IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013

ISSN: 2320 - 8791

[Link]

Face detection and its applications

Mrs. Sunita Roy1 and Mr. Susanta Podder2

1
Ph.D. scholar in the Dept. of Computer Science & Engineering, University of Calcutta, Kolkata, India.
2
Ph.D. scholar, CMJ University, Shillong, Meghalaya-793 003, India.

Abstract

In the past two decades, face detection has been proven Face detection is a procedure by which we can able
as the most interesting research field from the domain of to extract face region from a human body. Now, the
image processing. In this paper, we are going to describe concept can be implemented in various ways but
some important aspects of face detection, which are very mainly we use four steps for this implementation.
much useful in many applications like face recognition,
facial expression recognition, face tracking, facial feature
In the first step, we localize the face region that
extraction, gender classification, identification system, means we are anticipating those parts of an image
document control and access control, clustering, where a face may present. In the second step we
biometric science, human computer interaction (HCI) normalize the detected region, so that the
system, digital cosmetics and many more [1]. Before alignments of various facial features are in the
that, I would like to focus on some well-known face proper location. In the third step we extract various
detection techniques [2] and then some feature detection facial features like eyes, nose, mouth, etc. And in
methods because until and unless we extract the the forth step, we actually verify whether the
important features (eyes, nose, mouth) from a face we anticipated parts are actually carrying out a face or
won‟t be able to uniquely identify the person properly.
not. We are doing this using some rules, template
or image databases. The concept of extraction can
Keywords: Face detection, segmentation, facial
features, contour, feature restoration, fiducial point.
be implemented by various techniques. There are a
huge number of papers regarding the literature
survey of face detection [3]. Most of the earlier
1. Introduction work was on the frontal upright face, but recent
work is mainly focus on non-frontal face with
„Face detection‟ as the keyword itself reveals it‟s variation in their alignment. Also instead of still
meaning that it concerns about where a face is image, they are considering video stream images.
located in an image. Now it may seem very easy
but in reality we have to consider many constraints
like single face or multiple faces, image rotation,
pose etc. So there may arise some false detected
regions of an image, which do not contain any face.
In spite of all these problems there are lots of
techniques available. Here we start our journey
from the introduction of face detection and the
areas where it is used.
Figure 2: Steps of face detection

1.1 Localization

In this step we locate the region of an image where

the face is located [4]. A face region only contains
some facial features, now the number of features is
application dependent. In the localization process
we may go through some problems like false
detection due to the presence of some obstacles on
Figure 1: Result of a face detection algorithm the face, poor quality of picture, orientation of head
position, expression, etc. Hence we have to
consider all these limitations to make the face
detection process more powerful.

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
1
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

1.2 Normalization
2.2 Feature invariant approaches
After properly locating the region that contains a
face, we must normalize the face region [5]. Using These algorithms aim to find structural features that
normalization process, we align the face region in exist even when the pose, viewpoint, or lighting
such a way that all the facial features are in their conditions vary, and then use these to locate faces.
proper location. Not only this, we may have to To distinguish from the knowledge-based methods,
scale, rotate the image or some other the feature invariant approaches start at feature
transformation to correlate it with the entry in that extraction process and face candidates finding, and
database. later verify each candidate by spatial relations
among these features, while the knowledge-based
1.3 Facial feature extraction methods usually exploit information of the whole
image and are sensitive to complicated
In this step of face detection, we extract various backgrounds and other factors. Readers could find
facial features like eyes, nose, mouth, etc from the more works in [4][7][8]. Face detection based on
detected face region [4]. There are three types of color information, random labeled graph matching
feature extraction methods: Generic methods based fall in this category.
on edges, lines, and curves; Feature template based `
methods that are used to detect various facial 2.3 Template matching methods
features like eyes, nose, and mouth; Color
segmentation based methods that use face color In this category, several standard patterns of a face
instead of the intensity values; Appearance based are stored to describe the face as a whole or the
methods that are able to manage changes in facial feature separately. The correlations between
illumination conditions, shape, pose and reflectance an input image and the stored pattern are computed
and even to handle translation and partial for detection. These methods have been used for
occlusions. both face localization and detection. Deformable
template matching [9] falls in this category, where
1.4 Verification the template of faces is deformable according to
some defined rules and constraints.
In the verification process, we verify the
relationships between various features with some 2.4 Appearance-based methods
database entry containing a huge number of faces.
Now verification can not be done using only In contrast to template matching, the models (or
database entry, but also we can use some rule based templates) are learned from a set of training
techniques that uses the correlations of various images, which should capture the representative
facial features as its parameter, or by using so variability of facial appearance. These learned
template based methods where we use a specific models are then used for detection. More
template model and try to find out a face region significant techniques are included in [1][2].
that fits into this model. Examples of such type of methods are view-based
face detection, Haar features and the Adaboost
2. Face detection techniques algorithm.

