Transfer Learning for Fine-Grained Crop Disease Classification Based on Leaf Images

Abstract

In this paper, convolutional neural network models and transfer learning techniques were employed to perform automated early plant disease detection and diagnosis using simple leaf images of healthy and diseased plants taken in situ and in controlled environment. This paper presents a faster technique to perform fine-tuning of state-of-art models by using a concept of freezing blocks of layers instead of individual layers. Models were trained and tested on public dataset of 23,352 images from 28 classes incorporating 15 different crop species. Six different model architectures and their variants were trained, with the best performance attaining an accuracy of 99.74% obtained by fine-tuning deep learning model previously trained on ImageNet. The significantly high success rate and computationally efficient property on small amount of data makes this model a very useful early warning tool capable of being deployed as integrated plant disease identification system in portable and resource constrained devices to operate in real cultivation conditions. These techniques were tested on other datasets and was proven to be highly efficient for fine-grained dataset.