基于奇异值分解的人脸识别方法

Singular value decomposition-based approach for face recognition

提出了一种将傅里叶变换和奇异值分解相结合的人脸自动识别方法。首先对人脸图像进行傅里叶变换,得到其具有位移不变特性的振幅谱表征。其次,从所有训练图像样本的振幅谱表征中给定标准脸并对其进行奇异值分解,求出标准特征矩阵,再将人脸的振幅谱表征投影到标准特征矩阵后得到的投影系数作为该人脸的模式特征。然后,对经典的最近邻分类器算法进行了改进,并采用模式特征之间的欧式距离作为相似性度量,从而完成对未知人脸的识别。采用ORL(OlivettiResearchLaboratory)人脸库对本文提出的人脸识别方法进行验证,获得了100.00%的识别率。实验结果表明,本方法优于现有的基于奇异值分解的人脸识别方法,且对表情、姿态变换等具有一定的鲁棒性。

A method to extract algebraic features of a face image based on the Fourier transform and Singular Value Decomposition (SVD) is introduced, then the method with the algebraic feature is proposed to recognize faces. First, face images are processed by a 2D Fourier transform that has some effective properties such as a linear transform, and is invariant against spatial translation. The amplitudes of the transform coefficients are used to represent the image in the frequency domain. Second, the amplitude representation of the face image is projected onto the two compressed orthogonal matrixes, which come from the SVD of the standard face image obtained by averaging all training samples and then the projecting coefficients are used as the algebraic feature of the face image. The robustness of this feature is proved and used for face recognition. In the matching stage, the traditional Nearest Neighbor Classifier (NNC) is improved to recognize the unknown faces by using Euclidean distance as the similarity measurement. Finally, the standard face database from Olivetti Research Laboratory (ORL) is selected to evaluate the recognition accuracy of the proposed face recognition algorithm. This database includes face images with different expressions, small occlusions, different illumination conditions and different poses, etc. The recognition accuracy is up to 100.00% by selecting appropriate values of the parameters. The effectiveness of the proposed face recognition algorithm and its insensitivity to the facial expression, illumination and posture are shown in terms of both the absolute performance indices and the comparative performance against some popular face recognition schemes such as Singular Value decomposition-based method.