A huge number of representation techniques are 2.5 Part-based methods

available for face detection, including Knowledge-
based, Feature invariant based, Template matching With the development of the graphical model
method, Appearance-based methods, Part-based framework and the point of interest detection such
methods, etc. as the difference of Gaussian detector [10] (used in
the SIFT detector) and the Hessian affine detector,
2.1 Knowledge-based methods the part-based method recently attracts more
attention. Some well-known approaches like face
These rule-based methods encode human detection based on the generative model
knowledge [6] of what constitutes a typical face. framework, component-based face detection based
Usually, the rules capture the relationships between on the SVM classifier falls into this category.
facial features. These methods are designed mainly
for face localization, which aims to determine the 3. Face recognition
image position of a single face.

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
2
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

Face detection is the first step towards many were motivated to locate features within faces.
applications; one of them is face recognition. To Their approach utilized a priori information to
recognize a face, we first detect the face and then accurately find local features. Their implementation
we compare it with a set of known individuals consisted of two parts: first part was designed to
present in a database to verify the identity of the identify individual features, such as the eyes
person. The concept of face recognition can be (general location of the eyes), cheek, hair, jaw-line,
further extended to various biometric approaches mouth, mouth bits (edges and outline of the lips),
including fingerprint, iris/retina and voice head outline and the nose; the second part refined
recognition. the features found from the first part by using a
priori information to locate all pre-defined face
Face recognition techniques can be classified as features.
two main approaches: Geometric approach or
Feature-based approach where we analyze various 3.1.2 Elastic Bunch Graph Matching
features by means of their relationships [11] and
holistic approach [12] such as Eigenfaces, neural A very well known feature-based approach is the
networks [13]. Elastic Bunch Graph Matching method proposed
by Wiskott et al. [15]. This technique is based on
Dynamic Link Structures [16]. In these methods a
face is considered as a graph having several inter
connected nodes, which are called fiducial point.
Each node corresponds to a specific feature point.
For example, a node may corresponds to an eye,
another node may corresponds to nose etc. There is
also an arch between two fiducial points labeled
with the distance between the correspondent nodes.
Therefore the nodes for the input face graph are
interconnected to form a graph like data structure,
which is fitted to the shape of the face.

Here we recognize faces by matching the probe set

Figure 3: Face recognition using geometric represented as the input face graphs, to the gallery
approach set that is represented as the model face graph.
Hence, this model face graph can be conceptually
3.1 Geometric approaches or Feature-based thought of as a number of input face graphs stacked
approach to recognize a face on top of each other and concatenated to form one
model face graph, with the exception that this is
In Feature-based approaches [17] we first applied to the gallery set instead of the probe set.
preprocess the input image to remove the noise, Therefore, this would allow the grouping of the
and then we extract distinctive facial features such same types of face features from different
as the eyes, mouth, nose, etc., and then compute the individuals. For example, the eyes of different
geometric relationships among those facial points, individuals could be grouped together to form the
thus reducing the input facial image to a vector of eye feature point for the model face graph and the
geometric features. Standard statistical pattern noses of different individuals can be grouped
recognition techniques are then employed to match together to form the nose feature point for the
faces using these measurements. Most of the model face graph.
previous works was based on this technique.
3.2 Holistic based face recognition
3.1.1 General approaches
Unlike Feature based method, holistic based
The general approaches to feature based face approaches use the global information rather that
recognition are concerned with using a priori local feature information of the face. Here we
information of the face to find local face features. represent the entire image with some small key
Alternatively, another general approach is to find values, which are directly derived from the pixel
local significant geometries of the face that information of face images. This small key
correspond to the local features of faces. We will information is sufficient to uniquely differentiate
now discuss the general approaches that have been individual faces. Here we describe two holistic
applied to face recognition. Craw et al. (1992) [14] approaches to face recognition called statistical and

