Abordagens de aprendizado ativo e profundo para síntese e classificação de imagens de enteroparasitos

• Ano de defesa: 2019
• Autor(a) principal: Alves, Daniel Henrique Acorsi
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Cornelio Procopio; Brasil; Programa de Pós-Graduação em Informática; UTFPR
• Palavras-chave em Português: Aprendizagem ativa; Processamento de imagens; Helminto; Active learning; Image processing; Helminths; CNPQ::CIENCIAS EXATAS E DA TERRA; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/5168

Significant advances in the computer vision area have been boosted mainly by the use of deep learning techniques through convolutional neural networks. However, one of the main difficulties in using deep learning techniques is the availability of labeled data, which is indispensable when dealing with supervised learning for the task of classification. Other problems to be considered refer to: i-) scarcity of images, for example, images, given the difficulty of acquiring them in some applications; ii-) difficulty in obtaining images labeled by specialists. Thus, it becomes fundamental to develop mechanisms to obtain a larger set of images, as well as that they are labeled. Some applications may even present a reasonable amount or even large amounts of available images, as well as some efforts have been made in attempting to solve the mentioned problems. However, many of the image set samples may not be representative for learning, according to the application domain, and many of the samples considered in the learning can be redundant and unnecessary, negatively impacting the classifier performance. In addition, when dealing with a large set of data and applications that require interactive response times, the classification process may become inefficient and unfeasible to perform. Therefore, the present study presents the proposal for a new learning approach in order to obtain images synthesizers and more robust classifiers. In order to do so, it integrates active learning strategies in the synthesis processes (through the use of Generative Adversarial Networks) and images classification, generating and selecting informative synthetic images for classifier learning. In addition, it investigates and proposes new active learning strategies in order to select more informative samples (based on diversity and uncertainty criteria). To validate the proposed learning approach, experiments were performed using enteroparasites data. The results show that the inclusion of active learning strategies in different stages of the proposed approach is promising. Active learning strategies make it possible to select more informative samples for both image synthesis and classification processes.