引用本文格式逄正钧,董峦,温钊发,等.基于YOLOX的作物种子自动计数方法[J].农业工程,2023,13(1):29-35.DOI:10.19998/j.cnki.2095-1795.2023.01.005.PANGZhengjun,DONGLuan,WENZhaofa,etal.AutomaticcropseedcountingmethodbasedonYOLOXmodel[J].AgriculturalEngineering,2023,13(1):29-35.基于YOLOX的作物种子自动计数方法逄正钧,董峦,温钊发,张世豪,秦立浩(新疆农业大学计算机与信息工程学院,新疆乌鲁木齐830052)摘要:种子计数是获取作物种子千粒质量指标时关键而又烦琐的步骤。目前种子计数一般通过人工和千粒质量测量仪器实现,然而人工计数效率低,千粒质量测量仪器成本高、不易携带。以手机拍摄的6种常见作物种子图像构建数据集,在YOLOX模型的基础上引入注意力机制改进损失函数提出YOLOX-P模型,实现种子自动计数。结果表明,YOLOX-P相比YOLOX模型参数量仅增加0.09M,mAP改进0.74个百分点,达到99.38%;模型在显存6GB的NVIDIAGeForceRTX2060显卡上的推理时间为18.68ms,适宜部署在移动端。提出的模型显著改善千粒质量测定工作的效率和效果。关键词:千粒质量;种子计数;深度学习;目标检测;YOLOX中图分类号:S126文献标识码:A文章编号:2095-1795(2023)01-0029-07DOI:10.19998/j.cnki.2095-1795.2023.01.005AutomaticCropSeedCountingMethodBasedonYOLOXModelPANGZhengjun,DONGLuan,WENZhaofa,ZHANGShihao,QINLihao(CollegeofComputerandInformationEngineering,XinjiangAgriculturalUniversity,UrumqiXinjiang830052,China)Abstract:Seedcountingisthemostcriticalandtediousstepinacquisitionofkilograinweightindexofcropseeds.Atpresent,seedcountingisgenerallyachievedbymanualandspecializedequipment.Butmanualcountingefficiencyislowandspecializedequipmentareexpensiveandnoteasytocarry.Datasetswereconstructedusingimagesofsixdifferentcommoncropseedstakenbymobilephone.BasedonYOLOXmodel,attentionmechanismwasintroducedandlossfunctionwasimprovedtoobtaintheYOLOX-Pmodel,whichrealizedautomaticcountingofseeds.ResultsshowedthatYOLOX-Ponlyincreased0.09McomparedwithYOLOXmodelparameters,mAPimproved0.74percentagepointsandreached99.38%.ReasoningtimeofthemodelonNVIDIAGeForceRTX2060graphicscardwith6GBvideomemorywas18.68ms,whichwassuitablef...