摘要
谱聚类算法在聚类过程中要计算样本相似度矩阵,构造数据量大,并且要对拉普拉斯矩阵进行特征分解,计算比较耗时。Nystrm扩展方法通过部分采样数据来逼近原始特征空间,可以有效降低谱聚类算法的计算复杂度。采样点的选择是决定Nystrm扩展方法精度的重要因素,通过对Nystrm扩展方法的误差进行分析,结合图像特征信息,设计了一种新的采样方案。利用均匀采样方法对图像进行初步采样,并通过迭代的方法最小化采样点与像素点之间的误差,得到最终采样点特征值。通过在Berkeley图库上的图像分割实验表明了算法的可行性和有效性。
Spectral clustering is based on the similarity of data, but the similarity matrix is complex, and the calculation process of Laplacian characteristic decomposition is very time-consuming.The Nystrom extension method approximates the original feature space by sampling,which reduces the computational complexity of spectral clustering effectively. A new sampling method is presented in this paper,which is based on the features of image and Nystr^m error analysis. First, uniform sampling is used to generate a set of cluster centers;then, the error between data and centers is minimized by iteration.Finally, experiments verify the feasibility and effectiveness of the method.
作者
刘仲民
李博皓
李战明
胡文瑾
LIU Zhong-min LI Bo-hao LI Zhan-ming HU Wen-jin(College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou Gansu 730050, China College of Mathematic and Information Technology, Northwest University for Nationalities, Lanzhou Gansu 730000, China)
出处
《无线电工程》
2017年第4期20-23,共4页
Radio Engineering
基金
国家自然科学基金资助项目(64561042)
