Method and apparatus for creating foreground masks, or mattes, in images including complex images. A discriminative matting technique may generate accurate alpha mattes for textured images or objects with spatial-varying color distributions. Given an input image and a trimap defining an unknown region, a discriminative color analysis is applied to the unknown region, yielding estimated alpha values, estimated binary segmentation values, and a mixture probability map for the region. The map may be adaptively smoothed. The pixels in the unknown region are classified into boundary pixels and non-boundary pixels according to the probability map. The non-boundary pixels are classified as either foreground or background pixels using a differencing technique that compares multiple pixel features. The estimated alpha values for the boundary pixels are refined. An alpha matte for the image is output. The process may be repeated until convergence of the alpha matte.