26-11-2012, 11:44 AM
A Multi-Algorithmic Face Recognition System
1A Multi-Algorithmic Face.pdf (Size: 292.97 KB / Downloads: 41)
Abstract
The importance of utilising biometrics to establish personal
authenticity and to detect impostors is growing in the present
scenario of global security concern. Development of a
biometric system for personal identification, which fulfills the
requirements for access control of secured areas and other
applications like identity validation for social welfare, crime
detection, ATM access, computer security, etc. is felt to be the
need of the day. Face recognition has been evolving as a
convenient biometric mode for human authentication for
more than last two decades. Several vendors around the
world claim the successful working of their face recognition
systems. However, the Face Recognition Vendor Test
(FRVT) conducted by the National Institute of Standards and
Technology (NIST), USA, indicates that the commercial face
recognition systems do not perform up to the mark under the
variations ubiquitously present in a real-life situation.
Availability of a largely accepted robust face recognition
system has proved elusive so far. Keeping in view the
importance of indigenous development of biometric systems
to cater to the requirements at BARC and elsewhere in the
country, the work was started on the development of a facebased
biometric authentication system. In this paper, we
discuss our efforts in developing a face recognition system
that functions successfully under a reasonably constrained
set-up for facial image acquisition.
INTRODUCTION
Biometrics makes automated use of the unique personal
features to establish the identity of a person. It is a tool for
positive identification of a human subject as biometric
signatures cannot be stolen, forgotten, lost or communicated
to another, as is possible in the case of authentication
employing cards, keys or passwords, so common in day-today
use. In the present scenario of increased security concern,
the necessity and relevance of making use of biometrics to
establish personal identity and to detect impostors are
assuming significance. Biometric techniques are based on
either physiological characteristics (like finger print, iris, etc.)
or behavioral traits (like voice dynamics, gait, etc.).
Depending upon the suitability in a particular
application, one has to choose a particular biometric [15] to
be used as the basic signature for recognition. We selected
the face based approach on the considerations that facial
imaging, being non-intrusive, has easy client acceptance,
apart from the fact that face recognition is the most natural
means of biometric identification for human beings. The
present circumstances around us demand increased level of
security and, therefore, machine capability of personal
identification from the facial image is considered invaluable.
Face-based biometric systems have the potential to fulfill the
requirements for access control of secured areas, surveillance,
social welfare, law enforcement, etc. Although there are a
few commercial face recognition systems available around
the globe, their performance is not up to the mark under the
practical variabilities [5]. This encouraged us to take up face
recognition system development.
BACKGROUND OF THE WORK
For over two decades [4] face recognition has drawn attention
of the research community. Face identification from a single
image is a challenging task because of variable factors like
alterations in scale, location, pose, facial expression,
occlusion, lighting conditions and overall appearance of the
face. With the synergy of efforts from researchers in diverse
fields including computer engineering, mathematics,
neuroscience and psychophysics, different frameworks have
evolved for solving the problem of face recognition. Among
these, the prominent approaches are those based on Principal
Component Analysis (PCA), Local Feature Analysis (LFA),
Template Matching, Neural Network, Model Matching,
Partitioned Iterated Function System (PIFS), Wavelets and
Discrete Cosine Transform (DCT). The choice of a particular
solution is governed by its suitability in a particular
application.
In PCA method, also known as Eigenface method,
face images are projected onto the so called eigenspace [6]
that best encodes the variations among known facial classes,
and recognition is achieved by carrying out match of these
projected feature vectors. The advantage of this method is
real-time recognition, but the method in itself is sensitive to
change in illumination, facial orientation and its size. LFA [8]
method of recognition is based on analysing the face in terms
of local features, e.g., eyes, nose, etc. by what is referred to as
LFA kernels. LFA technique offers better robustness against
local variations on the facial image in carrying out a match,
but does not account for global facial attributes. Face
recognition based on template matching [13] represents a
face in terms of a template consisting of several masks
enclosing the prominent features e.g. the eyes, the nose and
the mouth.
MULTI-ALGORITHMIC FACE
RECOGNITION
Earlier we developed a technique [1] of verifying human
face by matching against templates retrieved from the
reference database created during registration process. In this
technique the matching is carried out in terms of a set of
correlation scores corresponding to different areas of interest
(rectangular bit-maps representing different regions of the
face) and their Euclidean distances measured in pixels. This
technique can be extended for identification by matching the
input face against every registered identity to choose the one
which gives the best correlation scores and minimum
distance error. Unfortunately, correlation is computeintensive,
and to match a face against a large number of
reference faces using correlation software takes unacceptably
long duration for a practical application.
A PCA-based Approach
A survey [2,4] on the available literature revealed the main
techniques of face recognition which are mentioned in
section II. We decided to implement the PCA method of
recognition first. In this method as well as those described in
subsections 3B and 3C, our prototype recognition system
(described in the later part of subsection 3C) makes use of
reasonably constrained imaging set-up with facial image
grabbed in frontal geometry under sufficient illumination
level in order to minimise the variations in the acquired
image.
CONCLUSIONS
Recognition of identity of the person based on facial
biometric signature can be used in many applications such as
access control of secured areas, tele-banking, surveillance,
law enforcement, etc. This paper describes a prototype
system for carrying out human face recognition based on a
combination of correlation with PCA or DCT technique used
for the purpose of deciding a facial match. The system has
been tested on a data set of 109 images belonging to 43
subjects.
Recognition of a person from a 2D projected image
of the 3D face is a challenging task because of pose
variations of the head along with change in facial appearance.
Our prototype recognition system makes use of reasonably
constrained imaging set-up with facial image grabbed in
frontal geometry under sufficient illumination level in order
to minimise the variations in the acquired image. The size of
the face in the image acquired during recognition varies due
to inadvertent and inevitable head movement occurring
between successive instants. Our algorithm tackles this
problem by computing the exact scale of the face from interocular
distance and normalizing with respect to the
registration data. Our technique performs successfully under
linear change of intensity level in the face image. However,
presently, any nonlinear intensity change cannot be
counteracted, although there are few measures suggested
regarding it in the literature [10].