=Paper=
{{Paper
|id=Vol-3421/short11
|storemode=property
|title=Methods of Face Recognition in Video Sequences and Performance Studies (short paper)
|pdfUrl=https://ceur-ws.org/Vol-3421/short11.pdf
|volume=Vol-3421
|authors=Mariia Nazarkevych,Vitaly Lutsyshyn,Hanna Nazarkevych,Liubomyr Parkhuts,Maryna Kostiak
|dblpUrl=https://dblp.org/rec/conf/cpits/NazarkevychLNPK23
}}
==Methods of Face Recognition in Video Sequences and Performance Studies (short paper)==
https://ceur-ws.org/Vol-3421/short11.pdf
Methods of Face Recognition in Video Sequences
and Performance Studies
Mariia Nazarkevych1, Vitaly Lutsyshyn1, Hanna Nazarkevych2, Liubomyr Parkhuts2,
and Maryna Kostiak2
1
Lviv Ivan Franko National University, 1 Universytetska str., Lviv, 79000, Ukraine
2
Lviv Polytechnic National University, 12 Stepan Bandera str., Lviv, 79013, Ukraine
Abstract
A method of capturing a person’s face in a video stream has been developed. The developed
methods of capturing the video stream are considered. Tracking methods are used in video
surveillance. Methods of video stream capture, image frame extraction, and face recognition
are considered. The method of flexible comparison on graphs, the principal component method,
The Viola-Jones method, Local binary patterns, and Hidden Markov models, which are used
for face recognition, are considered. The library in Python DeepFace was studied. Face
recognition experiments were conducted. Faces photographed in the genre of selfie, portrait,
and documentary photography were recognized. It has been found that the best recognition
methods are found in the genre of photography. The recognition results are somewhat worse
for selfies. The worst ones are for digital photography. Recognition was based on the
MediaPipe Face Detection library. The recognition time was from 10 to 22 mc.
Keywords 1
Face recognition, object tracking, machine learning
1. Introduction scene occlusions, and camera movement.
Tracking is usually performed in the context of
higher-level applications that require the
Tracking objects in surveillance camera
location and/or shape of an object in each frame.
footage is a challenging task. It is much more
Typically, assumptions are made to limit the
difficult to track objects in video sequences to
tracking problem in the context of a particular
improve their recognition. There are many
application. In this review, we classify tracking
existing object-tracking methods, but all have
methods based on the object and motion
some drawbacks. Some of the existing object-
representations used. Object tracking consists
tracking models are region-based contour
of using appropriate image features, selecting
models [1]. Tracking—tracking an object in a
motion models, and detecting objects [3]:
video sequence; and detection—detecting an
object in a video sequence. Tracking-by- • Target representation object.
detection—trackers first run a detector for each • Localization object.
frame, and then the tracking algorithm Difficulties arise when objects move fast
associates these detections to determine the compared to the frame rate or when the tracked
movement of individual objects and assign them object changes direction in time [4–6]. The
unique identification numbers [2]. sequential flow of object detection, object
Tracking objects is a complex problem. tracking, object identification, and object
Difficulties with object tracking can arise from behavior completes the tracking process [7].
abrupt object movement, changing appearance Video processing consists of the following steps:
patterns of both the object and the scene, non- video upload [8], prepro-cessing, a proposed
rigid object structures, object-object and object-
CPITS 2023: Workshop on Cybersecurity Providing in Information and Telecommunication Systems, February 28, 2023, Kyiv, Ukraine
EMAIL: mariia.a.nazarkevych@lpnu.ua (M. Nazarkevych); vitalylutsyshyn@gmail.com (V. Lutsyshyn); hanna.ya.nazarkevych@lpnu.ua
(H. Nazarkevych); liubomyr.t.parkhuts@lpnu.ua (L. Parkhuts); Kostiak.maryna@lpnu.ua (M. Kostiak)
ORCID: 0000-0002-6528-9867 (M. Nazarkevych); 0009-0008-1229-6706 (V. Lutsyshyn); 0000-0002-1413-630X (H. Nazarkevych); 0000-
0003-4759-9383 (L. Parkhuts); 0000-0002-6667-7693 (M. Kostiak)
©️ 2023 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
246
algorithm that includes video processing, then the A video stream is streamed in which a face
object capture step (Fig. 1). needs to be recognized [11]. We determine the
size of the face coordinates. The face contour is
Video frames aligned and the basic parameters are determined
(Fig. 2). As a result, a parametric vector is built.
