This dissertation presents new modular and integrative information methods and systems inspired by the way the brain performs information processing, in particular, pattern recognition. The proposed artificial systems use spiking neurons as basic elements, which are the key components of spiking neural networks. Of particular interest to this research are various spiking neural network architectures and learning procedures that permit different pattern recognition problems to be solved in an evolvable and adaptive way. Spiking neural networks are used to model human visual and auditory pathways and are trained to perform the specific task of person authentication. The systems are individually tuned and trained to recognize facial information and to analyze sound signals from spoken sentences. The modelling of the integration of different sources of information (multisensory integration) using spiking neural networks is also a subject of investigation. A network architecture is proposed and a model for audiovisual pattern recognition is designed as an example. The main original contributions of this thesis are: a) Evaluation and further extension of adaptive learning procedures to perform visual pattern recognition. A new learning procedure that enables the system to change its structure, creating/merging neuronal maps of spiking neurons is presented and evaluated on a face recognition problem. b) Design of two new spiking neural network architectures to perform person authentication through the processing of speech signals. c) Design and evaluation of a new architecture that integrates sensory modalities based on spiking neurons. The integrative architecture combines opinions from individual modalities within a supramodal layer, which contains neurons sensitive to multiple sensory information. An additional feature that increases biological relevance is the crossmodal coupling of modalities, which effectively enables a given sensory modality to exert direct influence upon the processing areas typically related to other modalities. The contributions were published in one journal paper and in four refereed international conference proceedings. The proposed system designs were implemented and, through computer simulations, demonstrated comparable performance with traditional benchmarking methods. The systems have some promising features: they can be naturally optimized in respect to different criteria: accuracy (when very accurate results are expected), energy efficiency (when management of resources play an important role), and speed (when a decision needs to be made within a limited time). In this thesis, most of the parameters have been exhaustively optimized by hand or by using simple heuristics. As a direction for future work, there is an opportunity to include automated, specially tailored parameters optimization procedures or even general-purpose optimization algorithms, e.g., Genetic Algorithms and Particle Swarm Optimization. Overall, the results obtained in this thesis clearly indicate that it is indeed possible to have fast and accurate adaptive pattern recognition systems scalable for multiple modalities computing with simple models of spiking neurons. However, it is important to advance the theory of spiking neurons to take advantage of its biological relevance to reach similar or better performance when compared to the human brain, for instance, exploring new neuron models, information coding schemes and network connectivity.

This dissertation presents new modular and integrative information methods and systems inspired by the way the brain performs information processing, in particular, pattern recognition. The proposed artificial systems use spiking neurons as basic elements, which are the key components of spiking neural networks. Of particular interest to this research are various spiking neural network architectures and learning procedures that permit different pattern recognition problems to be solved in an evolvable and adaptive way. Spiking neural networks are used to model human visual and auditory pathways and are trained to perform the specific task of person authentication. The systems are individually tuned and trained to recognize facial information and to analyze sound signals from spoken sentences. The modelling of the integration of different sources of information (multisensory integration) using spiking neural networks is also a subject of investigation. A network architecture is proposed and a model for audiovisual pattern recognition is designed as an example. The main original contributions of this thesis are: a) Evaluation and further extension of adaptive learning procedures to perform visual pattern recognition. A new learning procedure that enables the system to change its structure, creating/merging neuronal maps of spiking neurons is presented and evaluated on a face recognition problem. b) Design of two new spiking neural network architectures to perform person authentication through the processing of speech signals. c) Design and evaluation of a new architecture that integrates sensory modalities based on spiking neurons. The integrative architecture combines opinions from individual modalities within a supramodal layer, which contains neurons sensitive to multiple sensory information. An additional feature that increases biological relevance is the crossmodal coupling of modalities, which effectively enables a given sensory modality to exert direct influence upon the processing areas typically related to other modalities. The contributions were published in one journal paper and in four refereed international conference proceedings. The proposed system designs were implemented and, through computer simulations, demonstrated comparable performance with traditional benchmarking methods. The systems have some promising features: they can be naturally optimized in respect to different criteria: accuracy (when very accurate results are expected), energy efficiency (when management of resources play an important role), and speed (when a decision needs to be made within a limited time). In this thesis, most of the parameters have been exhaustively optimized by hand or by using simple heuristics. As a direction for future work, there is an opportunity to include automated, specially tailored parameters optimization procedures or even general-purpose optimization algorithms, e.g., Genetic Algorithms and Particle Swarm Optimization. Overall, the results obtained in this thesis clearly indicate that it is indeed possible to have fast and accurate adaptive pattern recognition systems scalable for multiple modalities computing with simple models of spiking neurons. However, it is important to advance the theory of spiking neurons to take advantage of its biological relevance to reach similar or better performance when compared to the human brain, for instance, exploring new neuron models, information coding schemes and network connectivity.

