空间一致性约束谱聚类算法用于图像分割  

IMAGE SEGMENTATION BY SPECTRAL CLUSTERING ALGORITHM WITH SPATIAL COHERENCE CONSTRAINTS

作　　者：贾建华[1] 焦李成[1] 

机构地区：[1]西安电子科技大学智能信息处理研究所和智能感知与图像理解教育部重点实验室,陕西西安710071

出　　处：《红外与毫米波学报》2010年第1期69-74,共6页Journal of Infrared and Millimeter Waves

基　　金：国家"863"计划(2008AA01Z125;2009AA12Z210);高等学校科技创新工程重大项目培育资金项目(706053);国家部委科技项目资助项目(XADZ2008159;51307040103);国家自然科学基金(60803097);教育部重点项目(108115);教育部长江学者和创新团队支持计划(IRT0645)

摘　　要：近来出现的谱聚类算法在模式识别和图像分割中得到了广泛应用.与传统的聚类算法相比,谱聚类算法能在任意形状的样本空间上聚类且收敛于全局最优解.本研究从谱聚类和权核K-均值的等价性出发,基于图像的空间一致特性,提出了一种基于空间约束特性的谱聚类算法.该算法通过对加权核K-均值的目标函数加上空间一致约束项,利用近似逼近将目标函数最小化与谱聚类算法等价起来.仿真实验表明,此算法在图像分割中取得了比原始谱聚类算法更好的分割效果.Recently, spectral clustering algorithm has a wide application in pattern recognition and image segmentation. Compared with traditional clustering methods, it can cluster samples in any form feature space and has a global optimal solution. By starting from the equivalence between the spectral clustering and weighted kernel K-means, a spectral clustering algorithm with spatial coherence property of images was proposed. By adding a term of spatial constraints to the objective function of weighted kernel K-means, the algorithm made the minimization of objective function be equivalent to the spectral clustering approximatly. Experimental results show that our proposed algorithm outperforms the traditional spectral clustering in image segmentation.

关 键 词：谱聚类 权核K-均值 空间一致特性 图像分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]