基于非线性迭代PLS的人脸识别算法

Nonlinear iterative PLS for face recognition

主成分分析(PCA)是一种无监督的线性降维方法,能有效地提取模式的类内特征,当样本之间出现高度相关性或多重相关性时,PCA提取的主成分解释能力不够。鉴于PCA的缺点,采用一种有监督的鉴别特征提取法——偏最小二乘(PLS),在保留输入变量的最大信息条件下,先在输入和输出变量组中建立模型,再用非线性迭代法提取类间特征,直至隐变量收敛。在ORL人脸库和Yale人脸库中实验结果表明,该算法具有有效性。

Principal Component Analysis （PCA）, a conventional dimension reduction method based on unsuper- vised learning, extracts components effectively, irrespective of the class information. As the samples have a high-correlation or multi-correlation, PCA method is invalid. Nonlinear iterative Partial Least Square （PLS） using supervised learning is proposed, wherein a technique for modeling a relationship between a set of input variables and output variables, while maintaining most of the information in the input variables, and then extract the discrimi- nation of classes features, until convergence of latent vector. Experimental results on the ORL face database and the Yale face database demonstrate the effectiveness of the presented scheme.