Dual-space Face Recognition Using Average Invariant Factor

In the modern information era, under the influence of science and technology devel-opment, face recognition has become the hot topics in the study of domestic and foreign researchers, moreover, it has been widely used in real applications.Face recognition makes face analysis by computer and extracts the effective feature information, then classify through the classifier, which becomes be an important part of pattern recognition.After the sample preprocess,the dimensions of the face images are usually high, They need to be indicated via making use of samples in low latitude space, to obtain the essential characteristics of classification, therefore, the effectiveness of feature extraction have the great influence on face recognition. Common feature extraction method is based on the linear subspace methods, typical method is LDA,which to take advantage of the label information, but it ignores the reality of the difference of face images,after people propose face recognition method based on the average invariant factors, the concept of average constant factor is proposed through the SVD、QR decomposition, each face can be divided into common facial difference, individuality difference, and intrapersonal difference,it usually makes a matrix singularity,owing to the number of sample is less than the sample dimension,which is also called Small Sample Size (SSS) problem. In order to solve the problem, the paper makes reference on the original methods, respectively extract regular discriminant features information and irregular discriminant features information from range space and null space matrix of within-class scatter matrix, and not only to solve SSS problem,but also to avoid the loss of effective information. the role of each discriminant feature information play different role in the classification, in order to give full consideration to two kinds of discriminant features, this paper proposes a new fusion rule again, the two discriminant information are weighted combination, the effectiveness and correctness of the algorithm are completed on the ORL and YALE face database to realize.