Explorando abordagens de aprendizado sequencial para floresta de caminhos ótimos

• Autor(a) principal: Nakamura, Rodrigo Yuji Mizobe [UNESP]
• Data de Publicação: 2014
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://hdl.handle.net/11449/122159
• topic: Computação; Processamento de imagens - Tecnicas digitais; Floresta de caminhos ótimos; Markov, Campos aleatórios de; Arvores (Teoria dos grafos); Computer science

The interpretation of classification problem as a graph search provides a rich framework with correct and efficient algorithms. The Optimum-Path Fo- rest classifier can reduce classification to the the computation of an optimum- path forest according to a connectivity function, which assigns a value to any path in the graph. Considering the maximum value among all possible paths with terminus at each node, the optimum path is trivial for some nodes, cal- led roots, and the remaining nodes will have an optimum path coming from their most strongly connected root, partitioning the graph into an optimum- path forest (disjoint sets of optimum-path trees). Notwithstanding, to clas- sify out-of-sample, we assume that each sample in the new dataset composes one node in the graph and we compute their most strongly connected root within all spanning trees. As one can see, this procedure do not take advan- tage of the problem structure information, which can be fundamental for a better precision of the results. In this context, the purpose of this work is to evaluate the contribution of contextual modelling techniques, such as Mar- kov random fields and stacked classifiers. The first approach, called Markov random fields, sumarizes the system overall behavior through its local inte- ractions. The second approach, based on combination of classifiers, model the interaction between samples in the space scale, which provides effici- ent implementations of long interaction by defining neighborly relations in multiple scales. The results for brain tissue segmentation of magnetic reso- nance images and land-cover classification of multi-spectral satellite images show that the contextual information can improve the effectiveness of the Optimum-Path Forest classifier