1 Introduction
Feature extraction by dimensionality reduction is a critical step in pattern recognition. Principal component analysis (PCA) is a classic method for dimensionality reduction in the field of face recognition, which was proposed by Turk and Pentland in Ref. Turk and Pentland [7]. Yang et al. [10] presented twodimensional PCA (2DPCA) to improve the efficiency of feature extraction, in which image matrices were used directly. Twodimensional weighted PCA (2DWPCA) was developed in Ref. Nhat and Lee [6] to improve the performance of 2DPCA. The complete 2DPCA method was presented in Ref. Xu et al. [9] to reduce the feature coefficients needed for face recognition compared to 2DPCA. In kernel PCA (KPCA) [11], samples were mapped into a high dimensional and linearly separable kernel space and then PCA was employed for feature extraction. Chen et al. [1] presented a pattern classification method based on PCA and KPCA (kernel principal component analysis), in which withinclass auxiliary training samples were used to improve the performance. Liu et al. [4]
proposed a 2DECA method, in which features are selected in 2DPCA subspace based on the Renyi entropy contribution instead of cumulative variance contribution. Moreover, some approaches based on linear discriminant analysis (LDA) were explored
[8, 12, 13].Contrast to the above L norm based methods, Kwak [3] developed LPCA by using L norm. Ding et al. [2] proposed a rotational invariant L norm PCA (RPCA). These none L norm based algorithms are less sensitive to the presence of outliers.
In this paper we propose 2DRPCA and 2DLPCA algorithms for face recognition by utilizing the advantages of L
norm method and 2DPCA. Instead of using image vectors in R
PCA and LPCA, we use image matrices in 2DRPCA and 2DLPCA directly for features extraction. Compared to the 1D methods, the corresponding 2D methods have two main advantages: higher efficiency and recognition accuracy. We extend RPCA and LPCA to their two dimensional case and the 2DRPCA and 2DLPCA methods are proposed.This paper is organized as follows: We give a brief introduction to the RPCA and LPCA algorithms in Section 2. In Section 3, the 2DRPCA and 2DLPCA algorithms are proposed. In Section 4, the mentioned methods are compared through experiments. Finally, conclusions are drawn in Section 5.
2 Fundamentals of subspace methods based on none L norm
In this paper, we use to denote the training set of 1D methods, where is a dimensional vector.
2.1 RPca
RPCA algorithm tries to find a subspace by minimizing the following error function
(1) 
where is the projection matrix, is defined as , and denotes the R norm, which is defined as
(2) 
In RPCA algorithm, the training set should be centered, i.e.,, where is the mean vector of , which is given by .
The principal eigenvectors of the R
covariance matrix is the solution to RPCA algorithm. The weighted version of Rcovariance matrix is defined as(3) 
The weight has many forms of definitions. For the Cauchy robust function, the weight is
(4) 
The basic idea of RPCA is starting with an initial guess and then iterate with the following equations until convergence
(5) 
The concrete algorithm is given in Algorithm 1.
2.2 LPca
The L norm is used in LPCA for minimizing the following error function
(6) 
where is the projection matrix, is defined as , and denotes the L norm, which is defined as
(7) 
In order to obtain a subspace with the property of robust to outliers and invariant to rotations, the L norm is adopted to maximize the following equation
(8) 
It is difficult to solve the multidimensional version. Instead of using projection matrix , a column vector is used in equation (8) and the following equation is obtained
(9) 
One best feature is extracted by the above algorithm. In order to obtain a dimensional projection matrix instead of a vector, an algorithm based on the greedy search method is given as follows
For to
Apply the LPCA procedure to to find
End
3 2drPCA and 2DLPCA algorithms
In 2D methods, is used to denote the training set, where is a matrix.
3.1 2drPca
In this paper we propose 2DRPCA algorithm, in which we iterate the projection matrix with an initial matrix until convergence.
First, the training set is centered, i.e., , where is the mean matrix of , defined as .
The R covariance matrix is defined as
(10) 
The Cauchy weight is defined as
(11) 
The residue is defined as
(12) 
After obtaining the eigenvectors of , the iterative formula is similar to which used in the RPCA algorithm
(13) 
The 2DRPCA algorithm is outlined in Algorithm 3.
3.2 2dlPca
Compared to LPCA, in the two dimensional case we want to find a column vector to solve the following problem
(14) 
In fact, is a row vector. The number of maximum absolute value in a vector contributes most to its L norm. Assume that the column index of the maximum absolute value in is , we can calculate by the th column of . The 2DLPCA algorithm is given in Algorithm 4.
Then we can obtain a dimensional projection matrix from the following algorithm.
.
For to
.
Apply the LPCA procedure to to find .
End
4 Experimental results and analysis
Three databases: ORL, Yale and XM2VTS are used to test methods mentioned above. The recognition accuracy and running time of extracting features are recorded.
The ORL database contains face images from 40 different people and each person has 10 images, the resolution of which is 92112. Variation of expression (smile or not) and face details (wear a glass or not) are contained in the ORL database images. In the following experiments, 5 images are selected as the training samples and the rest are selected as the test samples.
The Yale database is provided by Yale University. This database contains face images from 15 different people and each has 11 images. The resolution of Yale database images is 160121. In the following experiments, 6 images are selected as the training samples and the rest are selected as the test samples.
The XM2VTS[5] database offers synchronized video and speech data as well as image sequences allowing multiple view of the face. It contains frontal face images taken of 295 subjects at one month intervals taken over a period of few months. The resolution of XM2VTS is 5551. In the following experiments, 4 images are selected as the training samples and the rest are selected as the test samples.
