Research On Several Kinds Of Methods Of Dimension Reduction For Pattern Recognition

Dimensionality reduction methods in the field of face recognition of patternrecognition have a wide range of applications. As the representation of face imagesdata usually have very high dimension, and the high dimension will not only lead tothe “curse of dimensionality”phenomenon, but also makes it difculty to directlyunderstand and discovery the structure information of the data set. The majoralgorithms are used as the method of data preprocessing, on the one hand, it canovercome the “curse of dimensionality”phenomenon, on the other hand, it cangreatly reduce the calculation complexity and noise, so the data reduction algorithmsfor face recognition aroused the interest of researchers.In this paper, we firstly introduction the typical existing one-dimensional di-mensionality reduction methods and two-dimensional one side dimensionality re-duction methods in the field of face recognition, then we analysis the bidirectionalprojection method of tensor subspace analysis (TSA) and discriminant tensor sub-space(DTSA). In addition, as the two-dimensional discriminant supervised localitypreserving projection method uses only a single projection, and thus the generatedlow dimensional data matrices have still many features, so we propose the bidirec-tional discriminant supervised LPP (BDSLPP). Then to the problem of the left andright projection matrices are not usually orthogonal in TSA and DTSA, and the re-quirement of the orthogonality of the columns of projection matrices is common inthat orthogonal projection matrices preserve the metric structure of the facial imagespace we propose the orthogonal TSA (OTSA) and orthogonal DTSA (ODTSA), inwhich the left and right projection matrices are orthogonal. Finally we perform thealgorithms on ORL and Yale face data sets. Experimental results demonstrate therecognition algorithm is feasible and has good recognition capability.