Research on Feature Extraction Methods and Its Application to Face Recognition
|School||Nanjing University of Technology and Engineering|
|Course||Applied Computer Technology|
|Keywords||face recognition feature extraction Fisher linear discriminant analysis (LDA or FLD) fuzzy set theory feature fusion minimal criterion discriminant analysis two dimensional projection analysis maximum scatter difference criterion kernel methods manifold learning|
Feature extraction is an important research branch of pattern recognition field; it is the key to solve the face recognition problems. In this paper, we carried through some research to the current leading algorithms of feature extraction, and some more effective feature extraction methods are proposed. And, these methods are verified to be effective in the application of face recognition.Statistical independent component (ICA) is a generalization of PCA, which separates the high-order moments in addition to the second-order moments. But they are all the best reconstruction of the original data, not the optimal discriminant information. LDA can extract the optimal discriminant features. The classical LDA is based on binary classification criterion. While in face recognition, distribution of face samples are often affected by illumination, pose, and expression et al. Thus, fuzzy technology is introduced, and the within-class scatter matrix and between-class scatter matrix of the classical LDA method are redefined by using fuzzy membership degree, named fuzzy LDA. In order to incorporate class specific information into ICA, we employed fuzzy LDA and proposed a novel method, which is based on ICA combined with fuzzy LDA. Experimental results on AR, ORL, and NUST603 demonstrate the effectiveness of the proposed method.In face recognition, FLDA always encounter high dimensionality and small sample size problem, which usually leads to singularity of the within-class scatter matrix, which is a trouble for calculation of Fisher optimal discriminant vectors. Though, which can be avoided by scatter difference criterion, but they are all based on vectors. In this paper, scatter difference criterion based on vectors is generalized, and scatter difference criterion based on straightforward image projection techniques is proposed. The proposed one combines the ideas of two-dimensional principal component analysis and two-dimensional maximum scatter difference and which can directly extracts the optimal projective vectors from 2D image matrices rather than image vectors based on the scatter difference criterion. In addition, analyzes the generation matrix of the Generalized PCA (GPCA), and redefining the between-class scatter matrix by introducing a radical basis function, so classify features can obtained by adjusting the coefficient of the function. In particular, GPCA is a simplified version of the proposed method.Fisher minimal linear discriminant (FMLD) criterion avoids the singular problem of the within-class scatter matrix. ICA is a generalization of PCA, which separates the high-order moments in addition to the second-order moments. Inspired by this idea, in order to incorporate class specific information into ICA, we employed FMLD and proposed a novel method, named minimal linear discriminant analysis based on ICA. As the distribution of face samples are often affected by environment, and the kernel method is an effective strategy to solve nonlinear problem. In this paper, the Fisher minimal linear discriminant analysis is generalized to kernel Fisher minimal linear discriminant analysis. The experimental results show the effectiveness of the proposed one.Non-locality Preserving Projection (NLPP) is a linear manifold learning algorithm. This paper analyzes the modeling mechnism of LPP and NLPP, and generalized to kernel space, proposed kernel-based Non-locality preserving projection method. The experiments conducted on Yale, and ORL face image databases, and the experimental results show that the proposed one is effective.