Information Technology Reference
In-Depth Information
the nose tip and nose dip along y-axis and nose tip as the origin of the coordinate
system.
6.4.2 Recognition
Recent development in computer technology and call for better security applications
have brought biometrics into focus. The signature, handwriting, and fingerprint have
a long history. More recently voice, retinal scan, iris scan, and face information are
considered for biometrics. When deploying a biometrics-based system, we consider
its accuracy, cost, ease of use, whether it allows integrationwith other systems, and the
ethical consequences of its use. This chapter focus on 3D pattern classification using
neural network. Our method has successfully performed recognition irrespective
of variability in head pose, direction, and facial expressions. We present here two
illustrative examples to show how a neural network based on 3D real-valued neurons
can be used to learn and recognize point cloud data of 3D faces. A 1-2-1 network of
vector-valued neurons was used in following two experiments. The proposed pattern
classifier structure involves the estimation of learning parameters (weights) which
are stored for future testing. It is more compact and can be easily communicated to
humans than learned rules.
Example 6.4 The 3D vector-valued neural network was trained by a face (Fig. 6.3 a)
from first set of face data (Fig. 6.3 ). This face data contains five faces of different
persons, where each face contains 6397 data pints. Table 6.1 presents the testing error
yielded by trained network for all five faces (Fig. 6.3 ). The testing error for four other
faces is much higher in comparison to the face that is used in training. Thus, trained
network recognize the face, which is taken in training, and reject four faces of other
persons. Results bring out the fact that this methodology is able to learn and classify
the 3D faces correctly.
Example 6.5 In this example, the considered network was trained by first face
(Fig. 6.4 a) from the face set (Fig. 6.4 ). This face set contains five faces of same person
with different orientation and poses. Each face contains 4663 data pints. Table 6.2
presents the testing error yielded by trained network for all five faces (Fig. 6.4 ).
The testing error for four other faces is also minimum and comparable to the face,
which is used for training. Thus, trained network recognize all faces of same person.
Thus, considered methodology has successfully performed recognition irrespective
of variability in head pose and orientation.
 
Search WWH ::




Custom Search