Master Thesis - Department of Computer Science

face for each subject, and compare their performances. The performance of
the subband face representation on several linear subspace techniques: PCA,
LDA, 2D-PCA, 2D-LDA and Discriminative Common Vectors (DCV) with
Yale, ORL and PIE face databases show that the subband face based repre-
sentation performs significantly better than the proposed multiresolution face
recognition by Ekenel [36], for frontal face recognition in the presence of vary-
ing illumination, expression and pose. Peak Recognition Accuracy (PRA) for
all three databases are: 100% for Yale (using LDA), 95.32% for PIE (using
DCV) and 91.67% for ORL (using 2D-LDA).
• In chapter 4, we try to enhance classification performance by combining the
discriminative features from both range space and null space of within-class
scatter matrix Sw at (i) feature level and (ii) decision level.
(i) Feature Fusion Strategy: Feature level information fusion allows us to
utilize the whole set of discriminative directions present in an entire face space.
As every face has a unique decomposition in null space and range space, we
project all class means in both spaces to obtain two sets of projected means.
Now each of these sets are used separately to search for the directions that
discriminates them in that space. This step is equivalent to applying PCA
separately on the set of projected means. These two eigenmodels are then
combined using: 1) Covariance Sum and 2) Gramm-Schmidt Orthonormal-
ization [42]. These methods construct a new set of directions integrating the
information from both spaces. We then reorder and select the best combina-
tion among those directions using two techniques: 1) Forward Selection and
2) Backward Selection [39] on a validation set based on a class separability
criterion to obtain the best discriminability across classes.
(ii) Decision Fusion Strategy: For decision fusion, we extract two disjoint
sets of optimal discriminatory basis separately from null space and range space
of within-class scatter matrix Sw, to obtain two different classifiers. Then we
combine the classifiers obtained on null space and range space using sum rule
and product rule, the two classical decision fusion techniques developed by
Kittler et al. [60],[59]. We also exploit each classifier space separately using
119