Masters work
Kompanets Dmitry Olegovich
Abstract
Masters of Donetsk State Technical University
Faculty: Computer Information Technologies and Automatics
Speciality: Telecommunication Sistems and Networks
E-mail: dimadonetsk@ukrtop.com
Scientific headmaster: Turupalov Victor Vladimirovich
Development and research of network user
authentification/identification technics on a basis of facial asymmetry
characteristics
The technique of identification/authentication
of the user is developed on the basis of the person's characteristics
of asymmetry for databases of 100-400 users. The technique itself
is new and also is based on a phenomenon of human face asymmetry
and its ophtalmogeometrical characteristics uniqueness. The actuality
of work consists in revealing new ways of user's biometrical authentication.
Entrance data of a method is the motionless frontal image of the person. Its reception from the elementary Web-camera is possible. Target data are mathematical descriptions of identified images which can be processed, transferred, stored and compared in a consequence among themselves.
Figure 1 – Sequence of actions of recognition
system.
Advantages of biometric authentication[1]:
· Becomes possible to authenticate
the user, i.e. real acknowledgement of authenticity of the person
receiving access rights. Security of systems essentially raises
and, at the same time, process of identification of the user becomes
simpler – the user should not recollect, type and periodically change
passwords of access to various systems.
· The user may not remember and not enter the identification name (names).
· Authorization is carried out without dependence from language of an operational environment and codings of symbols.
· By virtue of simplicity of process, it can be initiated much more often, than traditional systems (requesting a name of the user and the password) allow.
· In all cases, except for cases of breaking of protection, it is possible to prove authorship of this or that electronic action confirmed by biometric authentication.
· Remote selection of the identifying information becomes harder.
· It is impossible to show the identifier by the third party.
Definition of the person on lineament is the closest to people as each person constantly solves for itself a similar problem. Besides universal distribution of digital videocameras (certainly, in the West) with which even house computers are equipped, do this method very attractive. However, really operating systems, suitable for use even in business, does not exist yet. Certainly, developments exist, but today users do not trust them their money (and it is modern criterion of safety). However, recently some countries including Russia, try to introduce at the airports and at railway stations systems of person detection, but the purpose of it is not restriction of access, but search of the criminal. For maintenance of safety of identification system on geometry of the face are still worked obviously insufficiently.
In machine vision often there are two updatings
of a problem of face detection – face localization and face tracking.
Localization of the person is the simplified variant of a problem
of detection as leans on knowledge that on the image there is one
and only one person. The tracking problem of moving face in a videostream
can be formulated as a problem of localization of the face on current
frame, leaning on the information on its position on the previous
frame.
The review of existing
researches:
Existing algorithms of persons detection is possible to break on two wide categories [2] . The first
category consists of the methods which are making a start experience of the person in
recognition of faces and doing attempt to formalize
this experience, having constructed on its basis automatic system of recognition.
The second category leans on toolkit of recognition of images,
considering a problem of detection of the face, as a special case of a recognition problem.
On the first category principles of patterns and other methods of recognition "top-down" were used, basically, in early works on detection of the person [3], [4], [5], [6], [7] . It were the first attempts of attributes formalization of face images, besides computing capacities of computers those years did not allow to use more complex methods of images recognition effectively. Despite of some naivety of algorithms, it is not necessary to underestimate value of these works as many techniques successfully applied now, have been developed or adapted for the given concrete problem in them.
Recognition "bottom-up" uses
invariant properties (invariant features) images of persons, leaning
on the assumption, that time the person can without efforts distinguish
the person on the image irrespective of its orientation, conditions
of illumination and specific features there should be some attributes
of presence of persons on images, invariant shootings concerning
conditions. Work algorithm of recognition"bottom-up" methods
can be briefly described as follows:
1. Detection of elements and features which are characteristic
for the image of the face.
2. The analysis of the found out features, decision on quantity
and arrangement of faces.
Second category of methods
approaches a problem on the other hand, and, not trying in an obvious
kind to formalize the processes occuring in a human brain, trying
to reveal laws and properties of the image of the face, applying methods
of mathematical statistics and machine training. Methods of this category
lean on toolkit of recognition of images, considering a problem of
detection of the person, as a special case of a problem of recognition.
To the image (or to its fragment) to some extent calculated vector
of attributes which is used for classification of images on two classes
– the face/not the face is put in conformity. The most widespread
way of reception of a vector of attributes this use of the image:
each pixel becomes a component of a vector, transforming the black-and-white
image n×m into spatial vector Rn×m. Lack of such
representation is extremely high dimension of space of attributes.
Advantage consists volume, that using the image entirely instead of
the characteristics calculated on its basis, from all procedure of
construction of the qualifier (including allocation of steady attributes
for recognition) is completely excluded participation of the person,
that potentially reduces probability of a mistake of construction
of wrong model of the image of the person owing to incorrect decisions
and errors of the developer.
