Algorithms Research On Face Image Feature Extraction And Recognition

Feature extraction is one of the most basic problems in face recognition. Many classical algorithms have been proposed to solve it, for example, the linear subspace methods, which including principal component analysis(PCA), linear discriminant analysis(LDA) and independent component analysis(ICA), are to solve linear problem, and kernel methods based on support vector machine (SVM) to solve nonlinear problem. Further study on feature extraction and settlement could greatly promote the development of the field of pattern recognition.This thesis pointed out the important significance of feature extraction technology in the science research and industry production based on the background of problem studied, and analyzed the current status and developing trend of the feature extraction. And then the basic principles, features and problems of feature extraction were summarized. The feature extraction was researched thoroughly and deeply.Three algorithms were put forward in this thesis for feature extraction: Firstly, discriminant margin maximum criterion(DMMC) based on margin maximum criterion was proposed. DMMC tries to find the intrinsic manifold that bestly discriminates different face classes by maximizing the between-class scatter, while minimizing the within-class scatter. Effectiveness and stability of algorithm are improved. Secondly, uncorrelated locality preserving projections (ULPP) was proposed based on locality preserving projections. ULPP was designed to achieve good discrimination ability by explicitly taking the local and global information of the samples into account. Besides, a simple uncorrelated constraint is introduced to generate statistically uncorrelated features, which further improves recognition performance. Finally, multiple locality preserving information projections(MLPIP) was proposed based on locality preserving projections. When constructing the graph, we did not adopt theε-neighborhood, but put an edge between nodes i and j if xi and xj both belong to the same class. It would reduce the computational complexity of our algorithm. Neighborhood points and non-neighborhood points were certained in MLPIP. As a result, it is easy to make neighborhood points as compact as possible, and it is easy to classify different manifolds, since even the originally close manifolds are projected far away. The discriminative information in the image space was taken full advantage in this method.The feasibility and effectiveness of three methods had been demonstrated through extensive experiments conducted on several face databases.