Vision Transformer for Plant Disease Detection: PlantViT

Abstract

With the COVID-19 pandemic outbreak, most countries have limited their grain exports, which has resulted in acute food shortages and price escalation in many countries. An increase in agriculture production is important to control price escalation and reduce the number of people suffering from acute hunger. But crop loss due to pests and plant diseases has also been rising worldwide, inspite of various smart agriculture solutions to control the damage. Out of several approaches, computer vision-based food security systems have shown promising performance, and some pilot projects have also been successfully implemented to issue advisories to farmers based on image-based farm condition monitoring. Several image processing, machine learning, and deep learning techniques have been proposed by researchers for automatic disease detection and identification. Although recent deep learning solutions are quite promising, most of them are either inspired by ILSVRC architectures with high memory and computational requirements, or light convolutional neural network (CNN) based models that have a limited degree of generalization. Thus, building a lightweight and compact CNN based model is a challenging task. In this paper, a transformer-based automatic disease detection model “PlantViT" has been proposed, which is a hybrid model of a CNN and a Vision Transformer. The aim is to identify plant diseases from images of leaves by developing a Vision Transformer-based deep learning technique. The model takes the capabilities of CNNs and the Vision Transformer. The Vision Transformer is based on a multi-head attention module. The experiment has been evaluated on two large-scale open-source plant disease detection datasets: PlantVillage and Embrapa. Experimental results show that the proposed model can achieve 98.61% and 87.87% accuracy on the PlantVillage and Embrapa datasets, respectively. The PlantViT can obtain significant improvement over the current state-of-the-art methods in plant disease detection.