DOI: 10.36347/sjet

Pages: 313-315

Sampled the prior distribution using the Metropolis Algorithm. We then estimated the percent of black pixels in each band and compared it to the percent of black pixels in the training set. Several other neighborhoods were considered for the binary prior. We found that with our banding we could accurately estimate cliques and separator distributions containing more than five pixels from the given training set. By that we mean that for cliques containing more than five pixels, we found that there existed at least one binary clique configuration that appeared less than thirty times in the training set. Furthermore, not all collections of cliques formed neighborhoods for which Graphical Model theory could be used. These constraints were the reason for the chosen small neighborhood.