Regularized Kernel Locality Preserving Discriminant Analysis for Face Recognition - Advanced Concepts for Intelligent Vision Systems - page 289

ORL

100

90

80

70

LDA
LPDA
KFDA
KCLPP
RKLPDA

60

50

40

30

0 20 40 60 80 100 120 140 160 180 200

number of features,TrNum=5

Fig. 2. Recognition accuracy (%) of different algorithms on ORL database

(KFDA and KCLPP) slightly outperform the corresponding linear methods
(LDA and LPDA), while in experiments on FERET, the performance of KCLPP
is dissatisfactory. This might be caused by the following two reasons: (1) Im-
proper kernel and kernel parameters are chosen; (2) Though KCLPP utilizes the
label information, it is not a discriminant analysis method in nature. In addi-
tion, as shown in Fig. 2, the LDA, LPDA, KFDA and KCLPP methods obtain
at most C 1

discriminant features. Meanwhile, the recognition accuracies increase with the
increasing of number of samples, and RKLPDA methods achieves a relative good
and stable performance with asmallernumberoffeatures.

− 1 discriminant features, while RKLPDA can obtain at most n −

5 Conclusions

This paper presents a regularized kernel locality preserving discriminant analy-
sis (RKLPDA) method. Kernel trick is introduced to extend LPDA to its kernel
formulation. To address the singularity problem of kernel locality preserving
within-class scatter matrix and utilize the discriminative information in both
the principal and null subspace of kernel locality preserving within-class scatter
matrix, the eigencvectors are regularized according to the predicted eigenvalues,
which de-emphasizes the eigenvectors susceptible to samples noises by raising
the eigenvalues, and heavily emphasizes the null space which contains abun-
dant of discriminative information. Extensive experiments on ORL database
and FERET subset show that RKLPDA consistently outperforms other linear
and kernel methods, which indicates the effectiveness of the proposed method.
However, the performance of kernel-based methods diversifies with different ker-
nel functions and kernel parameters, so more attentions on kernel function and
kernel parameter choosing should be paid in the future work. Also, the regular-
ization strategy is a key point to be considered.