Graph-Based Methods in Image Segmentation

Abstract

In this thesis, we focus on the problem of image segmentation using the methods from the graph theory. Image segmentation is a technique of computer vision for partitioning the image into multiple segments. Although this problem has been studied for a long time, the progress cannot be considered to be finished. A weighted graph is a common image representation that allows to apply the well-known techniques from the graph theory in image segmentation. In the state-of-the-art chapter of this thesis, we summarize the graph methods that are frequently used in the area of image segmentation; the shortest path, graph cuts, random walker, and minimum spanning tree. Many papers dealing with the mentioned methods have been published in recent years, which shows that the methods are reputable in this area. However, we illustrate that some of them do not perform well in more complicated images, which was the motivation to find new methods. Our contribution to this area can be split into three parts; (i) we introduce a new distance that is indented for graphs and images. In comparison with other distances, certain positive properties of the new distance are shown. (ii) We present two new forms of the data term for the energy function that is used in the graph cuts segmentation method. (iii) We introduce a new segmentation method whose principle is similar to the anisotropic diffusion process. The method is inspired by the flow of a liquid in a mesh of pipes.