Image-Based Plant Leaf Disease Recognition with InceptionV3 Network
Loading...
Date
2021-04
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Most traditional plant disease diagnosis strategies depend on human visual observation and inspection. However, this approach is time-consuming and requires significant human effort and expert knowledge. The recent advances in computer vision and deep learning provide a potential pathway to developing a plant disease diagnosis system that allows rapid detection of disease across large spatial areas with minimal human intervention. In this study, we developed a deep learning approach for plant leaf disease classification problems and conducted a range of experiments to quantify the performance of several state-of-the-art neural network architectures, including ResNet50, InceptionV3, and NASNet. All of the experiments were trained on the PlantVillage dataset with 54305 images in total, spanning over 38 plant disease classes. We evaluated four different performance metrics to assess each architecture: accuracy, precision, recall, and area under the curve (AUC). Our results showed that the InceptionV3 neural network architecture outperformed all other Convolutional Neural Network (CNN) architectures (ResNet50, NASNet-Large, NASNet-Mobile, MobileNet-v3-small, and MobileNet-v3-large) and produced a training accuracy of 94.14% and 97.94% over 6 epochs and 40 epochs of training, respectively. These results suggest that CNN architectures broadly, and the InceptionV3 model specifically, is capable of remote and automated plant disease detection. These results point to exciting future applications in lightweight mobile phone applications or backend workstation developments for plant leaf disease recognition problems.
Description
Engineering: Physical Sciences (The Ohio State University Denman Undergraduate Research Forum)
Keywords
Deep Learning, Transfer Learning, Plant Disease, Leaf Image