Abstract
Image co-segmentation is a relatively new branch of image processing. The goal of image co-segmentation is to separate common objects or similar objects among input image sets. Most existing co-segmentation approaches focus on two-class segmentation,. It means they aim for separating only one common object of a class, usually called foreground, and the remaining background in each of input images. From this binary labeling premise those methods inevitably suffer from some limitations. In this thesis, we propose a multi-class co-segmentation method which is in favor of coalition game theory. A game which is in a coalition form means that some players in the game can form an alliance to have a joint payoff. Our approach first over-segments input images and treat each region (superpixel) as a player; each region has the ability to ally with others regions to form a coalition. The payoffs of coalitions are determined by the scores of each class model. Therefore, we can take advantage of Shapley values to determine how the payoffs are being distributed among all coalition members. After assigning each region to one of the classes according to their scores, we can achieve the best segmentation result. Therefore, new class models can be trained by using previous region assignment. Performing those region assignment and model training iteratively, we can acquire better co-segmentation.