The parameters are compared. As a result,
recognition is performed.
Preprocessing
stream
video
Proposed algorithm
Facial recognition
Moving object detection and tracking
Figure 1: Scheme of video processing and object
outline selection
Face contour alignment
2. Object Recognition
alignmen
after
Face
t
The capture and encoding of digital images
should result in the creation and rapid
dissemination of a huge amount of visual
Determination of basic parameters
information. Hence, efficient tools for searching
and retrieving visual information are essential.
Although there are effective search engines for
text documents today, there are no satisfactory
systems for retrieving visual information.
Due to the growth of visual data both online Comparison of parameters. Result
and offline [9] and the phenomenal success of web
search, expectations for image and video search Figure 2: Face recognition algorithm
technologies are increasing. Object detection in offline video. This
However, with the evolution of video camera approach estimates the behavior of perceived
characteristics that can record at high frame rates objects and works best as a complement to other
in good quality, and with advances in detection, offline video-based object detection systems [12].
such as new approaches based on Convolutional In recent years, various other video object
Neural Networks (CNNs), the basis for Tracking- detection systems have emerged that have tried to
by-detection trackers [10] has become more use 3D convolutional networks that analyze many
robust. The requirements for a tracker in a images simultaneously.
tracking system have changed dramatically, Knowledge-based methods use information
allowing for much simpler tracking algorithms about the face, its features, shape, texture, or skin
that can compete with more complex systems color. In these methods, a certain set of rules is
requiring significant computational costs. distinguished that a frame fragment must meet to
Let’s analyze three ranking algorithms that be considered a human face. It is quite easy to
take into account the spatial, temporal, and define such a set of rules (Fig. 3). All rules are
spatiotemporal properties of geo-referenced video formalized knowledge that a person uses to
clips. determine whether a face is a face or not.
Object detection requires training machine For example, the basic rules are: the areas of
learning models, such as Recurrent Neural the eyes, nose, and mouth differ in brightness
Networks (RNNs) and CNNs, on images where from the rest of the face; the eyes on the face are
objects have been manually annotated and always symmetrically positioned relative to each
associated with a high-level concept. other. Based on these and other similar properties,
247
algorithms are built that check whether these rules systems, graphs can have a rectangular lattice and
are fulfilled in the image during execution. The a structure formed by characteristic
same group of methods includes a more general (anthropometric) points of faces.
method—the pattern-matching method. In this Graph edges are weighted by the distances [16]
method, a face standard (template) is determined between adjacent vertices. The difference
by describing the properties of individual face (distance, discriminative characteristic) between
areas and their specified relative position, with two graphs is calculated using a certain
which the input image is subsequently compared. deformation cost function that takes into account
both the difference between the feature values
Computing methodology calculated in the vertices and the degree of
deformation of the graph edges.
The graph is deformed by shifting each of its
Artificial Computer vertices by a certain distance in certain directions
inteligents graphics relative to its original location and choosing such
a position at which the difference between the
feature values in the vertex of the deformed graph
Computer and the corresponding vertex of the reference
image
vision manipulation graph is minimal. This operation is performed in
turn for all graph vertices until the smallest total
difference between the features of the deformed
Object and reference graphs is achieved. The value of the
recognition deformation cost function at this position of the
graph will be the measure of the difference
between the input face image and the reference
Detection
Face detection graph. This “relaxation” deformation procedure
should be performed for all reference faces in the
Figure 3: Classification of face detection system database. The result of the system’s
Face detection using such methods is recognition is the reference with the best value of
performed [13] by searching all rectangular the deformation cost function.
fragments of the image to determine which class The disadvantages of the method include the
the image belongs to. complexity of the recognition algorithm and the
Viola-Jones object detection [14]. The method complicated procedure for entering new templates
was proposed by Paul Viola and Michael Jones into the database.
and became the first method to demonstrate high The best results in face recognition were
results in real-time image processing. The method shown by the CNN or convolutional neural
has many implementations, including as part of network. The success is due to the ability to
the OpenCV computer vision library understand the two-dimensional topology of the
(cvHaarDetectObjects function). The advantages image, unlike the multilayer perceptron.
