Abstract:

To achieve automatic recognition of apple leaf disease images, this paper studied the recognition model of apple leaf disease. According to the data set of plant disease collected through the network open-source, three kinds of leaf images were obtained with apple scab, apple cedar rust and apple gray disease respectively, and one healthy leaf image was used as the study object. 4433 images were selected randomly to build the data set for model training, and all of these samples were expanded and averaged. The data were preprocessed with offline and online augmentation. On the basis of five pre-training models (Resnet50, Mobilenet v2, Vgg16, Vgg19, Inception v3), the parameters were adjusted and optimized for the migration model. Comparing the training results of the five models, the optimized Resnet50 model could achieve 0.9770 of accuracy. Featured by rapid recognition and high accuracy, the optimized training model could identify the disease types accurately and rapidly, and provide support for the automatic diagnosis of plant diseases.

Key words: recognition of apple leaf disease, deep learning, convolutional neural networks, transfer learning, data enhancement