Principal-component method [8]
is applied to decrease in dimension of space of attributes, not leading
to essential loss èíôîðìàòèâíîñòè a training set of objects (in this
case – images of persons). Application of a method of the main things
a component to a set of vectors of linear space Rn, allows
to pass to such basis of space, that the basic dispersion of a set
will be directed along the several first axes of basis named by the
main axes (or the main components). Thus, the basic variability of
vectors of a training set is represented several main components,
and there is an opportunity, having rejected remained (less essential)
to pass to space of essentially smaller dimension. Tense on the main
axes received thus subspace dimensions m << n is optimum among
all spaces of dimension m in the sense that in the best way (with
the least mistake) describes a training set of images.
Factor analysis (FA) [9],
As well as many methods of the analysis of multivariate data, leans
on a hypothesis that observable variables are indirect displays concerning
a small number of the certain latent factors. FA, thus, is set of
models and methods focused on revealing and the analysis of the latent
dependences between observable variables. In a context of problems
of the recognition, observable variables usually are attributes of
objects. The factor analysis is possible to consider as generalization
of a principal-component method.
The purpose of training of the majority of
qualifiers – to minimize a mistake of classification on a training
set (named by empirical risk). Unlike them, by means of a method of
basic vectors it is possible to construct the qualifier minimizing
the top estimation of an expected mistake of classification (including
for the unknown objects which are not entered into a training set).
Application of a method of basic vectors to a problem of detection
of the person consists in search of a hyperplane in ïðèçíàêîâîì the
space, a separating class of images of persons from images "not-persons".
Opportunity of linear division of so complex
classes as images of persons also it is not represented improbable.
However, classification by means of basic vectors allows to use the
device of nuclear functions for implicit displaying vectors-attributes
in space potentially much higher dimension in which classes can linearly
appear dividable. Implicit displaying by means of nuclear functions
does not lead to complication of calculations that allows to use successfully
the linear qualifier for linearly inseparable classes.
Neuronets are also successfully applied to
the decision of many problems of recognition for a long time. A plenty
of neural networks of various architecture ware applied to the decision
of a problem of detection of the person, in particular: multilayered
perseptrons, probabilistic decision-based neural networks (PDBNN)
[10], etc. Advantage of neuronet usage for
the decision of a problem of face detection is the opportunity of
reception of the qualifier well modelling complex function of distribution
of images of persons p(x | face). Lack is necessity for careful and
laborious adjustment íåéðîñåòè for reception of satisfactory result
of classification.
SNoW for face detection[11]
represents a two-layer network which entrance layer consists of units,
each of which corresponds to some characteristic of the entrance image
(generates 1 at presence of some feature and 0 in case of its absence
on the image), the day off consists all from two units, each of which
corresponds to distinguished classes of images ("person",
"non-person"). As characteristics of the image flags of
equality to the certain sizes of average value and a dispersion of
brightness in each of rectangular fragments of the image in the size
1x1, 2x2, 4x4 and 10x10 (all images has the size 20x20 pixels) are
used. It gives space of attributes of dimension 135424. At carrying
out of classification on entrance units the information on presence
of the certain characteristics in the processable image moves. Units
of a target layer calculate a linear combination of the signals generated
in entrance units. Factors of a linear combination are set by weights
of communications between entrance and target units. At excess of
the set threshold, the decision on presence of the person on the image
makes.
Solved problems:
Application of the given method in monitoring systems of access will allow to simplify procedure of identification/authentication of the user: not only for an input in a network, but also for authorization necessity of input of the password vanishes in any appendices. Also to become possible to lower expenses for safety of computer systems and networks. It, in turn, will allow to achieve a wide circulation of biometric systems and to raise the general level of safety of computer systems.
By development of algorithm available methods
of person's localization in the image and algorithms of definition
of eyes' position are considered. The algorithm «OFG» carries out
following functions:
- normalizes the received image (eliminates turns of a head, normalizes light parameters);
- visualizes ophtalmogeometric pattern
(construction of system of the coordinates, necessary points);
- carries out operations and calculations
above it;
- makes comparison.
Identification
of users on the basis of face asimmetry characteristics.
The chosen method is based on Czestochowa exact model of face asymmetry and on researches in the field of ophthalmogeometry by professor E.Muldashev.
Figure 2 – Representation of key parameters
of the face used in ophthalmogeometry.
Construction of ophtalmogeometrical pattern (Figure 3):
On the basis of available points and parameters
of face model of the person we shall construct ophtalmogeometrical
pattern in following sequence:
Animation – Sequence of the user authentication
procedure (press play).
