Translator Disclaimer
12 January 2012 Natural image matting through overlapping neighborhood propagation
Author Affiliations +
The Laplacian based matting methods are attracting a lot of attention due to their elegant and high quality closedform solution. In this paper, we develop an alternative Laplacian construction for matting task by using local linear learning model, and naturally derive its nonlinear extension by incorporating Kernel Ridge Regression algorithm. Our Laplacian matrix construction approaches are based on the assumption that the alpha matte of each pixel point can be reconstructed from its neighbors' alpha values in each of overlapping windows. In this way the induced Laplacians can better exploit neighborhood intrinsic structure to constrain the propagation of foreground and background labels. Experimental results demonstrate the proposed approaches produce very high accuracy matte values, of which our nonlinear method even outperforms other Laplacian based matting methods on many test images.
© (2012) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Hongjing Peng, Jiang Duan, Jianhua Yuan, and Dinghong Shao "Natural image matting through overlapping neighborhood propagation", Proc. SPIE 8350, Fourth International Conference on Machine Vision (ICMV 2011): Computer Vision and Image Analysis; Pattern Recognition and Basic Technologies, 83501C (12 January 2012);

Back to Top