The task of face recognition has been actively researched in
recent years. This paper provides an up-to-date review of major human face
recognition research. It first presents an overview of face recognition and its
applications. Then, a literature review of the most recent face recognition
techniques is presented. Description and limitations of face databases which
are used to test the performance of these face recognition algorithms are given.
Approach of face recognition aims to detect faces in still image and sequence
image from video have many methods such as local, global, and hybrid approach.
The main problem of face recognition is intensity, illumination, pose,
difficult to controlling and large occlusion. In 3D capture creates larger data
files per subject which applies significant storage requirements, slow
processing, most new devices can be capture in 3D. This is the problem for our
future work that want to solve and create accuracy gain for widely accept in 3D
face recognition system.
Face recognition is an important research problem spanning many fields
of study. This because face recognition is having many practical uses such as bank
card identification, access control, security identification and is a basic human behaviour that is extremely
important for effective communications andinteractions among people.
Progress has advanced to the point that face recognition systems are being proved in real-world settings. The fast development of face recognition is due to a combinationof factors: active development of sets of computer instructions,
the availability of a large of facial images,
and a method for the performance of face recognition sets
of computer instructions.
In the literatures, face recognition problem can be created as: given static (still) or video images
of a scene, check one or more people in the scene by comparing
with faces stored in a computer
file. When comparing person checking to face recognition, there are more aspects which differ. First a
client – an approved user of a personal identification system – is assumed to be
cooperative and makes an identity claim.
Face recognition is one of the few methods that possess
the good qualities of both high and low intrusiveness. It
has the accuracy of a physiological approach without being intrusive. Over past 30 years, many people have
proposed different face recognition ways of doing things, by the increased number
of real world computer programs needing the recognition of human faces. There are
several problems that make automatic face recognition a very
However, the face image of a person inputs to the database that is
usually received under different conditions. The important
of automatic face recognition is much be cope with numerous variations of
images of the same face due to changes in the following parameters such as
pose, illumination, expression, motion, facial hair, glasses, and background.
Computationally this means that it is not necessary to
consult the complete set of database images in order to verify a claim. An
incoming image is thus compared to a small number of model images of the person
whose identity is claimed and not, as in the recognition scenario, with every
image in a potentially large database. Second, an automatic authentication
system must operate in near-real time to be acceptable to users. Finally, in
recognition experiments, only images of people from the training database are
presented to the system, whereas the case of an imposter is of utmost
importance for authentication. Face recognition is a biometric approach that
employs automated methods to verify or recognize the identity of a living
person based on his/her physiological characteristics. In general, a biometric
identification system makes use of either physiological characteristics or
behaviour patterns to identify a person. Because of human inherent
protectiveness of his/her eyes, some people are reluctant to use eye
identification systems. Face recognition has the benefit of being a passive,
non-intrusive system to verify personal identity in a “natural” and friendly
All face recognition algorithms consistent of two major
parts: face detection and normalization and face identification. Algorithms
that consist of both parts are referred to as fully automatic algorithms and
those that consist of only the second part are called partially automatic
algorithms. Partially automatic algorithms are given a facial image and the
coordinates of the centre of the eyes. Fully automatic algorithms are only
given facial images. On the other hand, the development of face recognition
over the past years allows an organization into three types of recognition
algorithms, namely frontal, profile, and view tolerant recognition, depending
on the kind of images and the recognition algorithms. While frontal recognition
certainly is the classical approach, view-tolerant algorithms usually perform
recognition in a more sophisticated fashion by taking into consideration some
of the underlying physics, geometry, and statistics. Profile schemes as
stand-alone systems have a rather marginal significance for identification.
However, they are very practical either for fast coarse pre-searches of large
face database to reduce the computational load for a subsequent sophisticated
algorithm, or as part of a hybrid recognition scheme. Such hybrid approaches
have a special status among face recognition systems as they combine different
recognition approaches in an either serial or parallel order to overcome the
shortcoming of the individual components. Another way to categorize face
recognition techniques is to consider whether they are based on models or
exemplars. Models are used in to compute the Quotient Image, and in to derive
their Active Appearance Model. These models capture class information, and
provide strong constraints when dealing with appearance variation. At the other
extreme, exemplars may also be used for recognition. The ARENA method in simply
stores all training and matches each one against the task image. As far we can
tell, current methods that employ models do not use exemplars, and vice versa.