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
3
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

AI approaches. An overview of some of the recognizes faces of the same individual. A small
methods in these categories follows. number of features are found by maximizing the
Fisher Discriminant Criterion (Fisher 1936) [20],
3.2.1 Statistical approach which is achieved by maximizing the grouping of
individual faces whilst minimizing the grouping of
different individual faces. Therefore by grouping
In this simplest approach we represent the image as
faces of the same individual these features can be
a 2D array containing the intensity values and
used to determine the identity of individuals.
recognition is performed by direct correlation
comparisons between the input face and all the
As highly structured two-dimensional patterns,
other faces in the database. Though it seems very
human face images can be analyzed in the spatial
simple to implement but they often face some
and the frequency domains. These patterns are
common problems like variation in illumination,
composed of components that are easily recognized
scale, pose, etc. More precisely, it is
at high levels but are loosely defined at low levels
computationally very expensive and suffers from
of our visual system. Each of the facial components
the usual shortcomings of straightforward
(features) has a different discrimination power for
correlation-based approaches, such as sensitivity to
identifying a person or the person‟s gender, race,
face orientation, size, variable lighting conditions,
and age. There have been many studies of the
background clutter, and noise. The major
significance of such features that used subjective
encumbrance to these statistical approaches [18] is
psychovisual experiments. Using objective
that, they attempt to perform classification in a very
measures, in this section we propose a
high dimensional space. To defeat this type of high
computational scheme for evaluating the
dimensionality, some other schemes have been
significance of different facial attributes in terms of
proposed. These types of schemes have been
their discrimination potential. The results of this
proven to reduce the dimension but obtain and
analysis can be supported by subjective
retain the most meaningful feature dimensions
psychovisual findings. To analyze any
before performing recognition.
representation V, where V can be the original
3.2.2 Principal Component Analysis image, its spatial segments, or transformed images,
we provide the following framework. First, we
A very popular method based on the concept of need a training set composed of a relatively large
reduced dimensionality is Principal Component group of subjects with diverse facial characteristics.
Analysis (PCA) [19]. PCA is a technique based on The appropriate selection of the training set directly
the concept of eigenfaces and was first introduced determines the validity of the final results. The
by Kirby and Sirivich in 1988. PCA is also known database should contain several examples of face
as Karhunen Loeve projection. It is one of the more images for each subject in the training set and at
successful techniques of face recognition and easy least one example in the test set. These examples
to understand and described using mathematics. should represent different frontal views of subjects
This method uses Eigen faces. Eigen faces have with minor variations in view angle. They should
been used to track human faces. They use a also include different facial expressions, different
principal component analysis approach to store a lighting and background conditions, and examples
set of known patterns in a compact subspace with and without glasses. It is assumed that all
representation of the image space, where the Eigen images are already normalized to m  n arrays and
vectors of the training image set span the subspace. that they contain only the face regions and not
In this method, the input image and images of the much of the subjects‟ bodies.
eigenface gallery should be the same size and we
have to normalize the input image so that the eyes, Second, for each image and subimage, starting with
nose and mouth are properly lined up i.e. we only the two-dimensional m  n array of intensity values
consider the face region and that should be a frontal I(x, y), we construct the lexicographic vector
face. In case of a non-frontal face it may result poor expansion   R mn. This vector corresponds to
performance. The primary advantage of this the initial representation of the face. Thus the set of
approach is that, it reduces the dimension of the all faces in the feature space is treated as a high-
data by some compression technique. dimensional vector space. Third, by defining all
instances of the same person‟s face as being in one
3.2.3 Linear Discriminant Analysis class and the faces of different subjects as being in
different classes for all subjects in the training set,
Fisher‟s Linear Discriminant known as Linear
we establish a framework for performing a cluster
Discriminant Analysis finds a small number of
separation analysis in the feature space. Also,
features that differentiates individual faces but
having labeled all instances in the training set and

