基于随机纹理的代价滤波式抠图  

Cost Filtered Matting with Radom Texture Features

作　　者：陈秋凤[1] 申群太[1] 刘鹏飞[1,2] 

机构地区：[1]中南大学信息科学与工程学院 [2]中国人民解放军95856部队

出　　处：《电子与信息学报》2015年第11期2578-2586,共9页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61473318;60974048)~~

摘　　要：该文针对抠图中前背景颜色歧义这一难题,提出快速随机纹理算法来对颜色信息进行有效的补偿,先对原始图像进行稠密抽取得到初始纹理,后经随机投影降维,再根据前背景交叠度选择最优通道生成随机纹理图。结合生成的纹理信息,设计了空间、颜色、纹理联合样本选择指标。接着,综合考虑局部近邻和非局部近邻的作用,对样本选择代价进行滤波。最后论证近邻迭代滤波与全局能量方程平滑的关系,推导了后期迭代平滑公式。实验结果表明,基于随机纹理的代价滤波式抠图在前背景颜色分布近似时,能够取得视觉和定量上更好的结果。In order to deal with the color overlap problem in matting, a fast random projection method is proposed to complement the color information. First, the raw texture matrix is obtained through dense abstraction from color image. The random projection is performed and the best three texture channels are chosen by the foreground and background overlap factors. Combining the texture image, the new cost function takes into account texture, color, and spatial information. Second, the filtering process is carried out to the sample selection cost, including the effect of the local and nonlocal neighbors. Finally, the relationship between iterative filter and global energy smooth is proven, and the post filter formula is obtained. Experiments show that the cost filtered matting with random texture features produces both visually and quantitatively better results when the color distributions of the foreground and background are similar.

关 键 词：图像处理 纹理抠图 随机投影 代价滤波 迭代平滑 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]