As áreas de visão computacional e redes neurais artificiais (RNAs) e suas aplicações, tiveram um enorme progresso em pesquisa e aplicações práticas nos últimos anos. Sistemas de inspeção visual automática têm despertado muita atenção na indústria pois provêem meios econômicos, eficientes e precisos de obtenção de controle de qualidade. Porém, apesar do grande avanço tecnológico, a maioria dos sistemas existentes, com exceção de alguns poucos experimentais, são especializados e foram projetados para inspecionar um único objeto ou peça, de tipo previamente conhecido, e em posição, orientação e distância em relação à câmara altamente restritas. Este trabalho descreve um sistema de reconhecimento de imagens contendo múltiplos objetos de classes aleatórias e tolerante a ruído. Um estágio de pré-processamento filtra parte do ruído e segmenta regiões conectadas da imagem (RCI). A classificação dos padrões é feita com redes neurais de múltiplas camadas a partir de atributos invariantes calculados sobre as RCis. No final do processo temos uma listagem dos objetos contidos na cena, suas posições e orientações, os quais podem servir de entrada a um sistema de entendimento da cena, de mais alto nível, ou para outras máquinas, como um manipulador automático. Outros parâmetros podem ser utilizados para normalizar, em escala, orientação e posição, os padrões contidos na imagem, para efeito de comparações com imagens e parâmetros dos objetos previamente armazenados em bancos de dados. Dois métodos de treinamento de RNAs foram testados, o gradiente conjugado e o Levenberg-Marquardt, em conjunção com simulated annealing, para diferentes condições de erro e conjuntos de atributos. Imagens reais e sintéticas foram utilizadas para efeitos de testes de classificação correta e rejeição de padrões espúrios. Resultados são apresentados e comentados, destacando a capacidade de generalização do sistema mesmo com elevada degradação da imagem pelo ruído. Uma das vantagens do tipo de RNA empregado é a velocidade de execução, que permite ao sistema ser integrado a uma linha de montagem industrial. O sistema foi projetado com a utilização de recursos acessíveis e de baixo custo, sendo executado em computadores pessoais, e podendo ser facilmente adaptado para o uso em pequenas e médias empresas. / Computer vision (CV) and artificial neural networks (ANN) are important research fields of artificial intelligence. Visual pattern recognition (VPR) and object recognition (2 or 3-D) are central tasks in a high level computer vision system. Despite the great development in the recent years, most of the current automatic visual inspection systems work with only one kind of pattern at time which has pose highly restricted. This dissertation describes a system designed to recognize patterns and objects in a digital image which have unknown number object types and poses. Such image, which is also degraded by noise, serve as input for the system. After gray level change and filtering, the pixel connected regions (CR) are codified, and the remained noise is eliminated. lnvariant features, i.e., moment invariants, serve as inputs for artificial neural networks that perform pattern classification. An interpretation module decode the net\'s outputs and increases the correct assignment by testing the net\'s higher outputs values. After all identified patterns were classified, we have an object listing of the scene, their positions and other information, which can be the input for a higher level scene understanding system, that may check for objects relations and could send information for humans or for other machines. Two ANN learning methods were adopted for training the networks, the conjugate gradient and the Levenberg-Marquardt Algoritms, both in conjuction with siumlated annealing, for different error conditions and feature sets. Sinthetic and real images were utilized for testing the net\'s correct class assignments and rejections. Results are presented as well as comments focusing the system\'s generalization capability despite noise, geometrical transformations, object shadows and other degradations over the images. One of the advantages of the ANN employed is the low execution time allowing the system to be integrated to an assembly industry line. The system runs on low cost personal computers, therefore it can be easily adapted for the Brazilian reality and can even be used by little companies and industries.

