Detecção de Armas de Fogo em Imagens Baseada em Redes Neurais Convolucionais

• Ano de defesa: 2021
• Autor(a) principal: Cardoso, Guilherme Vinícius Simões
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Mestrado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Redes neurais convolucionais; visão computacional; detecção de objetos; YOLO; subject.br-rjbn; Engenharia Elétrica
• Link de acesso: http://repositorio.ufes.br/handle/10/15539

The demand for weapons has grown along with crime rates, being a contemporary problem that haunts several countries. In Brazil, possible changes are being discussed to make the ownership and possession of firearms more flexible, dividing opinions and generating a huge discussion on the subject. This has motivated scientists to devise solutions that can assist in public security in general. In an attempt to find ways to minimize this problem, a research was carried out on the main work related to the classification and detection of firearms, aiming to obtain information on the main techniques used. Thus, in this work is proposed a methodology for the detection of firearms in images using convolutional neural networks. Recent work has used object detectors based on these networks and presented relevant results. Therefore, this work proposes a methodology for detecting weapons using an object detector, called YOLO (You Only Look Once), and an architecture based on convolutional neural networks. Two approaches were taken to evaluate the proposed methodology, taking into account three threshold values for IoU. The first approach, compared with the results found in the literature, points to an improvement in the results, where an accuracy of 93, 67% and a F1 of 93, 05% was achieved, which represents a growth greater than 10% in accuracy and a slight improvement of almost 2% in the F1 metric. The second approach follows the same methodology, but applies a different initial step, where the object detector is modified and used to mark a database and compose a new labeled one. Such approach had a positive impact on the results, where there was an increase in accuracy and almost 4% in the F1 metric. In the three IoU values evaluated, the best one has an accuracy of 89, 91% and, in the same configuration, points to a F1 of 94, 54% with a confidence of 58%. These results show that the proposed methodology is promising to be applied for firearms detection in images.