Algorithm Research On Subspace Face Recognition

Face recognition is an important and difficult area in pattern recognition and image processing. Subspace face recognition has well developed as an efficient and popular method, which is qualified with low time consumption, perfect description and good separation. Face image dimensionality is often much larger than sample size, and then within-class scatter matrix becomes singular. However, inverse matrix of within-class scatter matrix needs to calculate to fulfill Fisher criterion which maximizes between-class distance and minimizes within-class distance. Such a failure of computing inverse matrix of within-class scatter matrix is the so-called small sample size problem. In this essay, we carry our researches on subspace face recognition. A serial of algorithms are proposed to overcome small sample size problem on Fisher criterion. The following gives the detailed contents:Two improved methods are proposed in this essay. The first is a modification for direct linear discriminant analysis. Complement subspace of within-class scatter matrix is used for a weight coefficient calculation, which is combined into similarity computation. The second is an algorithm based on generalized singular value decomposition. A subspace corresponding to trivial singular value of within-class scatter matrix and non-trivial singular value of between-class scatter matrix.A concept of intrinsic subspace is proposed for face space separation. The total face space is divided into four parts: pure between-class subspace, pure within-class subspace, common subspace and null subspace of two scatter matrixes. Utilize different methods in different intrinsic subspaces and make full use of different discriminatory features of intrinsic subspaces. In such a framework, we first analysis several classical recognition methods and estimate their recognition rates, experimental results prove our estimation. Face recognition methods based on intrinsic subspaces are also proposed. The simulated results tell us that recognition rates of intrinsic subspace methods are better than other methods and discriminatory information compactness of projective vectors is the highest.At last, kernel functions are used to extend our intrinsic subspace methods to nonlinear space. Polynomial kernel function and Gaussian RBF kernel with different parameters are used to map the original face image space to high dimensional feature space. The simulated result turns that kernel subspace separation methods are better than linear subspace separation on recognition rate.