Hybrid Face Recognition with Locally Discriminating Projection

Abstract

The face recognition task involves extraction of unique features from the human face. Manifold learning methods are proposed to project the original data into a lower dimensional feature space by preserving the local neighborhood structure. LPP should be seen as an alternative to Principal Component Analysis (PCA). When the high dimensional data lies on a low dimensional manifold embedded in the ambient space, the Locality Preserving Projections are obtained by doing the optimal linear approximations to the Eigen functions of the Laplace Beltrami operator on the manifold. However, LPP is an unsupervised feature extraction method because it considers only class information. LDP is the recently proposed feature extraction method different from PCA and LDA, which aims to preserve the global Euclidean structure, LDP is the extension of LPP, which seeks to preserve the intrinsic geometry structure by learning a locality preserving submanifold. LDP is a supervised feature extraction method because it considers both class and label information. LDP performs much better than the other feature extraction methods such as PCA and Laplacian faces. In this paper LDP along with Wavelet features is proposed to enhance the class structure of the data with local and directional information. In this paper, the face Image is decomposed into different subbands using the discrete wavelet transform bior3.7, and the subbands which contain the discriminatory information are used for the feature extraction with LDP. In general the size of the face database is too high and it needs more memory and needs more time for training so that to improve time and space complexities there is a need for dimensionality reduction. It is achieved by using both biorthogonal wavelet transform and LDP the features extracted take less space and take low time for training. Experimental results on the ORL face Database suggests that LDP with DWT provides better representation and achieves lower error rates than LDP with out wavelets and has lower time complexity. The subband faces performs much better than the original image in the presence of variations in lighting, and expression and pose. This is because the subbands which contain discriminatory information for face recognition are selected for face representation and others are discarded.