This is because these two approaches are by no means mutually exclusive. 2 Literature
2 Literature Review
Face recognition can be both in still image and video
sequence which comes from still-image face recognition. There are different
methods and techniques that apply to frontal faces such as: eigenfaces, neural
network, dynamic link architecture, hidden Markov model, geometrical feature
matching and template matching. The approaches are analysed by the facial
representation they used.
Eigenfaces is the most thoroughly investigated approach to
face recognition also known by the name Karhunen-Loeve expansion. It uses
principal component analysis to represent pictures of faces. The weights
describing every face is obtained by projecting the face onto the eigen
In mathematical terms eigenfaces are the most important
components of the distribution of faces.
The eigenvectors are ordered to represent different amounts
of the variation, respectively, among the faces. Each face can be represented
by a linear combination of the eigenfaces. It can also be approximated using
only the “best” eigenvectors with the largest eigenvalues. The best M
eigenfaces construct an M dimensional space, i.e., the “face space. As the
images include a large quantity of background area, the results are influenced
by background. There is a robust performance of the system under different
lighting conditions by correlation between images with changes in illumination.
However, showed that the correlation between images of the whole faces is not
efficient for satisfactory recognition performance. Illumination normalization
is usually necessary for the eigenfaces approach.
L. Zhao and Y.H. Yang proposed another strategy to figure
the covariance grid utilizing three pictures every one taken in distinctive
lighting conditions to represent subjective light impacts, if the object is
Lambertian. Reference expanded their initial work on eigen face to eigen
highlights comparing to face segments, for example, eyes, nose, and mouth. They
utilized a secluded eigenspace which was created of the above eigenfeatures
(i.e., eigen eyes, eigen nose, and eigen mouth). This strategy would be less
touchy to appearance changes than the standard eigenface strategy. The
framework accomplished an acknowledgment rate of 95 percent on the FERET
database of 7,562 pictures of roughly 3,000 people. In synopsis, eigenface
shows up as a quick, straightforward, what’s more, down to earth strategy. Be
that as it may, by and large, it doesn’t give invariance over changes in scale
and lighting conditions.
As of late, exploring different avenues regarding ear and
face recognition, utilizing the standard key segment investigation approach,
demonstrated that the recognition execution is basically indistinguishable
utilizing ear pictures or face pictures and joining the two for multimodal
recognition brings about a factually noteworthy execution change. For instance,
the contrast in the rank-one recognition rate for the day variety analyse
utilizing the 197-picture preparing sets is Worldwide Journal of Signal
Processing 2;2 2006 90.9% for the multimodal biometric versus 71.6% for the ear
also, 70.5% for the face.
2.2 Neural Network
The appeal of utilizing neural systems could be expected
to its non-linearity in the system. Henceforth, the component extraction step
might be more proficient than the straight Karhunen-Loève techniques. One of
the primary artificial neural systems (ANN) systems utilized for face recognition
is a solitary layer versatile arrange called WISARD which contains a different
for stored person. The route in building a neural system
structure is urgent for effective recognition.
It is particularly reliant on the proposed application.
For face location, multi-layer perceptron and ordinary neural system have been
connected. For face check, is a multi-determination pyramid structure.
Reference 37 proposed a half breed neural system which joins nearby
picture inspecting, a self-sorting out guide (SOM) neural
arrange, and a convolutional neural system. The SOM gives a quantization of the
picture tests into a topological space where inputs that are adjacent in the
first space are additionally adjacent in the yield space, along these lines
measurement diminishment and invariance to minor changes
in the picture test. The convolutional organize extricates progressively bigger
highlights in a various levelled set of layers and
gives incomplete invariance to interpretation, pivot,
scale, and deformation. The creators detailed 96.2% right recognition on ORL
database of 400 pictures of 40 people.