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
4
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

having defined all the classes, we compute the the system needs to be trained with several
within- and between-class scatter matrices as iterations. There are five types of feature extraction
follows: methods: Generic methods based on edges, lines,
and curves; Gabor wavelets transform based face
feature extraction; Feature template based methods
that are used to detect various facial features like
eyes, nose, and mouth; Color segmentation based
methods that use face color instead of the intensity
….(1) values; Appearance based methods that are able to
manage changes in illumination conditions, shape,
Here Sw is the within-class scatter matrix showing pose and reflectance and even to handle translation
and partial occlusions. Now apart from these
the average scatter i of the sample vectors (V) of
classification, there are some limitations also, those
different classes Ci around their respective mean,
which are very difficult to implement. It is very
vectors i :
difficult to design a robust facial feature detection
algorithm with so many challenges like variable
illumination, face orientation and facial expression.
For example, when we are extracting eyes, it often
….(2) happens that they are closed or open but covered
with glassed that‟s why they are called rigid
Similarly, Sb is the between-class scatter matrix, features. Extraction of mouth also faces some
representing the scatter of the conditional mean, problem due to its rigidity, because when we want
vectors (i) around the overall mean vector  PrCi to extract the lips for uniquely identify a person we
is the probability of the ith class. The won‟t be able due to the unclear picture of the
discriminatory power of a representation can be groove caused by the open mouth features. On the
quantified by using various measures. In this paper other hand nose is a non-rigid feature.
we use the separation matrix, which shows the
combination of within- and between-class scatters Recently, more reliable approaches have been
of the feature points in the representation space. developed to defeat the limitations stated above.
The class separation matrix and a measure of For example structural matching methods such as
separability can be computed as Active Shape Model, compared to earlier methods
these recent methods are much more robust in
terms of handling variations in image intensity and
feature shape. An even more challenging situation
is feature restoration, which tries to recover
….(3) features that are invisible due to some noise.
Jv is our measure of the discrimination power (DP)
of a given representation V. As mentioned above,
the representation may correspond to the data in its
original form (e.g., a gray-scale image), or it can be
based on a set of abstract features computed for a
specific task.

4. Facial features extraction

Facial features, such as eyes, nose, and mouth as

well as their spatial relationships are very important
in various applications like face recognition, facial
expression detection, face tracking, lip reading etc. Figure 4: Facial feature extraction
Hence, after detecting the face region from a
cluttered image we should extract various features 4.1 Generic methods
from the face, and then we verify them for our
recognition purpose [21]. Extraction property of The features are extracted by using relative
different features can be coded either by manually positions and sizes of the important components of
or by using some automatic learning procedure face. These group methods concentrate in two
from a set of data. In case of automatic learning directions. First, detecting edges, directions of
procedure if we want to increase the performance, important components or region images contain

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
5
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

important components, then building feature This method group will extract feature [26] of face
vectors from these edges and directions. Using such as eyes, mouth, etc. based on template
filters such as Canny filter to detect eyes or mouth function [27] and appropriate energy function. An
region of face image, or the gradient analysis image region is the best appropriateness with
method, which is usually applied in this direction. template for eye, mouth or nose, which will
Second, methods are based on the grayscales minimize the energy. The methods have been
difference of important components and proposed such as deformable template and genetic
unimportant components, by using feature blocks, algorithms. In the deformable template method
set of Haar-like feature [23] block in Adaboost [28], the feature of interest, an eye, for example, is
method[22] to change the grayscales distribution described by a parameterized template. An energy
into the feature. In LBP [24] method, it divides up function is defined to links edges, peaks, and
the face image to regions (blocks) and each region valleys in the image intensity with corresponding
corresponds with each central pixel. Then it properties of the template. Then the template
examines its pixel neighbors, based on the matching is done with the image, by altering its
grayscales value of central pixel to change its parameter values to minimize the energy function,
neighbor to 0 or 1. Therefore, every pixel will be thereby deforming itself to find the best fit. The
represented in a binary string. Since then, we build final parameter values can be used as descriptors
histograms for every region. Then these histograms for the features.
are combined to a feature vector for the face image.

