基于MobileNet V2的辣椒果实炭疽病识别及其应用

Identification of Anthracnose in Pepper Fruit Based on MobileNet V2 and the Application

【目的】构建基于MobileNet V2的辣椒炭疽病果实分类识别模型,开发移动端辣椒果实炭疽病识别系统,在田间离线状态下实现辣椒果实炭疽病的实时诊断,为及早防控辣椒炭疽病奠定基础。【方法】以朝天椒品种辣研101号为研究对象,首先对辣椒在体、离体状态下的健康果实和炭疽病果实图像进行数据扩充,然后将MobileNet V1、MobileNet V2和EfficientNet-B0模型参数大小相近的轻量级网络进行迁移学习,通过参数微调提取辣椒果实特征,对比图像分类效果,最后将识别精度最高的MobileNet V2网络移植到手机端。【结果】MobileNet V2网络在健康和炭疽病辣椒果实测试集上的查准率最高,达97.31%,单张图片识别时间最短,为75 ms,具有占用内存小、识别速度快及精准度高等优点。【结论】通过MobileNet V2网络模型的移植和应用,辣椒炭疽病识别系统可在田间离线状态下实现辣椒果实炭疽病的实时诊断。

【Objective】The classification and identification model of pepper fruit anthracnose based on MobileNet V2 was constructed,and the mobile terminal of identification system for pepper fruit anthracnose was developed,which aimed to realize the real-time diagnosis for pepper fruit anthracnose in an offline state in the field,and to lay the foundation for early preventing and controlling pepper fruit anthracnose.【Method】Pod pepper“Layan 101”was used as the research object.Firstly,the data of the images of healthy and anthracnose fruit on the plant and out of plant was expanded,and then the lightweight networks with similar size of MobileNet V1,MobileNet V2 and EfficientNet-B0 model parameters were used for transfer learning.Then,the pepper fruit characteristics were extracted through fine-tuning of parameters,and the classification effect of image was compared.Finally,the MobileNet V2 network with the highest identification accuracy was transplanted to the mobile terminal.【Result】MobileNet V2 network had the highest recognition accuracy for health and anthracnose pepper fruit in testing set,reaching 97.31%,and the recognition time of a single picture was 75 ms,which was the shortest time of three models.It was featured with the advantages of small memory occupation,fast recognition speed and high accuracy.【Conclusion】Through transplantation and application of MobileNet V2 network model,the identification system of pepper fruit anthracnose can realize real-time diagnosis of pepper fruit anthracnose in an offline state in the field.