Contemporary Problems of Social Work
INFORMATION AND TELECOMMUNICATION TECHNOLOGIES..
Neural networks apparatus in biometrics
Автор/Author: Ziroyan M.A., Tusova E.A., Hovakimian A.S., Sargsyan S.G.
1. Kucharev G.A. Biometrical systems: methods and means of identification. St.Petersbourg
2. Degtarev A. Methods of identification by iris. Computer graphics and multimedia. 2004. Issue
№ 2 (6) http://www.cgm.computergraphics.ru/content/view/61
3. Matveev I. Human identification by iris. Security systems– 2004. Issue №5.
5. Anil K. Jain, Jianchang Mao, K. M. Mohiuddin. Artificial Neural Networks: A Tutorial. Computer
– Special issue: neural computing: companion issue to Spring 1996 IEEE Computational
Science & Engineering, Volume 29, Issue 3, March 1996.
6. Samta Gupta, Susmita Ghosh Mazumda. Sobel Edge Detection Algorithm. International
Journal of Computer Science and Management Research. Vol 2, Issue 2, February 2013
7. Just Kjeldgaard Pedersen, Simon. Circular Hough Transform. Aalborg University, Vision,
Graphics, and Interactive Systems. November, 2007.
8. Jolliffe, I.T. Principal Component Analysis (Springer Series in Statistics), Springer; 2nd
edition (October 2, 2002)
9. Anna Hovakimyan , Siranush Sargsyan , Arshak Nazaryan. Self-Organizing Map Application for
Iris Recognition. Journal of Commun. & Comput. Eng. ISSN 2090-623, www.m-sciences.com,
Volume 3, Issue 2, 2013, pp 10-13
References in Roman script:
1. Kuharev, G.A. Biometricheskie sistemy: metody i sredstva identifikatsii lichnosti cheloveka.
2. Degtyareva A. Metody identifikatsii lichnosti po raduzhnoy obolochke glaza. Komp’yuternaya
grafika i mul’timedia2004. – Vyp. № 2 (6) http://www.cgm.computergraphics.ru/content/
3. Matveev, I. Raspoznavanie cheloveka po raduzhke . Sistemy bezopasnosti. – 2004.Vyp. №5.
5. Anil K. Dzheyn, Jianchang Mao, K. M. Mohiuddin. Iskusstvennye Neyronnye Seti: Uchebnoe
Posobie. Komp’yuter – Spetsial’nyy vypusk: neyronnye vychisleniya: kompan’on vopros do
vesny 1996 po standartu IEEE vychislitel’nyh nauk i inzhenerii, tom 29, Vypusk 3, mart 1996.
6. Samta Gupta, Nepal’skaya Gosh Mazumda. Sobel Algoritm Obnaruzheniya Kraya.
Mezhdunarodnyy zhurnal vychislitel’noy tehniki i menedzhmenta issledovaniy. Tom 2,
Vypusk 2, Fevral’ 2013
7. Prosto Kjeldgaard Pedersen, Saymon. Krugovoy Hafa. Ol’borgskogo universiteta, videnie,
grafiki i interaktivnyh sistem. Noyabr’, 2007.
8. Jolliffe, i. t. Analiz glavnyh komponent (Springer Seriya v statistike), Springer; 2-e izdanie (2
9. Anna Ovakimyan Siranush Sarkisyan , Arshak Nazaryan. Samoorganizuyushayasya karta dlya
identifikatsii po raduzhnoy obolochke glaza. Vestnik kommun. & Komp’yut. Angl. ISSN v
2090-623, www.m-sciences.com, tom 3, Vypusk 2, 2013, str. 10-13
Over the past two decades, biometric identification technologies have made a great step
forward and at present take a strong position in the security market, which are in great demand
by various institutions and organizations: banks, airports, libraries, and others. In biometric
identification system the identification is carried out of biometric parameters of a person and
not by a clue or card. According to experts, the share of biometric systems in the coming years
will become a significant part of the total security market ,
Biometric identification methods divert. They are based on static biometric characteristics
of humans, such as fingerprints, facial geometry, retina, iris, hand vein pattern, as well as the
dynamic characteristics, such as voice, handwriting, heart rate, gait. According to most experts,
the main biometric characteristics that continue to be used for identification, fingerprints, face
geometry and iris [2,3].
