Local Sub-pattern Feature Extraction For Face Recognition

As a key step of face recognition, feature extraction was and remains to be a hot topic in pattern recognition research. It is a process that extracts a new set of features of interest from the original data through some mapping. This thesis mainly focuses on Principal Component Analysis (PCA) based methods. Based on the previous research work, Local Sub-pattern PCA (L-SpPCA) and Sub-pattern Local Preserved projection (SpLPP) are proposed respectively and successfully applied to the task of face recognition.The traditional principal component analysis (PCA) is a very effective approach of extracting features and has successfully been applied in pattern recognition such as face classification. As an extension to PCA, Local PCA (LPCA) develops a local linear approach to dimensionality reduction that provides relatively accurate representation for the original data. However, LPCA is not so effective under different facial expression and illumination conditions due to only utilizing the global information of face images. Our previously-proposed sub-pattern PCA (SpPCA) operates directly on a set of partitioned subpatterns of the original pattern rather than the original pattern itself and acquires a set of projection sub-vectors for each partition to extract corresponding local sub-features and then synthesizes them into global features for subsequent classification. Although proved globally nonlinear, SpPCA still lacks the sufficient ability to describe complex sample distribution. In this paper, a novel L-SpPCA is proposed which benefits from both LPCA and SpPCA via inheriting the robustness of SpPCA as well as combining SpPCA with nonlinearity. Experiments on AR, ORL and a set of partly-missing faces show that the proposed method is both robust and effective. Further through incorporating a manifold formulation for face sub-pattern into SpPCA, a relatively simpler sub-pattern locality preserving projection (SpLPP) is proposed with aiming to avoid relatively complicated implementation of L-SpPCA without sacrificing its merits. Experiments show that the proposed methods are both robust and effective.