Exploring capsule networks for disease classification in plants
Shradha Verma, Anuradha Chug, Amit Prakash Singh
Abstract
The overarching goal of smart farming is to develop innovative solutions for future sustainability of humankind. Crop protection from biological/non-biological factors is a major hindrance for food security, plant diseases being one of the foremost challenges. Not only, plant diseases destroy the crops or diminish their overall quality, use of pesticides for their treatment renders the soil contaminated, which after a time becomes unsuitable for sowing and planting. Potato is a key crop and a major source of livelihood for a vast population. In this study, the authors have implemented capsule networks for classification of potato diseases, and compared the performance with few popular pretrained CNN models, namely ResNet18, VGG16 and GoogLeNet, implemented via transfer learning. Colored images of healthy as well as diseased leaf images were employed from the PlantVillage dataset and used for training the models. CapsNet proved to have achieved comparable performance to state-of-the-art CNN models, with 91.83% accuracy.