Recomendações de produtos em tempo real baseadas no reconhecimento de objetos em vídeo

• Ano de defesa: 2019
• Autor(a) principal: Souza Junior, Nelson Forte de
• Orientador(a): Marengoni, Mauricio
• Banca de defesa: Lopes, Fábio Silva , Roman, Norton Trevisan
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Programa de Pós-Graduação: Engenharia Elétrica
• Departamento: Escola de Engenharia Mackenzie (EE)
• País: Brasil
• Palavras-chave em Português: redes neurais profundas; redes convolucionais; visão computacional; recomendações; comércio eletrônico
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://dspace.mackenzie.br/handle/10899/24503

With the development of e-commerce in buying and selling relationships, and its constant overlap over the classic way of selling products, many computational and statistical methods were discovered and researched to recommend such products. Often the data used in recommendation methods involves user interactions and are saved in the form of structured, semistructured, or unstructured data, which images and video are a type of information seldom used. This work, which we call Xanathar, proposes to extend such paradigm by recognizing and recommending products in real time using video from several di erent sources. To achieve this, a convolutional neural network architecture known as ResNet-50 was used, and some changes in the layers of that network led to a top-5 error rate of 5.17% and recognition time of 0.016 seconds. For the network training, an annotated image database with 936563 samples was build, containing 25 di erent classes of objects.