Plant disease recognition based on parallel Vision Transformer

Abstract

Ensuring food security and increasing crop yields are significant challenges facing the world today as the global population grows. In this context, the seasonable and accurate perception of plant diseases is essential. This paper proposes to combine Vision Transformer (ViT) and Gpipe to identify disease pictures of plant leaves. Among them, ViT is used to ensure the identification accuracy of the model, and Gpipe is used to improve the running speed of the model. This experiment uses the PlantVillage dataset, which includes diseased pictures of various plant leaves, to evaluate model performance. It uses the method of controlled experiments to Identify the models performance and efficiency of parallelism in the pipeline. After many experiments, the recognition accuracy of the color picture part of ViT on this data set has reached 93%. In addition, the memory requirements for a single GPU are significantly decreased using pipeline parallelism. This study provides a low-cost and efficient plant disease identification tool for the agricultural field, which can detect plant diseases correctly and in time. It is beneficial to increase food production and ensure food safety.