University of Bucharest [612929]

University of Bucharest
Faculty of Physics
Theodora Daniela PREDA
Detection and recognition of human patterns
MASTERS PROGRAMME IN MEDICAL PHYSICS
Scientic Coordinator
Prof. Dr. Radu MUTIHAC
Bucucharest,
June 2017

Table of contents
1 Introduction 2
1.1 Pattern recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Techniques in pattern recognition . . . . . . . . . . . . . . . . . . . . 3
1.3 Pattern recognition system . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 The sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.3 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.4 Classication . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Chapter 2 6
2.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Basics of face detection . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Basics of face recognition . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Biometric technique . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Applications of face detection and recognition . . . . . . . . . . . . . 12
3 Chapter 3 Method 14
3.1 Face recognition analysis methods . . . . . . . . . . . . . . . . . . . 14
3.1.1 Linear Discriminant Analysis (LDA) . . . . . . . . . . . . . . 14
3.1.2 Principal Components Analysis (PCA) . . . . . . . . . . . . . 15
3.1.3 Elastic Bunch Graph Matching (EBGM) . . . . . . . . . . . . 16
3.2 Learning process in articial sistems . . . . . . . . . . . . . . . . . . . 19
3.2.1 Supervised learning . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.2 Unsupervised learning . . . . . . . . . . . . . . . . . . . . . . 20
3.2.3 Semi-supervised Learning . . . . . . . . . . . . . . . . . . . . 21
3.2.4 Reinforced learning . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 The geometric characteristics of the human face . . . . . . . . . . . . 22
4 Chapter 4 esxp 24
4.1 Viola-Jones Face Detection Algorithm . . . . . . . . . . . . . . . . . . 24
4.2 The diculties in face detection and recognition . . . . . . . . . . . . 26
4.2.1 The level of illumination . . . . . . . . . . . . . . . . . . . . . 26
4.2.2 Variation of position . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.3 Expressions faciales . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.4 The presence or absence of structural components . . . . . . . 27
4.2.5 The presence of structural components . . . . . . . . . . . . . 28
4.3 Results obtained . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3.1 Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2

4.3.2 Detecting and recognizing of faces patterns . . . . . . . . . . . 29
4.3.3 Face detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3.4 Training process . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3.5 Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5 Conclusions 35
6 Bibliography 36
1

Introduction
1.1 Pattern recognition
The word pattern means a repeated decorative design or a model but also
it is a design used as a guide, in this case, means a features. Pattern recognition is
the classication of input dates into knowns classes by the extraction of important
features or characteristics of the dates from a fund of irrelevant detail.The technique
of pattern recognition is very used for robotic technology and face recognition and
has many applications in security, science, engineering, medicine, or even business.
In pattern recognition, the feather represent the all distinctive characteristics
wich can be numeric or symbolic. A feature vector is a representation of column
vector and the combination of all features is the feature space. An object can be
dened as points in the feature space, representation is called a scatter plot. The
aspect of a feature vector, represents its capacity to discriminate samples from a
dierent category as one.
Figure 1.1: The feature
In gure 1.1 is represented the distinction between good and poor features
and the principles feature properties.
2

1.2 Techniques in pattern recognition
To create a pattern recognition system three approaches can be followed:
Statistical techniques are draws from concepts of statistical decision theory to
discriminate between dates taken from dierent classes based on the number
of features of the dates. The signicance of each characteristic is known from
its position within the vector: probability, similarity and grouping.
Quantitative techniques make the dierentiation task between the classes based
on the structural sub-patterns and their relation built within the dates bit dif-
cult. A development of a morphological approach was based that relies on
syntactic grammars to discriminate between dates from dierent classes using
the structural interrelationships present within the dates.
Hybrid approaches are the combination of two mechanisms. As a system,
hybrid features can be utilized in a structural or statistical classier.Hybrid
systems can combine the two mechanisms into a multiscale system using a
parallel or a hierarchical structure.
Figure 1.2: An example of the statistical approaches
Figure 1.2 represent the statistical approaches to pattern recognition used
to an ordinary identication problem with the purpose of discriminating among the
rectangle and the triangle.
1.3 Pattern recognition system
In the pattern recognition programs are three dierent operations impli-
cated:
3

