基于多标签核判别分析的人脸身份与性别识别方法

Face and gender recognition method based on multi-label kernel discriminant analysis

为解决多标签线性判别分析(MLDA)方法在非线性维数约简方面的局限性,提出了一种多标签核判别分析(MKDA)方法,并将其用于人脸的身份与性别识别中.该方法的基本思想是通过非线性映射将训练样本从输入空间映射到高维核特征空间中,并在该特征空间中进行基于MLDA的数据降维.在身份和性别识别中,首先采用MKDA方法对人脸图像特征向量进行降维,获取判别特征矢量集;其次,为每幅人脸图像赋予一个表征身份和性别的多标签类别矢量;最后,采用减秩回归模型(RRR)描述判别特征矢量与多标签类别矢量之间的回归关系,并利用该模型进行未知人脸的身份和性别识别.AR人脸数据库上的实验结果表明:在人脸身份和性别识别中,MKDA方法的识别率高于传统核判别分析(KDA)方法.

A multi-label kernel discriminant analysis（MKDA）method is proposed to overcome the limitation of multi-label linear discriminant analysis （MLDA）on nonlinear dimensionality reduction, and applied to the recognition of face and gender.The basic idea of the MKDA method is to map the training data samples from the input data space to a high-dimensional kernel feature space via a non-linear mapping and then to perform data reduction based on the MLDA method in the feature space. During the recognition of face and gender,the dimensionality of the face image feature vectors is firstly reduced by using the MKDA method and a set of discriminative feature vector set is obtained. Then,a multi-label class vector indicating the class membership of face and gender is assigned to each face image.Finally,a reduced-rank regression （RRR）model is built to describe the relationship between the discriminative facial feature vectors and multi-label class vectors,and is applied to the face and gender recognition of an unknown face image.The experimental results on AR face database show that the recognition rates of the MKDA method are higher than those of the traditional kernel discriminant analysis （KDA）in face and gender recognition.