of this method are high speed (due to the use of a The distinctive features of CNN are local
cascade classifier); high accuracy in detecting receptor fields (providing local two-dimensional
turned faces at an angle of up to 30 degrees. The connectivity of neurons), common weights
disadvantages include a long training time. The (providing detection of some features anywhere in
algorithm needs to analyze a large number of test the image), and hierarchical organization with
images. spatial subsampling. Thanks to these innovations,
The method of comparison on graphs (Elastic the CNN provides partial resistance to scale
graph matching) [15]. This method is related to changes, shifts, rotations, changes in angle, and
2D modeling. Its essence lies in the comparison of other distortions.
graphs describing faces (a face is represented as a CNN was developed in DeepFace, which was
grid with an individual location of vertices and acquired by Facebook to recognize the faces of its
edges). Faces are represented as graphs with social network users.
weighted vertices and edges. At the recognition Geometric face recognition method [17] is
stage, one of the graphs, the reference graph, one of the first face recognition methods used. The
remains unchanged, while the other is deformed methods of this type of recognition involve the
to best match the first graph. In such recognition selection of a set of key points (or areas) of the
face and the subsequent formation of a set of
248
features. The key points can include the corners of database containing images of faces with slight
the eyes, lips, the tip of the nose, the center of the changes in lighting, scale, spatial rotation,
eye, etc. The advantages of this method include position, and various emotions showed 96%
the use of inexpensive equipment. The recognition accuracy. The disadvantages of
disadvantages are as follows: low statistical methods based on neural networks include the
reliability, high lighting requirements, and addition of a new reference face to the database,
mandatory frontal image of the face, with small which requires complete retraining of the network
deviations. It does not take into account possible on the entire available set, and this is a rather
changes in facial expressions. lengthy procedure that, depending on the size of
The method of flexible comparison on the sample, requires hours of work or even several
graphs [18], the essence of which is to compare days.
graphs describing the image of a person’s face. Local Binary Patterns (LBPs) [15] were first
Some publications indicate 95–97% recognition proposed in 1996 to analyze the texture of
efficiency even in the presence of different halftone images. Studies have shown that LBPs
emotional expressions and changes in the angle are invariant to small changes in lighting
when forming a face image up to 15 degrees. conditions and small image rotations. LBW-based
However, it takes about 25 seconds to compare the methods work well when using images of faces
input face image with 87 reference images. with different facial expressions, different
Another disadvantage of this approach is the low lighting, and head turns. Among the
manufacturability of memorizing new standards, disadvantages is the need for high-quality image
which generally leads to a non-linear dependence preprocessing due to high sensitivity to noise, as
of the operating time on the size of the face the number of false binary codes increases in its
database. The main advantage is low sensitivity to presence.
face illumination and changes in face angle, but Hidden Markov models [16]. A hidden
this approach itself has lower recognition Markov model is a statistical model that simulates
accuracy than methods built using neural the operation of a process similar to a Markov
networks. process with unknown parameters. According to
The Principal Component Method (PCM) the model, the task is to find unknown parameters
[19] reduces the recognition or classification based on other observed parameters. The obtained
process to the construction of a certain number of parameters can be used in further analysis for face
principal components of images for an input recognition. From the point of view of
image. However, in cases where there are recognition, an image is a two-dimensional
significant changes in illumination or facial discrete signal. The observation vector plays an
expression in the face image, the effectiveness of important role in building an image model. To
the method is significantly reduced. avoid discrepancies in descriptions, a rectangular
The Viola-Jones method [14] allows you to window is usually used for recognition. To avoid
detect objects in images in real-time. The method losing data areas, rectangular windows should
works well when observing an object at a small overlap each other. The values for overlap, as well
angle, up to about 30°. The recognition accuracy as the recognition areas, are selected
using this method partially reaches over 90%, experimentally. Before use, the model must be
which is a good result. However, at a deviation trained on a set of pre-labeled images. Each label
angle of more than 30°, the recognition has its number and defines a characteristic point
probability drops sharply. This feature makes it that the model will have to find when adapting to
impossible to detect a face at an arbitrary angle. a new image.
Use of neural networks.
