A novel image matting approach based on naive Bayes classifier

Abstract

Image matting is a fundamental technique used in many image and video applications. It aims to softly extract foreground from the image accurately. In this paper, we propose a new matting approach based on naive Bayes classifier to produce matting results with higher accuracy. Spatially-varying probabilistic models for the classifier are established. Confidence values are defined to make better use of the classification results. The results are then refined and combined with closed-form matting to obtain the final alpha matte. We conduct qualitative and quantitative evaluations. Results show that our method outperforms many recent algorithms. © 2012 Springer-Verlag.