基金项目:国家自然科学基金(编号:６１６０１０７６)作者简介:刘敬宇,女,大连大学在读硕士研究生.通信作者:裴悦琨(１９８５—),男,大连大学讲师,研究生导师,博士.EＧmail:peiyuekun＠１２６．com收稿日期:２０２２Ｇ０５Ｇ０７改回日期:２０２２Ｇ０９Ｇ１５DOI:１０．１３６５２/j．spjx．１００３．５７８８．２０２２．８０３００[文章编号]１００３Ｇ５７８８(２０２３)０１Ｇ０１３９Ｇ０７基于改进YOLOX模型的樱桃缺陷及分级检测CherrydefectandclassificationdetectionbasedonimprovedYOLOXmodel刘敬宇１,２LIUJingＧyu１,２裴悦琨１,２PEIYueＧkun１,２常志远１,２CHANGZhiＧyuan１,２柴智１,２CHAIZhi１,２曹佩佩１,２CAOPeiＧpei１,２(１．大连大学辽宁省北斗高精度位置服务技术工程实验室,辽宁大连１１６６２２;２．大连大学大连市环境感知与智能控制重点实验室,辽宁大连１１６６２２)(１．BeidouHighPrecisionPositioningServiceTechnologyEngineeringLaboratoryofLiaoningProvince,DalianUniversity,Dalian,Liaoning１１６６２２,China;２．EnvironmentSensingandIntelligentControlKeyLaboratoryofDalian,DalianUniversity,Dalian,Liaoning１１６６２２,China)摘要:目的:实现工业化条件下樱桃的快速分级.方法:采用YOLOX网络对缺陷果进行检测,通过为特征金字塔网络设置适当的融合因子来提高不明显缺陷的检测精度,并将FocalLoss集成到损失函数中;使用YOLOX网络对完好果进行分级,引入注意力机制CBAM来加强网络特征提取.结果:樱桃表面缺陷的平均检测精度为９７．５９％,大小和颜色分级的平均检测精度为９５．９２％.结论:改进后的YOLOX网络可明显提升樱桃缺陷及分级检测的精度.关键词:樱桃分级;YOLOX;FPN;FocalLoss;注意力机制Abstract:Objective:Inordertoexpandthescopeofcherrysalesandachieverapidgradingofcherriesunderindustrialconditions．Methods:Firstly,theYOLOXnetworkwasusedtodetectthedefectivefruit,inordertosolvesomeproblemswherethedefectwasnotobvious．Thedetectionaccuracyoftheinconspicuousdefectwasimprovedbysettingtheappropriatefusionfactorforthefeaturepyramidnetwork,andinordertosolvetheproblemofimbalancebetweenvarioustypesofrealsamples,FocalLosswasintegratedintothelossfunction．Then,theintactfruitwasgradedusingtheYOLOXnetwork,andtheattentionmechanismCBAMwasintroducedtoenhancethene...
