基于迁移学习和残差网络的葡萄叶部病害识别  

Identification of Grape Leaf Diseases Based on Transfer Learning and Residual Networks

作　　者：谢圣桥 宋健[2] 汤修映[1] 白阳 Xie Shengqiao;Song Jian;Tang Xiuying;Bai Yang(College of Engineering,China Agricultural University,Beijing 100083,China;Weifang University,Weifang 261061,China)

机构地区：[1]中国农业大学工学院,北京100083 [2]潍坊学院智能农业装备实验室,山东潍坊261061

出　　处：《农机化研究》2023年第8期18-23,28,共7页Journal of Agricultural Mechanization Research

基　　金：山东省重点研发计划项目(2019GNC106144)。

摘　　要：为实现小样本数据下的葡萄病害图像识别,基于迁移学习和数据增强技术,实现了葡萄叶部病害图像的精准分类。在ResNet50模型的基础上,保留了卷积层并设计了全新的全连接层。对葡萄病害数据集原图进行亮度变换、旋转、添加高斯噪声等操作,以扩充数据集,防止模型过拟合,且将数据集按1:4划分为验证集与训练集,训练模型的全连接层。试验结果表明:ResNet50模型对葡萄黑腐病叶、褐斑病叶、轮斑病叶及健康叶的平均识别准确率达97.87,比VGG16和VGG19模型分别高出4.02%和1.86%,分类效果优于其他模型;葡萄叶片黑腐病、褐斑病、轮斑病和健康的召回率分别为97%、96%、99%及99%,均在96%以上。由此表明,模型具有较好的鲁棒性和识别性能,可为果园自动化喷药提供技术参考。To achieve grape disease image recognition with small sample data,this paper achieves accurate classification of grape leaf disease images based on transfer learning and data enhancement techniques.Based on the ResNet50 model,the convolutional layer is retained and a new fully connected layer is designed.Brightness transformation,rotation,and addition of Gaussian noise are performed on the original grape disease dataset to expand the dataset and prevent over-fitting of the model.The data set is divided 1:4 into a validation set and a training set,and only the fully connected layer of the model is trained.The experimental results showed that the average recognition accuracy of ResNet50 model for black rot leaves,brown spot leaves,verticillium leaves and healthy leaves of grapes was 97.87,which was 4.02%and 1.86%higher than VGG16 and VGG19 models,respectively,and the classification effect was better than other models.The recall rates for black rot,brown spot,verticillium disease and healthy leaves were 97%,96%,99%and 99%,respectively,all of which were above 96%.It is shown that the model has good robustness and recognition performance,and can provide technical reference for automated spraying in orchards.

关 键 词：葡萄 病害识别 迁移学习 残差网络 深度学习 

分 类 号：S126[农业科学—农业基础科学] S43