基金项目:国家重点研发计划项目重点专项项目子课题(编号:2018YFD0700305)收稿日期:2022-11-16基于Transformer模型的干辣椒等级分类方法研究*邱显焱,郭士超※,王俊杰(湖南工业大学机械工程学院,湖南株洲412007)摘要:针对生产出的干辣椒品相不一的问题,提出基于Transformer模型的干辣椒等级分类识别方法。根据干辣椒特征,确定干辣椒等级标准,包括优质干辣椒、合格干辣椒和不合格干辣椒3个等级。制备干辣椒数据集并划分为训练集和测试集。首先结合Vi‐sualTransformer(ViT)网络和分类网络,引入并改进ShiftedwindowsTransformer(SwinTransformer)网络,使用高斯误差线性单元(GELU)做激活函数,使用自适应矩估计(Adam)做优化函数。通过加载在ImageNet数据集训练的权重进行模型初始化,表明迁移学习方式可有效提高模型特征提取能力。通过对比ViT模型和SwinTransformer模型两种模型来训练干辣椒数据集,表明SwinTransformer模型测试准确率较高。实验表明,利用迁移学习的SwinTransformer模型准确率最高,达到95.26%,这为干辣椒等级分拣问题提供了新的解决办法,同时可作为其他果蔬品相识别的参考。关键词:干辣椒;品质分类;Transformer;迁移学习中图分类号:TP391文献标志码:A文章编号:1009-9492(2023)02-0034-04StudyonGradeClassificationMethodofDryChiliBasedonTransformerModelQiuXianyan,GuoShichao※,WangJunjie(SchoolofMechanicalEngineering,HunanUniversityofTechnology,Zhuzhou,Hunan412007,China)Abstract:Inviewoftheproductionofdrypepperproductsinconsistentsituation,aclassificationandrecognitionmethodofdrychilibasedontransformermodelwasproposed.Accordingtothecharacteristicsofdrychili,thegradestandardofdrychiliwasdetermined,includinghighqualitydrychili,qualifieddrychiliandunqualifieddrychili.Drychilidatasetwaspreparedanddividedintotrainingsetandtestset.Firstly,combinedwiththeVisualTransformer(ViT)networkandtheclassificationnetwork,thentheShiftedwindowsTransformer(SwinTransformer)networkwasintroducedandimproved,usingtheGaussianErrorLinearityUnit(GELU)astheactivationfunctionandtheAdaptiveMomentEstimation(Adam)astheoptimizationfunction.ByloadingweightstrainedinImageNetdatasetformodelinitialization,itisconcludedthatthetransferlearningmet...