Dissertation > Excellent graduate degree dissertation topics show

Improvement of Manifold Learning for Face Recognition

Author: GaoShiQiang
Tutor: JingXiaoYuan
School: Nanjing University of Posts and Telecommunications
Course: Pattern Recognition and Intelligent Systems
Keywords: Face Recognition Discriminant analysis Manifold Learning Sparse Representation
CLC: TP391.41
Type: Master's thesis
Year: 2011
Downloads: 70
Quote: 0
Read: Download Dissertation

Abstract

In the field of pattern recognition, feature extraction process, manifold learning is an effective dimension reduction technique. It is designed to protect a given sample concentration of each neighborhood structure. This paper focuses on the more popular in recent years, feature extraction techniques, such as: discriminant analysis techniques, manifold learning techniques and sparse representation of technology and other. In the learning process, we found that the traditional manifold learning methods, such as local retention projection (locality preserving projection., LPP) algorithm, is to protect the sample for the center of the neighborhood structure. Neighborhood structure can truthfully reaction close neighbor of the relationship between the initial sample, but may in its neighborhood there is a plethora of heterogeneous samples, which is not conducive to the final classification and recognition. Therefore, we propose a new method based on the center of the sub-class neighborhood structure, namely the sub-class center manifold retained projection (subclass-center manifold can preserving projection. SMPP). This approach to the points in the sample set of each sub-class center for the neighborhood center, the establishment of the structure of the neighborhood. We proved by theoretical data samples usually show a Gaussian mixture distribution, while the center of the distribution is subclass center, so within the structure of this neighborhood there will be more of the class sample. To be able to use the category information of the sample, and further improve the identification results, we then made to identify the sub-class center projection algorithm (Discriminant of SMPP, DSMPP), reserved. Unlike identification manifold algorithm, DSMPP establish a dual-objective optimization model and solve linear weighted method. Inherent category information of subclass manifold structural information and samples together to achieve both projection can preserve the structure of the sample manifold, but also be able to make a similar sample near, heterogeneous sample separately. Feature extraction techniques want to be able to as much valid information extracted in this process, the article sparse representation prompt us to contain such mutual sparse representation of the relationship between each sample data set. Accordingly, in order to be able to make full use of the samples inherent information, we propose a new sparse representation method, i.e. with a sparse stream shaped information mapping (sparsity embedding with manifold information, SEMI) algorithm. This algorithm first original sample sparse reconstruction, the reconstructed sample set. Then retain the the manifold structure between the old and new samples in the projection process, so as to achieve the purpose of the sparse relationship between the sample and the manifold structure while retaining. We proved in experiments on the AR, FERET and CAS-PEAL database on the classification results, our proposed method with respect to the correctness and validity of other methods.

Related Dissertations

CLC: > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device
© 2012 www.DissertationTopic.Net Mobile