Study On Tensor For Face Recognition

In today's society we pay more attention on the security of people's identity information,but with the informatization people's identity have been more and more digital and secret,it has been an important problem.With the development of computer technology and biological recognition technology,biological characteristics such as face,fingerprint and iris has been widely applied to information security certification.So to speak,the security informat ion verification technology base on biological characteristics has the ability of commercialization and replace the traditional validation(id verification,etc.)gradually.As the best one,face recognition has been widely used by various government departments,banks and companies because the advantages of uniqueness,non-contact,no copy,concealment,and simplicity.Feature extraction as a core step in face recognition,not only extract the features which is useful for classification from the original high-dimensional face images,also reduces the original face image data,it improves the accuracy and speed of face recognition.This paper pays attention on feature extraction which based on subspace.At the same time,the tensor representation is used to put forward the color face recognition algorithm based on tensor,the main research works are as follows:At present,we use vector and matrix to represent a face data almost.Vectoring means transforming the face matrix into a one-dimensional column or row vector,and it can cause the high-dimensions of the sample data and destroy the face image spatial structure,finally reduce the effect of face recognition.Although 2D-ways can use face matrix directly,it ignores the correlation between RGB color components and in some degree destroy the face image spatial structure.Many of the objects can be naturally expressed as a tensor actually.So,we use a tensor to represent a RGB color face image,1-mode means horizontal direction of image,2-mode means column direction,3-mode means RGB color components direction.Tensor model not only can keep the correlation of color information component and the spatial relationship of information,can also be integrated into the color information,which means do not need to make gray face image operation.Although 2D-PCA color face recognition has obtained some achievements,its color information matrix model misses part of the original data information.Firstly,we propose a new color information matrix model based on tensor 1-mode(vertical direction)unfolding which only destroy the correlation of R,G,B components,keep the spatial relationship of each pixels,and retain more information of original image data,the experimental results also show that the new model has higher recognition rate.Secondly,we extend the 2D-PCA to tensor space and use the tensor model of RGB color face,propose Tensor PCA.In order to achieve the best classification,it seeks three projection matrices which consist of the eigenvectors correspo nding to the largest eigenvalues of the n-mode total scatter matrix to maximize the distance of the projected samples,and constructs an ALS iterative procedure to optimize the projection matrices.As is shown in the results,in contrast with the process o f 2D-PCA the recognition rate increases and the training time decreases.Aim at LPP don't use the discriminant information of face image effectively,we improve the classification effect of(2D)~2-LPP by joining discriminant information,and then propose(2D)~2-DLPP.It uses a criterion function with discriminant information do row and column 2D-LPP operation respectively,and gets two projection matrixes.Meanwhile,in order to keep the image space structure better and make use of the color information which can improve the classification accuracy,we take advantage of 3~rd-tensor to represent a RGB color face image,propose TDLPP.It builds within-class and between-class similar matrix with discriminant information directly by tensor distance and KNN.Then decompose the tensor matrix criterion function by one iteration,and obtain three optimal projection matrixes corresponding to the three modes.The experimental results show that,compared with(2D)~2-LPP,(2D)~2-DLPP and TDLPP improve the classification effect,and TDLPP is the best.