Layered Depth Image Segmentation Using Graph Cuts

There are image segmentation algorithms that using just color information or edgeinformation. Graph cut approaches provide a possible way of combining both the color andedge information by using energy minimization to get the global optimal cut. The purpose ofthis paper is to study the effective depth image segmentation method with the improvedGraph cut adapted to a depth image. Also，in order to reduce the executing time of Graph cut,a layered method is developed to achieve the goal.Firstly, this paper introduces the theory of Graph cut, and uses label to change the imagesegmentation problem into a math problem, also uses probability theory to infer the changedlabel problem. Besides these two, Markov random ifeld is also introduced as a priorhypothesis and the energy function is used to describe the problem. After building the graph,the network lfow theory is adopted to solve the problem.Secondly a probabilistic model is designed，which is the foundation of our Graph cutapproach. The probabilistic model is used to describe what users want to segment and whatthey do not. There are usually two tools used in Graph cut, one is color histogram, and theother is Gaussian mixture model.To accelerate the speed of the Graph cut algorithm, this paper develops a layered strategyby combining the image pyramid and the Graph cut together.Finally, this paper implements a depth image segmentation approach by combiningdifferent models together to create a platform for depth images segmentation, and theexperimental results show it works very well.