一种新的基于Gabor特征的零空间人脸识别算法

A Novel Gabor Feature Based Null Space Algorithm for Face Recognition

为了克服人脸识别中的小样本集(small sample size,3S)问题,本文首先利用人脸图像距离模型揭示了线性鉴别分析的物理过程,指出了零空间鉴别分析相对于主元空间鉴别分析可以提取出更有利于人脸识别的鉴别信息.在此基础上,提出了一种有效的基于Gabor特征的零空间人脸识别算法,该算法采用一种新的基于邻域保持的鉴别嵌入准则(neighbor-preserving based discriminant embedding,NDE),并利用Gabor小波变换减轻人脸图像中光照和表情变化等因素的影响.在ORL,FERET和AR等人脸数据库上的实验结果表明本文算法具有较优的人脸识别性能.

Recently, linear discriminant analysis （LDA） has been widely used in the field of face recognition. However, in many real applications, LDA suffers from the small sample size （3S） problem, where training samples are limited so that LDA cannot be directly used. To overcome the 3S problem, in this paper we first reveal the mechanism of LDA to show how it extracts the most discriminative features according to an image distance model, and then identify that the null space based LDA is much more efficient than the principal space based LDA for the extraction of discriminative features. Based on this identification, we propose an effective Gabor feature based null space algorithm for face recognition, which exploits a new neighborhood-preserving based discriminant embedding （NDE） criterion to overcome the drawbacks of the traditional Fisher criterion, and during the process of the extraction of discriminative features, the Gabor wavelet transform is incorporated to further reduce the influences of illumination and expression changes in the face images. Experimental results on several public face databases, such as ORL, FERET and AR, show that the proposed null space based NDE algorithm outperforms the state-of-the-art algorithms, such as LDA, NDP, NDA, and LDE, and it can achieve the encouraging face recognition performance.