One of the best results in face recognition is 3. Face Detection
achieved by using CNNs, which are a logical
development of such architectures as cognition
MediaPipe Face Detection is a face detection
and recognition. The success is due to the ability
to take into account the two-dimensional topology software product that includes 6 landmarks and
of the image, unlike the multilayer perceptron. support for multiple faces. It is based on
Thanks to these innovations, the ANN provides BlazeFace [17], a lightweight and high-
partial resistance to scale changes, shifts, performance face detector specifically designed
rotations, changes in perspective, and other for mobile GPUs. The detector’s real-time
performance allows it to be applied to any real-
distortions. Testing of the ANN on the ORL
249
time video stream that requires an accurate face A collection of detected faces, where each face
region to be used as input to other task-specific is represented as a proto-message containing a
models, such as 3D face keypoint estimation (e.g., bounding box and 6 key points (right eye, left eye,
MediaPipe Face Mesh), facial features, or facial nose tip, the center of the mouth, right ear tragion,
expression classification, and face region and left ear tragion). The bounding box consists of
segmentation. BlazeFace utilizes a simplified xmin and width (both normalized to [0.0, 1.0] by
feature extraction network inspired by the width of the image), and ymin and height (both
MobileNetV1/V2, but distinct from it, a GPU- normalized to [0.0, 1.0] by the height of the
friendly binding scheme modified from Single image). Each key point consists of x and y, which
Shot MultiBox Detector (SSD). are normalized to [0.0, 1.0] by the width and
height of the image, respectively (Fig. 4).
Figure 4: Face capture in video
4. Face Mash 3D primitives, including a face pose
transformation matrix and a triangular face mesh
[21]. A lightweight statistical analysis method
MediaPipe Face Mesh is a solution that
called Procrustes Analysis is used to drive robust,
estimates 468 3D facial landmarks in real-time,
efficient, and portable logic. The analysis is
even on mobile devices [18, 19]. The program
performed on the CPU and has a minimal speed
uses machine learning to determine the 3D surface
footprint.
of the face, requiring only a single camera input
The machine learning pipeline consists of two
without the need for a special depth sensor. Using
real-time deep neural network models that work
a simplified modeling architecture along with
together [22]: a detector that works on the full image
GPU acceleration throughout the pipeline, the
and calculates the location of the face, and a 3D
solution delivers real-time performance that is
facial landmark model that works on these locations
critical.
and predicts an approximate 3D surface using
Additionally, the solution comes with a face
regression. Accurate face cropping significantly
transformation module that bridges the gap
reduces the need for conventional data
between facial landmark estimation and useful
augmentation.
real-time Augmented Reality (AR) applications
The pipeline is implemented as a MediaPipe
[20]. It establishes a metric 3D space and uses the
graph that uses a face landmark subgraph from the
positions of facial landmarks on the screen to
face landmark module and visualizes using a special
estimate facial transformations in that space. The
face renderer subgraph. The face landmark subgraph
face transformation data consists of conventional
250
internally uses the face_detection_subgraph from 5. Model Development
the face detection module.
The face detector is the same BlazeFace model
There are two models in this solution: general
used in MediaPipe Face Detection.
and landscape. Both models are based on
For 3D facial landmarks, we applied transfer
MobileNetV3 with modifications to make them
learning and trained the network with multiple
more efficient. The general model works with a
objectives: the network simultaneously predicts 3D
256×256×3 (HWC) tensor and outputs a
landmark coordinates on synthetic visualized data
256×256×1 tensor representing the segmentation
and 2D semantic contours on annotated real-world
mask. The landscape model is similar to the
data. The resulting network provided us with
general model but works on a 144×256×3 (HWC)
reasonable predictions of 3D landmarks not only on
tensor. It has fewer FLOPs than the regular model
synthetic but also on real-world data [23, 24].
and is therefore faster. MediaPipe Selfie
The 3D landmark network receives a cropped
Segmentation automatically resizes the input
video frame as input without additional depth
image to the right tensor size before feeding it to
input. The model outputs the positions of the 3D
the ML model [27].
points, as well as the probability of the presence
The general model also supports ML Kit, and
and proper alignment of a face in the input data
the landscape model option supports Google Meet
[25, 26]. A common alternative approach is to
(Fig. 6).
predict a 2D heat map for each landmark, but it
does not lend itself to depth prediction and has
high computational costs for so many points. We
further improve the accuracy and reliability of our
model by iterative loading and refining the
predictions. In this way, we can increase our
dataset to increasingly complex cases such as
grimaces, obliques, and occlusions.
