Neurocomputing in Space - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 122

Example 6.3 A neural network based on 3D vector-valued neurons has been trained
with line for the composition of scaling and translation. Figure 6.2 b presents the gen-
eralization ability of this trained network. There are 451 test data points on cylinder.
All points in 3D are contracted by factor 1 . Results

bring out the fact that given neural network is able to learn and generalize 3Dmotion.

/ 2 and displaced by ( 0 , 0 , 0 . 2 )

6.3 Point Clouds of Objects in Practical Application

The CVNN was applied to the various benchmarks and mapping problems in the
previous chapters. In the present chapter, the problem of 3D object classification is
performed through the sorting of point clouds of objects using vector-valued neural
network. The wide spectrum of problems in science and engineering deal in con-
structing and analyzing surfaces from clouds of points. A typical problem in the area
has a data set at hand obtained by running a scanning-equipment across the object
of interest. The scan must be performed in an orderly fashion to ensure the data
points are well organized (and do not appear at irregular intervals or appear disor-
derly). After the operation, a point cloud of the object gets generated that is subject
to further analysis. So the algebraic functions (polynomials) may also be employed
in many cases to fit point clouds. It must be acknowledged, however, that the actual
form of the surface of object if known would give a more accurate picture of the
surface and the following analysis more correctly placed. The problem of object
classification is a problem of mapping, which refers to sorting a set of object into
categories with predefined characteristics, in which the number of classes is usually
fixed. To start with, the 3D vector-valued back-propagation algorithm (3DV-BP) may
be run to train the neural network architecture of appropriate size. To keep the study
uniform, a same size of neural network architecture is chosen with uniform learn-
ing rate and the training epochs. To test the neural network another sets of object's
surfaces are constructed.

6.4 3D Face Recognition

Biometrics can be defined as the automated use of physiological (face, finger prints,
periocular, iris, Oculomotor plant characteristic, and DNA) and behavioral (signature
and typing rhythms) characteristics for verifying the identity of living person. The
physiological features are often non-alterable except severe injury, while behavioral
features may fluctuate due to stress, fatigue, or illness. Face recognition is one of
the few biometric methods that possess the merits of both high accuracy and low
intrusiveness. It is also one of the most acceptable biometrics because a human face
is always bare and often used in their visual interactions. It is a potential identify of
a person without document for identification.