Research On Natural Image Matting Method

Digital image matting is to extract the specified foreground object from a natural image. It is first used in special effect production of film industry, and creates a lot of commercial value for the industry. And now, with the development of technology, image matting has permeated in our daily life and developed from fixed background image matting to natural background image matting.Natural image matting has good prospects in many areas, but there are still many problems to be solved. The present natural image matting methods can be divided into two groups, one is natural image matting based on trimap, and the other is natural image matting based on lines. While the first ones require the users to carefully partition the natural image into three regions, so the user interaction of these natural image matting approaches are considerable complicated. These matting methods are inconvenient for using. And the second ones are extremely sensitive to the location of users'input, these natural image matting approaches will generate large errors when the users'input is insufficient.By studying and analyzing present natural image matting methods, to solve the problems of present natural image matting, this thesis proposes a new natural image matting method which combines the existing algorithm models of present computer vision. The new approach requires user to input lines marked foreground in the foreground region which is nearby the color variation of foreground contour and the same as in background region. It uses region growing approach to grow these lines to regions. In order to obtain a good result, this thesis adopts color similarity measurement of Munsell color space as growing principle of region growing. Then by finding contours of foreground region and background region and filling the inside region of this two regions, the task of partitioning image is completed. When dealing with the pixels of unknown region, it uses optimized color sampling method to optimize the foreground sample clusters and background sample clusters, this helps to improve the precision of matting result.Experiments show that, compared with traditional natural image matting methods, the one proposed by this thesis improves the way of partitioning a image, so it not only needs less interworking, performs more convenient, but also obtains a better trimap. By using optimized color sample method to optimize the samples, the matting result is more precise.