4.2 Gabor Wavelets Transform Face Feature

Extraction

Here, given the predicted feature positions, the Figure 5: Template based feature detection
multi-scale and multi-orientation Gabor wavelet
matching method [25] is used to detect each facial 4.4 Color Segmentation Based Techniques
feature in the vicinity of the predicted locations. In
addition, the robust matching in the Gabor space This approach makes use of skin color to isolate the
also provides an accurate and fast solution for face. Any non-skin color region within the face is
tracking multiple facial features simultaneously. viewed as a candidate for eyes or mouth [29].
Color based feature extraction use color models
During tracking, in order to adaptively compensate such as RGB, YCbCr or HSV with certain range of
for the facial feature appearance changes, the color pixels, skin region is detected [30]. After
Gabor wavelet coefficients are updated getting the skin region, facial features viz. Eyes and
dynamically at each frame to serve as the tracking Mouth are extracted. The image obtained after
template for the subsequent image frame. This applying skin color statistics is subjected to
updating approach works very well when no binarization. it is transformed to gray-scale image
significant appearance change happens for each and then to a binary image by applying suitable
facial feature. However, under the large face threshold. This is done to eliminate the hue and
orientations in which an arbitrary profile is saturation values and consider only the luminance
assigned to the occluded feature during self- part. This luminance part is then transformed to
occlusions, as well as the significant facial binary image with some threshold because the
expressions in which the facial feature appearance features for face are darker than the background
varies significantly, the tracker often fails. colors. After thresholding, opening and closing
operations are performed to remove noise. These
Therefore, a shape-constrained correction are the morphological operations, which are used to
mechanism is developed to tackle the above remove holes. Then eyes, ears, nose can be
problems and to refine the tracking results. As a extracted from the binary image by considering the
result, via our proposed technique, a set of twenty- threshold for areas which are darker in the mouth
eight facial features can be detected and tracked than a given threshold. So triangle can be drawn
robustly in real-time under significant appearance with the two eyes and a mouth as the three points in
changes in various facial expressions and face case of a frontal face. And it is easy to get an
orientations. isosceles triangle (i j k) in which the Euclidean
distance between two eyes is about 90-110% of the
4.3 Feature template based methods Euclidean distance between the center of the
right/left eye and the mouth. After getting the

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
6
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

triangle, it is easy to get the coordinates of the four Human facial expression recognition by a machine
corner points that form the potential facial region. can be described as an interpretation of human
Since the real facial region should cover the facial characteristics via mathematical algorithms
eyebrows, two eyes, mouth and some area below [32]-[36]. Gestures of the body are read by an input
the mouth, this coordinates can be calculated. The sensing device such as a web-cam. It reads the
performance of such techniques on facial image movements of the human body and communicates
databases is rather limited, due to the diversity of with computer that uses these gestures as an input.
ethnical backgrounds. These gestures are then interpreted using algorithm
either based on statistical analysis or artificial
4.5 Appearance Based Techniques intelligence techniques. The primary goal of
gesture recognition research is to create a system
This type of methods is using linear transformation which can identify specific human gestures and use
and statistical methods to find the basic vectors to them to convey information. By observing face,
represent the face. Methods have been proposed in one can decide whether a man is serious, happy,
the literature for this aim such as PCA and ICA thinking, sad, feeling pain and so on. Recognizing
[31]. In detail, goal of PCA method is to reduce the the expression of a man can help in many of the
number of dimensions of feature space, but still to areas like in the field of medical science where a
keep principle features to minimize loss of doctor can be alerted when a patient is in severe
information. PCA method uses second-order pain. It helps in taking prompt action at that time.
statistic in the data. However, PCA method has still In this paper the main focus is to define a simple
disadvantages. High order dependencies still exist architecture that recognizes the human facial
in PCA analysis, for example, in tasks as face expression.
recognition, much of the important information
may be contained in the high order relationships
among the image pixels, not only second-order.
While other method ICA uses technique
independent component analysis. it is an analysis
technique not only use second-order statistic but
also use high order statistic. PCA can be derived as
a special case of ICA, which uses Gaussian source
models. PCA is not the good method in cases non
Gaussian source models. it has been observed that Figure 6: Facial expression detection
many natural signals, including speech, natural
images, are better described as linear combinations 5.1 Neural Network based facial expression
of sources with super Gaussian distributions . In recognition
that case, ICA method better than PCA method
because: I) ICA provides a better probabilistic Here we describe a neural network based facial
model of the data. II) It uniquely identifies the expression detection technique. This system is
mixing matrix. III) It finds an unnecessary divided into five modules as shown in the
orthogonal basic, which may reconstruct the data following figure.
better than PCA in the presence of noise such as
variations lighting and expressions of face. IV) It is
sensitive to high-order statistics in the data, not just
the covariance matrix but it require that the image
matrices must be first transformed into vectors,
which are usually of very high dimensionality. This
causes expensive computational cost and
Figure 7: Simple architecture of facial expression
sometimes the singularity problem. The appearance
recognition
based method group has been found the best
performer in facial feature extraction because it
keeps the important information of face image, Each box shown in the above figure is treated as
rejects redundant information and reflects face one module. The first module captures the image
global structure. using the web cam. Second module is for face
detection, which can detect the human face from
the captured image. A set of modules bounded by a
boundary line represent pre-processing block. It
5. Facial Expression Recognition consists of histogram equalization, edge detection,
thinning, and token generation modules. The next

