Automatic Grouping Method of Flue-cured Tobacco Based on MobileViT
Aiming for solving the problems of difficult feature extraction,low grouping accuracy,large number of model parameters and difficult deployment in automatic tobacco leaf identification and grouping process,an automatic tobacco grouping method based on the advantages of lightweight network MobileNetV2 and ViT model was developed in this study.First,pre-processing of the foreground and background of the acquired tobacco leaf images was conducted to solve the problem of extracting tobacco leaf features;And a complete tobacco leaf image data set was established based on the preprocessed image;Finally,the established data sets were grouped using a lightweight MobileViT model.The model was used to group 5871 mixed flue-cured tobacco leaves purchased from multiple places in Yunnan Province.The results showed that using the MobileViT's model,the grouping accuracy rates of positive group,deputy group and primary and secondary group reached 79.44%,83.83%and 79.22%respectively,and increased 9.5%,17.66%and 17.56%respectively compared with that of MobileNetV2;13.8%,30.83%and 34.12%increase in grouping accuracy rates of the three groups were observed using the MobileViT's model compared to the ViT model.Compared with the mobile NetV3,Efficient and Resenet50 models which are widely used currently,the packet accuracy of the mobile ViT model on the mixed group was increased by 11.86%compared with the lightweight mobile NetV3 network;Compared with the Effcient and Resnet50 models represented by CNN,the model size was increased by 9.6%and 6.55%respectively,and the model size was reduced by 24.62%and 79.1%.The lightweight MobileViT model could combine the advantages of the lightweight network MobileNetV2 and the ViT models,and has a high packet accuracy while reducing the size of the model,making it easier to be deployed in industrial equipments and meeting the actual industrial application requirements.
automatic grouping of flue-cured tobaccolightweightMobileVIT model
国家科技期刊平台
NSTL
万方数据