The iris is a unique characteristic of man: it does not change throughout his life. It is believed
that the method of identification by the iris is the most accurate biometric technologies [1,2,3].
But today in the international market this method is used by only 6.9%, while the technology of
biometric fingerprint recognition forms more than half of the market.
CONTEMPORARY PROBLEMS OF SOCIAL WORK VOLUME 1, No. 2, 2015
Currently, the information technologies are of great help to achieve the objectives of
biometric identification ensuring efficient and effective processing of biometric information.
Biometric identification problem can be related to the recognition pattern, among which a
special place occupies the problem of classification. In these problems, the classification under
way to understand its relevance to one of the previously known classes of objects, formed on
the basis of data on the master image.
There are many ways to solve the problems of classification of images, among which we can
note the technology of artificial neural networks .
The paper considers the problem of constructing and implementing a software system that
takes as input a picture of the human eye, performs pre-processing of the input data on the
allocation of the iris using Sobel operator and Huff transformation and represents the image of
the iris as a binary vector. Before we apply the input vector to the neural network classification
the system selects the essential components in the vector, thereby reducing its dimension. The
designed neural network belongs to the type of self-organizing maps, coached by the method of
“learning without a teacher” and recognized as enough satisfactory for solving image pattern
classification [5, 6, 7]. Software implementation of the system has been implemented in Java.
The system has been tested on a number of test cases for the purpose of the experimental setup
used by the neural network.
1. Neural networks and self-organizing maps
Artificial neural network (ANN) is a mathematical model of the biological neuron. It consists
of neurons relating to each other. A neural network is subjected to learning by providing
her input information in the form of numerical sequences. In the training process internal
connections between neurons are set up through which the network endows the ability to
recognize unfamiliar images.
There are different types of neural networks that differ in both the topology and the learning
algorithms. In classification problems, some of the images are the most effective self-organizing
maps (SOC), which are trained to algorithms unsupervised . SOC consists of two layers: the
input and functional. In the functional layer the neurons are located on the grid composed of
cells. Each neuron takes a single cell and is connected to all neurons of the input layer (Figure
Fig.1. The structure of the self-organizing map
SOC operates on the principle of “winner takes all”. That is, when a sequence of input vectors
are applied to the network, all the neurons of the functional layer cause activation function,
which reflects the relationship between the input data and the weights of these neurons.
The winner becomes the neuron for which the activation function receives the optimal value,
say maximum or minimum. In this case, as the activation function is selected the Euclidean
distance between the input data and the weight of neuron function where an optimum value
is considered its minimal value. If V_1, V_2, V_3, ..., V_n is the input vector, and the functional
layer SOC consists of k neurons with weighting coefficient
W_11, W_12, W_13, ..., W_1n
W_21, W_22, W_23, ..., W_2n
W_k1, W_k2, W_k3, ..., W_kn
where W ij weighting coefficient of i-th neuron of functional layer associated with the j-th
neuron of the input layer, then the winner will be the neuron for which
At the stage of learning the network some training sequence of data is presented. The essence
of teaching is that at each step the neuron-winner is determined, and the neurons, located in a
neighborhood of the neuron-winner move closer to him. As a result of training neurons closely
located collect in the grid area, which are determined by the training sequence (Figure 2).
Fig.2. Neighborhood of the neuron-winner
The radius of the neighborhood of the neuron-winner varies according to the formula:
Experiments show that as the initial neighborhood radius value it is appropriate to take
where the width is the width of the functional layer mesh, height, λ is time constant
calculated by the formula
where n is the number of training iterations.
Thus, the radius of the neighborhood is changing exponentially (Figure 3)
CONTEMPORARY PROBLEMS OF SOCIAL WORK VOLUME 1, No. 2, 2015
Fig 3. Changing the radius of neighborhood of the neuron-winner
During the training, a network of neurons and neurons winner of its σ- neighborhood change
their weights as follows:
W (t + 1) = W (t) + θ (t) L (t) (V (t) -W (t))
where t is the number of iteration, L (t) = L_0 exp (-t / λ), t = 1,2,3, ...,
θ (t) = exp (- 〖〗 dist ^ 2 / (2σ ^ 2 (t))), t = 1,2,3, ..., dist –is the distance between neuron
and neuron-winner on the grid of the functional layer. The function L (t) is called the learning
rate, θ (t) reflects the influence of the location of the neuron on the grid. It is obvious that
most of the weight change of the winning neuron (Fig. 4).
