Development of image analysis techniques using machine learning.

Abstract

Image analysis and computer vision play a crucial role in many real-world applications such as smart agriculture and smart vehicles. For plant diseases diagnosis, one of the most recent challenges is the improvement of the plant diseases classification on Low-Resolution (LR) images. The farmer is supposed to obtain High-Resolution (HR) images of plant leaves from the field. In this thesis, we propose two contributions. We first propose to fine-tune a low-complex classifier named Dense Convolutional Network (DenseNet) on HR plant leaves images to detect tomato leaves diseases. Because of the small size of plant leaves and other limitations, the obtained HR images can miss some detailed information that results in blurred LR images of leaves with fewer details. As a second contribution, we introduce a novel Super-Resolution (SR) algorithm named Wideractivation for Attention-mechanism based on a Generative Adversarial Network (WAGAN) to improve the diseases classification of the tomato LR images. To evaluate the potential of the proposed SR method in plant diseases recognition, we first recovered SR plant diseases images from LR images using the WAGAN. Next, we compared the performance of the diseases classification using LR, SR, and HR images. The classification results proved the efficacy of the proposed SR method (97.63%) with ×3.6 lower complexity than the state-of-the-art method and very close to the reference HR (97.81%) accuracy. Due to the efficient design of proposed SR model, the WAGAN focuses more on edges, textures, and other valuable information, which are the key information, needed for the DenseNet to recognize the diseases.