Iterative Refinement of Alpha Matte using Binary Classifier

Abstract

Image matting is a technique used to extract a foreground object from an input color image by estimating its opacity mask, which is called an alpha matte. Many previous methods have estimated a poor-quality alpha matte when a foreground object and its background have similar colors. To overcome this problem, we treat an image matting problem as a binary classification problem to classify unknown pixels into foreground and background pixels. In this study, we propose a learning-based matting method based on binary classification. In our method, a binary alpha matte is first obtained by a binary classifier. The binary alpha matte is then iteratively refined to a high-quality alpha matte by repeat use of the classifier. Although we used a support vector machine classifier and the closed-form matting method, an important merit of our method is that it provides a general mechanism to refine an alpha matte by combining various binary classifiers and matting methods. The excellent performance of our method was revealed in several experiments, especially, for input images having similar foreground and background colors as well as objects with small holes.