融合核主成分分析和最小距离鉴别投影的人脸识别算法

Face Recognition Algorithm Fused Kernel Principal Component Analysis and Minimum Distance Discriminant Projection

针对人脸识别问题,在原有的最小距离鉴别投影算法的基础上,根据核主成分分析(KPCA)方法,提出一种新的融合核主成分分析和最小距离的鉴别投影算法。运用KPCA对高维样本空间进行降维,通过核技巧将样本映射到高维非线性空间,继而在降维后的核子空间上通过鉴别投影方法计算其相应的投影矩阵,采用最近邻分类方法对样本进行分类并最终实现人脸识别。在ORL,FERET和YALE人脸库上的实验结果表明,该算法的识别率优于其他算法,可避免高维矩阵的计算复杂问题,同时定义的核子空间相似度权重也较好地保持了样本之间的近邻关系。

By fusing Kernel Principal Component Analysis（KPCA） and Minimum Distance Projection （MDP）, a new method based on the original minimum distance differential projection is developed to address the face recognition problem. Different from the classical minimum-distance discriminant projection algorithm,the new algorithm is based on dimensionality reduction algorithm. By using KPCA to reduce the dimension of these samples, they are mapped to a high- dimensional nonlinear space. The corresponding projection matrix is calculated in the dimension-reduced nuclear space by differential projection method. The nearest neighbor classification method is adopted to classify these samples and the face recognition is succeeded eventually. Experimental results on the ORL,FERET, and YALE face databases show that the proposed algorithm can outperform other algorithms. It can be avoided by using the nuclear technique to extract the nonlinear feature of the face image, and on the same time, the neighbor relationship of the samples is also kept by the defined similarity weight of the nuclear space.