近邻自适应局部尺度的谱聚类算法

提出一种在LUV空间中基于多层次化结构Nystrm方法的自适应谱聚类算法。首先引入LUV色彩空间,避免了RGB色彩空间中色彩辨别阈对分割的影响,在纹理、边缘区域取得了更好的分割效果;其次将谱聚类算法中基于多层次化结构的方法和基于Nystrm采样的方法结合起来,有效减少了运算时间、解决了数据量较大时计算过程中内存溢出的问题;最后在K均值聚类中通过对特征间隙(eigengap)的分析,自适应地选择K值的大小,解决了自动确定聚类数目的问题。将提出的方法在LUV色彩空间中和RGB色彩空间中分别进行图像分割实验,结果表明在LUV色彩空间中取得效果更加理想。同时也将提出的算法与基于Nystrm方法的谱聚类算法(spectral clustering-Nystrm,SC-N)进行比较。实验结果表明,该算法在数据运算量、运行时间和分割结果上都优于SC-N方法。

Spectral clustering based on neighboring adaptive local scale

Considering the performance of traditional spectral clustering using Gaussian kernels,a new spectral clustering based on neighboring adaptive local scale is presented in this paper.Based on clustering consistency characteristics,the proposed method first emphasizes the flexibility of the local scale,which means each sample has a corresponding scale parameter.Furthermore,it overcomes the limitations of traditional methods in all samples with the same global scale parameter.Hence,it can depict the intrinsic structure of data sets better.Second,it stresses the convenience of parameter selection.It can determine the value of a local scale for one sample by computing the sum of weighted distances of N neighbors.Therefore,it can determine the scale parameter automatically.This paper illustrates the proposed algorithm not only has inhibition for certain outliers but is able to cluster the data sets with different scales.Finally,experiments on both,artificial data and UCI data sets,show that the proposed method is effective.