欢迎访问《中国农学通报》,

中国农学通报 ›› 2021, Vol. 37 ›› Issue (7): 138-143.doi: 10.11924/j.issn.1000-6850.casb2020-0430

所属专题: 植物保护 园艺

• 农业信息·科技教育 • 上一篇    下一篇

基于迁移学习的园艺作物叶部病害识别及应用

李博1(), 江朝晖1,2(), 谢军1, 饶元1,2, 张武1,2   

  1. 1安徽农业大学信息与计算机学院,合肥 230036
    2智慧农业技术与装备安徽省重点实验室,合肥 230036
  • 收稿日期:2020-09-07 修回日期:2020-12-29 出版日期:2021-03-05 发布日期:2021-03-17
  • 通讯作者: 江朝晖
  • 作者简介:李博,男,1995年出生,安徽淮南人,硕士研究生,研究方向:智能信息处理。通信地址:230036 安徽省合肥长江西路130号 安徽农业大学信息与计算机学院,E-mail: iotboy@163.com
  • 基金资助:
    智慧农业技术与装备安徽省重点实验室自主创新研究基金“基于边缘智能的病虫害在线检测”(APKLSATE2019X002);安徽高校自然科学研究重大项目“视频感知+迁移学习的茶树病害智能监测”(KJ2019ZD20);安徽省科技攻关项目“面向皖南茶园的水肥精准调控关键技术研发与应用”(1804a07020108);安徽省科技攻关项目“大别山区云农场智慧服务平台关键技术研发与集成示范”(201904a06020056)

Leaf Disease Recognition of Horticultural Crops Based on Transfer Learning and Its Application

Li Bo1(), Jiang Zhaohui1,2(), Xie Jun1, Rao Yuan1,2, Zhang Wu1,2   

  1. 1College of Information and Computer Science, Anhui Agricultural University, Hefei 230036
    2Anhui Key Laboratory of Intelligent Agricultural Technology and Equipment, Hefei 230036
  • Received:2020-09-07 Revised:2020-12-29 Online:2021-03-05 Published:2021-03-17
  • Contact: Jiang Zhaohui

摘要:

为给农户提供质优价廉的园艺作物叶部病害识别服务,提出基于迁移学习的模型训练及基于Flask的Web部署方法。对PlantVillage数据集进行预处理,分别使用ResNet18、ResNet50和ResNet152 3种模型进行迁移学习训练,得到3种识别模型。利用Flask将模型部署到服务器上。3种识别模型对苹果等14类园艺作物26种叶部病害的平均识别准确率分别是95.61%、96.63%和97.33%,识别单张图像的时间分别是10.9、17.9、33.7 ms。综合考虑模型特点和用户期望,设计快速、标准和准确3种识别模式,实现深度模型在服务器中稳定运行,具有一定的实用价值。

关键词: 园艺作物, 病害识别, ResNet, 迁移学习, Web部署

Abstract:

To provide farmers with recognition services for horticultural crop leaf disease more conveniently and economically, a model training method based on transfer learning and a Web deployment method based on Flask are proposed. The PlantVillage dataset is preprocessed, and three recognition models are obtained by transfer learning using ResNet18, ResNet50 and ResNet152 models. Then, these three models are deployed to the server using Flask. The average recognition accuracy of the three models for 26 leaf diseases of 14 horticultural crops, such as apple, is 95.61%, 96.63% and 97.33%, respectively, and the recognition time of a single image is 10.9, 17.9 and 33.7 ms, respectively. Considering the characteristics of the model and users’ expectation, three fast, standard and accurate recognition patterns are designed to realize the stable operation of the deep model in the server, which has some practical value.

Key words: horticultural crop, disease recognition, ResNet, transfer learning, web deployment

中图分类号: