNetv2用于图像的特征提取，通过调整输入图像的分辨率以及扩大部分基本单元中深度卷积核尺寸，确保算法既具有轻量级网络结构又具有高精度的检测水平；采集不同工况下注油孔的图像，将其分类并标注，采用YOLOv5-S，YOLOv5，YOLOv3-tiny算法对其进行训练实验，验证提出算法的有效性。结果表明：YOLOv5-S在注油孔数据集上的检测精度保持在99.4%，与原算法相比，其模型容量压缩了77%，检测速度提升了11.7F/s。本文提出的检测算法在工控机算力和存储资源有限的条件下具备良好的识别准确率和检测速度。关键词：轻量级卷积神经网络；机器视觉；目标检测；注油孔中图分类号：O436文献标志码：Adoi：10.12415/j.issn.1671−7872.22230OilInjectionHoleDetectionAlgorithmBasedonLightweightConvolutionNeuralNetworkMENGRui1a,1b,LUZongyuan1a,CONGYinghao2(1.a.SchoolofMechanicalEngineering,b.EngineeringResearchInstitute,AnhuiUniversityofTechnology,Maanshan243032,China;2.ShanghaiMeishanIronandSteelLimitedbyShareLtd,Nanjing210039,China)Abstract：Inordertosolvetheproblemsofweakadaptabilityandlowrecognitionrateofcurrenttargetdetectionalgorithmforoilinjectionrobots,anewoilinjectionholedetectionalgorithmYOLOv5-SwasproposedbyimprovingYOLOv5algorithm.ShuffleNetv2wasusedasthefeatureextractionoftheimage,theresolutionoftheinputimagewasadjusted,andthedepthconvolutionkernelsizeinsomebasicunitswasexpandedtoensurethatthealgorithmhasbothalightweightnetworkstructureandahigh-precisiondetectionlevel.Thecollectedoilinjectionholeimagesunderdifferentworkingconditionswereclassifiedandlabeled,andYOLOv5-S,YOLOv5,YOLOv3tinyalgorithmswereusedtotrainthemtoverifytheeffectivenessoftheproposedalgorithm.TheresultsshowthatthedetectionaccuracyofYOLOv5-Sontheoilinjectionholedatasetremainsat99.4%.Comparedwiththeoriginalalgorithm,itsmode...