1. To define diameter of the iris and
to equate to 1 Muld (fi = 1 Muld). It is defined by three points
Ar, Br, Cr. Concerning them then the iris circle is constructed,
its diameter is found and is equated to 1 Muld. Muld is the size
equal to diameter of an iris. This size equal 10±0,56 mm also is
a unique constant from biometric characteristics of the person,
since 4-5 - years age. Considering this circumstance, all ophtalmogeometrical
patterns can be normalized, i.e. to lead to one scale for convenience
of processing.
2. To measure distance between B1r
and B1l Base1 [Muld]. The distance between centres of irises of
frontally looking person forms base of the face. It is measured
also in Mulds and serves for the coordination of the sizes of the
images received from the camera and stored in the base.
3. To define the middle between irises O*. We find a point necessary
for construction of system of coordinates {Y *, X *}.
4. To define points 1* and 2*. These
points are necessary for construction of a perpendicular to axis
X*.
5. To lead lines {Y *, X *}, Y* – a line which is passing through
points
1* - O* - 2 *, X* – a line which is passing through points B1r -
O* - B1l.
6. To construct system of coordinates (lines) {X1, X2, Y} according
to figure 3. The given system of coordinates is constructed with
the purpose of a finding of the least changing base. Concerning
it the further readout and measurements are made. Its basis is made
with two points Î1 and Î2. Point Î1 is as a midpoint, connecting
points Îcr and Ocl. These points are one of the most motionless
on the face. They do not change the position with the years, with
an emotional condition. Axis X of system of coordinates is led through
them. Axis Y is a perpendicular to axis Õ1 in point Î1. Point Î2
is on axis Y on distance 4 Mulds from axis Õ1. This point also should
satisfy to a principle of the least susceptibility to changes. It
to be in the geometrical middle between a line connecting tips lobes
of ears and a line passing through corners of a mouth with the center
in the geometrical middle of an oval of the person. Axis Õ2 is parallel
to axis Õ1 and passes through point Î2.
7. To measure a corner α1, a measure of deviation Y* from Y.
Figure 3 - Construction of axes of coordinates.
Definition of model:
1. To define points in corners of eyes
4,8,8* and 2,6,6*.
2. In middle point across and on a vertical between 4,8* and 2,6*
to find the geometrical center of an eye (can not coincide with
the center of an iris of the eye).
3. Of the received center to construct 16 concentric beams, and
one of beams should pass obligatory through a point 8 * (6 *).
Figure 4 – Construction of concentric beams of an eye.
4. To construct the straight lines which
are passing through a point 4 (2) and the nearest points of concentric
beams crossing with a contour of an eye. On crossing of these straight
lines we shall receive points 1 and 3.
5. To construct the straight lines which are passing through a point
8 (6) and two nearest crossings with concentric beams, except for
the beams which are passing through points 8 and 6 (on two crossings
from above and from below) and to receive points 7 and 5.
6. To construct a tangent to the top points of eyebrowes. Its crossing
with the straight lines which are passing through points 7,8 both
7,6 will give points 14 and 16.
7. To construct a point 9 on crossing of tangents in two extreme
points of a fold above eyes (points 12 and 10).
8. To find the geometrical middle between points 8* and 6* (point
Î1).
Figure 5 – Construction of ophtalmogeometric pattern.
Thus, we receive individual ophtalmogeometrical
portrait of each person.
 
Figure 6 – Visualization of an image
and allocation of meaning points.
Identification of a human face should pass
whenever possible in a neutral emotional condition. In case of incorrect
relieving of an image from the camera (the big inclination or turn
of a head) it is necessary for user to repeat procedure of reading
of an image.
Reception of numerical characteristics of an identified pattern.
The point 0 is the beginning of system
of coordinates. From it readout of two-dimensional coordinates of
points is made. A unit of measure of system of coordinates is Muld.
Further, having two-dimensional system of coordinates and universal
unit of measurements, to all received points of model can appropriate
two coordinates (X, Y), concerning O. Significant points in the
developed system are 1, 5, Î1, 9, 7, 3. Obtained data is easily
tabulated.
Then the database of images of network
users (login=pattern) is created. In it all the images are stored
in the coordinated size and identical quality. During procedure
of authentication the user enters a login, or otherwise defines
itself. The system finds an image corresponding a login and makes
comparison (so-called comparison one to one). It makes the decision
on conformity of an image to the entered login. During procedure
of identification the login is not required. The system compares
an entrance image to all base and then itself defines the user.
Such method borrows more time, but has also a number of advantages:
noncontactness and anonymity (the user even can not know about spent
procedure). Procedure of comparison of an entrance image from base
occurs to image on comparison of values and signs on each coordinate.
