Novel Color Models and Sampling Methods for Image Matting

Abstract

In this thesis, we propose novel matting algorithms that can extract high-quality and robust mattes from various regions of an image. Although many matting algorithms have been proposed, most of them cannot always produce satisfactory matting results for various regions of images, such as those that are smooth, contain holes or have complex color. To provide a comprehensive understanding about image matting, we first introduce a review of previous matting algorithms which are divided into propagation-based approaches and sampling-based approaches. We analyze each algorithm and find characteristics. Next, we propose a new matting algorithm using local and nonlocal neighbors. We assume that K nearest neighbors satisfy the color line model, that RGB distribution of the neighbors is roughly linear, and combine this assumption with the local color line model that RGB distribution of local neighbors is roughly linear. Our assumptions are appropriate for various regions mentioned above. Experimental results show that the new model-based matting algorithm outperforms previous propagation-based matting methods. Finally, we propose a new alpha matting algorithm based on clusters. To reduce sampling complexity and obtain high-quality matting results, a new sampling strategy selects the unknown pixels to apply a sampling method sparsely. The proposed cluster-based sampling algorithm uses a superpixel algorithm to perform clustering and uses the sampling method to estimate the alpha of one unknown pixel per cluster. The estimated alpha is used as prior information for a propagation-based matting. The alpha information can be propagated well in each cluster because all pixels that belong to the same cluster have similar colors and locations. The cluster-based sampling algorithm also uses the superpixel algorithm to reduce the redundancy of sample candidates that have similar colors and locations, although the sample candidates are gathered from a large range such as near the entire boundary of the unknown regions. Experimental results show that the cluster-based sampling algorithm performs the sampling method sparsely on images, but is as effective as the sampling method that is applied to all unknown pixels. A quantitative evaluation result on a benchmark data set shows that the cluster-based sampling algorithm estimates alpha mattes more accurately than the state-of-the-art matting algorithms.