Otimizando CNNs Com Aprendizado Acumulativo Via Múltiplas Redes Neurais.

• Ano de defesa: 2020
• Autor(a) principal: SCAVONE, Joaquim Martins
• Orientador(a): ALMEIDA NETO, Areolino de
• Banca de defesa: ALMEIDA NETO, Areolino de , BRAZ JUNIOR, Geraldo , OLIVEIRA, Alexandre César Muniz de , ROCHA, Marcelo Lisboa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: reconhecimento e classificação de semáforos; aprendizagem profunda; rede neural convolucional; múltiplas redes neurais autocoordenadas
• Palavras-chave em Inglês: Traffic Lights Recognition; Deep Learning; Multiple Self-coordinated Neural Networks
• Área do conhecimento CNPq: Modelos Analíticos e de Simulação
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/tede/3290

The number of fatalities in traffic accidents is staggering. Many of these accidents result from disrespect for signaling, which often happens involuntarily, due to distractions, for example. This issue has been treated with great attention in the scientific community. This led to the emergence of Advanced Driver Assistance Systems (ADAS), which are systems that can interpret signaling and flow on the road and, based on this information, issue alerts to the driver or even intervene in driving. Convolutional networks are already widely used in ADAS and are promoting real progress in this area. Thus, this work presents a strategy that uses neural networks in this type of problem. The developed research made a union of the techniques of multiple self-coordinated neural networks and convolutional neural networks, which demonstrated its efficiency when applied to already trained networks. The proposed technique achieved 95.33% accuracy, the possibility of reducing training time and a new strategy to escape local minimums, which opens up a range of new research that can be carried out.