Algoritmo ejeção-absorção metropolizado para segmentação de imagens

• Ano de defesa: 2014
• Autor(a) principal: Calixto, Alexandre Pitangui
• Orientador(a): Milan, Luis Aparecido
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Estatística - PPGEs
• País: BR
• Palavras-chave em Português: Estatística; Segmentação de imagem; Potts, Modelo de; Distribuição de Gibbs; Algoritmo de Swendsen-Wang; Atualização split-merge
• Palavras-chave em Inglês: Reversible jump
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA::ESTATISTICA
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/4496

We proposed a new split-merge MCMC algorithm for image segmentation. We describe how an image can be subdivided into multiple disjoint regions, with each region having an associated latent indicator variable. The latent indicator variables are modeled with a prior Gibbs distribution governed by a spatial regularization parameter. Regions with same label define a component. Pixels within a component are distributed according to a Gaussian distribution. We treat the spatial regularization parameter and the number of components K as unknown. To estimate K, the spatial regularization parameter and the component parameters we propose the Metropolised split-merge (MSM) algorithm. The MSM comprises two type of moves. The first one, is a data-driven split-merge move. These movements change the number of components K in the neighborhood K _ 1 and are accepted according to Metropolis-Hastings acceptance probability. After a split-merge step, the component parameters, the spatial regularization parameter and latent allocation variables are updated conditional on K by using the Gibbs sampling, the Metropolis- Hastings and Swendsen-Wang algorithm, respectively. The main advantage of the proposed algorithm is that it is easy to implement and the acceptance probability for split-merge movements depends only of the observed data. The performance of the proposed algorithm is verified using artificial datasets as well as real datasets.