Abstract: Aiming at the problem of small number of  samples of apple leaf disease and the lack of sufficient labels, which led to the difficulty of the early detection and classification of apple leaf diseases, we proposed a few-shot learning method for image classification of apple leaf diseases. First, according to the distance of feature vectors  between the images and the density of sample points, we found out the outlier factors in the samples, after the outlier factors were eliminated, the average value of the support set in the embedding space was calculated, and then the classification was performed according to the relationship between the search sample and the value. The experimental results show that the proposed few-shot learning method can significantly improve the classification performance and convergence speed of the model, and has high classification accuracy, its  average classification accuracy is 97.62%, and the classification accuracy of apple rust, scab, mixed diseases and healthy leaves is 98.01%, 97.32%, 96.30%, 98.85%, respectively. It has  strong robustness to data sets such as sample imbalance and uneven background, etc.

Key words: apple leaf disease, image classification, few-shot learning, outlier removal