关键词: 小麦；叶部病害, 卷积神经网络；迁移学习；图像识别；MobileNetV2；计算机视觉

Abstract: In order to realize the convenient recognition of wheat leaf disease image on mobile terminal,image recognition model of wheat leaf disease was established based on lightweight convolutional neural network (CNN) and transfer learning.Firstly,a sample data set of 224 pixel×224 pixel was established,which was composed of images of wheat powdery mildew,stripe rust and leaf rust. Secondly,the deep learning framework Tensorflow 2.0 was used to build the wheat leaf disease image recognition model based on MobileNetV2, and the trained parameters on ImageNet data set were used as the initial parameters of the model. Finally,the effects of transfer learning method,sample size,adding Dropout layer before global average pooling and initial learning rate setting on the model performance were analyzed.The results showed that the average recognition accuracy of the images of three wheat diseases was as high as 99.96% by setting all layers of the model as trainable,selecting suitable data enhancement method to increase the sample size, adding Dropout layer before global average pooling and setting the initial learning rate of 0.000 01.Based on MobileNetV2 and transfer learning, a wheat leaf disease image recognition model with high recognition accuracy and strong generalization ability could be constructed for mobile terminal.

Key words: Wheat, Leaf diseases;Convolutional neural network;Transfer learning;Image recognition;MobileNetV2;Computer vision