基于轻量化MobileViT深度学习模型的烤烟自动分组方法

Automatic Grouping Method of Flue-cured Tobacco Based on MobileViT

针对烟叶自动识别分组过程中特征提取困难、分组准确率低、模型参数量大、部署困难等问题,本文结合轻量化网络MobileNetV2与ViT模型的优点提出了一种轻量化MobileViT模型的烤烟自动分组方法。首先对所采集的烟叶图像进行前景和背景的预处理,解决提取烟叶特征困难的问题;然后将预处理后的图像建立成完整的烟叶图像数据集;最后针对建立的数据集使用基于轻量化MobileViT模型进行分组。使用该模型对云南省多地收购的5871张混合烤烟叶片进行分组试验,结果表明MobileViT对正组烟叶、副组烟叶和正副混合组烟叶分组准确率分别达到了79.44%、83.83%、79.22%,与MobileNetV2相比分别提升了9.5%、17.66%、17.56%;与ViT模型相比分别提升了13.8%、30.83%、34.12%。与当前应用较多的MobileNetV3、Effcient、Resnet50三种模型相比,MobileViT模型在混合组上分组准确率较轻量化MobileNetV3网络提升了11.86%;较CNN为代表的Effcient、Resnet50模型分别提升了9.6%、6.55%,模型大小降低了24.62%、79.1%。轻量化MobileViT模型同时结合了轻量化网络MobileNetV2与ViT模型的优点,在降低模型大小的同时具有较高的分组准确率,更容易部署在工业设备中,符合实际工业应用需求。

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.