Deep Convolutional Neural Network to Recognize Plant Leaf Images

Abstract

This thesis focuses on three main types of research. - (1) data augmentation ensemble learning (2) learning rate schedules improved and (3) plant leaf disease recognition performance Chapter 1 provides a brief general introduction to deep learning for plant leaf disease recognition and uses deep learning techniques for detecting and diagnosing diseases in plants, followed by the research questions. The objectives of the dissertation and its contributions are described. In Chapter 2, Two deep convolutional neural networks (CNNs): MobileNetV2 and NasNetMobile are proposed to recognize plant leaf disease. I have experimented with training techniques; online, offline, and mixed training techniques on two plant leaf diseases which were the leaf disease dataset, and the iCassava2019 dataset. I have also used data augmentation techniques combining rotation, scrolling, and zooming techniques to further enhance the recognition performance. Chapter 3 presents the stacking ensemble of lightweight convolutional neural networks to improve the performance of the recognition of plant leaf disease images. I proposed a stacking ensemble of deep CNNs to evaluate three plant leaf disease datasets; PlantDoc, Crop-PlantDoc, and iCassava2019. We experimented with five classifiers that were logistic regression, support vector machine, K-nearest neighbors, random forest, and long short-term memory network. The random forest method achieved a more accurate performance. Chapter 4 proposes fusion and ensemble CNN to improve the performance of plant leaf disease recognition. The work reported in this thesis experimented with a new learning rate schedule, called equal learning rate range (ELRR) and step decay equal learning rate range (SD-ELRR), which is proposed and compared with two baseline learning rate [AP1] schedules. The proposed learning rate schedule was evaluated on two datasets: Cropped-PlantDoc and Plant Pathology. The results showed that the ELRR and SD-ELRR equations improved the efficiency of plant leaf disease recognition significantly better than the basic equations in the entire E plant disease dataset. Chapter 5 comprises two main sections: - 1) Answers to the research questions 2) Future work. This chapter briefly explains the proposed approaches and answers three main research questions in plant leaf disease recognition using deep learning techniques. Two main approaches are planned to be the focus of future work, as follows. For the data augmentation techniques, I plan to study and apply other data augmentation techniques such as AutoAugment and neural style transfer. For the ensemble learning techniques, I will focus on experiments with the other CNN frameworks, such as snapshot ensemble CNN and 1D-CNN.

-