鉴别局部特征分析及其在人脸识别中的应用  

Discriminant local feature analysis with applications to face recognition

作　　者：杨琼[1] 丁晓青[1] 

机构地区：[1]清华大学电子工程系智能技术与系统国家重点实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2004年第4期530-533,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家"八六三"高技术项目(2001AA114081);国家自然科学基金资组项目(69972024)

摘　　要：由于传统的鉴别主分量分析(DKL)算法中,主分量分析(PCA)基于全局特征,难以提取人脸的局部特性,该文提出鉴别局部特征分析算法。该算法中,局部特征分析(LFA)代替PCA作为线性鉴别分析(LDA)的前端。一方面,LFA在保留大部分全局信息的同时提取局部特征。另一方面,它为信号提供一种有效的低维表示,增强LDA在小样本问题中的数值稳定和推广性能。文中结合开集模式的人脸认证领域,在PoliceFace、OCRLab人脸库和它们的组合库上对新算法和DKL算法进行实验比较。实验表明,通过结合LFA和LDA,新算法明显降低认证错误率:在PoliceFace库上,等错误点错误率降低43.10%;在OCRLab库上错误率降低25.87%;在组合库上错误率降低33.16%。Principal component analysis (PCA) was performed before linear discriminant analysis (LDA) in the traditional discriminant karhunen loeve (DKL) algorithm. However, principal component analysis is based on global information, which ignores significant local characteristics. This paper proposes discriminant local feature analysis (DLFA) algorithm which uses local feature analysis (LFA) instead of PCA before the LDA. LFA captures the local characteristics with little loss of global information and it provides a low-dimensional representation of the signals, which reduces the dimensionality for LDA. The new algorithm was compared with the DKL algorithm on three databases for open-set face verification. The tests demonstrated that by combining LFA and LDA, the DLFA algorithm outperforms the DKL algorithm, reducing the error rate by 43.10% on the PoliceFace database, by 25.87% on the OCRLab database and by 33.16% on the combined database.

关 键 词：模式识别 人脸认证 局部特征 线性鉴别 主分量分析 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]