Tomato Leaf Disease Classification with Vision Transformer Variants

Waheed Moonwar, Mohammad Rakibul Hasan Mahin, Yasmin Nadia, Fahmid Bin Kibria, Aurnab Biswas, Fatiha Ishrar Chowdhury, Sifat E. Jahan, Annajiat Alim Rasel

Abstract: Agriculture has played a significant role for many years. Its growing significance is attributable to the money it has generated. The full advantages of crop cultivation are, however, prevented by several circumstances. Organic plant diseases have a role in this case. For an agriculture dependent country like Bangladesh, extreme weather and heavy pesticide use are accountable for its economic crisis. This work aims to offer farmers visual information to facilitate the implementation of preventive measures beforehand. This work proposes four different transformer models for tomato leaf disease classification, which includes Vision Transformer (ViT), Swin Transformer (SwT) and Compact Convolutional Transformer (CCT). In addition to that, a distinct variation of the ViT algorithm was incorporated into the categorization process. It centers on employing a comparative analysis of various transformer models, which represents a novel contribution to the existing literature. The models have been trained and validated on the PlantVillage dataset which resulted in test accuracy of 95.22%, 82.61%, 82.82%, and 92.78% for ViT, SwT, CCT, ViT with Shifted Patch Tokenization and Locality Self Attention respectively.