Fig. 4. Effect of the distance from the neuron-winner at changing the balance
After training, the network is able to recognize the data, similar to those on which it is
2. Recognition of the iris
Iris recognition is the task of biometric identification. The iris is unique to each person and
remains unchanged throughout his life. It is reliable characteristics of identification for it can’t
be a subject to distortions caused by time or by direct contact with the capture of the sample.
Currently, there are many software systems that recognize human by iris [2,3].
In this paper the problem is solved by using a neural network self-organizing map type.
The problem of iris pattern recognition is divided into three subtasks:
– detection of the circle framing the iris on the eye,
– allocation of bandwidth to the iris and its pretreatment in order to isolate the iris essential
– iris recognition using a neural network type of self-organizing maps.
There are various algorithms to detect contours and shapes in the image. The most effective
algorithms among these are based on the transformation of Huff (Huff Transformation) . In
the current task considered, we apply to Huff transformation circle as far as the desired images
in the image are the circles; one that limits the iris, and the other that limits the pupil.
Huff transformation is performed more effectively when algorithms for detection and
isolation of the desired shapes on the borders of the image are previously applied. In this case
we used Sobel operator to solve this problem, . This is a discrete differential operator that
computes the approximate value of the gradient function, reflecting the brightness of the
image at each point. Thus the magnitude and direction of increasing the brightness in each
pixel of the image is detected, and hence marking the probability that the pixel is on a border of
the image. Fig. 5 shows the result of the Sobel operator applied to the image of the human eye.
The figure shows the selected borders of the eye.
Fig. 5. The result of the Sobel operator
You can improve the efficiency of the Sobel operator, considering the fact that the iris
region is surrounded by white, and it itself is of darker color. Consequently, there is a threshold
luminance pixel value, where the pixels with greater brightness value lie outside the region of
the iris, and pixels with lower luminance value in the iris area. Thus, if the pixel value is less
than the threshold, it is replaced with zero, otherwise it is replaced with the maximum possible
value. As a result of such replacement the image will be split into two classes of pixels - dark
and light, converting the image from color to black and white and the increasing contrast and
brightness of the pixels, owing to which Sobel operator will work more efficiently.
Huff transformation is a method of finding the pieces on the image. This is a common
mechanism by which you can find all kinds of shapes known in advance: straight lines, ovals, circles,
etc. In this paper we apply the transformation krugovoya Huff (Circular Huff Transformation),
designed to detect circles in the image that has been processed by Sobel operator. The data is
stored in a three-dimensional matrix, where the first two dimensions define two-dimensional
image, and the third dimension should take the value of the radius, where the circle is sought
in the image . Through all the boundary points, i.e. the points where the pixel value is
not zero, the circles are conducted, and the values in the matrix for the appropriate pixels are
incremented by one. As a result of the algorithm the point at which the circles intersect will be
the center of the desired circle. We take the circle, the radius of which is the maximum in the
matrix. In practice, usually circles with a fixed radius in advance are looked for .
After detection of circles, limiting the iris and the pupil, the band is allocated between
the circles,which contains information only on the iris. This information is represented as
n-dimensional vector where n is sufficiently large. The aim is to transform the vector in order to
isolate the main components that contain the most important information about the subject of
pattern recognition. We have applied the method of principal component analysis, the essence
of which is to identify similarities and differences in the values of attributes characterizing the
CONTEMPORARY PROBLEMS OF SOCIAL WORK VOLUME 1, No. 2, 2015
set of study data . The method works with a certain set of the objects represented by vectors
of n-dimensional space. By analysis of covariance matrix and finding its own vectors and own
values, n-dimensional space is replaced by m-dimensional space (m≤n), elements of which
contain the most important information about the object. Thus we can discard appropriately
selected (n - m) components of the input vector, without losing the basic characteristics of the
Vector processed with typical for iris information is input to the neural network,that solves
the problem of recognition through classification. Neural network of software is designed and
implemented in the type of neural network self-organizing map. The isolation of the major
components in the input vector encoding the iris, has also led to a decrease in the input layer of
neurons network that contributes to its effectiveness. The network is trained on a representative
set of images and tested on test cases. Originally selected network topology and the values of
its parameters have been refined as a result of experimental evaluation of the network.
