A system for face recognition in colored frontal face
images is proposed. This study presents a new scheme for feature extraction
based on deformable models and morphs using the 2D wavelet multi-resolution
decomposition, then, a classical semi-parametric system of a multi layered
perceptron neural network is employed for classification purpose. A heuristic
is designed for defining the face only bounding rectangle thus excluding
most of the facial hair, ears and neck and then we defined depending on
this computed rectangle four regions of interest, representing the forehead,
eye`s sockets, nose and chin regions. This heuristic also gives our system
invulnerability against both translations and z-axis limited rotations.
Next, 2D wavelet coefficients for the three channels of red, green and
blue are computed for the whole image using the multi-resolution decomposition.
A classification system of back propagation multi layered perceptron neural
network is designed and trained with momentum learning and the cross validation
during network training in the search procedure and the hyper tangent
nonlinearity as an activation function. Our system is experimented with
colored faces from the Stirling University database and the preliminary
results we obtained show an 88% success rate.