Multi-class discriminant analysis based on support vector machine ensembles

• Ano de defesa: 2019
• Autor(a) principal: Filisbino, Tiene Andre
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Laboratório Nacional de Computação Científica; Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA); Brasil; LNCC; Programa de Pós-Graduação em Modelagem Computacional
• Palavras-chave em Português: Probabilidades; Estatística; Análise multivariada; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::MATEMATICA DA COMPUTACAO
• Link de acesso: https://tede.lncc.br/handle/tede/311

Many areas such as pattern recognition and analysis of image databases require the managing of datasets originally represented in high dimensional spaces. Besides, the original data representation suffers, in general, of redundancy and noise. Thus, we must compute a more suitable feature space, reducing both the dimension and redundancy of representation in order to minimize the computational cost of further operations. Once a feature space has been defined there is the necessity of determining the most important discriminant features for pattern recognition tasks, like classification. Discriminant analysis techniques, which in the literature are known as discriminant functions, seek to solve this type of problem. Thus, the goal of the proposed thesis is to develop discriminant analysis methods for multi-class classification problems. The key idea is to combine N classifiers to form a global discriminant function, which allows to rank the components of the space according to the importance of each feature to the classification. To achieve this goal, we use separate hyperplanes computed by linear support vector machines (SVMs) or defined by a Kernel SVM (KSVM) decision boundary, and use the ensemble methodology known as AdaBoost.M2 to combine the weak linear classifiers. More specifically, our proposed techniques seek to generate multi-class versions of the Discriminant Principal Component Analysis (DPCA), which was originally developed for binary problems. In this work, principal components analysis (PCA), Convolutional neural networks (CNNs) and texture descriptors, are used to create feature spaces that serve as input to perform discriminant analysis. In terms of application for validation of the proposed techniques our focus are human face and texture images obtained from granite tiles. Our experimental results have shown that the features selected by our proposal allow higher recognition rates using less features when compared with related methods as well as robust reconstruction and interpretation of the data. Further works will be undertaken by exploring deep learning methods, color images, tensor subspaces as well as to improve performance.