Image Segmentation with Joint Regularization and Histogram Separation
Abstract: In this thesis optimization methods for image segmentation are studied. The common theme of all the methods is that we have a histogram model for appearance terms that we optimize jointly with smoothness. Recently it has been shown that if one assumes a histogram model for appearance, it is possible to optimize an approximation of the energy using only one graph cut, by ignoring the non-submodular volumetric penalty term. We show how to include the volumetric term using the Fast trust region framework. Fast trust region is a recently proposed method that is able to handle a large class of non-submodular energies by solving a sequence of graph cut problems. A comparison of these methods shows that Fast trust region typically obtains a lower energy value and higher segmentation quality, at the cost of requiring multiple graph cuts. Furthermore, we extend the simple histogram term to the multi-class setting and show that it is possible to optimize it with alpha-expansions. This is applied to the problems of stereo depth estimation and geometric model fitting.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)