Color and depth based image segmentation using a game-theoretic approach

In this thesis a new game theoretic approach to image segmentation is proposed. It is an attempt to give a contribution to a new interesting research area in image processing, which tries to boost image segmentation combining information about appareance (e.g. color) and information about spatial arrangement. The proposed algorithm firstly partition the image into small subsets of pixels, in order to reduce computational complexity of the subsequent phases. Two different distance measures between each pair of pixels subsets are then computed, one regarding color information and one based on spatial-geometric information. A similarity measure between each pair of pixel subset is then computed, exploiting both color and spatial data. Finally, pixels subsets are modeled into an evolutionary game in order to group similar pixels into meaningful segments. After a brief review of image segmentation approaches, the proposed algorithm is described and different experimental tests are carried up to evaluate its segmentation performance