3. Experimental setup parameters of self-organizing map
A neural network for biometric identification on iris has been implemented in Java. A library
of special classes has been developed, which provides parallel processing . The network is
trained on a certain training set of images, and tested on a number of test cases for the purpose
of the experimental setup of its parameters characterizing the topology of the network and the
details of its training.
The self-organizing map has four parameters, the values of which depend on the efficiency
of the network. These parameters are the number of internal neural network, the characteristic
of the grid which expresses the ratio of the width of the grid on which the neurons are allocated,
relation to its height, to the initial learning rate, the number of training iterations. On the basis
of experiments to classify iris the values of network parameters have been found under which
the network works most effectively. Efficiency is defined correctly with sharing images in the
set of test cases. Fig .6-9 demonstrate the results of assessment of network parameters.
Figure 6 shows the dependence of the efficiency of the network on the number of neurons
in the functional layer. The dependence is expressed in a percentage. The number of neurons
is taken of multiple number of training vectors. The abscissa represents the values of the
coefficient of multiplicity, the vertical axis shows the percentage of correctly recognized images.
Figure 6. The dependence of the efficiency of the network by the number of neurons
As the graph shows, network efficiency first increases sharply with increasing frequency
factors. Then the efficiency of the network does not grow. This value limit is equal to the
multiplicity factor 10. In other words, the network is most effective when the number of neurons
is approximately 10 times greater than the amount of training data.
Another parameter is the shape of a rectangle network mesh the functional layer. This
parameter is expressed by the ratio of net height to its width. Figure 7 shows the efficiency of
the network on the shape of a rectangle grid. The abscissa represents the value of the ratio of
height to its width grid, the vertical axis shows the percentage of correctly recognized images
Figure 7. Dependence of the efficiency on the shape of the mesh network
As can be seen from the graph, the network efficiency grows with an increasing net height to
its width, and at a value of 100, it almost stops growing. This value is chosen as the best value
for the parameter network.
Two other network parameters are associated with network training. The first is the initial
learning rate. Figure 8 shows the dependence of the efficiency of the network on this factor.
Figure 8. The dependence of the efficiency of the network from the initial learning rate
As seen on the chart, at first the effectiveness of the network is growing, but after reaching
the value of the learning rate of 0.1, it begins to decrease. Thus, the value of 0.1 for the initial
learning rate is considered optimal for the problem.
The last parameter of the network, which determines its effectiveness - is the number of
iterations training schedule. Fig. 9 illustrates this relationship. The abscissa represents values
of the number of iterations, and the ordinate shows percentage of correctly recognized images.
CONTEMPORARY PROBLEMS OF SOCIAL WORK VOLUME 1, No. 2, 2015
Figure 9. The dependence of the efficiency of the network on the number of training iterations
By plotting one can conclude that in the beginning, when the number of iterations increases,
the network efficiency grows. But when it reaches a value of 100, it almost remains unchanged.
We can therefore conclude that the number of iterations as the best can assume a value of 100.
Thus, on the basis of experiments testing the self-organizing maps and analysis of the
parameters under study, it can be stated that the efficiency of neural network like selforganizing
map for the task of recognition of the iris with an appropriate choice of parameter
values can form an average of 95%.
The network was also tested for other classification tasks, and it has been found that the
values of the parameters of the network obtained are also very satisfactory.
The study and practical implementation of a software system to solve the problem of
biometric identification of the iris using neural network technology have shown that the
proposed mechanism can be applied successfully in practice. It is highly desirable that the
network was trained on a larger number of images. This would provide an opportunity to better
It can be asserted with a certain degree of confidence that the proposed mechanism is quite
versatile and can be used to solve a wide class of problems for data mining.