The Research Of Face Feature Extraction And Recognition Based On Multi-manifold

Face recognition technology is one of pattern recognition and computer vision, the study belongs to the field of biometrics. Among them, feature extraction is the most important in the area of pattern recognition, the key of the face recognition technology research is how to extract discriminate in favor of classification features. Traditional global feature extraction methods can not be extracted face image local features, local traditional extraction method can not take into account the global feature face image, and there is data dimensions are too high, fewer samples and identify the effect is not ideal and other issues. In this paper, based on multi-feature extraction manifold theory and method to do the following related research, the main work is divided into the following sections:(1) The overview of face recognition, research background and content, applications and difficulties are all discussed. We will give a brief introduction about some typical face database in the last.(2) Detailing description face recognition algorithms of the current classical feature extraction method: principal component analysis(PCA), linear discriminant analysis(LDA), locality preserving projection(LPP) and locally linear embedding(LLE) method. The advantages and disadvantages of these four methods were systematically explained.(3) Based on maximum marginal criterion(MMC) algorithm, the maximum margin criterion locally graph embedding feature extraction methods based on multi-manifold(MLGE/MMC) was proposed by introducing multi-manifold thought. Firstly, the algorithm constructed the external divergence of the multi-manifold. Then, the internal divergence of the multi-manifold was constructed by a multi-manifold internal reconstruction weighting matrix. Finally, the purpose to be achieved is to maximize the interval of external manifold and to minimize the changes of internal manifold. This algorithm can be more effectively used for feature extraction and classification. In this method, the objective function is constructed by employment of MMC. It effectively solved due to less training samples resulting in the serious decline of the ability to distinguish.(4) Unsupervised linear differential projection(ULDP) algorithm construct the local neighborhood graph and the global variogram through manifold distance between the sample data to characterize the structure of the local neighborhood and the global structural information. However, ULDP method has the following shortcomings: 1. Number of over-reliance in the learning process of training samples, when faced with a one-sample problems, it severely limits the application of this method; 2. Among the many feature extraction, we can not reveal which characteristics of classification and prediction play a leading role. To this end, we propose a multi-manifold unsupervised linear differential projection(MULDP) algorithm. This algorithm can get the low-dimensional manifolds embedded in a high-dimensional space and maintain the local and global structural information effectively.(5) At last, we created a face recognition system using MATLAB system, in order to verify the classical methods and the proposed method.