Research On Face Recognition Method Based On Statistical Theory

The main task of pattern recognition is dividing the samples into corresponding categories of model with the characteristics of samples. Feature extraction is the key to face recognition. An effective feature extraction method not only helps to simplify the classification of follow-up design, but also can enhance the recognition rate. Feature extraction and description forms of Face recognition can be divided into two broad categories, which are the methods of based on the geometrical features and statistical features based on. Early studies of face recognition are mainly based on the geometric characteristics, and the major methods proposed in recent years are based on the statistical characteristics.Face recognition methods based on statistical characteristics are mainly template matching, subspace method, support vector machine, etc. This article focuses on the subspace method and the main contributions are given as follows:1) The GLRAM (Generalized low rank matrix approximation) and LDA (Linear Discriminant Analysis) method that based on statistical characteristics were studied, a method combined GLRAM with the LDA for face recognition was proposed. First of all, effective feature could be obtained using GLRAM , and then LDA was used to depress the feature dimension and acquire the best classification feature .This enhance the discriminatory power of extracted features. Experimental results demonstrate that higher recognition rate can be achieved in shorter time. This proposed method outperforms the traditional GLRAM methods2) With the study of the two-dimensional locality preserving projections (2DLPP), A method of two-dimensional locality preserving projections (2DLPP) for face recognition is proposed, based on modular image. Firstly, the original images are divided into modular images in presented approach. Secondly, the 2DLPP method can be used to the sub-images obtained from the previous step. Therefore, dimension reduction could be performed. This approach could distill the local features of the images effectively. The experimental results indicate that the proposed method is superior to that of 2DLPP in recognition performance.