The thesis deals with variational approaches to the segmentation of images with particular emphasis on geometry-based algorithms. Criteria for the choice of segmentations are given by minimization of a suitable functional ranging over a certain prescribed segmentation class. The arising minimization problem is in general intractable. Therefore various geometric restrictions on the class of segmentations are formulated and characterized. Main contribution of the thesis is the design of fast and flexible algorithms for the computation of wedgelet-type approximations, based on a particularly efficient solution of a local regression problem. Here, the angular resolution can be prescribed in an arbitrary manner. For dyadic partitions, the implementation gives fast access to the entire scale of minimizers. Consistency results for the wedgelet estimators are formulated and proved. These results describe the asymptotic behavior of the wedgelet-minimizers of discretized continuous mages, as the pixel size tends to zero. They provide heuristics for the choice of the scaling parameter and yield convergence rates that are close to theoretically optimal. Experimental results substantiate the efficiency of the algorithms, they rate the dependency of the angular resolution and give hints for the choice of the scaling parameter.
«