1. data pre-processing
2. feature extraction
3. classication
To understand the issues of creating such a system, we need to know the
issues that each of these components needs to solve. In gure 1.3 is a schematic
representation of the pattern recognition system.
Figure 1.3: A pattern recognition system
1.3.1 The sensor
The sensor is a technical device that reacts qualitatively or quantitatively
to a stimulus. The sensor in pattern recognition system it is a camera or something
else. The problems can be due to limitations of the sensor like signal-to-noise ratio,
sensitivity, resolution, distortion etc.
1.3.2 Pre-processing
Segmentation is part pre-processing process and is one of the biggest prob-
lems in pattern recognition. The segmentation diculty is closely related to the
issue of recognizing or grouping the dierent features of an object.
1.3.3 Feature extraction
The traditional objective of the feature extractor is to describe an entity to
be recognized by measurements whose estimation are very similar for an entity in
the same groups, and very distinct for an entity in dierent groups. The features
that characterize properties like shape or many types of texture do not change to
translation or rotation.Some of the fundamentals of pattern classication can be
utilized in the creation of the feature extractor.
1.3.4 Classication
The classier component utilizes the feature vector created by the feature
extractor to sent the object to a category. The development of a domain independent
classication theory is created using vector data representations. One problem that
4

appear in practice is that it may not be possible every time to know the values to
any of the features of a particular date.
5

Chapter 2
2.1 History
The rst thing you recognize to other person is his face. Progress in com-
puter science in the past few decades is similar to progress in face recognition. Face
recognition algorithms works with easy geometric models, but the recognition pro-
cedure has now transformed into a science of complex mathematical representations
and tting processes. Major progression and initiatives in this time have stimulated
face recognition technology into the center of studies. Face recognition can work for
both identication and verication.
Face recognition is a computer application used to identify or verify a per-
son from an image or a video. The ways to do this are by comparing selected facial
images from the photos with the face database. The rsts who developed the au-
tomatic face recognition are Charles Bisson, Helen Chan Wolf and Woody Bledsoe
between 1964 and 1965.
The experiment was to extract the coordinates of a set of parameters from
the photo, which were then are utilized by the computer for recognition. From these
coordinates were computed a list of 21 distances, such as the width of mouth as can
be seen in Fig 2.1. In a database, the name of the person in the image was associated
with the list of computed distances calculate. In the recognition phase, the set of
distances was compared with the corresponding distance for each photograph. The
closest records are returned.
They use a large database of images and a photograph. The diculty was
to select from the database a small set of les such that one of the image records
matched the photograph. The eciency of the method could be set in function of
the report of the answer list to the number of records in the database.
Kirby and Sirovich, in 1988, applied principle component analysis, a stan-
dard linear algebra technique, to the face recognition. This was considered very
important as it showed that less than 100 values were needed to accurately code.
In 1991, Turk and Pentland discovered that the residual error could be used to
detect faces in photo a discovery that enabled reliable real-time automated face
recognition systems. Although the approach was limited by environmental factors,
it created a signicant interest in the development of automated face recognition
technologies.This fact that simple, real-time face recognition techniques could be
combined to build a useful system create an interest in the study of face recognition.
6

Figure 2.1: The values of a set of parameters from the image
In 1997, the system developed by Christoph von der Malsburg outperformed
most systems. The Bochum system was developed through funding by the United
States Army Research Laboratory. The software was used by banks, airports and
other busy locations. The software was good enough to make identications from
bad face views. It can also identify through such impediments of identication as
mustaches, beards or even sunglasses.
Face recognition technology, today, is being used to combat passport fraud,
support law enforcement, to stole identity, identify missing children, and to counter
terrorism.
2.2 Basics of face detection
Face detection is the rst stage in face recognition and is the most important
part of the recognition process because it is can recognize if can not detect it.
The process of face detection has been intensely studied in dierent elds,
demonstrating its important role in understanding most aspects of the detection
and recognition processes in the brain. Images through the detection process are
separated into two classes
7

