摘要
基于核的Fisher线性鉴别分析(KFDA)已成为抽取非线性特征的最有效方法之一。但是,针对必然面临的奇异性问题,如何抽取非线性最优鉴别特征还没有得到很好的解决。基于同构映射的思想,我们提出了一种最优的核Fisher鉴别分析(OKFDA)方法,从理论上巧妙的解决了奇异情况下最优鉴别矢量集的求解问题。在FERET人脸库的子库上的实验结果验证了OKFDA方法的有效性。
It is well known that kernel-based Fisher linear discriminant analysis (KFDA) has become one of the most effective techniques for nonlinear feature abstraction. But, in this method, the kernel within-class scatter matrix is always singular, and thus how to extract the optimal nonlinear Fisher discriminant features remains unsolved. Based on the idea of isomorphic mapping, we proposed an optimal kernel Fisher discriminant analysis (OKFDA), from which we acquire a general algorithm for the computation of the optimal discriminant vectors in the singular cases. Experimental results on a subset of the FERET face image database indicate that the OKFDA is valid.
出处
《系统仿真学报》
EI
CAS
CSCD
2004年第12期2864-2868,共5页
Journal of System Simulation
基金
国家自然科学基金资助项目(60072034)
关键词
核FISHER鉴别分析
最优鉴别矢量集
特征抽取
人脸识别
kernel Fisher discriminant analysis
optimal discriminant vectors
feature extraction
face recognition
