Explorando abordagens de classificação contextual para floresta de caminhos ótimos

• Autor(a) principal: Osaku, Daniel
• Data de Publicação: 2016
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/8098
• topic: Classificação contextual de imagens; Sensoriamento remoto; Ressonância magnética; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO

Pattern recognition techniques have been widely studied and disseminated in order to develop ways to improve the e ectiveness of the pattern classi ers using labeled samples. However, such techniques usually work following the premise that the samples are independent and identically distributed in the feature space, taking into account only the local properties of the image and no information about the correlations between neighboring pixels are employed. The Optimum-Path Forest (OPF) classi er models the instances as the nodes of a graph, being the problem now is reduced to a partition of this graph. Although there are approaches that consider the context in the pattern recognition process, there is no such version for Optimum-Path Forest up to date. Thus, one of the main goal of the presented thesis is to propose a contextual version for the OPF classi er, which would employes contextual informations to support the data classi cation task using methods based on information theory and Markov Random Fields for such purpose. Since the Markov models are parameter-dependent and it is not a straightforward task to nd out the optimal values for such parameters because can assume in nite solutions, another contribution of this work is to propose an approach for modeling the process of nd out the parameters as a optimization problem, being the tness function to be maximized is the OPF accuracy over a labeled set. The results obtained by contextual classi cation were better than traditional classi cation results, as well as the optimization methods applied seemed to be a good alternative to ne-tune parameters of the Markov models as well.