with faces (targets)
with background (clutter)
Face detection is a very dicult mecanism because are a lot of characteristics
that in
uence faces such as skin color, hair or just facial expressions and also the
features of the camera as image quality and image resolution.
The primary stages of face detection are:
1. INPUT
2. PRE PROCESS
3. CLASSIFIER
4. OUTPUT
In face detection from images or video are locates face areas from these in-
puts. This is made by separating face zone from non-face regions. An ideal face
detector is that able to detect as many faces under dierent circumstances and very
fast.
The localization of facial feature is important for detected face. Feature are
extracted then simplies face region where is detected the face. Then is aligned to
coordinate framework in order to reduce the large number of variables produce from
dierent face posturing and sizes.For face identication, the input parameters are
the value of feature from the shape of face characteristics as mouth, nose and eyes.
The output of face recognition is the identity of the input face image. A fea-
tures array is obtained from the input face and compared with faces from a database.
If a similar image is found, the program returns the associated image.
It is important to choose a proper face classication technique because is
need to provide a good method of separation between dierents human faces. Face
recognition is used in many applications because it is a non-invasive identication
method and is used in the security application.
2.3 Basics of face recognition
The most general technique of face recognition include the next steps: face
detection, feature extraction and face recognition, each step consisting of more other
parts.
Face recognition require a set of visual tasks for the soft to work well. These
include:
1 Acquisition: the detection and trace of face-like in patches in a video,
2 Normalisation: the segmentation and alignment of the face images,
8

3 Recognition: the representation and building of face form as identities, and
the linking of new face images with known models.
These tasks seem to be successive and frequently have been discussed as
such.But, it is both computer and human more appropriate to see them as a set
which works together with closed-loop feedbacks.To make this dicult system, an
approach has been adopted which will perform recognition, normalization, acquisi-
tion, and in a fast way.
In Fig. 2.2 it was represented the system design. Images from the record
video are processed in real-time to obtain normalized and aligned of face sequences.
This proceeding is a closed-loop prosses that include the computation and existing
of three dierent visual marks: face appearance, color, and motion models.In ad-
dition to recognition, it have been developed mechanisms to estimate in real-time
face pose, which can be work with to improve the quality of alignment and detection
Figure 2.2: A framework for face recognition in dynamic scenes.
A lot of research eort has been used on face recognition tasks. Only a single
photo or just a few images of all person are available. A bigger concern has been to
connect to large databases which include thousands of people. A form of biometric
face recognition utilizing the iris is better tted to populations.
The tasks include recognition of a smaller number of people with more im-
ages, this might appear to be simpler initially. In comparison, the tasks used here
needed recognition to be work using sequences processed and normalised automati-
cally. These are described by low resolution. Recognition based on isolated images
of this tip is highly unsubstantial and unreliable.
9

Dierent face recognition tasks works, given a database containing of a set,
T , of P known people. Four tasks are used here as follows:
1 Face classication: identication of the subject presume that the subject is a
member of set T.
2 Known/Unknown: determination if the subject belong of set T
3 Identity verication: identication of the subjects is given by some other
means.
4 Full recognition: determination if subject is a member of set T or not, and to
determine the identity of subject if is the case.
When considering a basic aspect to these tasks it is bettering to know some
of the forms of classes of face images in an photos space. The faces set become a
small number of developed connected. A face which passes by transformations like
as rotation and translation results in a linked but powerful non-convex subarea in the
image space.While transformations might be almost corrected using linear transfor-
mations of the image plane, rotations in depth, facial expressions, and illumination
changes can't be so simply to detect.
Figure 2.3: The tasks
In g 2.3 is illustrates the recognition tasks builded in a hypothetical face
space P , where P include all face images.
2.4 Biometric technique
Now to increase the security systems is used biometrics. Biometrics is a
method which is referring to identify an individual using a part of his body.Biometric
10

techniques not only identify a particular person but lower the risk that someone else
may use a stolen identity. Biometric authentication also supports identication and
authentication aspects in information security.
Dierent biometric techniques can be generaly classied as:
1. Physical biometrics – involves physical measurements and includes the descrip-
tion of the face, ngerprints, iris/retina scans, hand geometry.
2. Behavioral biometrics – involves measurements of a person's behavior while is
making certain simple tasks such as speech, signature, walking.
3. Chemical biometrics – involves the measurement of chemical indicators such
as the chemical structure of human perspiration.
Figure 2.4:Biometric techniques
Numerous authentication techniques have been developed using body pa-
rameters. A biometric system has two functions verication and authentication.
The verication is related to the search of the database. Seven factors determine
physical, behavioral and chemical characteristics for biometric systems. These are:
Measurability,
Acceptability,
Permanence
Uniqueness
Universality
11

