Transfer Learning for Plant Disease Detection on Complex Images

Abstract

In the last few years, the use of deep Convolutional Neural Networks (CNN) for the detection and classification of plant diseases from leaf images became an active research area that shows very good results. However using deep learning classifiers requires large labelled datasets. The large and freely available dataset of plant diseases widely used by researchers is the PlantVillage dataset. The main issue related to this dataset is that it contains only laboratory images, which reduces the classifier performance when tested on complex field images. In this paper we study the use of both laboratory and field images for training deep CNN on classifying healthy and unhealthy plants. We combine laboratory and complex field images taken respectively from the PlantVillage and the EdenLibrary datasets, to build a dataset containing 54,000 images equally distributed on two classes: ’healthy’ and ’unhealthy’. This dataset is used to fine-tune three state-of-the-art image classifiers that are pre-trained on the ImageNet dataset: AlexNet, ResNet34 and DenseNet121. The experiment results show that using combined dataset significantly improves the classification accuracy for complex images.