Redes neurais convolucionais e energia de dobramento multiescala na descrição e análise de formas

• Ano de defesa: 2019
• Autor(a) principal: Lopes, José Gerardo Fonteles
• Orientador(a): Medeiros, Fátima Nelsizeuma Sombra de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Teleinformática; Redes neurais (Computação); Processamento de imagens; Imagens digitais; Shape analysis; Convolutional neural networks; Shape descriptor; Content based image retrieval
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/47319

The automatic recognition of pills through image can reduce wrong drug administration in elderly patients and also enable forensic intelligence to establish links among illegal drug marketing. Considering the remarkable progress and performance of the convolutional neural networks (CNNs) in pattern recognition problems, this study proposes the application of neural network, namely LeNet and Inception ResNet as well as the normalize normalize multiscale bending energy (NMBE) for feature extraction of licit and illicit pills in order to develop algorithms for automatic recognition of pills. We assess the generalization of the algorithms in an image database from the international alphabet of sign language (Hands). We conduct classification and content-based image retrieval (CBIR) experiments and evaluate the results quantitatively using the Accuracy and mean average precision (MAP) measures. We perform the qualitative analysis through the U matrix for visualization of the cluster arrangement. The results showed that LeNet outperformed the Inception Resnet and NMBE for image databases with a great amount of images for training as the Licit Pills and Hands databases.