Uma abordagem não-supervisionada para segmentação de cenas naturais coloridas

• Ano de defesa: 2011
• Autor(a) principal: Komati, Karin Satie
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Processamento de imagens; Sistemas de reconhecimento de padrões; Visão por computador; Fractais; Engenharia Elétrica; 621.3
• Link de acesso: http://repositorio.ufes.br/handle/10/9701

Segmentation of natural scenes is an important task in image processing and computer vision, with applications in several areas such as robot navigation and object recognition. However, the segmentation can become extremely complex due to the huge variability of color, lighting and textures found in an image. In other words, it is very difficult to develop an approach that can successfully segment all changes in a scene. This Thesis proposes a new unsupervised and fully automatic method for boundary detection in natural color images, consisting of two levels of integration, or two-stage sequential processes. In the first stage, two different techniques to measure color-texture homogeneity in a region-growing method are combined by two different control algorithms. One control algorithm is based on a local function and the other is based on a global statistical property (the shape of the power spectrum of the image being analyzed). One homogeneity measure is the J-value (provided by the JSEG algorithm) and the second measure is a multifractal descriptor. This first stage performs region extraction. In the second integration, edge information is extracted by a classical method, and integrated with region information. This process eliminates false boundaries in the region map, guided by the edge map, and reduces the noise in the edge map as well, now guided by the region map, thus taking advantage of their complementary nature. Furthermore, it integrates the two maps into a single final result, enhancing the coincident information of both maps. Each phase of integration improves, progressively, the detection of the boundaries. Experiments on a large and consolidated dataset of natural color images (“The Berkeley Segmentation Dataset and Benchmark”) suggest that the results for the approach here proposed are closer to the human perception than the individual methods, quantitatively and qualitatively speaking.