Um algoritmo rápido e modelos de propagação para técnica de level set aplicados na segmentação hierárquica de imagens

• Ano de defesa: 2020
• Autor(a) principal: Braga, Alan Magalhães
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Segmentação de imagem; Método de contornos ativos; Método Level set; Neoplasias do Colo do Útero; Imagens digitais; Image segmentation; Active contouring method; Level set method; Cervical Neoplasms; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/78329

Image segmentation methods are crucial in computer vision systems, since the results of these methods are inputs to the following steps, such as feature extraction and classification. The level set methods have been used successfully in many digital image segmentation applications. In order to apply these methods for region segmentation, an initial curve evolves according to a propagation model, in which the choice of this model must be directly related to the segmentation problem addressed. Thus, different types of digital images, such as cervical cell images and synthetic aperture radar (SAR) images, can be segmented using level sets. The traditional proposal of this approach provides binary segmentation and it has a high computational cost. Thus, in this thesis, we proposed a fast binary level set algorithm, implemented in narrow band and regularized with a median filter. In addition, we proposed propagation models for digital image segmentation and hierarchical implementations for nuclei segmentation on cervical cell images and for SAR image segmentation. The proposed hierarchical approaches, using the proposed binary level set algorithm, recursively segment a region into two new regions, starting from the whole image, and the process stops when all regions cannot be further divided. For performance assessment of the proposed approaches, we carried out experiments on three public image databases with cervical cells and experiments on synthetic and real SAR images, following the models and . For the quantitative evaluation of the segmentation results on cervical cell images, we used the pixel-based precision and recall measures, the Zijdenbos similarity index (ZSI) and the object-based precision and recall measures. These measures indicated that the proposed hierarchical implementation performed well concerning the number of correctly segmented nuclei and the Zijdenbos Similarity Index achieved values equal to or higher than 0.90. For the quantitative evaluation of the segmentation results in SAR images, we used the cross-region fitting (CRF), error of segmentation and ZSI measures. Based on these measures, the proposed approach achieved good segmentation results in images modeled with the distributions and .