Redes neurais profundas para reconhecimento facial no contexto de segurança pública

• Ano de defesa: 2020
• Autor(a) principal: Silva Júnior, Jones José da
• Orientador(a): Soares, Anderson da Silva
• Banca de defesa: Soares, Anderson da Silva, Calixto, Wesley Pacheco, Costa, Ronaldo Martins da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RG)
• País: Brasil
• Palavras-chave em Português: Visão computacional; Reconhecimento facial; Redes neurais
• Palavras-chave em Inglês: Computer vision; Face recognition; Neural networks
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/10567

Face recognition is an important tool for law enforcement. Bein able to compare a face image of a suspect filmed at a crime scene with a database of millions of photos and thus find his true identity represents a significant increase in crime resolution rates. Although this task has been researched since the 1970s, it was with the use of Convolutional Neural Networks (RNCs) from 2014 that a relevant advance was achieved that allowed some to reach 99.63% accuracy in the benchmark Labeled Faces in the Wild (LFW). Despite different architectures and cost functions, a common feature of the papers published since then is the fact that they are trained in a supervised manner, thus requiring large collections of facial images previously labeled. Even state of arts models in public benchmarks, they may not achieve the same results in the real world. The main reason is the lack of demographic data distribuition of these public datasets, which results in models with greater accuracy in specific demographic subgroups and worst accuracy in other subgroups, such as afrodescendant women. This work aims to investigate the fine tuning training strategies of deep neural network architectures for facial recognition in public safety context, using a dataset with the Brazilian faces in order to generate a more accurate model for a investigations police department. We managed to improve accuracy on test set with samples representative of the context of this work training a model with private dataset with a very small number of samples compared to the public ones.