Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
plored both <strong>of</strong> the spaces to capture and then combine discriminative directions for<br />
enhancing discriminability across classes. Our approach efficiently exploit both the<br />
spaces by combining them in two different levels: (i) Feature level and (ii) Decision<br />
level.<br />
For feature level fusion, we develop a dual space by combining the discriminative<br />
features from both range space and null space <strong>of</strong> within-class scatter matrix Sw.<br />
This allows us to utilize the whole set <strong>of</strong> discriminative directions present in an<br />
entire face space. As every face has a unique decomposition in null space and range<br />
space, we project all class means in both spaces to obtain two sets <strong>of</strong> projected<br />
means. Now each <strong>of</strong> these sets are used separately to search for the directions that<br />
discriminates them in that space. This step is equivalent to applying PCA on the<br />
set <strong>of</strong> projected means separately. These two eigenmodels are then combined using:<br />
1) Covariance Sum method and 2) Gramm-Schmidt Orthonormalization [42]. These<br />
methods construct a new set <strong>of</strong> directions integrating the information from both<br />
spaces. Then we reorder and select the best combination among those directions to<br />
obtain the best discriminability across classes. The feature reordering and selection<br />
is performed using two techniques: 1) Forward Selection and 2) Backward Selection<br />
[39] on a validation set, based on a class separability criterion.<br />
For decision level fusion, we extract two disjoint sets <strong>of</strong> optimal discriminatory<br />
basis separately from null space and range space to obtain two different classifiers.<br />
Then we combine the classifiers obtained on null space and range space using sum<br />
rule and product rule, two classical decision fusion techniques developed by Kittler<br />
[60],[59]. We also exploit each classifier separately using LDA and nonparametric<br />
LDA to enhance class separability at classifier response level and then combine them<br />
using sum rule. We denote the class scores provided by a classifier on a sample as<br />
response vector. Basically, we use response vectors as features vectors at decision level<br />
and employ LDA and nonparametric LDA to enhance class separability at classifier<br />
output space. Response vectors on a validation set (disjoint from training and testing<br />
sets <strong>of</strong> a database) is used as training data at decision level. Then the response<br />
vectors on testing set <strong>of</strong> the database are recalculated in the eigenmodel to improve<br />
combined classification accuracy.<br />
73