Multi-granularity Feature Extraction Based on Vision Transformer for Tomato Leaf Disease Recognition
Shupei Wu, Youqiang Sun, He Huang
Abstract
At present, the task of identifying crop diseases is mainly to simply distinguish the types of different crop diseases. However, the current classifiers cannot solve problems, such as accurate identification of similar disease categories. Compared with convolutional neural network (CNN), the recent vision transformer (VIT) has achieved good results on image tasks. Inspired by this, this paper proposed a multi-granularity feature extraction model based on vision transformer. By combining image block information of different scales, the model can learn image information from different granularities. At the same time, in order to further grasp the important areas, this paper developed a feature selection module. Through experimental comparison, the scheme has an accuracy improvement of nearly 2% compared with other classification models, and the model parameters have not improved much.