On segmentation with Markovian models

FLESIA ANA GEORGINA; GIMENEZ ROMERO, JAVIER ALEJANDRO; BAUMGARTNER JOSEF

Córdoba

Simposio; XIV Argentine Symposium on Articial Intelligence (ASAI 2013); 2013

SADIO

Markovian models are ubiquitous in image segmentation. Nevertheless, model selection is an extremely difficult problem to solve. Markovian models depend on several hypothesis that determine the number of parameters and general complexity of the estimation and prediction algorithms. The Markovian neighborhood hypothesis, order and isotropy, are the most conspicuous properties to set. In this paper, we study goodness of fit of two different Markov Field environments for image segmentation, considering the labels of the image as hidden states and solving the estimation of such labels as a solution of the equation s* = argmax_s P(s|O; heta) where O are the observed image and heta the model. The emission distribution are assumed the same in both models, and the difference lays in the Markovian hypothesis made over the labeling random Field. Estimation over all possible combination of states values s is well known to be unfeasible, so we compare and contrast two approximated solutions, 2d path-constrained Viterbi training (PCVT) and Potts-ICM, in an integrated Matlab environment, using simulated and real images. Our emphasis is to investigate goodness of fit by studying statistical complexity, computational gain, extent of automation, and rate of classification measured with kappa statistic. All code written is provided in a toolbox available for download from our website, following the Reproducible Research Paradigm.