基于IRAM和半监督的谱聚类图像分割  

Image segmentation by spectral clustering based on IRAM and semi-supervised

作　　者：龚文文 葛玉荣[1] 

机构地区：[1]中国海洋大学信息科学与工程学院,山东青岛266100

出　　处：《电子设计工程》2016年第17期156-159,共4页Electronic Design Engineering

基　　金：国家国际科技合作专项(2012DFG22080)

摘　　要：谱聚类判别割(Dcut)计算正则化相似度矩阵及其特征向量比较耗时,对于大规模矩阵特征值问题,隐式重启动Arnoldi方法(IRAM)能够快速收敛到模值最大的k个特征值(即主导特征值)。因此本文采用IRAM算法计算相似度矩阵的主导特征值,减小运算时间。为解决谱聚类敏感于尺度参数的问题,提出利用少量监督样本求取每幅图像特有的尺度参数,进行半监督图像分割。通过对UCI数据集和自然图像的仿真实验表明,本文算法能有效提高Dcut的运算速度,同时性能稳定,分割结果优于Dcut和PCA-Dcut。Spectral clustering discriminant cut（Dcut） is time consuming to calculate the normalized similarity matrix and its eigenvectors. For large scale matrix eigenvalue problem, implicitly restarted Arnoldi method（IRAM） is easy to converge to the k largest magnitude eigenvalues（i.e., dominant eigenvalues）. Therefore, the paper uses the IRAM algorithm to calculate the dominant eigenvalues of similarity matrix to reduce runtime. To solve the problem of spectral clustering sensitive to the scale parameter, the paper proposes to calculate the scale parameter unique to each sample set by using a few supervised samples,and do semi-supervised image segmentation. The experiments on UCI data set and natural images show that the algorithm of this paper can effectively improve the running speed of Dcut with stable performance, and the results are better than Dcut and PCA- Dcut.

关 键 词：判别割 隐式重启动Arnoldi 尺度参数 半监督 图像分割 

分 类 号：TN911.73[电子电信—通信与信息系统]