摘要
提出一种基于多尺度张量类标子空间的人脸特征提取算法,提高人脸识别对光照的鲁棒性,同时不破坏原始数据固有的高阶结构和数据之间的相关性。采用多尺度小波变换组建人脸三维张量样本,将三维人脸张量空间投影到低维张量子空间,对高维人脸进行降维和特征提取,应用多线性主成分类标算法对样本进行类标号,同时使用最近邻算法完成人脸识别。利用CAS-PEAL-R1东方人脸库进行评测,实验结果表明,该识别算法比经典的主成分分析、线性判别分析和多尺度Gabor识别算法具有更好的识别效果。
The paper proposes a face feature extraction algorithm based on multi-scale tensor class label subspace, which improves the robustness of the light in face recognition without damaging the inherent higher order structure and the correlation between the original data. Multi-scale wavelet transform were used to form 3D face tensor sam pies, which were then projected onto a low dimensional tensor subspace for dimensionality reduction and feature ex traction. The multiple linear principal component class-label algorithm was proposed to label the samples and the nearest neighbor algorithm was utilized to complete face recognition. CAS-PEAL-R1 oriental face database was used for evaluation. The experimental results show that this recognition algorithm has better recognition results than classical recognition algorithms （principal component analysis,linear discriminant analysis, multi-scale Gabor recognition algorithm） ,and has better feasibility.
出处
《山东科技大学学报（自然科学版）》
CAS
2015年第4期55-61,共7页
Journal of Shandong University of Science and Technology(Natural Science)
基金
国家自然科学基金项目(61462042
61462045)
同济大学嵌入式系统与服务计算教育部重点实验室开放基金项目
关键词
人脸识别
多尺度变换
张量子空间
多线性主成分分析
类标
face recognition
muhi-scale transform
tensor subspace
multiple linear principal component analysis
class-label
