基于深度学习的芒果病虫害分类识别

来源：计算机技术与发展杂志2023年第10期北京时间：

作者：曹欢;方睿

单位：成都信息工程大学 计算机学院,四川 成都 610225 Author(s): CAO Huan;FANG Rui School of Computer,Chengdu University of Information Technology,Chengdu 610225,China 关键词: 芒果病虫害识别;轻量级卷积神经网络;MobileViT;迁移学习;MangoLeafBD Keywords: mango diseases and pests identification;lightweight convolutional neural network;MobileViT;transfer learning;MangoLeafBD 分类号: TP391. 41;S41 DOI: 10. 3969 / j. issn. 1673-629X. 2023. 10. 018 摘要: 传统芒果病虫害防治,需要人工进行识别,现引入深度学习技术,可快速准确地对芒果病虫害进行识别。 以攀西地区芒果的 12 种病虫害为研究对象,采用的数据集一部分来自公开数据集 MangoLeafBD,另一部分由爬虫技术获得的网络图片组成,共获取图片 6 769 张,其中 4 879 张为训练集,1 220 张为验证集,670 张为测试集。 为迎合实际应用的需要,选择了 MobileNetV3、MobileViT 等 4 种不同规模的轻量级深度学习网络模型,结合迁移学习训练策略进行对比实验,比较了各个模型的参数量、精确率、召回率等参数。 实验结果显示,MobileViT 模型用于芒果病虫害分类识别效果最佳,该模型的精确率为 96. 31% ,召回率为 96. 12% ,F1 为 96. 20% ,均优于其他模型。 由此表明,模型具有较好的鲁棒性和识别性能,可为芒果病虫害分类识别提供技术参考。 Abstract: Traditional mango pest control requires manual identification. Now deep learning technology is introduced to quickly andaccurately identify mango pests and?diseases. In this paper,12 diseases and insect pests of mango in Panxi area are taken as the researchobject. One part of the data set is from the public data set MangoLeafBD,　and the other part is composed of network pictures obtained bycrawler technology. A total of 6 769 pictures were obtained,of which 4 879 were training sets,1 220 were validation sets,and 670 weretest sets. In order to meet the needs of practical applications,four lightweight deep learning networks with?different scales,such as MobileNetV3 and MobileViT,were selected. Combined with the transfer learning training strategy,comparative experiments were carried outto compare the parameters,accuracy and recall rate of each model. The experimental results show that the MobileViT model has the besteffect on the classification and recognition of mango diseases and insect pests. The accuracy rate of the model is 96. 31% ,the recall rateis 96. 12% ,and the F1 is 96. 20% ,which are better than that of other models. It is showed that such model has excellent robustness andrecognition performance,which can provide technical reference for mango pest classification and recognition.

摘要：传统芒果病虫害防治,需要人工进行识别,现引入深度学习技术,可快速准确地对芒果病虫害进行识别。 以攀西地区芒果的 12 种病虫害为研究对象,采用的数据集一部分来自公开数据集 MangoLeafBD,另一部分由爬虫技术获得的网络图片组成,共获取图片 6 769 张,其中 4 879 张为训练集,1 220 张为验证集,670 张为测试集。 为迎合实际应用的需要,选择了 MobileNetV3、MobileViT 等 4 种不同规模的轻量级深度学习网络模型,结合迁移学习训练策略进行对比实验,比较了各个模型的参数量、精确率、召回率等参数。 实验结果显示,MobileViT 模型用于芒果病虫害分类识别效果最佳,该模型的精确率为 96. 31% ,召回率为 96. 12% ,F1 为 96. 20% ,均优于其他模型。 由此表明,模型具有较好的鲁棒性和识别性能,可为芒果病虫害分类识别提供技术参考。

关键词：芒果病虫害识别;轻量级卷积神经网络;MobileViT;迁移学习;MangoLeafBD