Improving plant disease recognition with generative adversarial network under limited training set
Date
Authors
Major Professor
Advisor
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Authors
Research Projects
Organizational Units
Journal Issue
Is Version Of
Versions
Series
Department
Abstract
This thesis introduces a generative adversarial network (GAN) based method to classify diseased images using a limited training set. A general introduction of machine learning applications in the agriculture domain is provided. The issue of plant disease recognition has been investigated in this thesis.
First, the successful applications of convolutional neural networks (CNNs) to plant disease classification have been reviewed. It is found out that most of the methods are built under the assumption that there is enough training set. The issue of limited training data is overlooked. Thus, the over-fitting problem caused by a limited training set is discussed.
Second, a new approach is proposed to solve the limited training set problem. The proposed method consists of four parts: CNN, data augmentation, GAN and label smoothing regularization (LSR). CNN is used to classify plant diseases and species. Data augmentation and GAN are used to generate additional samples for training. LSR technique can help the model avoid the over-fitting problem.
Finally, three comparison experiments have been designed. The analysis proves the effectiveness of the proposed method. Compared with using the real dataset only, the proposed method improves the prediction accuracy by 6%.