融合显著性分析与图割的姿态无关服装区域分割算法  

Pose-Invariant Clothing Segmentation Based on Saliency Detection and Graph Cuts

作　　者：黄磊[1,2] 夏添[1] 张勇东[1] 林守勋[1] 

机构地区：[1]中国科学院计算技术研究所前瞻研究中心,北京100190 [2]中国科学院研究生院,北京100049

出　　处：《计算机辅助设计与图形学学报》2012年第5期620-627,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"九七三"重点基础研究发展计划项目(2007CB311100);国家"八六三"高技术研究发展计划(2009AA01A403);国家自然科学基金项目(61100087);北京市自然科学基金项目(4112055);北京市科技新星计划项目(2007B071);北京市教育委员会共建项目专项资助

摘　　要：提出了一种针对多姿态人的服装区域分割算法,通过融合显著性分析和图割方法有效地提高了服装区域分割的性能.首先,提出一种基于滑动窗口的视觉显著性区域分析方法,计算前景?背景种子区域初始定位,实现种子区域定位的姿态无关性;然后,通过基于图的分割方法对初始种子区域进行矫正;最后,通过将种子区域作为输入的迭代图割方法——GrabCut获得服装区域分割.实验结果表明,文中算法具有较好的分割性能,具有应用前景.A novel algorithm of clothing segmentation,which can handle variable human poses,is proposed in this paper.The segmentation performance is significantly improved by combining saliency detection and graph cuts methods.Firstly,we propose a sliding window based visual saliency detection method to locate initial foreground background seed regions,which are invariant to poses.Secondly,the seed regions are further adjusted by a graph based segmentation method.Finally,the segmentation results of clothing regions are acquired by an iterated graph cuts method called GrabCut,with the seed regions as the input.The experimental results demonstrate that our proposed algorithm can achieve promising results and is feasible in practical applications.

关 键 词：服装区域分割 显著性分析 GRABCUT 姿态无关性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]