Análise de estruturas de rede neocognitron para aplicação no reconhecimento facial

Detalhes bibliográficos
Ano de defesa: 2008
Autor(a) principal: Santana, Cristiane Oliveira de
Orientador(a): Saito, José Hiroki lattes
Banca de defesa: Não Informado pela instituição
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
Departamento: Não Informado pela instituição
País: BR
Palavras-chave em Português:
Área do conhecimento CNPq:
Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/380
Resumo: The present work consists in the use of the combination of neocognitron networks for the face recognition tasks integrating a system of face recognition, categorized as holistic method, by the fact of to approach all the face in the extration of the characteristics of the input image. The use of the face as biometric attribute in the recognition of individuous has grown mainly to its characteristic that dispense their cooperation for analysis. The great challenges in the area of face recognition are about the application in non-controlled environments in which the variation of illumination and pose can decrease the performance of the system. In order to work the challenges (variation of illumination and distortion in the patterns) it was used frontals images of CMU-PIE (University Carnegie Mellon - Pose, Illumination and Expression) database using the advantage of illumination and expression variations. The neural network model applied in the work is the neocognitron network that is capable to recognize patterns without its capacity of recognition been affected by deformations, changes in the size or position, of the input pattern. With the goal of to get the structure more adjusted of the neocognitron for the face recognition task, inside of a set of structures, it was analyzed some structures of neocognitron network with one output class and with different images resolutions. The best result corresponds to a recognition rate of 78% for a set with 30 classes with thirty and six patterns each one in the recognition phase. To improve the results it was applied combination of classifiers using: Decision Templates and the Modified Decision Templates method developed in this work. The performance of the classifiers was analyzed through the error estimation using the hold-out method and the Kappa coefficient. The final results pointed that the combination of classifiers applied to this model did not result in significant improvements due the inherent characteristics of the applied model.