The characterization time is under 0.5 second, yet the
preparing time is the length of 4 hours. Reference 39 utilized probabilistic
choice based neural system (PDBNN) which acquired the secluded structure from
its ancestor, a choice based neural system (DBNN). The PDBNN can be connected
1) face locator: which finds the area of a human face in a
2) eye localizer: which decides the places of the two eyes
keeping in mind the end goal to produce significant element vectors
3) face recognizer.
PDNN does not have a completely associated arrange topology.
Rather, it separates the system into K subnets. Every subset is committed to
remember one individual in the database. PDNN employments the Guassian
initiation work for its neurons, and the yield of each “face subnet”
is the weighted summation of the neuron yields. At the end of the day, the face
subnet gauges the probability thickness utilizing the prominent blend
of-Guassian show. Contrasted with the AWGN plot, blend of Guassian gives a
substantially more adaptable and complex model for approximating the time
probability densities in the face space.
The learning plan of the PDNN comprises of two stages, in
the main stage; each subnet is prepared by its own particular face pictures. In
the second stage, called the choice based taking in, the subnet parameters
might be prepared by a few specific examples from other face classes. The
choice based learning plan does not utilize all the preparation tests for the
preparing. Just misclassified designs are utilized. On the off chance that the
example is misclassified to the wrong subnet, the legitimate subnet will tune
its parameters with the goal that its choice locale can be drawn nearer to the
PDBNN-based biometric recognizable proof framework has the
benefits of both neural systems and factual methodologies, and its conveyed
figuring rule is generally simple to execute on parallel PC. In 39, it was
accounted for that PDBNN face recognizer had the ability of perceiving up to
200 individuals and could accomplish up to 96% right acknowledgment rate in around
1 second. In any case, when the quantity of people builds, the processing cost
will turn out to be all the more requesting. When all is said in done, neural
system approaches experience issues when the quantity of classes (i.e., people)
increments. Additionally, they are not appropriate for a solitary model picture
acknowledgment test on the grounds that different show pictures per individual
are essential all together to train the frameworks to “ideal”
2.3 Graph Matching
Graph matching is another way to deal with face recognition.
Reference 41 introduced a dynamic connection structure for distortion
invariant object recognition which utilized versatile diagram coordinating to
locate the nearest stored chart. Dynamic connection design is an expansion to
traditional manufactured neural systems. Remembered objects are spoken to by
scanty diagrams, whose vertices are named with a multiresolution depiction
regarding a nearby power range and whose edges are marked with geometrical
separation vectors. Object recognition can be planned as versatile diagram
coordinating which is performed by stochastic advancement of a coordinating
expense work. They announced great outcomes on a database of 87 individuals and
a little arrangement of office things containing extraordinary articulations
with a revolution of 15 degrees.
The coordinating procedure is computationally costly, taking
around 25 seconds to contrast and 87 stored objects on a parallel machine with
23 transputers. Reference 42 broadened the system and coordinated human
appearances against an exhibition of 112 impartial frontal view faces. Test
pictures were misshaped because of pivot top to bottom and changing outward
Empowering comes about on faces with substantial revolution
points were gotten. They announced recognition rates of 86.5% and 66.4% for the
coordinating trial of 111 countenances of 15-degree pivot and 110 countenances
of 30-degree pivot to an exhibition of 112 impartial frontal perspectives. By
and large, dynamic connection design is unrivalled to other face recognition
strategies as far as pivot invariance; be that as it may, the coordinating
procedure is computationally costly.
2.4 Hidden Markov Models (HMMs)
Stochastic displaying of nonstationary vector time
arrangement in view of (HMM) has been extremely effective for discourse applications.
Reference 43 connected this technique to human face recognition. Countenances
were naturally isolated into locales for example, the eyes, nose, mouth, and so
on., which can be related with the conditions of a shrouded Markov demonstrate.
Since HMMs require a one-dimensional perception grouping and pictures are
two-dimensional, the pictures ought to be changed over into either 1D fleeting
arrangements or 1D spatial successions.