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
7
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

module is the training module to store the token 5.3 Support Vector Machine (SVM)
information that comes from the image pre-
processing module. This training has been done Support Vector Machine [39] is a popular
using back propagation neural network [37]. And technique for classification [40]. SVM performs an
the last module is the token matching and decision implicit mapping of data into a higher dimensional
making called recognition module, which produces feature space, where linear algebra and geometry
the final result. The following flow chart represents can be used to separate data that is only separable
how all the modules work. with nonlinear rules in the input space.

Face detection is a process that aims to locate a Given a training set of labeled examples T = {(xi,
human face in an image. The process is applied on yi), i = 1,……l} where xi  Rn
stored image or images from a camera. Human and yi  {1, -1}, the new test data x is classified by
face varies from one person to another. This the following function:
variation in faces could be due to race, gender, age,
and other physical characteristics of an
individual. Therefore face detection becomes a
challenging task in computer vision. It becomes
more challenging task due to the additional ….(5)
variations in scale, orientation, pose, facial where i are Lagrange multipliers of a dual
expressions, and lighting conditions. Many optimization problem, and K(xi, x) is a kernel
methods have been proposed to detect faces such as function. Given a nonlinear mapping  that embeds
neutral networks [38], skin locus, and color input data into feature space, kernels have the form
analysis. Since these detected faces become an of
input to the recognition of the gestures, it is K(xi, xj) = <  (xi) .  (xj) > . SVM finds a linear
important to get rid of non-facial information in the separating hyperplane with the maximal margin to
image. In this paper, the technique proposed by separate the training data in feature space. b is the
Viola and Jones is used to detect the face. The parameter of the optimal hyperplane.
main reason for using this technique is that its
implementation is feature based and relatively fast 6. Conclusions
compared to other available techniques. The
following figure represents the detected faces with
their corresponding input images. In this paper, we have covered a detail discussion
on the various stages of any face detection
5.2 Template Matching technique. Also, some popular well-known face
detection techniques are described very briefly.
For expression recognition, we first adopt template Recently, face detection techniques have been
matching for its simplicity. In training, the LBP employed in different applications such as face
histograms of face images in a given class are recognition, facial feature extraction, detection of
averaged to generate a histogram template for this facial expression, which are also the subjects to be
class. In recognition, a nearest-neighbour classifier focused of this paper. Hence, before developing
is adopted: the LBP histogram of the input image is any kind of method of your choice, if you go
matched with the closest template. We select Chi through this paper, you will definitely get an
square statistic (X2) as the dissimilarity measure for overview various ways and applications used in
histogram. It is observed that facial features face detection process.
contributing to facial expressions mainly lie in
some regions, such as eye area and mouth area; 7. References
these regions contain more useful information for
facial classification. Therefore, a weight can be set [1] M. H. Yang, D. J. Kriegman, and N. Ahuja,
for each face region based on the importance of the “Detecting face in images: a survey,” IEEE Trans. Patter
information it contains. Our weighted (X2) statistic Analysis and Machine Intelligence, vol. 24, pp. 34–58,
2002.
is
[2] E. Hjelmas and B. K. Low, “Face detection: A
survey,” Computer Vision and Image Understanding,
vol. 83, pp. 236–274, 2001.
….(4)
[3] Mrs. Sunita Roy [Link]., "A Tutorial Review on Face
where S and M are two LBP histograms, wj is the Detection", International Journal of Engineering
weight for region j.