As áreas de visão computacional e redes neurais artificiais (RNAs) e suas aplicações, tiveram um enorme progresso em pesquisa e aplicações práticas nos últimos anos. Sistemas de inspeção visual automática têm despertado muita atenção na indústria pois provêem meios econômicos, eficientes e precisos de obtenção de controle de qualidade. Porém, apesar do grande avanço tecnológico, a maioria dos sistemas existentes, com exceção de alguns poucos experimentais, são especializados e foram projetados para inspecionar um único objeto ou peça, de tipo previamente conhecido, e em posição, orientação e distância em relação à câmara altamente restritas. Este trabalho descreve um sistema de reconhecimento de imagens contendo múltiplos objetos de classes aleatórias e tolerante a ruído. Um estágio de pré-processamento filtra parte do ruído e segmenta regiões conectadas da imagem (RCI). A classificação dos padrões é feita com redes neurais de múltiplas camadas a partir de atributos invariantes calculados sobre as RCis. No final do processo temos uma listagem dos objetos contidos na cena, suas posições e orientações, os quais podem servir de entrada a um sistema de entendimento da cena, de mais alto nível, ou para outras máquinas, como um manipulador automático. Outros parâmetros podem ser utilizados para normalizar, em escala, orientação e posição, os padrões contidos na imagem, para efeito de comparações com imagens e parâmetros dos objetos previamente armazenados em bancos de dados. Dois métodos de treinamento de RNAs foram testados, o gradiente conjugado e o Levenberg-Marquardt, em conjunção com simulated annealing, para diferentes condições de erro e conjuntos de atributos. Imagens reais e sintéticas foram utilizadas para efeitos de testes de classificação correta e rejeição de padrões espúrios. Resultados são apresentados e comentados, destacando a capacidade de generalização do sistema mesmo com elevada degradação da imagem pelo ruído. Uma das vantagens do tipo de RNA empregado é a velocidade de execução, que permite ao sistema ser integrado a uma linha de montagem industrial. O sistema foi projetado com a utilização de recursos acessíveis e de baixo custo, sendo executado em computadores pessoais, e podendo ser facilmente adaptado para o uso em pequenas e médias empresas. / Computer vision (CV) and artificial neural networks (ANN) are important research fields of artificial intelligence. Visual pattern recognition (VPR) and object recognition (2 or 3-D) are central tasks in a high level computer vision system. Despite the great development in the recent years, most of the current automatic visual inspection systems work with only one kind of pattern at time which has pose highly restricted. This dissertation describes a system designed to recognize patterns and objects in a digital image which have unknown number object types and poses. Such image, which is also degraded by noise, serve as input for the system. After gray level change and filtering, the pixel connected regions (CR) are codified, and the remained noise is eliminated. lnvariant features, i.e., moment invariants, serve as inputs for artificial neural networks that perform pattern classification. An interpretation module decode the net\'s outputs and increases the correct assignment by testing the net\'s higher outputs values. After all identified patterns were classified, we have an object listing of the scene, their positions and other information, which can be the input for a higher level scene understanding system, that may check for objects relations and could send information for humans or for other machines. Two ANN learning methods were adopted for training the networks, the conjugate gradient and the Levenberg-Marquardt Algoritms, both in conjuction with siumlated annealing, for different error conditions and feature sets. Sinthetic and real images were utilized for testing the net\'s correct class assignments and rejections. Results are presented as well as comments focusing the system\'s generalization capability despite noise, geometrical transformations, object shadows and other degradations over the images. One of the advantages of the ANN employed is the low execution time allowing the system to be integrated to an assembly industry line. The system runs on low cost personal computers, therefore it can be easily adapted for the Brazilian reality and can even be used by little companies and industries.