Face Recognition Using Contourlet Transform and Multidirectional Illumination from a Computer Screen - Advanced Concepts for Intelligent Vision Systems - page 339

Table 1. Experiment 2 results (in %) using our database

Training error recog. verif. rate
images

rate rate at 0.1% FAR

8 0.13 99.87 99.87

7 0.18 99.82 99.87

6 0.66 99.33 99.67

5 0.63 99.37 99.37

4 0.89 99.11 98.76

3 1.32 98.68 98.76

2 3.10 96.90 95.69

1 14.41 85.59 79.46

100 100

99.8 99

99.6 98

99.4 97

99.2 96

99.0 95

98.8 94

98.6 93

98.4 92

8 training images
7 training images
6 training images
5 training images

8 training images
7 training images
6 training images
5 training images

98.2 91

98.0 90

1 2 3 4 5 6 7 8 9 10

10 −3 10 −2 10 −1 10 0

Rank

False Accept Rate (log scale)

(a) (b)
Fig. 6. Experiment 2 results for the extended Yale B database. (a) CMC curves for 5
to 8 training images and the corresponding (b) ROC curves.

were left with image 5 only. This order was chosen so that there is at least one
frontal lit image in the training data and the lateral images are the ones that
make a smaller angle with the optical axis. This is sensible from a practical stand
point because placing lights at smaller angles requires less space.
Fig. 6 shows the CMC and ROC curves of our algorithm on the extended Yale
B database and a summary of the results is presented in Table 2. Fig. 7 shows
plots of error rates for different subsets of the database for direct comparison
with [13]. Our algorithm achieved 100% recognition rate and 100% verification
rate at 0.001FAR on this database. On the CMU-PIE database [20], with just five
images used for training and the remaining for testing, our approach achieved
100% recognition rate and 100% verification rate at 0.01% FAR. Using three
training images, the recognition rate dropped to 94.6%.

4.3 Experiment 3

In this experiment, we study the effects of time lapse between training and test
images. This experiment cannot be performed on the Yale B and CMU-PIE
databases because they were acquired in a single session per subject.Therefore,
we perform this experiment on our database only. The setup of this experiment
is similar to experiment 2 except that the system is trained using images from