GPU-based nonlocal smoothing for alpha matting
Thanathorn Phoka, Attawith Sudsang · 2016
This paper addresses the problem of improving an extracted alpha matte from any matting method. One conventional method is to smooth the alpha matte using Gaussian filter. However, this sometimes induces over smoothing which does not preserve edge features and the background region on the boundary of the alpha matte is turned to semi transparent. The main contribution of this paper is to propose a matte smoothing method based on nonlocal method which is broadly used in image denoising to avoid over smoothing. Another contribution is an implementation of the method on GPU which is plausible to run in real time.