Performance
Circumvention
The commonly used biometrics have many disadvantages. Iris recognition
is extremely accurate but expensive. Fingerprints are reliable as biometric and no
painless but are not useful if people do not collaborate.Three of the biometric tech-
niques are illustrated in the Fig. 2.4 ngerprints, iris scan, and face recognition.
Facial recognition is the most commonly used biometric technique because
it is accurate and can be used in highly crowded places. Face recognition method
are typically based on the location and shape of facial features, such as eyes, lips,
ears and distances between them. In order to be able to detect and recognize with
the same accuracy as the human brain, there are numerous studies in the eld of
image processing.
2.5 Applications of face detection and recognition
The face recognition and detection are a part of information security, it is
responsible for security and protection the information in any forms. Another way
we can identify in the system is using personal identication number or a password,
but this method is no longer eective because this data can be easily stolen.
Figure 2.5:Face recognition at an airport
Given these issues, developing biometric approaches, such as iris/retina, n-
gerprint, face recognition and voice recognition, is a better way to identify people
than PINs.
In this days, the face recognition is being used in parks, airports, train sta-
tion and other popular locations. This technology was used during an experiment.
12

The role of this test was to analysis the ability to recognize felons or potential ter-
rorist. There were some correct identications, however, it had a few issues. Many
airports across the world use face recognition at the security checkpoints.
Face recognition has used iIn many dierent elds as follows
Civil applications and law enforcement
{passport, national ID
{Surveillance of crowded spaces
{forensic applications
Security applications for electronic device
{secure access to networks
{ebanking
intelligent devices
{smart homes
{humanmachine interaction
This software will be the rst line defense in against the terrorist attacks by
stopping then to enter in airports or other public places and leaving the country.
This system is ecient not only to combat possible acts terrorist but can be used
for missing persons, to identify refugees and escapees.This technology will work and
against the identity theft.
13

Chapter 3 Method
3.1 Face recognition analysis methods
Human face images are formed by thousands of pixels.The utilization of
subspace for the description of face images helps to reduce dimensionality in later
classication. Reducing the size of a data set can be obtained by extracting charac-
teristics, with condition that the new features include most of the information from
the original data set. The subspace representation is used to decrease the changes
of light level and facial expressions.The most studied analysis methods from face
recognition research are Linear Discriminant Analysis (LDA), Principal Components
Analysis (PCA) and Elastic Bunch Graph Matching (EBGM).
3.1.1 Linear Discriminant Analysis (LDA)
Face image is created of a big number of pixels in face recognition. LDA is
used with the purpose to reduce the number of characteristics to a small number
before classication.To determinate the elements classes unknown it is to apply LDA
which is based on statistics, using elements with known classes.this algorithm has
the role to magnify the variance among the class and to reduce variance in the class.
In Fig. 3.1 each block symbolize a class, there are big variances between
classes, but little variance inside classes. When the dimensional of face data are
high, this technique confronts with the problem of the small sample size that ap-
pears where the dimensionality of the sample space and number o samples are too
dierent. The number o samples is smaller than the dimensionality of the experi-
ment space.
Figure 3.1: Example of LDA algorithm
14

3.1.2 Principal Components Analysis (PCA)
PCA is associated with eigenfaces, is the technique used and created by
Kirby and Sirivich. When is used PCA the face image and the set need to be the
same size. The eyes and mouth from the images rst need to be normalized and
align. The PCA have to decrease the dimension of the data by removeing the in-
formation that is not valuable by data compression basics. This process is the most
eective low dimensional conformation of the facial image. The vectors are named
eigenfaces because they had the same height as the initials face images.
Figure 3.2: Eigenfaces – Feature vectors
Every face image may be characterized as vectors of eigenfaces, which are
stocked in an array of one dimension. In Fig. 3.2 can be seen an example of using
the feature vectors. A sample image is compared with an image from the set by
comparing the feature vectors distance.
The PCA projection of an face image Y0of dimension t1t2pixels is rep-
resented as a point in <t1t2space. If the average image of N number of the training
imageyiis
m=NX
i=1yi (3.1)
and the equivalentn mean centred image is
vi=yim (3.2)
15

