Research Of Graph-Based Semi-supervised Face Recognition Method

Since the19th century, face recognition algorithms have already been studied, it hasmany advantages that other biometric technology don’t possess. However, there are manydifficulties taking advantage of this technology. Currently, the mainly difficulty of thistechnology lies in the section of feature extraction. Researchers have proposed various featureextraction methods. These feature extraction methods simply divided into three categories:unsupervised learning methods, supervised learning methods, semi-supervised learningmethods. Unsupervised learning doesn’t have to require the labeled samples, such as thePrincipal Component Analysis (PCA), Locality Preserving Projection (LPP), NeighborhoodPreserving Embedding (NPE). The process of supervised learning must have labeled samples,such as Linear Discriminant Analysis (LDA), Marginal Fisher Analysis (MFA). If there aresufficient labeled samples, supervised learning methods perform better compared tounsupervised learning algorithms. In addition, in order to get more useful information fromthe samples, the input of many algorithms is encoded2nd or higher order, such as MultilinearPrincipal Component Analysis of Tensor Objects (MPCA)，Discriminant Analysis with TensorRepresentation (DATER) and so on. It avoids the curse of dimensionality problem and thesmall sample size problem that is caused when transforming the input image data into avector.Supervised learning algorithm achieve good results compared to unsupervised learningalgorithms, but in practical applications, because of the need to join the label which makes thesamples will consume quite a lot of labor, material and equipment, professional technicalpersonnel. So researchers begin to pay close attention to a semi-supervised learning algorithm,it is hoped that the sample space distribution information from face samples set can be obtainby the analysis of mass unlabeled samples and a small amount of labeled samples. With thedevelopment of semi-supervised learning algorithms, graph-based learning algorithmattractive more and more attention, because of the advantages of simple, intuitive.In this paper we focus on graph-based semi-supervised learning algorithm, based on someof the existing framework, we will add a lot of unlabeled data to the learning algorithm, in theprocess, we construct graph in a better way to try to avoid other information caused by noise of changing of illumination and expression and obtain the more accurate frameworkembedding in original face images, meanwhile, in order to obtain more information from theoriginal sample, we encode an object as a general tensor of2nd or higher order. Finally wepropose the semi-supervised discriminant analysis with tensor representation (SDATER)algorithm, we carried out experimental on three classic data sets, experimental results validatethat compared to the previous algorithms, the new algorithm is able to extract more effectivefeatures, and the recognition rate has been significantly improved.