Identification of monocotyledons and dicotyledons leaves diseases with limited multi-category data by few-shot learning

In order to achieve good identification performance, the existing identification method of crop diseases based on deep learning needs considerable annotated images to train. However, collection of crop leaves disease images in field is time-consuming and laborious, so it is urgent to propose a timely and effective identification model for limited labeled crop disease images. This paper proposed a Few-shot Learning method based on Siamese Network to identify crop leaves diseases, which used randomly generated image sample pairs as input. In experiments, since crops are divided into monocotyledons and dicotyledons, this paper used Few-shot Learning method to train monocotyledon plant diseases model named SiamNet2 and dicotyledon plant diseases model named SiamNet1, which were used to identify monocotyledon and dicotyledon plant diseases. The results of 5-way 5-shot dicotyledon plant disease identification and 10-way 10-shot monocotyledon plant disease identification showed that the identification accuracy of SiamNet2 was 8.7% and 12.4% higher than that of SiamNet1, respectively.