Background Separation In The Image Before Realization

Matting is the technique of extracting some interesting part out of the original image or video, whose main functionality is to composite the extracted foreground object into images with new backgrounds perfectly. Matting has been already widely used in digital image and video editing and almost all the image and video editing software, film special effects and synthesis of television scenes. Matting is a challenging task with three difficulties. First, non-uniqueness of solutions:for any interesting part of the image, different degrees of human interaction, different prior information, different assumptions and constraints on all kinds of features of natural image, different user needs, different samplings, statistics, analysis, modeling methods can result in totally different solutions. Second, possible huge space or time complexity:being particular about extraction of details, many matting algorithms may calculate the statistics of image. A natural image with rich details may have huge space and time complexity. Third, the establishing of the evaluation criteria:even for human vision, given the same image with the same interesting part, different person could have different solution for details. A matting algorithm can be evaluated from different angles, such as accuracy, efficiency and memory usage. So far no united and authoritative evaluation criteria have been established for matting yet.This paper has two contributions. One is its model of acquiring depth information of video integrated with edge information, which has better experimental result than existing methods of acquiring depth information of video. The other is a two-scale matting model based on both color and depth, comprehensively considering intrinsic color information and spatial information of objects in the image. Color features are roughly estimated by some existing matting methods using alpha matte of the image. Depth features, represented by depth maps, are acquired by adopting fast stereo matching algorithm with double horizontally-paralleled viewpoints for images, while by depth information model integrated with edge information for videos. When colors are similar at the border of foreground and background, only considering color features cannot give exact solution, but with depth map providing clearer spatial information at the border, matting result can be improved a lot. However, depth map is dependent on accuracy of other algorithms and models to some extent. So if we want to do matting to extract the alpha value of each single pixel of the image accurately, color features are still necessary for final solution. We combine two-scale information, color and depth into one matting model and give detailed algorithm, which proves to have better experimental results of matting.