Font Size: a A A

Face Recognition Algorithm Based On Two-dimensional Subspace Analysis Methods And Improved LPP

Posted on:2016-05-27Degree:MasterType:Thesis
Country:ChinaCandidate:Q Q LiFull Text:PDF
GTID:2308330470960340Subject:Electronic Science and Technology
Abstract/Summary:PDF Full Text Request
Face recognition, a friendly identity recognition technology in the biometrics verification, has been attracting worldwide research interests. Feature extraction is the key component of face recognition technology, and subspace analysis methods have become active research area for feature extraction. Locality Preserving Projection(LPP) and Two-Dimensional Locality Preserving Projection(2DLPP) are nonlinear subspace analysis method, which can obtain local manifolds information of training set. The main idea of LPP or 2DLPP is that if two pictures of training set are neighbor in high-dimensional image space, they are also neighbor in low-dimensional nonlinear manifold.LPP and 2DLPP can preserve nonlinear manifold structure of image, but they have unsupervised learning problem and only contain the local information, some new face recognition algorithms based on improved LPP are presented:1. Firstly, the weight matrix of LPP is optimized with the within-class and between-class label information, which make the improved LPP(ILPP) become a supervised nonlinear learning method. By combining ILPP with Two-Dimensional Principal Component Analysis(2DPCA) and Two-Dimensional Linear Discriminant Analysis(2DLDA), a new face recognition algorithm based on 2DPCA+2DLDA and improved LPP is presented. 2DPCA+2DLDA+ILPP not only can retain the spatial information and class label information, but also can obtain local manifolds information.2. Using the similarity between different images and the within-class and between-class label information to optimize the weight matrix of LPP, the local manifolds information and class label information of images are obtained by optimized LPP(OLPP). While the algorithms of Diagonal Principal Component Analysis(DiaPCA) and Diagonal Linear Discriminant Analysis(DiaLDA) not only can retrain spatial information and class label information of images, but also can reserve the images correlations between variations of rows and those of columns. So if we train the images with the algorithm of DiaPCA+DiaLDA firstly, the images spatial information and class label information will be retrained, then we train the images with OLPP algorithm, the local manifolds information of images also can be obtained.3. Firstly, a face recognition algorithm of Two-Dimensional Adaptive Locality Preserving Projection(2DALPP) is proposed by using the within-class and between-class label information, distance between two images and the average distance between images to optimize the weight matrix of 2DLPP and to make parameters such as the number of nearest neighbor and the weight of weight matrix become adaptive. The image gradient reflects the change of image information in the row or column direction, if two images of training set in the high dimensional data space are neighbor, gradient in the row or column direction of the two images are also similar. So, we use the information of images gradient to improve the 2DLPP algorithm, and put forward a new method based on gradient of Two-Dimensional Adaptive Locality Preserving Projection(G2DALPP).
Keywords/Search Tags:Face Recognition, Principal Component Analysis, Linear Discriminant Analysis, Locality Preserving Projection, Feature Extraction
PDF Full Text Request
Related items