基于非负矩阵分解的类标签映射集成学习  

Class Labels Mapping Ensemble Learning for Nonnegative Matrix Factorization

作　　者：高亮[1] 潘积远[2] 于佳平[3] 

机构地区：[1]西安理工大学自动化与信息工程学院,西安710048 [2]中国电子科技集团公司第二十研究所数据链技术重点实验室,西安710068 [3]东华大学理学院,上海201620

出　　处：《科学技术与工程》2018年第1期59-65,共7页Science Technology and Engineering

基　　金：国家自然科学基金(11501097)资助

摘　　要：非负矩阵分解已广泛应用于人脸识别,但因无监督、子空间线性表示、基特征局部次优等特点,它识别光照复杂、表情丰富的人脸图像的能力有限。为优化非负矩阵分解的人脸识别能力,分析并建立了非负矩阵分解的集成分类框架,整合多组基特征的弱类别结构信息,在无监督情形下利用偏最小二乘回归建立符合统计属性的集成标签映射,突显正确的类结构。通过多组人脸数据集的试验结果表明,基于非负矩阵分解的集成分类能力显著提高,适用光照复杂、表情丰富的人脸图像识别。Nonnegative matrix factorization( NMF) has been used for face recognition successfully. But,because of the facts on unsupervised model,low dimensional subspace representation and sub-optimal basis features,it is not good at illumination and expression face recognition. To enhance its face recognition ability by additional category,ensemble face recognition framework based on NMF is built,which is called ECNMF for short. ECNMF integrates the weak category structure information,and establishes an ensemble label mapping based on partial least squares regression to manifest correct class. The results on several face datasets show that the recognition rate of ECNMF is the best compared with some unsupervised NMF models on illumination and expression face recognition.

关 键 词：非负矩阵分解 人脸识别 集成分类 偏最小二乘回归 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]