Abordagens em aprendizagem estatística para computar componentes tensoriais em subespaços multilineares com aplicações em reconhecimento de expressões e gênero em imagens de face

• Autor(a) principal: Filisbino, Tiene Andre
• Data de Publicação: 2015
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/213
• topic: Reconhecimento de padrões; Pattern perception; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::PROCESSAMENTO GRAFICO (GRAPHICS)

The main contributions of this dissertations are divided in two parts. The main goal of the first one is to investigate the efficiency of ranking techniques for tensor components in gender and facial expression (happiness versus neutral) experiments for classification and reconstruction problems in multilinear subspace learning. We will verify the consequences of ranking techniques in subspace learning through Fisher criterion, by estimating the covariance structure of the database and by using weights generated through separating hyperplanes, such as support vector machine (SVM) and linear discriminant analysis (LDA). The latter is named tensor discriminant principal component analysis (TDPCA). This analysis will be performed in the context of multilinear principal components analysis (MPCA) and concurrent subspace analysis (CSA), which are known techniques for multilinear subspace learning. The former follows the principal components analysis methodology, that centers the samples before the subspace learning computation while the latter performs the learning procedure using raw data. In the second part, we will analyze the influence of weights computed using SVM, LDA and Fisher criterion as a prior information to generate new tensor components in the context of MPCA. This new supervised subspace learning technique, named weighted MPCA (WMPCA), is combined with the ranking methods to re-orient weighted tensor components generating the WTDPCA subspaces. We apply the WMPCA and WTDPCA frameworks, for face image analysis. The results show the efficiency of WMPCA and WTDPCA subspaces to distinguish sample groups in classification tasks as well as some drawbacks between global and local patterns for reconstruction. We also address theoretical issues related to the connection between MPCA and CSA, as well as foundations in multilinear subspace learning problems; that is, the corresponding covariance structure and discriminant analysis.