Regularized Kernel Locality Preserving Discriminant Analysis for Face Recognition - Advanced Concepts for Intelligent Vision Systems - page 288

Fig. 1. Preprocessed sample images of the two databases: (a)ORL database; (b)FERET
database

and FERET face database [13] are used for testing. The ORL database contains
400 images from 40 individuals. The FERET database contains 14126 images
from 1199 individuals. From FERET database, a subset containing 1131 frontal
images from 229 individuals with at least 4 images per individuals, is selected
in this work. All the images are scaled to 32 32 pixels and represented by

1024-dimensional vectors. Fig. 1 shows the preprocessed sample images from the
two face databases.

×

4.2 Recognition Experiments and Discussion

In this section, the recognition performances of the proposed RKLPDA with
LDA [1], LPDA [2], KFDA [7] and KCLPP [8] are compared. In experiments,
cosine polynomial kernel function is chosen, and the parameters are set the same
as [14]. For all algorithms, we randomly select i ( i =2 , 3 , 4 , 5 for ORL database
and i =2 , 3 for FERET subset) images of each individual for training and the
remaining images for testing. The nearest-neighbor classifier based on cosine
distance metric is used for classification. The recognition results from 20 runs
are given in Table 1 and 2. Also, an illustration of the recognition accuracies
against the number of features on ORL database for i = 5 is given in Fig. 2.
From Table 1 and 2, the proposed RKLPDA method consistently and re-
markably outperforms the other 4 methods, which validates the effectiveness
of the proposed method. In experiments on ORL, the kernel-based methods

Table 1. Recognition accuracy (%) and corresponding number of features on ORL
database

TrNum LDA LPDA KFDA KCLPP RKLPDA

2 75.5 ± 2.62(39) 58.1 ± 3.05(30) 78.0 ± 2.47(39) 78.4 ± 2.51(39) 79.1 ± 2.55(39)

3 84.5 ± 2.44(39) 77.5 ± 2.10(35) 86.7 ± 2.16(39) 86.0 ± 2.79(39) 89.0 ± 2.29(39)

4 90.5 ± 2.10(39) 87.2 ± 2.49(35) 91.9 ± 1.90(39) 89.5 ± 2.07(39) 94.2 ± 1.62(39)

5 91.9 ± 1.98(39) 91.4 ± 1.85(35) 93.7 ± 1.91(39) 92.6 ± 1.59(39) 96.1 ± 1.08(60)

Table 2. Recognition accuracy (%) and corresponding number of features on FERET
database

TrNum LDA LPDA KFDA KCLPP RKLPDA

2 68.5 ± 1.74(30) 68.2 ± 1.69(30) 68.4 ± 1.64(228) 40.5 ± 1.77(228) 74.3 ± 1.68(80)

3 79.4 ± 1.34(20) 79.4 ± 1.51(20) 79.0 ± 1.88(228) 58.3 ± 1.97(457) 85.2 ± 1.31(90)