Performance of Bayesian-based digital matting techniques

Abstract

Statistical-based digital matting of natural background is considered in this paper. Performance of the original Bayesian-based matting method has been studied. It is shown that the performance depends on the algorithm initial condition as well as a noise variance parameter. Two initialization methods have been considered. The first one obtains initial value from Ruzonpsilas maximum likelihood method, while the second one employs the centers of foreground and background color clusters as the initial value. Both methods were found to offer improved performance. In addition, optimum choice of the noise variance was found to vary from one image to another. An alternative formulation of a Bayesian-based digital matting technique which does not contain the noise variance parameter is described. It was found to yield results comparable to the original method for most of the test cases.