Mean square value is the final value.
| |
Pattern 2 |
Pattern 1 |
|
| |
Value |
|
Value |
|
Delta |
| Xo1 |
-0,18
|
- - |
0,04
|
+ |
0,22 |
| Yo1 |
0,22 |
+ |
-0,35 |
- |
-0,57 |
| X3 |
-0,3 |
- |
0,33 |
+10% |
0,64 |
| Y3 |
-1,35 |
|
-1,62 |
|
-0,17 |
| X1 |
-0,47 |
- |
0,64 |
+ |
1,13 |
| Y1 |
3,77 |
|
2,88 |
51,67% |
-1,09 |
| X5 |
-0,18 |
|
0,36 |
|
0,54 |
| Y5 |
0,21 |
|
0,25 |
|
0,04 |
| X7 |
-0,03 |
|
0,17 |
|
0,20 |
| Y7 |
-1,2 |
|
-0,17 |
|
0,20 |
| X9 |
0,16 |
|
0,13 |
|
-0,03 |
| Y9 |
-0,99 |
|
-0,76 |
|
0,33 |
| Mean square value
|
0,61 |
Table 1 – Comparison of patterns of two
different people of one age.
Mean square value of a deviation of coordinates
of points is the defining parameter at comparison of images. But
a sign on coordinate of a point which defines its position concerning
an axis of coordinates also it is necessary to consider.
Main results of the work:
1. Exact two-dimensional asymmetric model
of the person has been developed. Two basic problems which have
arisen during creation of an image: a finding of a "truthful"
vertical axis on the face of the person (axis Y) and creation of
the special combined system of coordinates {Y-O1-X1-O2-X2} with
points Î1 and Î2 for displaying of asymmetric images of the face
and normalization of two compared images or their fragments.
2. The basic points, facial elements and
their significant combinations from the anthropological point of
view which can be applied in the automated methods have been defined.
The basic innovation is application of unit of measure Muld for
normalization of the two-dimensional image of the person. Muld is
defined by diameter of an iris of the eye of and it is equal 10±0,56
mm. Diameter of an iris of the eye is a unique constant from biometric
characteristics of the person, since age of 4-5 years.
3. The algorithm of measurement of pseudo-information
similarity of two asymmetric faces which are nonmathematical objects
has been developed. Visualization and approbation of ophtalmogeometrical
pattern has been made.
4. Some interesting features of face asymmetry
and ophtalmogeometrical pattern have been noticed: uniqueness of
face asymmetry and ophtalmogeometrical characteristics; a topological
invariance of the curve which is passing through significant points
of ophtalmogeometrical pattern, without dependence from age and
an emotional condition; uniqueness of a way of normalization of
the image of the person, using as unit of measure Muld.
Future trends: on the
basis of the further researches it is necessary to find out a threshold
of sensitivity of a method and to define optimum value of a threshold
of operation. It depends on concrete conditions of application of
system, the size of base of images and quality of used equipment.
Literature:
1. V. Zadorojnij
Biometrics in the general words. [http://www.biometrics.ru/document.asp?group_id=11&nItemID=9&sSID=3.7]
2. V. Vejentsev, À. Degtjareva
Detection and localization of the face on the image. [http://cgm.graphicon.ru:8080/issue2/face_detection/index.html
]
3. G. Yang and T. S. Huang, “ Human Face
Detection in Complex Background,” Pattern Recognition, vol. 27,
no. 1, pp. 53-63, 1994.
4. C. Kotropoulos and I. Pitas, “
Rule-Based Face Detection in Frontal Views ,” Proc. Int’l
Conf. Acoustics, Speech and Signal Processing, vol. 4, pp. 2537-2540,
1997.
5. T. Sakai, M. Nagao, and S. Fujibayashi,
“Line Extraction and Pattern Detection in a Photograph ,” Pattern
Recognition, vol. 1, pp. 233-248, 1969.
6. I. Craw, H. Ellis, and J. Lishman, “Automatic
Extraction of Face Features,” Pattern Recognition Letters, vol.
5, pp. 183-187, 1987.
7. V. Govindaraju, “ Locating Human Faces
in Photographs ,” Int’l J. Computer Vision, vol. 19, no. 2, pp.
129-146, 1996.
8. "Method of the main components"
Digital library of computer grafics and multimedia laboratory at
faculty VMiK of the Moscow State University http://library.graphicon.ru/catalog/19.
9. “Factoral analysis” Digital library
of computer grafics and multimedia laboratory at faculty VMiK of
the Moscow State University
http://library.graphicon.ru/catalog/217.
10. S.-H. Lin, S.-Y. Kung, and L.-J. Lin,
“ Face recognition/detection by probabilistic decision-based neural
network ,” IEEE Trans. Neural Networks 8, pp. 114-132., 1997
11. N.
P. Costen , M. Brown "Exploratory Sparse Models
for Face Recognition" British Machine Vision Conference,
2003
Kompanets
D. O.
Development and research of network user
authentification/identification technics on a basis of facial asymmetry
characteristics
|