Automatic defocus spectral matting

Abstract

Alpha matting for single image is an inherently under-constrained problem and thus normally requires user input. In this paper, an automatic, bottom-up matting algorithm using defocus cue is proposed. Different from most defocus matting algorithms, we first extract matting components by applying unsupervised spectral matting algorithm on single image. The defocus cue is then used for classifying matting components to form a complete foreground matte. This approach gives more robust result because focus estimation is used in component level rather than pixel level.