Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 147

Face recognition is one of the few biometric methods that possesses the merits of
both high accuracy and low intrusiveness. It is also one of the most acceptable bio-
metrics because a human face is always bare and often used for its visual interactions.
It cannot be forgotten or mislaid like a password and it has the potential to character-
ize a person without any document for identification. The success of intelligent face
recognition systemmainly depends on the development of a computational model for
facial feature representation and classifiers for identification of an unknown person.

7.4 Recognition with Higher-Order Neurons
in Complex Domain

Classification needs to be performed on mathematical models (extracted features)
given by aforementioned statistical techniques. A neural network is an artificial intel-
ligence technique, which has many advantages for nonlinear classification because
of fast learning, better generalization, efficiency and robustness toward noise and
natural environment [ 50 , 51 ]. In case of learned networks, the weights can be eas-
ily communicated to humans for generalization than learned rules (in traditional
AI techniques). The efficient learning ability and incredible generalization of single
RSP and RSS neurons in complex domain have motivated to use them in design-
ing the OCON-based machine recognition system. In this chapter, an ensemble of
complex-valued neurons have been used to develop a OCON classifier.
The training of ANN classifiers involves the estimation of learning parameters
(weights) only, which are stored for future testing. Hence, it is most desirable to
search a structure for classifier which require minimum weights and yield best accu-
racy. The beauty of proposed classifier structure for image classification system is
that it uses an ensemble of proposed single neurons in complex domain instead of
ensemble of multilayer network. Only single RSS or RSP neuron has given solution
to many benchmark problems in previous chapters. Similar computational power
is also observed in image classification, where every single neuron is dedicated to
recognize images of its 'own class' assigned to it. The output of the ensemble is
forming an aggregate output of the classifier.
The number of neurons in ensemble is set to the number of image classes (sub-
jects). Each neuron is trained to give output '1

j ' for its own class and 0 for other
class. The number of nodes in input layer is equal to the number of elements in the
feature vector. The data values in the feature vectors are normalized to lie with in the
first quadrant of a unit circle. An input face image presented to a neuron associated
to its own class is considered as positive example, while images of other classes to
this neuron is considered as negative example. It yields better confidence in classifier
decision. The classifiers may be trained with error back-propagation learning algo-
rithm. For testing the identity claim of an image, a M dimensional feature vector
is extracted from the image and this vector is given to every neuron or network of
classifier. The discriminant function applied in the proposed system calculates the

+