基于分块集成的图像聚类算法  

Block Integration Based Image Clustering Algorithm

作　　者：刘淑君 魏莱[1] LIU Shu-jun;WEI Lai(School of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《计算机科学》2020年第6期170-175,共6页Computer Science

基　　金：国家自然科学基金(61203240);上海市科研创新项目(14YZ102)。

摘　　要：基于谱聚类的子空间聚类算法已经显示出良好的效果,但是传统的子空间聚类算法需要将图像进行向量化处理,而这种向量化会导致图像本身携带的二维结构信息的丢失。为了减少这种信息的丢失,文中提出了基于分块集成的图像聚类算法(Block Integration Based Image Clustering,BI-CI)。首先,将图像数据分为若干矩阵块;然后,利用核范数矩阵回归构造基于某一矩阵块的系数矩阵,同时提出了一种依据矩阵块秩信息设定各个矩阵块的权重方法;最后,通过每一系数矩阵及其所对应矩阵块的权重,得到整体系数矩阵。在此系数矩阵上,利用谱聚类算法得到最终的聚类结果。在4个图像数据集上的实验表明,相比现有算法,所提算法具有更强的鲁棒性,可以获得更优的聚类效果。Spectral based subspace clustering algorithms have shown good results.But the traditional subspace clustering algorithms need to vectorize the image,which will lead to the losses of the two-dimensional structure informations carried by the ima-ge itself.In order to reduce the losses,block integration based image clustering(BI-CI)algorithm is proposed.First,the images are divided into several matrix blocks.Then,the nuclear norm based matrix regression is used to get the coefficient matrix of one block,and a method is proposed to set the weight for each matrix block according to the rank information of matrix blocks.Finally,based on each coefficient matrix and according to the rank of the corresponding matrix block,the integral coefficient matrix is obtained.The final clustering results are obtained by using spectral clustering performed on the coefficient matrix.Experimental results show that the proposed metho d is more robust than the existing algorithms and can achieve more accurate clustering results.

关 键 词：子空间聚类 矩阵块 核范数 矩阵回归 秩 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]