融合LBP和表观流形鉴别分析的人脸识别算法

Fusion of LBP and Appearance Manifold Discriminant Analysis for Face Recognition

流形学习方法可以有效地发现存在于高维图像空间的低维子流形,但是流形学习是一种非监督学习方法,其鉴别能力反而不如传统的维数约简方法,且对人脸图像的光照、姿态等局部变化敏感.针对这两个问题,本文提出一种基于人脸表观流形鉴别分析的识别方法,该方法利用局部二元模式(Local binary pattern,LBP)对人脸图像进行局部特征描述,提取对局部变化不敏感的特征,然后使用有监督的核局部线性嵌入算法(Supervised kernel local linear embedding,SKLLE)对由局部特征构造的全局特征进行维数约简,提取低维鉴别流形特征进行人脸识别.该方法不仅对局部变化不敏感,而且将人脸表观流形和类别信息进行有效的结合,同时对新样本有较好的泛化性.实验结果表明该算法能有效的提高人脸识别的性能.

Manifold learning method can discover intrinsic low-dimensional submanifold embedded in the high-dimensional image space, but manifold learning is an unsupervised learning method, the discriminative ability of the low-dimensional feature obtained by the algorithm is often lower than those obtained by the conventional dimensionality reduction methods. Furthermore, manifold learning methods are sensitive to the variation of lighting conditions, pose and expression. To address the two prob- lems, this paper introduces a novel appearance manifold discriminant analysis method for face recognition, it first uses LBP operator to obtain the local features of face image, then fuses prior class-label information and nonlinear submanifold of face im- ages to extract discriminative features from the global features which formed bY the local features. This method can not only gains a perfect approximation of face appearance manifold, but also enhances local within-class relations. It also does well on the new samples. Experimental results show that the proposed method can improve face recognition performance effectively.