Abstract
In recent years, figure-ground segmentation has been a popular research topic in a number of different types of image segmentation methods. The goal of the figure-ground segmentation is to divide an image into two regions, which are foreground and background. There are many methods which have been proposed for solving figure-ground segmentation problems, but these methods are usually supervised approaches. In other words, the procedures of those methods need some interactions of users. It makes those methods unfavorable. Also, there are some disadvantages in traditional unsupervised image segmentation methods. In this thesis, we propose an unsupervised figure-ground segmentation method based on an object proposal generation algorithm to generate a small number of regions in an image, such that each object is well-represented by at least one region. Then, we combine the saliency map which measures the saliency likelihood of the image, color information, and gradient information to construct an objective function for the situation that only single foreground object exists in an image. Otherwise, the objective function is combined with an overlap constraint to handle the situation that multiple foreground objects appear in an image. Then we use a simple and efficient optimization method to get the initial object-wise segmentation results, and then refine the results by using pixel-wise graph cut. Comparing to other unsupervised figure-ground segmentation approaches, our method in MSRA-1000 database can get good experimental results.