直推式多视图协同分割  

Transductive Co-segmentation of Multi-view Images

作　　者：朱云峰[1] 章毓晋[2] 

机构地区：[1]清华大学信息科学与技术国家实验室,北京100084 [2]清华大学电子工程系,北京100084

出　　处：《电子与信息学报》2011年第4期763-768,共6页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(60872084)资助课题

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。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 segmentation and qusi-sparse 3D points segmentation,they are used to combine the sub problems with a global label consistent constrains.Finally,the experiments in several images show the error rate of ITC-Seg is 3.4%,discussions and future improvements of the method are also given.

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 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]