Then can get a set of elements piwhich has the greatest possible extrapola-
tion onto each of the vi. The purpose is to
nd a class of Northonormal vectors pifor which the quantity
i=1
NNX
k=1(piTvK)2(3.3)
is may be write with the orthonormal constraint
piTpK=lk (3.4)
The eigenvectors piand eigenvalues iof the covariance matrix C=VTV
are calculated, where Vis a matrix formed of the vectors viplaced in line. The
weight vector of the face images is calculated as
wi=piTvi (3.5)
piis the eigenvector is calculated through PCA and Trepresent the transpose op-
eration.
The principal advantage of this technique is the decrease of the data needed
to identify at 0.1 percent from all the set of data. The eigenface is a lter for data,
which may check if every image is good correlated with itself. The images with a
small correlation can be refused. These two aspects (correlation and face lter) are
approached by classication following regions:
1. known faces
2. unknown faces
3. non-faces
3.1.3 Elastic Bunch Graph Matching (EBGM)
EBGM is based on the concept there a lot of nonlinear characteristics in the
real face images, that can not be resolved by the methods discussed earlier, such as
the dierence in level of light and expression. A Gabor wavelet transforms projects
the face from the image into an elastic grid. The convolutions of that
ducial points form the Gabor jet. In Fig. 3.3 are represented the Gabor jet by a
circle on the elastic grid, which represents the image around a specical pixel.
It is used this Gabor lter to extract and detect the forms of the face using
the convolution of the image. The Gabor lter is based on the processes performed
in the visual cortex of human.
For starting the face detection and recognition is needed to detect the facial
landmarks. By convolving the pictures around of certain
ducial points are stores the dates about the faces with 2D Gabor wavelets.The model
used for identi
cation is named the face bunch graph (FBG).
16

Figure 3.3: The Gabor wavelet transform
Gabor wavelets
The EBGM algorithm is a 2D form of Gabor wavelets. Wavelets are used
as Fourier transforms, to calculate space properties of an image. The dier-
ence is that the wavelets operate on a spacce of image , while the Fourier
transforms changes the entire image.
The wavelet characteristics follows an equation:
V(x;y;;
;;; ) =ex02+
y02
22cos(2x0
+) (3.6)
where
x0=xcos() +ysin() (3.7)
y0=xsin() +ycos() (3.8)
The wavelet specications are given by the nextsets of parameter:
1.give the orientation of the Gabor wavelet, gure 3.4 reprezent Gabor
wavelets for four dierent orientation of .
Figure 3.4: Gabor wavelets for dierent orientation of 
2.give the wavelength of the sine wave, gure 3.5 reprezent Gabor wavelets
for four dierent values of .
3.describe the phase of the sine wave
17

Figure 3.5: Gabor wavelets for dierent values of 
4.
describe the aspect ratio of the Gaussian.
5.describe the radius of the Gaussian, gure 3.6 reprezent Gabor wavelets
for four dierent values of .
Figure 3.6: Gabor wavelets for dierent values of 
Gabor jets
The transformation of Gabor wavelet oers a value for every wavelet
at all points of the image. The Gabor jet is the set of values that result from
the transformation of Gabor wavelet for a single pixel position. The jet con-
tains values of dierent frequency and orientation, and it looks as the Fourier
transform. The wavelets are formats from pairs of a real and an imaginary
part. Each jet is very sensitive to a small change.
The amplitude and phase are calculated as can be seen below:
a=2p
a2real+a2imag
= arctanaimag
areal(areal0)
=+ arctanaimag
areal(areal0)
=
2(areal= 0;aimag0)
=
2(areal= 0;aimag0)
By combining those three methods LDA, PCA and EBGM can be obtained
the most useful analysis method in face detection and recognition.
18

