20.01.2013 Views

Master Thesis - Department of Computer Science

Master Thesis - Department of Computer Science

Master Thesis - Department of Computer Science

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

From Tables 4.1-4.3, we conclude the following: A large quantity <strong>of</strong> training sam-<br />

ples drive discriminative information to the range space <strong>of</strong> within-class scatter. For<br />

a database with a large number <strong>of</strong> subjects, the number <strong>of</strong> training samples will not<br />

allow null space to perform well. The main aim <strong>of</strong> our work is to capture the dis-<br />

criminative informations available in range space and combine them efficiently with<br />

the discriminative informations obtained from null space. Thus, we exploit the whole<br />

gamut <strong>of</strong> discriminative informations present in the entire face space and utilize them<br />

for enhancing classification performance.<br />

4.5.2 Performance <strong>of</strong> Dual Space Face Recognition Approach<br />

We split a image database into three disjoint sets called as training, validation and<br />

testing set. Feature fusion technique requires validation set for selecting optimal<br />

feature, W Dual<br />

opt<br />

and the performance <strong>of</strong> selected features is observed on the testing<br />

set. For decision fusion, training set along with validation set is required for generating<br />

“training response vector set” that is used for learning each classifier and to construct<br />

LDA or nonparametric LDA based eigenmodel at classifier output space. The sample<br />

distribution for a single subject (which is same for feature fusion and decision fusion)<br />

over the three sets for all three databases are given in Table 4.4.<br />

Table 4.4: Sample distribution (per subject) in training, validation and testing sets<br />

for Yale, ORL and PIE databases.<br />

Set Yale ORL PIE<br />

Training 4 3 4<br />

Validation 3 4 12<br />

Testing 4 3 26<br />

The unique decomposition <strong>of</strong> a face into null space and range space is displayed<br />

in Fig 4.6, with three faces drawn from Yale, ORL and PIE databases respectively.<br />

Fig. 4.7 shows the images used for training for a single subject on Yale, ORL and<br />

PIE databases. Table 4.5 shows the performance <strong>of</strong> null and range spaces for Yale,<br />

ORL and PIE databases. The performances <strong>of</strong> our dual space based methods using<br />

96

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!