Crop disease long-tail image recognition method based on multi-stage training

The invention relates to a crop disease long-tail image recognition method based on multi-stage training and belongs to the field of deep learning and image recognition. The method comprises the following steps that a convolutional neural network model is built to recognize crop diseases and insect pests, and training is carried out by adopting a multi-stage training method to improve robustness of the model and the recognition capability of unbalanced data; in a first stage of training, original unbalanced data is adopted to carry out model training so that the model learns original data distribution; in a second stage of training, model training is carried out by adopting a CutMix enhanced data set so that robustness of the model is improved; in a third stage of training, a data set which is distributed in a balanced mode after balanced sampling is adopted for model training, parameter updating of a convolution module is frozen during training, only parameters of a full-connection layer are updated, and the