Combinação de múltiplos classificadores para identificação de materiais em imagens ruidosas

• Ano de defesa: 2004
• Autor(a) principal: Ponti Junior, Moacir Pereira
• Orientador(a): Mascarenhas, Nelson Delfino d'Ávila
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• País: BR
• Palavras-chave em Português: Processamento de imagens; Reconhecimento de padrões; Combinação de classificadores; Tomografia computadorizada
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/ufscar/493

Material identification in images has been explored in multiple areas and very interesting applications are arising in this field. This work uses noisy multispectral images from a computerized tomograph scanner acquired with multiple energies for soil sciences applications and developes a recognition system to identify materials on the scanned body. Techniques of statistical classification were used. The individual classifiers: Parzen, k-nearest neighbors, logistic and linear Bayesian were combined in order to study the behavior of classifier combination techniques. For this task, we used the fixed rules combiners: majority voting, maximum, minimum, median, sum and product. Also, a second stage of combination was considered and used, the majority voting of combiners. The performance of the classifiers was analyzed through the leave-one-out cross-validation error estimation method and the Kappa coefficient. The advantages of the use of multiple energies in the problems of identification of images and the behavior of each combination method are also demonstrated. The results pointed out that the combination of classifiers gives better capacity of generalization and more stable results than the individual classifiers, using information supplied for all individual classifiers, including the weakest one, being recommended in classification of scarce, difficult discrimination data, on the presence of ambiguity or high`noise levels.