ABSTRACT
Human face images are usually represented by thousands of pixels encoded in high-dimensional array; however, they are intrinsically embedded in a very low-dimensional subspace. The use of subspace for representation of face images helps to reduce the so-called curse of dimensionality in subsequent classification. Dimension reduction of a dataset can be achieved by extracting the features, provided that the new features contain most of the information of the given dataset. To put it in a different way, the dimensionality reduction can be done if the mean squared error (MSE) or the sum of variances of the elements (which are going to be eliminated), are minimum. Subspace-based methods are such techniques of dimensionality reduction. Further the subspace representation helps in suppressing the variations of lighting conditions and facial expressions. Two of the most widely used subspace methods for face detection and face recognition are the principal component analysis (PCA) [95] and the Fishers linear discriminant analysis (LDA), though there are dozen of dimension reduction algorithms are available for selecting effective subspaces for the representation of face images [63].