This method can be used for a variety of face
masking applications (Fig. 5).
Figure 6: Landscape model—segmentation mask
During this experiment, the issue of
recognizing objects in a video stream was
considered. The main Python libraries that can be
used to recognize and classify objects from video
are highlighted. MediaPipe methods for achieving
a particular result in recognition are clearly
described (Fig. 7).
Figure 5: Creating a face mask in a video track
img = cv2.imread(img_path)
cv2.imshow('image', img)
cv2.waitKey(0)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
faces = face_cascade.detectMultiScale(gray,1.1,5)
faces_detected = "Знайдено обличчя: " + format(len(faces))
print(faces_detected)
for (x, y, w, h) in faces:
cv2.rectangle(img, (x, y), (x+w, y+h), (255, 0, 0), 2)
cv2.imshow('img', img)
cv2.waitKey()
Figure 7: Fragment of the face detection program
251
6. Experimental Results photo was taken in the style of a selfie, a
documentary photo, and a portrait. In addition,
these images had different variations in quality
About 50 images of graduating master’s
and contained several facets of variation in color,
students were used. The images were taken from
position, scale, rotation, pose, and facial
mobile phones. Subsequently, after taking photos,
expression. We present the detection results in
they recorded videos in MPEG7 format. In the
Tables 1 and 2 for the HHI MPEG7 image set. The
experiment, the placement of the face to the plane
face was fascinated by tracking.
of the photograph was taken into account. The
Table 1
FP: False Positives, DR: Detection Rate
Out of man Frontal Close to the Semi-profile Profile
frontal
Selfi
Number of images 12 10 7 15
Image size
FP: False Positives 6204 5205 3090 2580
DR: Detection Rate 87% 85% 90% 99%
Time (sec) 10мс
Portrait
FP: False Positives 5290 5005 2590 2800
DR: Detection Rate 93% 92% 85% 95%
Time (sec) 18 mc
Documentary photography
FP: False Positives 3458 FP: False 3458 FP: False
Positives Positives
DR: Detection Rate 85% DR: Detection 85% DR: Detection
Rate Rate
Time (sec) 22 mc
7. Acknowledgments 8. References
Research on face detection in a video stream [1] A. Yilmaz, O. Javed, M. Shah, Object
has been conducted. This is done using the Tracking: A Survey. ACM Computing
MediaPipe Face Detection library. The results of Surveys (CSUR), 38(4) (2006) 13-es.
face detection are shown in Fig. 4. Frames in the [2] T. Huang, Computer Vision: Evolution and
form of photos are recorded in the video stream. Promise, CERN School Comput., 1996,
The DeepFace library is used to capture faces. If 21–25. doi: 10.5170/CERN-1996-008.21
there are several faces in the frame, then [3] Z. Pang, Z. Li, N. Wang, Simpletrack:
DeepFace captures several faces. Experiments Understanding and Rethinking 3D Multi-
were carried out when taking pictures in the genre Object Tracking, ECCV 2022 Workshops:
of the selfie, portrait photography, and Tel Aviv, Israel, October 2022, 680–696.
documentary photography. The photo was taken doi:10.1007/978-3-031-25056-9_43
from the frontal, close to the frontal and profile. [4] O. Iosifova, et al., Analysis of Automatic
The performed recognition showed a high Speech Recognition Methods, in:
percentage of face recognition. Workshop on Cybersecurity Providing in
Information and Telecommunication
Systems, vol. 2923 (2021) 252–257.
[5] K. Khorolska, et al., Application of a
Convolutional Neural Network with a
Module of Elementary Graphic Primitive
252
Classifiers in the Problems of Recognition [16] E. Rica, S. Álvarez, F. Serratosa, Learning
of Drawing Documentation and Distances Between Graph Nodes and
Transformation of 2D to 3D Models, Edges, Structural, Syntactic, and Statistical
Journal of Theoretical and Applied Pattern Recognition: Joint IAPR Int.
Information Technology 100(24) (2022) Workshops, S+SSPR 2022, Montreal, QC,
7426–7437. Canada, August 2022, 103–112.
