Deep learning based segmentation and interpretable features for leaf image analysis

Details

Supervisors

Lee, John Aldo

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée

Abstract

In recent decades, agriculture has experienced an unprecedented increasing prevalence of plant diseases, which pose a significant threat to global food security and sustainable agriculture practices. To overcome these problems, new efficient and fast methods should be developed to better monitor crop health. In this master thesis, we propose to monitor crop health using leaves' images. We develop a deep learning based segmentation combined with interpretable features extraction method to analyze plant's leaves' images. The goal is to develop a framework for helping researchers study the potential of image analysis for this problematic. We succeeded to build a segmentor based on a Unet architecture reaching a test accuracy of 0.97. We show the high potential of colors, shape and LBP features for classification purpose, especially when combined with PCA and LDA as features extraction techniques. Finally, we reached the best test scores using SVM to classify leaves into healthy/not healthy, species and disease classes (resp. 0.94, 0.92 and 0.86).