Professional Documents
Culture Documents
RBF nets are generally called upon Available results on neural network based
for use in function approximation problems, approaches come from experiments with
particularly for interpolation. few individuals, what makes it difficult to
FACE RECOGNITION compare with other reported approaches. All
3. SYSTEM those works rely on a preprocessing to
Previous work on face recognition: detect a face in a scene and to compensate
Earlier face recognition systems were for variation of lighting,position,rotation and
based mainly on geometric facial features scale.
and template matching. In those works a The work reported here studies a face
face was characterized by a set of features recognition system consisiting of a standard
such as mouth position, chin shape, nose PCA used for dimensionality reduction,
width and length which are potentially followed by RBF network acting as a
insensitive to illumination conditions. classifier. As in the most approaches
Brunelli et al. (1993) compared this mentioned before, the database used for the
approach with a traditional template evaluation contains face images with
matching scheme which produced higher moderate lighting, rotation, scale and
recognition rates for the same face database viewing variation .A previous work has
(90% against 100%) Cox,Ghosn and indicated that a RBF network performs
Yianilos (1996) [11] proposed a mixture better than conventional distance classifiers
distance technique which achieved the best The present work focus on the study of
alternative network optimization to common one is the Gaussian function
maximize the recognition rate. defined by,