Natural Image Matting Based On Image Structure Information

Natural image matting technology is one of the basic technologies in computer vision.It is mainly applied to image and video editing,or the preprocessing of advanced image processing.The existing natural image matting algorithms are mainly divided into three categories: sampling-based,propagation-based,and machine-learning-based.The sampling-based algorithms need to select candidate sample sets for unmarked pixels,and select suitable foreground and background sample pairs.The propagationbased algorithms need to select adjacent points for unmarked pixels and establish relational weights between adjacent points.The process of selecting pixels in these algorithms can be referred as to the selection of correlation points.This process has an important influence on the accuracy of the matting.The traditional matting algorithms only use a single pixel feature to select the correlation point,but the information carried by a single pixel feature is limited,which is not enough to distinguish some highly similar pixels.Different from these methods,this paper designs two correlation point selection methods based on image structure information.The first method is based on prior knowledge of local self-similarity of images and establishes a candidate set of related points through similar image blocks,thereby using the image structure information of similar blocks to distinguish highly similar pixels,and searching for related points in this set.Then we obtain the matting results through a propagation-based approach.The second method replaces the similar image blocks in the first method to similar superpixels.In order to use the image structure information of superpixels more effectively,a new metric function between superpixels is designed.In addition,a simple and effective sampling-based method is designed based on this metric function,and it is combined with the propagation-based method.Experimental results show that the proposed methods based on image structure information can effectively distinguish some highly similar pixels and obtain more accurate results.Among them,the pixel-based algorithm based on superpixels ranks high on the standard dataset,and it has achieved a leading effect on some images whose background is segmented by foreground.