LLE Algorithm in Natural Image Matting
Junbin Gao, David Kho Ching Tien, James Tulip · Charles Sturt University Research Output (CRO) · 2011
Accurately extracting foreground objects in images and video has wide applications in digital photography. These kind of problems are referred to as image matting. The most recent work [1] in natural image matting relies on local smoothness assumptions about foreground and background colours on which a cost function has been established. The closed-form solution has been derived based on a certain degree of user inputs. In this paper, we present a framework for formulating a new cost function from the manifold learning perspective based on the socalled Locally Linear Embedding [2] where the local smoothness assumptions have been replaced by an implicit manifold structure defined in local colour spaces. We illustrate our new algorithm using the standard benchmark images and very comparable results have been obtained.