Segmentação de pele em imagens digitais para a detecção automática de conteúdo ofensivo

• Ano de defesa: 2010
• Autor(a) principal: Kuiaski, Diogo Rosa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial
• Palavras-chave em Português: Processamento de imagens - Técnicas digitais; Sistemas de reconhecimento de padrões; Mineração de dados (Computação); Image processing - Digital techniques; Pattern recognition systems; Data mining
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/1338

This work presents a study of suitable approaches for automatic detection of offensive content (pornography) in digital images. Extensive experiments were conducted for skin pixel segmentation, colour spaces and content descriptors. This work focus its efforts on skin pixel segmentation, since this segmentation is the pre-processing stage for almost every content-based offensive image classification methods in the literature. Four skin skin segmentation methods were tested in six colour spaces. Also, a structured image database was built to help improve studies in skin segmentation, with the possibility of adding meta-information to the images in the database, such as illumination conditions and camera standards. With the help of meta information from the image database, experimets involving illumination conditions and skin colour segmentation were also done. Finally, some feature extraction algorithms were implemented in order to apply content-based image retrieval (CBIR) algorithms to classify offensive images.