Image Matting via LLE/iLLE Manifold Learning

Main Article Content

David Tien, Junbin Gao, Jim Tulip


Accurately extracting foreground objects is the problem of isolating the foreground in images and video, called image matting which has wide applications in digital photography. This problem is severely ill-posed in the sense that, at each pixel, one must estimate the foreground and background pixels and the so-called alpha value from only pixel information. The most recent work in natural image matting rely on local smoothness assumptions about foreground and background colours on which a cost function has been established. In this paper, we propose an extension to the class of affinity based matting techniques by incorporating local manifold structural information to produce both a smoother matte based on the socalled improved Locally Linear Embedding. We illustrate our new algorithm using the standard benchmark images and very comparable results have been obtained.

Article Details