| In the modern highly information technology, networking and digital social environment, information security issues have attracted wild public attention. For the merits of universal, unique, stable, safe and reliable, biometric has become one of the most important topics in the current research area of information security. Among various biometric technologies, face recognition is a most development potential technology due to its character of natural, direct, friendly, secure and practical. Moreover, the study of face recognition has both important theory significance and broad application prospects and potentiality.During the past few decades, face recognition has gained great progress, and a lot of theories and excellent face recognition algorithms have been studied and proposed. Although the existing face recognition algorithms can achieve superior performance, face recognition is still a complex problem since human face images are easily affected by the variations caused by illumination, face expression and pose, etc. Therefore, there are still many problems and key technologies which need to be further addressed and improved. Based on the analysis and summary of the existing face recognition methods, we propose three novel face recognition methods in this thesis and compare them with some state-of-the-art algorithms. Extensive experimental results on several publicly available standard databases show the feasibility and effectiveness of our proposed methods. The main research works and contributions of this thesis can be summarized as following:1. As an effective feature extraction method, Non-negative Matrix Factorization(NMF) has been widely used in various fields. However, NMF is an unsupervised algorithm without taking the class information of the samples into account during feature extraction process. Therefore, NMF may probably discard much useful discriminative information and weaken the recognition accuracy. In order to take full advantage of the information of both labeled samples and unlabeled samples to improve the classification performance of NMF algorithm, a novel feature extraction algorithm is proposed by jointing Label Propagation(LP) and Matrix Factorization(MF) techniques, which named Label propagation based Semi-supervised Non-negative Matrix Factorization(LpSNMF). Firstly, by introducing LP technique, LpSNMF can effectively overcome the limitation of Semi-supervised NMF(SNMF) methods, which cannot explore the distribution relationships between the labeled and unlabeled samples. Secondly, LpSNMF integrates LP with NMF into a unified framework to achieve the purposes for extracting low-dimensional feature and predicting unknown sample’s labels. In addition, we propose an alternating iterative optimization algorithm for solving the objective function, and also prove the convergence of the algorithm from the aspects of the theoretical analysis and numerical experiments, respectively. Finally, experimental results validate the feasibility and effectiveness of the proposed method.2. In order to alleviate the effects of external factors including lighting, facial expressions and pose on the performance of global information based face recognition methods, researchers have proposed a number of local information based face recognition methods. Recently, Structure Preserved Projections(SPP) as an effective local face recognition method is presented. However, the SPP algorithm does not fully take into account the label information of samples and the two-dimensional spatial structure information of face image during the local facial feature extraction, which may reduce the discriminative ability of the low-dimensional features and weaken classification performance to some extent. In order to overcome the shortcomings of SPP algorithm, a novel supervised local face recognition algorithm termed Spatially Smoothed Discriminant Structure Preserved Projections(SS-DSPP) is proposed. Firstly, for the sake of improving the discriminative ability of low-dimensional local facial features, SS-DSPP makes full use of label information of samples to construct the within-class neighborhood graph and between-class neighborhood graph. Secondly, for maintaining the two-dimensional spatial structure information of face images during the local facial feature extraction, a simple and flexible constraint called Spatially Smooth Constraint(SSC) is also addressed, which utilizes the dispersion of nearby element values of the basis matrix obtained by SS-DSPP. Finally, the feasibility and effectiveness of the proposed algorithm are evaluated on four standard face databases. Experimental results demonstrate that SS-DSPP significantly outperforms several state-of-the-art algorithms.3. Although the performance of supervised locally recognition methods is better than global methods and unsupervised local methods, they require a lot of labeled samples to get a better performance, and cannot handle the training set containing unlabeled samples. Moreover, they also ignore the complementary information of different sub-patterns from the same face image. To tackle above problems, we integrate Adaptive Weighted Multiple Graphs based Label Propagation(AWMGLP) and Local Ridge Regression(LRR) to present a new semi-supervised local face recognition method called Semi-supervised Local Ridge Regression(SSLRR). SSLRR utilizes the information of both the unlabeled samples and labeled samples, and it also takes the relationship among sub-pattern sets into consideration. Moreover, SSLRR can also avoid the "Out-of-Sample" problem. In particular, SSLRR can achieve a dual purpose of unlabeled sample’s label prediction and classifier learning. Furthermore, a simple yet efficient iterative optimization algorithm is proposed to update objective function of SSLRR for obtaining local optimal solution. At the same time, we show the convergence of the algorithm from theoretical analysis and numerical experiments. Experimental results demonstrate the performance of the proposed SSLRR method is significantly better than other comparison methods.
|
PDF Full Text Request