图像分割的谱聚类集成算法

A Spectral Clustering Ensemble Algorithm for Image Segmentation

针对谱聚类算法对尺度参数敏感的问题,利用集成学习算法良好的鲁棒性和泛化能力,提出了一种无监督集成学习算法——谱聚类集成算法.该算法先利用谱聚类的内在特性产生集成学习所需的多个聚类个体,再采用Hungarian算法对生成的聚类个体进行重新标记,计算每个样本点关于每一个类别所占的比例,得到一个成分向量,然后运用对数比变换将所得的成分向量映射到另一个空间,去除成分数据的不适定性,最后对映射后的数据进行聚类,从而得到最终的集成结果.通过对UCI数据集和纹理图像的仿真实验表明,所提算法的聚类准确率与常用的共识函数具有一定的可比性,且运算代价较小,所需时间大约为MCLA算法的一半,同时避免了精确选择谱聚类算法的尺度参数.

An unsupervised ensemble learning algorithm,spectral clustering ensemble,is proposed to solve the sensitivity of scaling parameter of spectral clustering.The algorithm utilizes the good robustness and generalization ability of ensemble learning.The property of spectral clustering is exploited to generate the components for integrating learning,and the Hungarian algorithm is used to relabel the resulting component clusterings.Then a compositional data vector is obtained by computing the ratio of each label for each sample.The compositional data vectors are mapped into another space via log contrast transform to solve the ill-posed characteristic of compositional data.The final aggregated results are generated by clustering the mapped data.Experiments on UCI data and texture images show that the proposed algorithm is comparable with some common consensus functions in accuracy,its computational cost is about half of that of MCLA,and it avoids the selection of the accurate parameter in spectral clustering.