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
8
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

Research & Technology (IJERT), Vol. 1 Issue 8, October [18] A. K. Jain, R. P. W. Duin, and J. C. Mao,
- 2012, ISSN: 2278-0181. “Statistical pattern recognition: a review,” IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 22, no. 1,
[4] K. Sobottka and I. Pitas, “Face localization and pp. 4–37, 2000.
feature extraction based on shape and color [19] [Link] Rao ,Srinivasulu Asadi,
information,”Proc. IEEE Int‟l Conf. Image Processing, [Link] Rao “A Comparative study of Face
pp. 483-486, 1996. Recognition with Principal Component Analysis and
Cross-Correlation Technique”, International Journal of
[5] T. Sasaki, S. Akamatsu, and Y. Suenaga. Face image Computer Applications (0975 – 8887),Volume 10– No.8,
normalization based on color information. Tech. Rep. November 2010.
I.E.I.C.E., IE91-2, pp. 9–15 (1991). [20] M. H. Yang, “Kernel Eigenfaces vs. Kernel
Fisherfaces: Face recognition using kernel methods,”
[6] C. Kotropoulos and I. Pitas, “Rule-based face AFGR, pp. 205–211, 2002.
detection in frontal views,” Proc. Int‟l Conf. Acoustics, [21] K. J. Liao, Face detection by outline, color, and
Speech and Signal Processing, vol. 4, pp. 2537-2540, facial features, Mater thesis, GICE, NTU, Taipei, 2010.
1997. [22] D. Cristinacce, T. Cootes, “Facial Feature Detection
[7] C. Lin, K.C. Fan, “Human face detection using Using AdaBoost With Shape Constrains”, British
geometric triangle relationship,” Proc. 15th ICPR, pp. Machine Vision Conference, 2003.
945–948, 2000. [23] Adolf, F. How-to build a cascade of boosted
classifiers based on Haar-like features.
[8] E. Hjelmas and B. K. Low, “Face detection: A [Link]
survey,” Computer Vision and Image Understanding, ectDetection_HowTo.pdf, June 20 2003.
vol. 83, pp. 236–274, 2001. [24] T. Ojala, M. Pietikinen, and T. Menp,
“Multiresolution grayscale and rotation invariant texture
[9] Alan L. Yuille, “Deformable Templates for Face classification with local binary patterns,” IEEE PAMI,
Recognition”, Journal of Cognitive Neuroscience vol. 24, no. 7, July 2002.
Volume 3, pp. 59-70,Number 1991. [25] S. J. McKenna, S. Gong, R. P. Wurtz, J. Tanner, D.
Banin “Tracking Facial Feature Points with Gabor
[10] D. G. Lowe, “Distinctive image features from scale- Wavelets and Shape Models” 1st Int. Conf. on Audio-
invariant keypoints,” Int‟l Journal of Computer Vision, and Videobased Biometric Person Authentication,
vol. 60, no. 2, pp. 91-110, 2004. Lecture Notes in Computer Science 1997.
[26] E. Bagherian, R. Wirza, N.I. Udzir “Extract of
[11] G. Yongsheng and M.K.H. Leung, “Face Facial Feature Point” IJCSNS International Journal of
recognition using line edge map”, IEEE Transactions on Computer Science and Network Security, VOL.9 No.1,
Pattern Analysis and Machine Intelligence, , Vol.24, January 2009.
Iss.6, Pages:764-779, Jun 2002. [27] N. Bhoi, M. Narayan Mohanty “Template Matching
based Eye Detection in Facial Image” International
[12] M. Turk and A. Pentland. “Eigenfaces for Journal of Computer Applications (0975 – 8887) Volume
recognition”. Journal of Cognitive Neuroscience, 3, 71- 12– No.5, December 2010.
86, 1991. [28] Paul Kuo, John Hannah, “An Improved Eye Feature
Extraction Algorithm Based On Deformable Templates”,
[13] H.A. Rowley, “ Neural Network-Based Face IEEE International Conference on Image Processing,
Detection”, PhD thesis, Carnegie Mellon Univ, 1999. Volume 2, pp. 1206-1209,2005.
[29] [Link], [Link] ,A. Andreeva ,” A Survey
[14] Craw, I., Tock, D. & Bennett, A. (1992), Finding on Pixel-Based Skin Color Detection Techniques”,
Face Features, in „European Conference on Computer Graphics and Media Laboratory, Moscow State
Vision‟, pp. 92–96. University, Moscow, Russia.
[30] K. Sobottka, I. Pitas, "Face Localization and Facial
[15] L. Wiskott, J.-M. Fellous, N. Krüger, and C. von der Feature Extraction Based on Shape and Color
Malsburg, "Face Recognition by Elastic Bunch Graph Information", 1996.
Matching," IEEE Transactions on Pattern Analysis and [31] Bruce A. Draper, Kyungim Baek, Marian Stewart
Machine Intelligence, Vol.19, pp.775-779, 1997. Bartlett, J. Ross BeveRidge, “Recognizing Face with
PCA and ICA”, Computer Vision and Image
[16] M. Lades, J. C. Vorbrüggen, J. Buhmann, J. Lange, Understanding 91, pp. 115–137, 2003.
C. v. d. Malsburg, R. P. Würtz, and W. Konen, [32] Y. Tian, T. Kanade, and J.F. Cohn, Facial
"Distortion invariant object recognition in the dynamic Expression Analysis, Handbook of Face Recognition,
link architecture," IEEE Trans. Computers, Vol.42, Springer, October 2003.
pp.300-311, 1993. [33] B. Fasel and J. Luettin, “Automatic facial expression
analysis: a survey,” Pattern Recognition, vol. 36, pp.
[17] Manjunath, B., Chellappa, R. & von der Malsburg, 259–275, 2003.
C. (1992), „A Feature Based Approach to Face [34] M. J. Lyons, J. Budynek, and S. Akamatsu,
Recognition‟, IEEE Conference Proceedings on “Automatic classification of single facial images,” IEEE
Computer Vision and Pattern Recognition pp. 373–378. PAMI, 1999.

