Deep Learning for Plant Disease Classification

Abstract

Deep learning using convolutional neural networks (CNNs) has become increasingly successful in plant disease diagnosis, and consequently offers new promise in the ongoing battle against crop disease. While many studies continue to demonstrate its continued success in plant disease diagnosis, the complexity of the CNN architecture remains largely unexplored. The aim of this thesis is to shed more light onto the architectures of the CNNs for plant disease classification specifically, ensuring its reliability and authenticity by human intervention. These results can then be utilized to further improve the prediction accuracy of deep models by optimizing the underlying architecture. Four models are trained and tested on the PlantVillage dataset: InceptionV3, AlexNet, GoogLeNet, and ResNet50. To further improve the CNN model, attention modules are applied to the baseline CNN model. Last but not least, several visualization techniques (feature visualization) are explored. These visualizations will help to uncover the CNN “black box,” and shed light on features learned by the CNNs.