A Bayesian Approach to Clustering Matting Components in Spectral Matting

This paper proposes to apply Bayesian principle to clustering matting components in spectral matting. Spectral matting is a useful and effective technique for digital image matting. A crucial issue of spectral matting is how to cluster the computed matting components, which compose the final alpha matte. In this paper, a new clustering strategy based on Bayesian decision theory is proposed to solve this problem. In our algorithm, given the input scribbles as a trimap, the foreground and background information is propagated outward into unknown region iteratively, which makes up the calculated foreground/ background distribution function. Then the Bayesian decision theory is adopted to cluster the matting components. The matting components which are clustered into foreground are summed up to generate the final alpha matte.