Semantic segmentation for plant disease using deep learning

• Main Author: Shu, Shi Hao
• Format: Thesis
• Language: English
• Published: 2020
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utm.my/93030/1/ShuShiHaoMSKE2020.pdf

In the era of artificial intelligence, various applications are using machine learning to solve some of the engineering problems and ease the people work. Plant disease is important to be detect as early as possible because the early we found out there is a disease, the lesser in loss of crops from plant disease which may impact to the economic and increase of price for crops. Numerous researches have been reported in the literature that employed different aspects of machine learning methods for plant disease detection, with the majority of them are focusing on the plant leaves. There are two common and destructive foliar diseases that are early blight and late blight. Early blight infection usually associated with plant physiological maturity and fruit load and late blight can infect and devastate the plants at any developmental stages. Since the leaves are found to be the most commonly observed part for detecting an infection. Segmentation technique is a basic and easy way to classify and estimate the severity of the diseases because it works well with plant disease detection since the infected leaf area shows significant color differences from its original color. Feature extraction is an essential step before the segmentation process that determines the applicability of every machine learning model. Convolution Neural Network (CNN) is a method that is gaining popularity to solve the feature extraction problem since it can automatically extract the features directly from the input images. Hence, this project aims to utilize the semantic segmentation with CNN through transfer learning from the VGG16 network to segment the plant leaf into healthy, necrotic and symptomatic regions. There are 1200 samples to be labeled as healthy, necrotic and symptomatic and classified into three severity levels to serve as the training material in the supervise training of CNN. To observe the performance of the training, normal training and optimized training will be compared in terms of accuracy and efficiency. Lastly, a deep learning model will be developed and is capable of recognizing and labeling the given leaves regions whether it is healthy, necrotic and symptomatic. The deep learning model able to achieve the global accuracy of 93% and IoU of 67.83% with 1200 data samples which segmented into 4 different classes.