4.1 RPCA and 2DRPca
The experimental results of RPCA and 2DRPCA are shown in Table 1, and the number of iterations of RPCA and 2DRPCA is 120.
ORL  Yale  XM2VTS  
Recognition accuracy  
PCA  0.90  0.77  0.71 
RPCA  0.88  0.77  0.71 
2DRPCA  0.90  0.80  0.78 
Running time  
PCA  1.37  0.36  17.27 
RPCA  914.21  411.06  1409.30 
2DRPCA  403.90  372.76  619.78 
The initial projection matrix is obtained by PCA (2DPCA) at the beginning of RPCA (2DRPCA). The final projection matrix is obtained by an iterative method starting with . As a result of the iteration, the computational complexity is high. Meanwhile, they have nearly the same recognition accuracy.
In the experiment of RPCA algorithm tested on the ORL database, the convergence process is shown in Fig. 1 (a), in which the coordinate denotes the norm of projection matrix and the coordinate denotes the number of iterations. The norm of a projection matrix is used to observe its convergent process. After iterating at least 100 times the projection matrix converges. As a comparison, 2DRPCA just needs less than 30 iteration to obtain a convergent projection matrix, which is shown in Fig. 1 (b). Image matrices used in 2DRPCA leads to a faster convergence.
The convergence illustration tested on the Yale database is shown in Fig. 2. The convergent speed of RPCA is similar to that of 2DRPCA. In the experiment tested on the XM2VTS database, the convergent speed of 2DRPCA is much faster than that of RPCA shown in Fig. 3. In other words, the efficiency of 2DRPCA is higher than that of RPCA.
4.2 LPCA and 2DLPca
The experimental results of LPCA and 2DLPCA are shown in Table 2.
ORL  Yale  XM2VTS  
Recognition accuracy  
PCA  0.90  0.77  0.71 
LPCA  0.90  0.76  0.71 
2DLPCA  0.91  0.80  0.76 
Running time  
PCA  1.37  0.36  17.27 
LPCA  15.96  5.15  83.52 
2DLPCA  3.52  3.62  40.43 
From Table 2 we can see that the performance of 2DLPCA is better than that of LPCA and PCA. In 2DLPCA, image matrices are used directly for feature extraction. Features extracted by 2DLPCA is less than features extracted by LPCA.
We implement another experiment on the ORL database. Different number of features is extracted by PCA, LPCA and 2DLPCA, respectively. Then these features are used for face recognition. The experimental result is shown in Fig. 4, from which we can see that less features extracted by 2DLPCA achieves a higher recognition accuracy.
5 Conclusions
In this paper we proposed 2DRPCA and 2DLPCA for face recognition. We extend RPCA and LPCA to their 2D case so that image matrices could be directly used for feature extraction. Compared to the L norm based methods, these L norm based methods are less sensitive to outliers. We analyze the performance of 2DRPCA and 2DLPCA against RPCA and LPCA algorithms based on experiments. The experimental results show that the performance of 2DRPCA and 2DLPCA is better than that of RPCA and LPCA, respectively.
Acknowledgements.
This work was partially supported by the National Natural Science Foundation of China (Grant No.61672265 and U1836218) and the 111 Project of Ministry of Education of China (Grant No. B12018).References
 [1] (2017) KPCA method based on withinclass auxiliary training samples and its application to pattern classification. Pattern Analysis and Applications 20 (3), pp. 749–767. Cited by: §1.

[2]
(2006)
R1pca: rotational invariant l1norm principal component analysis for robust subspace factorization.
In
Proceedings of the 23rd international conference on Machine learning
, pp. 281–288. Cited by: §1.  [3] (2008) Principal component analysis based on l1norm maximization. IEEE transactions on pattern analysis and machine intelligence 30 (9), pp. 1672–1680. Cited by: §1.
 [4] (2011) ECA and 2deca: entropy contribution based methods for face recognition inspired by keca. In 2011 International Conference of Soft Computing and Pattern Recognition (SoCPaR), pp. 544–549. Cited by: §1.
 [5] (1999) XM2VTSDB: the extended m2vts database. In Second international conference on audio and videobased biometric person authentication, Vol. 964, pp. 965–966. Cited by: §4.
 [6] (2005) Twodimensional weighted pca algorithm for face recognition. In 2005 International Symposium on Computational Intelligence in Robotics and Automation, pp. 219–223. Cited by: §1.
 [7] (1991) Eigenfaces for recognition. Journal of cognitive neuroscience 3 (1), pp. 71–86. Cited by: §1.
 [8] (2004) A new direct lda (dlda) algorithm for feature extraction in face recognition. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., Vol. 4, pp. 545–548. Cited by: §1.
 [9] (2006) Complete twodimensional pca for face recognition. In 18th International Conference on Pattern Recognition (ICPR’06), Vol. 3, pp. 481–484. Cited by: §1.
 [10] (2004) Twodimensional pca: a new approach to appearancebased face representation and recognition. IEEE transactions on pattern analysis and machine intelligence 26 (1), pp. 131–137. Cited by: §1.
 [11] (2000) Face recognition using kernel eigenfaces. In Proceedings 2000 International Conference on Image Processing (Cat. No. 00CH37101), Vol. 1, pp. 37–40. Cited by: §1.
 [12] (2006) Nearest neighbour line nonparametric discriminant analysis for feature extraction. Electronics Letters 42 (12), pp. 679–680. Cited by: §1.
 [13] (2006) A reformative kernel fisher discriminant algorithm and its application to face recognition. Neurocomputing 69 (1315), pp. 1806–1810. Cited by: §1.
Comments
There are no comments yet.