3.2 Learning process in articial sistems
Learning process, or automatic learning, is part of one of the approaches
to articial intelligence. The learning process is, therefore, a scientic discipline
centered on the development, analysis, and implementation of automated methods,
which allow a machine to evolve through a learning process. And thus to accomplish
tasks that would have been dicult, if not impossible, to realize with more conven-
tional algorithms. In the gure 3.7 is represented the schema of the the learning
process.
Figure 3.7: The learning process
3.2.1 Supervised learning
Supervised learning is an automatic learning technique where one seeks to
automatically generate rules from a learning database containing "examples" (in
general, cases already dealt with and validated).
When the system learns to classify according to a predetermined classica-
tion model as well as known examples.
Supervised learning is divided into two parts:
The rst corresponds to determining a labeled data model.
The second consists in predicting the label of a new datum, knowing the model
previously learned.
For example, discriminant analysis is a typical example. It is a matter of
explaining and predicting an individual's belonging to a predened class based on
its characteristics, as measured by predictive variables.
A learning database is a set of input-output pairs ( xm;ym)1mMwith
xm2Xandym2Y, which are considered to be drawn according to a law on xed
19

and unknown XY, for example xmfollows a uniform law and ym=f(Xm) +wm
wherewmis a centered noise.
The purpose of a supervised learning algorithm is, therefore, to generalize
for unknown entries what he has been able to learn from the data already dealt with
by experts in a reasonable manner. It is said that the prediction function learned
must have good guarantees in generalization.
Application
{Computer vision
{Patterns recognition
{Recognition of handwriting
{Speech recognition
{Bioinformatics
3.2.2 Unsupervised learning
In the computer eld, unsupervised learning (sometimes referred to as clus-
tering) is an automatic learning method. the aim is for a software program to divide
a heterogeneous group of data into subgroups so that the data considered to be the
most similar are associated within a homogeneous group and on the contrary, the
data considered dierent are found in other distinct groups. The objective being to
allow an extraction of organized knowledge from these data.Another form of unsu-
pervised learning is data partitioning, which is not always probabilistic.
This is when the system has only examples, and the number of classes and
their nature has not been predetermined, we talk about unsupervised learning or
clustering. The algorithm must discover the structure by itself according to the
data. In the gure is represented the schema of nsupervised learning process.
This method is distinguished from supervised learning by the fact that there
is no a priori output. In unsupervised learning, there is a collection of data collected.
Then the program treats these data as random variables and constructs a model
of densities for this dataset.In the gure 3.7 is represented the schema of the the
unsupervised learning process.
Figure 3.8: The unsupervised learning process
20

3.2.3 Semi-supervised Learning
It uses a set of labeled and untagged data. This is the case between super-
vised learning that uses only labeled data and unsupervised learning that uses only
unlabeled data. It has been shown that the use of unlabeled data, in combination
with labeled data, signicantly improves the quality of learning.
Another interest arises from the fact that data labeling requires the inter-
vention of a human user. When the datasets become very large, this operation can
be tedious. In this case, semi-supervised learning, which requires only a few labels,
is of obvious practical interest.
3.2.4 Reinforced learning
Learning by reinforcement refers to a class of automatic learning problems,
the aim of which is to learn from experiments what to do in dierent situations, so as
to optimize a quantitative reward time. The reinforcement learning corresponds to
the case where the algorithm learns a behavior given an observation. The action of
the algorithm on the environment produces a feedback value that guides the learning
algorithm.The most important steps are outlined in the Figure 3.9.
Figure 3.9: The reinforced learning
A classical paradigm for presenting reinforcement learning problems is to
consider an autonomous agent, immersed in an environment, and making decisions
based on its current state. In return, the environment provides the agent with a
21

reward, which may be positive or negative.
The agent seeks, through experiments, an optimal decision-making behavior
(called strategy or policy, which is a function associating the current state with the
action to be executed), in that it maximizes the sum of rewards over time.
3.3 The geometric characteristics of the human
face
The human face can be built as a 2D smooth surface S and represented by
coordinates from a subset
R2toR3.
x(
) = (x1(1;2);x2(1;2);x3(1;2)) (3.9)
Supposedly that the x1;x2andx3appetreinCrand@ix=@
@1x(i= 1;2)
are linearly independent.
The Riemannian metric is a characteristic of the subset and it is used to
measure the distances on S. The Riemannian complex is a pair ( S;g). A distance
element can be written as the metric tensor gij(x).
ds2=gijdidj;i;j= 1;2; (3.10)
The denition of the metric tensor gijis:
gij=@ix@jx;i;j = 1;2: (3.11)
A vector orthogonal can be written as:
N(x) =@1x@2x
jj@1x@2xjj2(3.12)
The second fundamental form, is write in coordinate:
bij=@ix@jN (3.13)
The second fundamental equation is written in the normal eld N. This
characteristic has a crucial role. The principal curvatures are maxandminof the
tensorbij=bikgkj. The mean curvature is value of H and the Gaussian curvature
is value of K, where
H=1
2trace (bij) =1
2(max+min); (3.14)
K=det(bij) =maxmin; (3.15)
In Fig. 3.7 can see the local coordinate system and the parametric manifold
.
22

