Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
the main problem; whereas the problems for a machine face recognition system are:<br />
1. Facial expression change<br />
2. Illumination change<br />
3. Aging<br />
4. Pose change<br />
5. Scaling factor (i.e. size <strong>of</strong> the image)<br />
6. Frontal vs. pr<strong>of</strong>ile<br />
7. Presence and absence <strong>of</strong> spectacles, beard, mustache etc.<br />
8. Occlusion due to scarf, mask or obstacles in front.<br />
The problem <strong>of</strong> automatic face recognition (AFR) is a composite task that involves<br />
detection <strong>of</strong> faces from a cluttered background, facial feature extraction, and face<br />
identification. A complete face recognition system has to solve all subproblems, where<br />
each one is a separate research problem. This research work concentrates on the<br />
problem <strong>of</strong> facial feature extraction and face identification.<br />
Most <strong>of</strong> the current face recognition algorithms can be categorized into two classes,<br />
image template based and geometry feature-based. The template based methods [9]<br />
compute the correlation between a face and one or more model templates to estimate<br />
the face identity. Brunelli and Poggio [16] suggest that the optimal strategy for face<br />
recognition is holistic and corresponds to template matching. In their study, they<br />
compared a geometric feature based technique with a template matching based system<br />
and reported an accuracy <strong>of</strong> 90% for the first one and 100% for the second one on a<br />
database <strong>of</strong> 97 persons. Statistical tools such as Support Vector Machines (SVM) [92,<br />
127], Principal Component Analysis (PCA) [114, 124], Linear Discriminant Analysis<br />
(LDA) [12], kernel methods [109, 136], and neural networks [50, 74, 97] have been<br />
used to construct a suitable set <strong>of</strong> face templates. Other than statistical analysis<br />
and neural network approach there are other approaches known as hybrid approaches<br />
which use both statistical pattern recognition techniques and neural network systems.<br />
Examples for hybrid approaches include the combination <strong>of</strong> PCA and Radial Basis<br />
Function (RBF) neural network [37, 122]. Among other methods, people have used<br />
10