| Image segmentation is the partitioning of an image into meaningful regions or pixel groups and is a necessary prerequisite for many higher level computer vision tasks, such as object recognition, scene interpretation, and content-based image retrieval. However, the segmentation problem is inherently ill-posed due to the large number of possible partitionings for any single image. Much effort in image segmentation research is devoted to making the problem more tractable by constraining the solution space using prior information. Commonly, the optimality criteria used to compute a preferred partitioning are formulated base on measures that account for contour smoothness, regional coherence, and visual homogeneity.;In this thesis, we present a set of novel image segmentation algorithms that utilize high-level semantic priors available from specific application domains. These priors are incorporated into the segmentation framework to further constrain the results to a more semantically meaningful solution space. Our algorithms are formulated using Random Field models and employ combinatorial graph cuts for efficient optimization. For many instances, they guarantee the globally optimal solutions, and our experiments demonstrate that the algorithms are applicable to a wide range of segmentation tasks. |