Figure 3.10: The coordinate system
Is used the isometric model to describe facial expressions. The presume is
that natural deformations of face are isometries models.The isometric model is just
an approximation and structure of natural expressions other than pathological cases
like the open mouth.
23

Chapter 4 esxp
4.1 Viola-Jones Face Detection Algorithm
The Viola{Jones is the rst algorithm for face and object detection in real-
time made by Paul Viola and Michael Jones. In face detection has to prove if an
image contains a face and where it is localized in the image.A issue for considering
this tasks is that of binary classication. This classication is used to decrease the
incorrect classication risk.
The objective is to minimize the both rates( false negative and false pos-
itive) to achieve a good performance. A classier is not good if, in general, it can
not touch a predened classication target.To discriminate human faces from other
objects it is used an algorithm called Adaboost.
The algorithm works in four steps:
1. Haar Feature Selection
The simple features of Haar basis selection:
two-rectangle feature;
three-rectangle feature;
four-rectangle feature.
In Fig. 4.1 it can see two-rectangle features (image A), three-rectangle
feature (image B) and four-rectangle feature(image C).
Figure 4.1: The Haar basis selection
24

2. Creating an Integral Image
Rectangle features can be formate very rapidly using an integral image. The
integral image at position x, y whit sum of the pixels:
ii(x;y) =X
x0x;y0y;i(x0y0) (4.1)
whereii(x;y) is the integral image and i(x;y) is the original image. Using the
next pair:
S(x;y) =S(x;y1) +i(x;y) (4.2)
ii(x;y) =ii(x1;y) +S(x;y) (4.3)
whereS(x;y) is the cumulative row sum, if S(x;1) = 0 and ii(1;y) = 0, the
integral image can be compared to the original image in one step.
3. Adaboost Training
AdaBoost is a machine learning algorithm used to make a strong classier
assembled from a combination of weak classiers. A weak classier is mathe-
matically write as:
h(x;f;p; ) =1 for pf(x) >p
0 for pf(x)p
where f is the applied feature, p the polarity and determined if x is a positive
(a face) or a negative (a non-face).
4. Cascading Classiers
The structure of the cascade show the that within an single image
Figure 4.2: The cascaded classier
a lot of sub-windows are negative. Simpler classiers are used to reject the
25

most of subwindows and complex classiers are used to obtain low rates of
false positive. In Fig 4.2 is presented a schematic description of the cascading
classiers.
4.2 The diculties in face detection and recogni-
tion
In the human brain, the face recognition process is a high-level visual task.
Although people can detect and identify faces in a scene without too much trouble,
building an automated system that performs such tasks is a serious challenge.This
challenge is all the more important when the conditions for acquiring images are
very variable.
There are two types of variations associated with face images: inter and
intra-subject. Inter-subject variation is limited due to the physical resemblance
between individuals. On the other hand, the intra-subject variation is higher. This
can be attributed to several factors that we are discussing below.
4.2.1 The level of illumination
The appearance of a face in an image varies very much depending on the
illumination of the scene when shooting (see Figure ::::). Variations in lighting make
the task of face recognition very dicult. Indeed, the change in the appearance of
a face due to illumination is sometimes more critical than the physical dierence
between individuals and may lead to mis-classication of the input images.
This fact was experimentally observed in Aden et al, in that case, the au-
thors used a database of 25 people. Face recognition in an uncontrolled environment
remains an open research area. FRVT ratings have shown that the issue of illumi-
nation variation is a major challenge for facial recognition.
4.2.2 Variation of position
The rate of face recognition drops considerably when there are variations in
the pose in the images. This diculty has been demonstrated by evaluation tests
developed on the basis FERET and FRVT. The variation of pose is considered a
major problem for facial recognition systems.
When the face is prole in the image plane (orientation <30o), it can be
standardized by detecting at least two facial features (passing through the eyes).
However, when the rotation is greater than 30o, geometrical normalization is no
longer possible (see gure :::).
26