[6] V. Sokolov, P. Skladannyi, A. Platonenko, [17] X. Qi, et al., A Convolutional Neural
Video Channel Suppression Method of Network Face Recognition Method Based
Unmanned Aerial Vehicles, in: IEEE 41st on BiLSTM and Attention Mechanism,
International Conference on Electronics Computational Intelligence and
and Nanotech-nology (2022) 473–477. doi: Neuroscience (2023).
10.1109/ELNANO54667.2022.9927105 [18] Y. Yasuda, et al., Flexibility Chart 2.0: An
[7] I. Delibaşoğlu, Moving Object Detection Accessible Visual Tool to Evaluate
Method with Motion Regions Tracking in Flexibility Resources in Power Systems.
Background Subtraction, Signal, Image and Renewable and Sustainable Energy
Video Processing, (2023) 1–9. doi: Reviews, 174 (2023) 113116.
10.1007/s11760-022-02458-y [19] G. Ramadan, et al., Impact of PCM type on
[8] X. Yu, Evaluation of Training Efficiency of Photocell Performance Using Heat Pipe-
Table Tennis Players Based on Computer PCM Cooling System: A Numerical Study,
Video Processing Technology, Optik, 273 J. Energy Systs. 7(1) (2023) 67–88.
(2023) 170404. [20] S. Sut, et al., Automated Adrenal Gland
[9] L. Nixon, How Do Destinations Relate to Disease Classes Using Patch-Based Center
One Another? A Study of Destination Symmetric Local Binary Pattern Technique
Visual Branding on Instagram, ENTER with CT Images, J. Digital Imaging (2023)
eTourism Conference, 2023, 204–216. 1–14.
[10] C. Xiao, Z. Luo, Improving Multiple [21] R. Glennie, et al., Hidden Markov Models:
Pedestrian Tracking in Crowded Scenes Pitfalls and Opportunities in Ecology.
with Hierarchical Association, Entropy, Methods in Ecology and Evolution, 14(1)
25(2) (2023) 380. (2023) 43–56.
[11] S. Garcia, et al., Face-To-Face and Online [22] N. Bansal, et al., Real-Time Advanced
Teaching Experience on Experimental Computational Intelligence for Deep Fake
Animals and Alternative Methods with Video Detection, Appl. Sci. 13(5) (2023)
Nursing Students: A Research Study, BMC 3095.
Nursing, 22(1) (2023) 1–10. [23] B. Deori, D. Thounaojam, A Survey on
[12] M. Lee, Y. Chen, Artificial Intelligence Moving Object Tracking in Video. Int. J.
Based Object Detection and Tracking for a Inf. Theor. (IJIT), 3(3) (2014) 31–46.
Small Underwater Robot, Processes, 11(2) [24] M. Medykovskyy, et al., Methods of
(2023) 312. Protection Document Formed from Latent
[13] A. Boyd, et al., CYBORG: Blending Element Located by Fractals, in: X
Human Saliency Into the Loss Improves International Scientific and Technical
Deep Learning-Based Synthetic Face Conference “Computer Sciences and
Detection, IEEE/CVF Winter Conference Information Technologies,” 2015, 70–72.
on Applications of Computer Vision, 2023, [25] M. Logoyda, et al., Identification of
6108–6117. Biometric Images using Latent Elements,
[14] B. Hassan, F. Dawood, Facial Image CEUR Workshop Proceedings, 2019.
Detection Based on the Viola-Jones [26] M. Nazarkevych, B. Yavourivskiy,
Algorithm for Gender Recognition, Int. J. I. Klyuynyk, Editing Raster Images and
Nonlinear Analysis Appls. 14(1) (2023) Digital Rating with Software, The
1593–1599. Experience of Designing and Appl. of CAD
[15] E. Hartman, et al., Elastic Shape Analysis Systems in Microelectr., 2015, 439–441.
of Surfaces with Second-Order Sobolev [27] V. Hrytsyk, A. Grondzal, A. Bilenkyj,
Metrics: A Comprehensive Numerical Augmented Reality for People with
Framework, Int. J. Comput. Vision, 2023, Disabilities, in: X Int. Sci. and Technical
1–27. Conf. “Computer Sciences and Information
Technologies,” (2015) 188–191.
253