Lightweight Grape Leaf Disease Recognition Method Based on Transformer Framework

Abstract

Abstract Grape disease image recognition is a crucial part of agricultural disease detection, and accurate identification of grape leaves plays a vital role in agricultural production. This study proposes a deep learning-based method for grape disease classification and recognition to address issues such as the complexity of grape disease features, uneven distribution of disease features, and data imbalance. First, the adversarial generative network FastGAN is used to generate grape disease images to enrich the sample information for different categories in the dataset and address the data imbalance problem. Then, a novel Transformer structure called LVT Block and a CNN structure called MARI Block are proposed to process the global and local information of images, respectively. Dense connections between these structures result in the DLVT Block, which leads to the lightweight neural network model DLVTNet. Additionally, a lightweight self-attention mechanism combined with CNN called CLSHSA is introduced , which maintains high recognition performance while reducing the model size. Moreover, a multi-scale attention mechanism (MELA) is proposed, which combines positional and multi-scale information to obtain attention weights. Experimental results show that this method achieves an average recognition accuracy of 98.48% in grape leaf disease detection, outperforming mainstream CNN and Transformer models, and effectively focuses on disease areas in leaf images. The method also demonstrates high recognition accuracy in tomato disease detection, indicating good generalization ability and suitability for detecting and recognizing various leaf diseases. The proposed method provides an effective solution for detecting and recognizing grape and other plant leaf diseases, offering a new research approach that combines CNN and Transformer structures.

Code and data availability

The paper's grape (and tomato) leaf disease image inputs come directly from the public New Plant Diseases Dataset on Kaggle, which is the image dataset used for the paper's phenotyping measurements. No author code, trained models, or generated FastGAN dataset is publicly deposited; the data availability statement only'