The Research On Face Feature Extraction Based On Manifold Learning

Feature extraction is one of the most basic researches in face recognition. The essence of it is to reduce a high-dimensional original sample space to a low-dimensional subspace that is benefit to analysis. There are many classical algorithms like principal component analysis(PCA), Locality Preserving Projections(LPP), Marginal Fisher Analysis(MFA), and subspace method which is widely used for its simple calculation and effective. And manifold learning is usually adopted by subspace feature extraction methods to explore intrinsic structure of data.According to the analysis and summary of feature extraction technology research status and development trend, the manifold learning method for feature extraction was researched thoroughly and deeply. In this thesis, three kinds of feature extraction algorithm were put forward.Firstly, Discriminant Neighborhood Preserving based on NPE was proposed, and Orthonormalized. DNP considers both the between-class scatter and the within-class scatter, and keeps geometrical relationship and the distance measurement among data unchanged. This algorithm is benefit to classification and highly improves the precision and efficiency of recognition. Secondly, Locality Margin Criterion(LMC) was proposed based on locality preserving projections. When Adjacency graph is constructing, only treat a part of the points in the same class as neighborhood points, and only a part of the points in other classes as non-neighborhood points. in this way each sample point can be made close to the points of the same class, and away from the points of different classes, that can form highly efficient clustering, while also reducing the algorithm computational complexity. Finally, Uncorrelated Multiple Information Projection(UMIP) was proposed based on locality preserving projections. With this method the mapping obtained is linear, so linear structure of the global data set is kept and embodied. It has the function of auto-discovery the intrinsic characteristics or non-linear structure of high-dimensional data.The feasibility and effectiveness of these three methods had been demonstrated by extensive experiments on several face databases.