Caracterização de texturas com o auxílio de redes complexas, projeções topológicas e padrões semânticos

• Ano de defesa: 2019
• Autor(a) principal: Ribeiro, Thiago Pirola
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Caracterização de Texturas; Characterization of Texture; Redes Complexas; Complex Networks; Padrões Semânticos; Semantics Patterns; Projeções Topológicas; Topological Projections; Computação; Processamento de imagens; Modelagem de processos; Topologia; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::PROCESSAMENTO GRAFICO (GRAPHICS)
• Link de acesso: https://repositorio.ufu.br/handle/123456789/25340; http://dx.doi.org/10.14393/ufu.te.2019.4

Image processing is a computing area that is present in many segments of society. This area is in constant development and, through the improvement of several techniques, it can help resolving many different problems which are based on images. There are several methods in the literature that deal with the modeling and the characterization of textures. This work is focused on natural and artificial textures. The proposed approach uses complex networks to represent textures. Energy histograms generated from maps of degrees of complex networks were used, selecting the their peaks and valleys for defining thresholds to be adopted to generate the feature vector based on the target images to be classified. We also used alternative forms for image representation: first, a method that analyzes the data modeling using semantic patterns for pollen grain image analysis was proposed, and, second, the analysis of influence of characteristics of the feature vector composition using the Klein bottle topology for designing image patches in topological spaces. The computation of the K-Fourier Estimated Coefficients was explored, the composition of the feature vector as well as new combinations of sizes, which would result in a reduction of the amount of descriptors and improvement in the characterization of texture datasets. Several experiments were carried out on some public domain datasets already classified: Brodatz, CUReT, Outex_TC_00013, KTH-TIPS, UIUCTex, VisTex, and a pollen grains image dataset. This work demonstrates that the proposed methods were able to reach out the expectations, being able to stand out for the reduced amount of used descriptors and the accuracy obtained in the experiments.