直推式多视图协同分割transductiveco-电子与信息学报.pdfVIP

第33 卷第4 期 电 子 与 信 息 学 报 Vol.33No.4 2011 年4 月 Journal of Electronics Information Technology Apr. 2011 直推式多视图协同分割 朱云峰* 章毓晋 (清华大学信息科学与技术国家实验室 北京 100084) (清华大学电子工程系 北京 100084) 摘 要：多视角拍摄条件下获取到含有同一刚性或静态目标多幅图像后，快速准确地在所有图像中分割出该目标是 一个尚未被关注的问题。该文首次给出这个问题的数学描述，并借助图模型的描述方法，推导了它与传统的基于图 割分割问题之间的关系。在求解该问题时，该文提出了迭代式协同直推优化算法，算法思想是将优化目标函数拆分 为图像空间和3 维空间分割两个子优化问题，利用图割和谱聚类分割的方法迭代求解两个子优化问题，并在迭代过 程中设计传播、过滤、投票处理引入两个空间标签一致性约束，算法收敛时同时得到图像空间和3 维空间的分割结 果。最后，使用实际图像序列实验得出算法的平均误判率为3.4%，比较讨论和改进思路一并给出。 关键词：图像处理；直推式学习；图割；谱聚类分割；多视角协同分割 中图分类号：TP391.41 文献标识码： A 文章编号：1009-5896(2011)04-0763-06 DOI: 10.3724/SP.J.1146.2010.00839 Transductive Co-segmentation of Multi-view Images Zhu Yun-feng Zhang Yun-jin (Tsinghua National Laboratory for Information Science and Technology, Tsinghua University, Beijing 100084, China) (Department of Electronic Engineering, Tsinghua University, Beijing 100084, China) Abstract: Fast and efficient segmentation of rigid or stable object in multi-view images is still a unsolved problem. In this paper, the problem is formulated and the relationships between it and traditional min-cut based segmentation problems are also deduced with Graph representation. To minimize the energy function, a novel algorithm named Interactive Transductive Co-Segmentation (ITC-Seg) is proposed. In ITC-Seg, the function is divided into two sub problems which are solved with graph cuts and spectral segmentation methods. Moreover, propagation, filter, voting methods are introduced into the iteration between multi-view images segm