A robust double-layer affinity based matting method

In traditional matting algorithms, sampling and affine methods are the primary approaches. The sampling method provides initial alpha values for each pixel but often results in discontinuities and noise, necessitating affine methods for smoothing. Affine methods, which involve both short- and long-range searches, enhance alpha value continuity between neighboring pixels but have varying effects on different regions. To address this, we propose a dual-layer post-processing matting algorithm that combines robust priors with long-range and shortrange modeling. The robust prior step independently samples each unknown pixel, selecting optimal foreground and background samples to compute initial alpha values. Then, a dual-layer affine framework, utilizing the KNN method and a color linear model, further refines these alpha values. Experiments show that this dual-layer algorithm improves both the visual quality and the accuracy of alpha values for foreground, background, and mixed regions.