基于改进MobileNetV3—Small的甘薯外部品质分类方法  

External quality classification method of sweet potato based on improved MobileNetV3-Small

作　　者：马瑞峻[1] 丁世春 陈瑜[1] Ma Ruijun;Ding Shichun;Chen Yu(College of Engineering,South China Agricultural University,Guangzhou,510642,China)

机构地区：[1]华南农业大学工程学院,广州市510642

出　　处：《中国农机化学报》2025年第4期211-217,共7页Journal of Chinese Agricultural Mechanization

基　　金：广东省科技计划项目(2021B1212040009)。

摘　　要：传统图像处理技术依靠人工提取特征,费时费力且难以提取到准确的特征。为准确实现对甘薯发芽、霉腐、损伤和正常品质的分类,提出一种改进的MobileNetV3—Small(M3S)分类方法。使用高效通道注意力(ECA)模块替换M3S中的压缩激励(SE)模块,构建ECA—M3S模型结构;基于迁移学习训练模型,并对比不同学习率组合的训练效果;测试甘薯品质分类模型的性能,同时和多种模型进行对比,并使用Flask设计网页界面展示测试结果。结果表明,初始学习率为0.01,学习率衰减速率为0.5时,模型整体性能最优,验证准确率为92.82%,训练损失为0.0492;和其他10种不同复杂度的模型进行对比,该模型对4类甘薯品质的召回率均高于90%,测试平均准确率为92.43%,仅比最高的模型低0.79%,比未改进的M3S高3.59%,且模型尺寸仅为4.18 MB,仅比尺寸最小的SqueezeNet模型大1.34 MB。Traditional image processing technology relies on manual feature extraction,which is time-consuming,labor-intensive and difficult to extract accurate features.In order to accurately realize the classification of sweet potato germination,mildew,damage and normal quality,an improved MobileNetV3—Small(M3S)online classification method is proposed.The efficient channel attention(ECA)module was used to replace the squeeze and excitation(SE)module in M3S,the ECA—M3S model structure was constructed.The model was trained based on transfer learning,and the training effects of different learning rate combinations were compared.The performance of the trained sweet potato quality classification model was tested and compared with a variety of models,and the web interface was designed by Flask to display the test results.The results show that the overall performance of the model is the best when the initial learning rate is 0.01 and the learning rate attenuation rate is 0.5,the verification accuracy is 92.82%,and the training loss is 0.0492.Compared with 10 other models of different complexity,the recall rate of the four types of sweet potato quality was higher than 90%,the average test accuracy was 92.43%,only 0.79%lower than the highest,3.59%higher than the unimproved M3S,and the model size was only 4.18 MB,only 1.34 MB larger than the smallest SqueezeNet model.

关 键 词：甘薯 外部品质 MobileNetV3—Small 高效通道注意力 迁移学习 

分 类 号：S531[农业科学—作物学] TP391.4[自动化与计算机技术—计算机应用技术]