Abstract: Accurate identification of crop diseases can help improve the yield and quality of crops. As image training sample data of crop disease is limited, this paper adopts the transfer learning method combine with the step-by-step identification model to identify a host of crop disease types. Taking PlantVillage public data set of ten types of crop leaf images of three crops as training samples, the paper uses the direct recognition method to train the original model of VGG16 and ResNet and the model of transfer learning respectively, and obtain the classification results of the model. This paper proposes a step-by-step identification method, the training samples are classified according to the types of crops and disease types, the models are trained separately, and a step-by-step identification model is constructed. The experimental results show that the transfer learning method can increase the recognition accuracy by more than 20% on the basis of the original model. On the basis of this, the step-by-step identification method is introduced and compared with the direct recognition method, the accuracy of the VGG16 and ResNet models is increased by 14% and 8%, respectively. The transfer learning step-by-step identification method proposed in this research can realize the accurate identification of crop diseases in the case of small samples of training data, and can provide effective technical support for crop disease prevention and control.

Key words: transfer learning, few-shot learning, convolutional neural network, step-by-step identification, crop diseases