In 44, a spatial perception grouping was extricated from a
face picture by utilizing a band testing strategy. Each face picture was spoken
to by a 1D vector arrangement of pixel perception. Every perception vector is a
piece of L lines and there is a M lines cover between progressive perceptions. An
obscure test picture is first inspected to a perception arrangement. At that
point, it is coordinated against each HMMs in the show face database (each HMM represents
an alternate subject). The match with the most astounding probability is
considered the best match and the important model uncovers the personality of the
The recognition rate of HMM approach is 87% utilizing ORL database
comprising of 400 pictures of 40 people. A pseudo 2D HMM was accounted for to
accomplish a 95% recognition rate in their preparatory analyses. Its
arrangement time and preparing time were not offered (accepted to be extremely
costly). The selection of parameters had been founded on subjective instinct.
2.5 Geometrical Feature Matching
Geometrical feature matching systems depend on the calculation
of an arrangement of geometrical highlights from the photo of a face. The way
that face recognition is conceivable even at coarse determination as low as 8×6
pixels when the single facial highlights are barely uncovered in detail, infers
that the general geometrical design of the face highlights is adequate for recognition.
The general setup can be depicted by a vector representing to the position and
size of the fundamental facial highlights, for example, eyes and eyebrows,
nose, mouth, also, the state of face layout.
One of the pioneering works on automated face recognition by
utilizing geometrical highlights was finished by T. Kanade in 1973. Their framework
accomplished a peak execution of 75% recognition rate on a database of 20 individuals utilizing
two pictures for each individual, one as the model and alternate as the test
picture. References 47,48 demonstrated that a face recognition program
furnished with highlights removed physically could perform recognition clearly
with agreeable outcomes. Reference 49 consequently removed an arrangement of
geometrical highlights from the photo of a face, for example, nose width and
length, mouth position, and jaw shape. There were 35 highlights removed frame a
The recognition was
at that point performed with a Bayes classifier. They announced a recognition rate
of 90% on a database of 47 individuals. Reference 50 presented a blend remove
procedure which accomplished 95% recognition rate on an inquiry database of 685
people. Each face was spoken to by 30 physically removed separations. Reference
51 utilized Gabor wavelet deterioration to recognize highlight focuses for
each face picture which significantly decreased the capacity necessity for the database.
Normally, 35-45 highlight focuses per face were created. The coordinating
procedure used the data introduced in a topological realistic portrayal of the
element focuses. In the wake of adjusting for various centroid area, two cost
esteems, the topological cost, and closeness cost, were assessed. The recognition
exactness regarding the best match to the perfect individual was 86% and 94% of
the right individual’s faces was in the best three hopeful matches.
2.6 Template matching
A basic adaptation of template matching is that a test
picture spoken to as a two-dimensional exhibit of power esteems is analysed
utilizing a reasonable metric, for example, the Euclidean remove, with a
solitary format representing the entire face.
There are a few other more complex renditions of template
matching on face recognition. One can utilize more than one face format from
various perspectives to represent to a person’s face. A face from a solitary
perspective can likewise be represented by a set of numerous unmistakable
littler formats. In 49, Bruneli and Poggio consequently chosen an arrangement
of four highlights formats, i.e., the eyes, nose, mouth, and the entire face,
for the majority of the accessible countenances. They looked at the execution
of their geometrical matching algorithm and template matching algorithm on the
same database of faces which contains 188 pictures of 47 people. The template matching
was unrivalled in recognition (100 percent recognition rate) to geometrical matching
(90 percent recognition rate) and was additionally less difficult.
Since the primary segments (otherwise called eigenfaces or eigenfeatures)
are straight blends of the layouts in the information premise, the strategy
can’t accomplish preferred outcomes over correlation, yet it might be less
computationally costly. One downside of template matching is its computational intricacy.
Another issue lies in the depiction of these formats. Since the recognition
framework must be tolerant to certain inconsistencies between the format and
the test picture, this resilience may normal out the distinctions that make singular
When all is said in done, template based methodologies
contrasted with highlight matching are a more intelligent approach. In
synopsis, no existing system is free from constraints. Promote endeavours are required
to enhance the exhibitions of face recognition systems, particularly in the
extensive variety of situations experienced in certifiable.