Color range determination and alpha matting for color images

In this paper, a novel matting method is proposed to automatically detect and separate foreground, background and transitional (unknown) regions in a color image. In order to detect the background color, K-means clustering in YC_bC_r color space is firstly used to classify the background colors into a limited number of clusters. Then the spatial information is further used to refine the background and minimize the unknown regions. In this case, an image can be automatically segmented into three hard regions: foreground, background and unknown regions. For transitional (unknown) regions, the alpha matting based on Wang´s robust matting algorithm is utilized to refine the accuracy of the separation results. By combining an automatical background determination metric and Wang´s robust matting, the proposed matting method can handle images with single-colored or gridded background. The required user input is significantly simplified compared to conventional alpha matting schemes which require users to provide a hard image segmentation manually. The experimental results show that improved matting results can be achieved for complex unknown regions which contain semi-transparent materials or tiny objects such as hair stripes.