SEViT: a large-scale and fine-grained plant disease classification model based on transformer and attention convolution

Qingtian Zeng, Liangwei Niu, Shansong Wang, Weijian Ni

Abstract: Plant diseases are the leading cause of crop yield reduction. Rapid diagnosis using deep learning-based methods can effectively control the deterioration and spread of diseases. Convolutional Neural Network (CNN)-based methods are the current mainstream disease classification solution. However, most methods based on CNN are aimed at different diseases of a single crop, and they are difficult to distinguish similar diseases, which does not perform well in large-scale and fine-grained disease diagnosis tasks. In this paper, an image classification model for large-scale and fine-grained diseases named Squeeze-and-Excitation Vision Transformer (SEViT) is proposed to solve the above problems. SEViT uses ResNet embedded with channel attention module as the preprocessing network, ViT as the feature classification network. It aims to improve the model’s classification accuracy in the case of many types of diseases and high similarity of disease features. Experimental results show that the classification accuracy of SEViT in the test set achieves 88.34%, higher than comparison models. Compared with the baseline model, the classification accuracy of SEViT is improved by 5.15%.