核稀疏概念编码算法及在图像表示中的应用  

Kernel sparse concept coding algorithm and its application for image representation

作　　者：舒振球[1] 赵春霞[2] 

机构地区：[1]江苏理工学院计算机工程学院,常州213001 [2]南京理工大学计算机科学与工程学院,南京210094

出　　处：《系统工程理论与实践》2016年第5期1331-1339,共9页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(61472166;61503195;61302124;11274091)~~

摘　　要：稀疏编码算法是一种常用的图像数据表示方法.为了处理高度非线性分布的数据,文中提出了一种核稀疏概念编码算法,并应用于图像表示.该算法首先对邻域图进行谱分析,提取数据的几何流形结构信息;然后将原始特征空间数据映射到高维特征空间中,利用谱回归在高维特征空间中来计算基向量;最后在高维特征空间中对每个样本逐个进行表示.文中算法不仅能有效地处理非线性结构数据,而且只需求解一个稀疏特征值问题和两个回归问题,计算简单有效.在Yale、ORL和PIE图像库的聚类实验表明,文中算法的准确率和归一化互信息均优于其它几种对比算法.Sparse coding algorithm is a popular data representation method. In order to deal with the high nonlinear data, in this paper, a kernel sparse concept coding （KSCC） algorithm is proposed for image representation. Our algorithm performs spectral analysis on nearest neighbor graph and captures the geometric manifold structure of the data. Then the data in the origin feature space is mapped into the high-dimensional feature space and the basis vector in high-dimensional space is obtained using spectral regression. Finally, the samples are individually represented in high-dimensional feature space. Therefore, the proposed algorithm not only effectively handles the nonlinear structure data, but also needs to solve a sparse eigen-problem and two regression problems, which is very simple and effective. The experiments on Yale, ORL and PIE image datasets demonstrate that the accuracy and normalized mutual information of our proposed algorithm are superior to other comparison algorithms.

关 键 词：基向量 数据表示 核 非线性 稀疏编码 谱回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]