Arquitetura de redes neurais para o reconhecimento facial baseado no neocognitron.

• Ano de defesa: 2001
• Autor(a) principal: Bianchini, Ângelo Rodrigo
• Orientador(a): Saito, José Hiroki
• 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; Redes neurais (computação); Reconhecimento facial; Rede neural neocognitron
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/ufscar/391

In this dissertation it is presented a face recognition system based in the neocognitron (NEOPC). This system is divided into three major stages: (a) pré-processing: which normalizes the facial samples utilized in the neural network architectures; (b) Neural Network for Detection of Control Points (NNDCP): which obtains the control points to be used in the non supervised training structure of the neocognitron and (c) Neural Network for the Facial Recognition (NNFR), which carried out the recognition of face samples.The main characteristic of NEOPC is the use of control points for the extraction of patterns strategically located, such as eyes, noses and lips, used on non supervised training of the NNFR. The tests of the NEOPC were carried out, considering facial samples having variations of angles and expressions, and different amounts of classes and samples to the training and recognition stages. For the attainment of the results, a base of images released by the University of Cambridge and three bases of images developed as part of this work. In order to compare the several results obtained with the four bases of images investigated in this dissertation we present in all of them relevant information, such as, the thresholds used for the excitation of neurons, rightness and error rates. The results shows that the NEOPC performance increases with the number of samples used for the training, until an optimal point, and then decreases. It is explained by the increase of the cell numbers with the features from new samples, difficulty the recognition.