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.
on three face databases and compared with the state-<strong>of</strong>-art closest competitor [36] <strong>of</strong><br />
it. Section 3.5 concludes the chapter.<br />
3.1 Introduction and Motivation<br />
How do humans identify individuals with remarkable ease and accuracy? This ques-<br />
tion has haunted psychologists, neurologists and, recently, engineers in biometry for a<br />
long time. The human face is different from any other natural or man-made objects,<br />
but has similar structural features across different races, sex and regions. The subtle<br />
variations in the face structure are captured by the human brain and help in discrim-<br />
inating one face from the another. The human brain is able to filter out the common<br />
visual features <strong>of</strong> a face and retain only those suitable to exhibit the unique charac-<br />
teristics (discriminatory evidence) <strong>of</strong> an individual. An efficient face-based biometric<br />
system must possess these properties (as in those in the retina and visual cortex), to<br />
perform efficiently like a human being.<br />
Wavelet-based features have already been used to obtain a better face representa-<br />
tion in [36, 40, 26, 140]. The use <strong>of</strong> wavelet packet for face recognition was reported in<br />
[40]. Here, 2D-DWT is used to fully decompose the face image and simple statistical<br />
features such as mean and variance are extracted from the decomposed coefficients,<br />
and used as a feature vector for representation. The use <strong>of</strong> wavelet subband represen-<br />
tation along with kernel associative memory for face recognition was reported in [140].<br />
Here wavelet decomposed faces are used to build an associative memory model for<br />
each class and kernel methods are used to exploit higher order relations which cannot<br />
be captured by linear transformations. Face images with illumination variations can<br />
be modeled by low-dimensional linear subspaces. The existence <strong>of</strong> single light source<br />
directions that are effective for face recognitions was discussed in [75]. Ekenel [36] et<br />
al. proposed multiresolution face recognition with fusion at data, feature and deci-<br />
sion levels to test on face images that differ in expression or illumination separately,<br />
obtained from CMU PIE, FERET and Yale databases. Significant performance gains<br />
are reported against illumination perturbations. They selected a number <strong>of</strong> subbands<br />
on the basis <strong>of</strong> performance on testing set and termed them as successful subbands.<br />
Data, feature and decision level fusion are performed on these successful subbands to<br />
45