Master Thesis - Department of Computer Science

each subject, 42 samples (flashes from 21 different directions with and without
the room lights on) were used for our study. Henceforth we use the following
notations for each of the 21 flashes with and without room lights on:
(i) 21 flashes with room lights on: lights flash02 - lights flash22,
(ii) 21 flashes without room lights on: illum flash02 - illum flash22.
The face samples were extracted from the original gray level images for the
PIE database, based on the specification of eye locations given in [5].
• The ORL face database has 40 subjects with 10 samples each. There is no
change in illumination but significant (near frontal, no profile views) change in
the face pose.
3.4.2 Performance Analysis on Three Standard Face Databases
To select subject-specific optimal subbands we split the image database into three
disjoint sets, namely training, validation and testing set. Training set along with
validation set are used for subband selection, and then the performance of selected
subbands is observed on the testing set. The number of images used for a subject
over the three sets, for all three databases are given in Table 3.1. Fig. 3.9 shows
the images in training set for a single subject on Yale, PIE and ORL databases. For
subband selection, the image size has been kept as 64*64 for all subspace methods
except DCV where it is 25*25. The maximum values for l and k are 2 and 4 for DCV,
while for other subspace methods they were chosen as 4 and 7 , respectively. For the
PIE database, which has 42 frontal samples, only 4 face images corresponding to 2
frontal flashes (illum flash08, lights flash08, illum flash11, lights flash11) are used for
training.
Table 3.2 shows the performances of subspace methods on original images for
three databases. Table 3.3 shows the verification (in terms of Equal Error Rate or
EER) and recognition (in terms of Peak Recognition Accuracy or PRA) performance
of subband face representation integrated with subspace methods (PCA, LDA, 2D-
PCA, 2D-LDA and DCV) on Yale database. The results in the column labeled
“Subband Face(C1)” corresponds to that obtained using optimal subband faces for
subjects as selected by minimizing the first criterion discussed in Section 3.3.1. Same
61