Aplicação de deep learning para análise de imagens de microscopia eletrônica, varredura e campo

• Main Author: Migliari, Sara Lopes
• Publication Date: 2024
• Format: Master thesis
• Language: por
• Source: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Download full: http://repositorio.utfpr.edu.br/jspui/handle/1/36842
• topic: Aprendizagem; Processamento de imagens; Agricultura; Learning; Image processing; Agriculture; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Genética

It is undeniable that technologies, such as those derived from programming languages, applied to machine learning and image processing, have been crucial for scientific research. Alongside these, open and higher­quality databases that feed algorithms and optimize the application of deep learning codes. It is in this context of innovative methodologies and high­level information that the present work was carried out. We sought to identify patterns in images using different segmentation methods such as binary thresholding, inverted binary, Gaussian, adaptive mean, and overlay. Another method was the use of simple and scanning electron microscopy (SEM) to visualize plant morphological structures such as trichomes and endospores over periods of 24 and 72 hours. It is also important to mention that the field images generated by the segmentation method were used to assess possible events such as diseases caused by bacteria, and fungi. As for the results obtained from our study, the segmentation method is more reliable when the output information is extracted from convolutional networks applied to transfer learning in the Xception and InceptionV3 networks. Likewise, the segmentation method proves to be significant when it is used to increase the quality of images. Therefore, based on the analyzes conducted and the observation of the images resulting from the proposed algorithm, we understand that a deep learning application for the analysis of images by segmentation can help professionals both in detecting anomalies and in identifying patterns in images based on the quantity of pixels. We add that deep learning automates analyzes of microscopic or stereoscopic view counts with greater precision, selecting regions of interest to be examined as well as can indicate whether the analytical tools in operation provide promising outputs elaborated on reliable statistical and computational analyses.