Tomato Disease Detection Using Vision Transformer Model

The agricultural sector is also attracting attention as an important industry, and crop management and facility management using AI are in progress, but crop disease management is still showing limitations by relying on experience and intuition rather than data-based analysis using AI. In this paper, we conduct a study on tomato disease detection using the ViT model. We would like to construct an appropriate tomato image dataset, conduct AI-based image data learning, and confirm the learning results using the model. After that, the research is conducted by comparing the performance of the CNN model and the ViT model such as accuracy and precision. The accuracy of the learning results through the ViT model is 88.9%, the precision is 86.7%, the reproduction rate is 84.2% F1-score 85.4%, and the AUC-ROC 0.905 log loss is 0.276, showing excellent performance compared to the CNN model. Through this study, it was confirmed that the ViT model is also a competitive model in the field of image data analysis. In addition, it is expected that it can minimize crop damage by early detection of diseases in the smart farm environment and help reduce labor and improve productivity of farmers.