Research On Subspace Dimansion Reduction Method For Face Recognition

In recent years, face recognition has become a popular field of research of biological feature recognition. Today’s society, face recognition has been widely used in military, economic, transportation, public safety and other fields, and achieved nice effect, caused widespread concern at home and abroad. In view of this, face recognition has great theoretical significance and practical value, future prospects are very good. A face recognition algorithm is effective and practical, the key lies in whether the algorithm can quickly and accurately identify an effective feature extraction, and in the process of data dimensionality reduction how to maintain the same internal structure and sample data are mapped to low-dimensional space. On the basis of the research and development of face recognition, through the study of data reduction and subspace analysis method, this paper presents some face recognition algorithms and achieved good results on databases of face recognition.On the basis of today’s linear discriminant algorithms, the paper made some improvements and promotion, focusing on the following aspects:(1) Based on modular sorting PCA of linear discriminant analysis algorithm (Modular Sorting Principal Component Analysis, MSPCA)MSPCA method divided image matrix into blocks on the first, sub-image matrix for all sub-blocks obtained using PCA method to obtain all eigenvalues and corresponding eigenvectors be identified, then find the k largest eigenvalues all these eigenvalues corresponding to eigenvectors, then use these feature vectors to extracted the characteristics of the sub-image which it belongs to respectively. Finally, based on the MSPCA, then merger the feature matrix which is obtained by extracting sub-image matrix, and use the combined features of this matrix as a new sample to PCA+LDA. Due to the use of sub-image matrix, MSPCA avoid the use of singular value decomposition theory, make the algorithm easier,it strengthen the role of local information in the recognition algorithm, compared to PCA method it has made substantial progress in the classification and recognition performance. Experimental results and analysis on the ORL face database show that MSPCA algorithm is fast and effective in the feature extraction and classification of the sample data, compared with the PCA method,this algorithm improves the recognition efficiency, the recognition rate has also been greatly improved.(2) Locality preserving projection based on weighted maximum margin criterion(Locality Preserving Projection Based On Weighted Maximum Margin Criterion, WMMCLPP)LPP algorithm is an unsupervised algorithm essentially, it is no category information of samples used in the algorithm. Based on this, this chapter presents a locality preserving projection based on weighted maximum margin criterion. First, it use PC A method to reduce the dimension of original sample, and then define the weight matrix of each sample point by calculating the distance that each sample point to the centre point of its category, category to build the membership matri, then redefined the mean of each category of sample matrix and between-class scatter matrix and within-class scatter matrix by the weight matrix. Then add the objective function of weighted maximum margin criteria redefined as a regulatory factor to the LPP algorithm to solve the optimization problem. Finally, it use the ultimate criterion function to build the symmetric matrix to solve the problem that optimal projection vector is nonorthogonal. Experimental results and analysis on the Yale and ORL face database show that, compared with the traditional method of feature extraction, recognition performance of WMMCLPP algorithm is more favorable.(3) Linear discriminant analysis based on neighborhood locality preserving projection (Linear Discriminant Analysis Based On Neighborhood Locality Preserving Projection, NLPPDA)Classical manifold learning method can not obtain the low dimensional projection of the new samples as linear feature extraction method. The linearization of manifold learning method is a perfect solution to this problem. Based on this, this chapter presents a linear discriminant analysis based on neighborhood locality preserving projection. First, it use PCA method to reduce the dimension of original sample, and then use the idea of LPP algorithm to calculate the neighbor points of samples, the divide into two categories:one is a collection include all sample points whose classification is the same as the points,another one is a collection include all sample points whose classification is not the same as the points. Then redefined the within-class scatter matrix by the former and the between-class scatter matrix by latter. Then improve the object function of NRLDA algorithm, add a constraint condition,standardize the distance of the sample to within-class neighbor points. NLPPDA algorithm maintaining the local geometry structure information and supervised information of the sample. Experimental results and analysis on the Yale and ORL face database show that recognition performance of NLPPDA algorithm is more favorable.