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
9
IJREAT International Journal of Research in Engineering & Advanced Technology, Volume 1, Issue 2, April-May, 2013
ISSN: 2320 - 8791
[Link]

[35] I. Cohen, N. Sebe, Garg A., L. Chen, and T. Huang, for facial verification systems”, Sage Publications Inc.,
“Facial expression recognition from video sequences: USA,1992.
Temporal and static modeling,” CVIU, vol. 91, pp. 160– [38] Nigrin A., “Neural Networks for Pattern
187, 2003. Recognition”, Cambridge, The MIT Press, 1993.
[36] M.S. Bartlett, G. Littlewort, I. Fasel, and R. [39] Hsu, C.W., Chang, C.C., and Lin, C.J.: „A practical
Movellan, “Real time face detection and facial guide to support vector classification‟. Department of
expression recognition: Development and application to Computer Science, National Taiwan University, 2005.
human computer interaction,” in CVPR Workshop on [40] V. N. Vapnik, Statistical Learning Theory, Wiley,
CVPR for HCI, 2003. New York, 1998.
[37] Solheim, Inger, Payne, Tanya L., Castin, and Ralph,
“The potential in using back propagation neural networks

[Link]
Published by: PIONEER RESEARCH & DEVELOPMENT GROUP ([Link])
10