Master of Science
Image segmentation is a fundamental problem in computer vision that has drawn intensive research attention during the past few decades, resulting in a variety of segmentation algorithms. Segmentation is often formulated as a Markov random field (MRF) and the solution corresponding to the maximum a posteriori probability (MAP) is found using energy minimiza- tion framework. Many standard segmentation techniques rely on foreground and background appearance models given a priori. In this case the corresponding energy can be efficiently op- timized globally. If the appearance models are not known, the energy becomes NP-hard, and many methods resort to iterative schemes that jointly optimize appearance and segmentation. Such algorithms can only guarantee local minimum.
Here we propose a new energy term explicitly measuring L1 distance between the object and background appearance models that can be globally maximized in one graph cut. Our method directly tries to minimize the appearance overlap between the segments. We show that in many applications including interactive segmentation, shape matching, segmentation from stereo pairs and saliency segmentation our simple term makes NP-hard segmentation functionals unnecessary and renders good segmentation performance both qualitatively and quantitatively.
Tang, Meng, "Color Separation for Image Segmentation" (2013). Electronic Thesis and Dissertation Repository. 1834.