Plant Disease Detection on Edge Devices

Protecting plants in general and against disease is vital to meeting the demand for food globally in both quality and quantity. Although the disease problem is vastly known, their rapid identification remains a problem, especially in parts of the world where necessary infrastructure is lacking. One solution can be using edge devices for disease diagnosis. As a proof of concept, Raspberry Pi 4 is used to showcase the capability of smaller models on low-powered and computation devices. One such device is smartphones. Already 83.72% of the world uses smartphones. Leveraging smartphones for this task would be beneficial; for that, a framework is proposed in this project. Using a public dataset, the PlantVillage dataset, added images from Google Images, inaturalist website, and clicked photos in real-time, these images are added to represent real-world scenarios. Two deep learning models are used SqueezeNet and EfficientDet-Lite0. SqueezeNet is used for image classification, i.e., plant and disease identification, and EfficientDet-Lite0 is used for object detection, i.e., leaves. The trained model can achieve 96.88% accuracy with inference time under one second on Raspberry Pi 4. This should showcase the potential of smartphones and other low-powered edge devices for this problem.