Figure 4.3:Variation of position
4.2.3 Expressions faciales
Another factor that aects the appearance of the face is the facial expres-
sion (see gure ::::::). The facial deformity that is due to facial expressions is located
mainly in the lower part of the face. The facial information located in the upper
part of the face remains almost invariable. It is usually sucient to carry out an
identication.
Figure 4.4:Expressions faciales
However, since facial expression changes the appearance of the face, it in-
evitably results in a lower recognition rate. Facial identication with facial ex-
pression is a dicult issue that is still valid and remains unresolved. The time
information provides signicant additional knowledge that can be used to solve this
problem.
4.2.4 The presence or absence of structural components
The presence of structural components such as beard, mustache, or spec-
tacles can dramatically alter facial features such as shape, color, or face size. In
addition, these components may hide the basic facial features, thus causing a failure
of the recognition system. For example, opaque glasses do not make it possible to
distinguish clearly the shape and the color of the eyes, and a mustache or beard
alters the shape of the face.
27

Figure 4.5:The partial coverage of the face
4.2.5 The presence of structural components
The face can be partially masked by objects in the scene, or by the port
of accessories such as glasses or scarf . In the context of biometrics, the proposed
systems must be non-intrusive, that is to say, that The active co-operation of the
subject should not be expected. Therefore, it is important to recognize partially
obscured faces.
Figure 4.6:The partial coverage of the face
Gross et al studied the impact of wearing sunglasses, and of the mask con-
cealing the lower face on facial recognition. Their experimental results seem to
indicate that, under these conditions, the performance of the recognition algorithms
remains low.
Conclusions
In this section, it has presented the technologies used in biometric systems
for the identication of individuals. It gives an overview of techniques for measuring
their performance. The face recognition is attracting more and more interest from
the scientic community, because it presents several technological challenges.
28

Finally, we have highlighted the dierent diculties inherent in the automatic recog-
nition of faces, which allowed us to clearly dene the problems dealt with in this
thesis.
4.3 Results obtained
The stages of the facial recognition process are the follows steps:
1. Create a database
2. Detecting and recognizing of faces patterns
3. Face detection
4. Training process
5. Recognition process
4.3.1 Databases
4.3.2 Detecting and recognizing of faces patterns
The process begins with identifying the face of the entire image, which usu-
ally contains a background and sometimes even other faces. If a human being is
very easy to distinguish what an individual's face is in a photo, the computer must
decide which pixels are on the face and which are not.
1. Nose detection
2. Mouth detection
3. Eyes detection
4.3.3 Face detection
1. Test 1
2. Test 2
Caz1
Caz 2
4.3.4 Training process
The histogram of oriented gradients (HOG) is a characteristic used in com-
puter vision for object detection. The technique calculates local histograms of the
orientation of the gradient on a dense grid, that is to say on areas regularly dis-
tributed over the image. The method has proved particularly eective in detecting
people.
29

Figure 4.7:The detection of the nose
The important idea behind the HOG descriptor is that the appearance and
local shape of an object in an image can be described by the distribution of the
intensity of the gradient or the direction of the contours. This can be done by
dividing the image into adjacent small-sized regions, called cells, and calculating for
each cell the histogram of the gradient directions or outline orientations for the pixels
within that cell. The histogram then forms the HOG descriptor. For best results,
local histograms are standardized in contrast, calculating a measure of intensity over
areas wider than cells, called blocks, and using this value to normalize all cells in
the block. This normalization allows better resistance to changes in illumination
and shadows.
4.3.5 Recognition
The facial recognition system will standardize – as much as possible – the
image so that it has the same dimensions, rotation, brightness with images in the
image from database. The standardized image is taken over by the face recognition
system.
30

Figure 4.8:The detection of the mouth
Figure 4.9:The detection of the eyes
31

Figure 4.10: The face detection
Figure 4.11: The detection of grup
32

Figure 4.12: The detection of grup2
Figure 4.13: HOG
33

Figure 4.14: Rezults
34

Conclusions
35

Bibliography
36