Список использованной литературы:
1 Назарбаев Н. План нации 100 шагов по реализации пяти институциональных реформ //
www.inform.kz/rus/article/2777943;
2 Амандыкова С.К. Становление доктрины конституционализма в Казахстане. Караганда:
Издво КарГУ, 2002. С.382;
3 Абдрасулов Е.Б. Толкование закона и норм конституции : теория, опыт, процедура.
Алматы: Оркениет, 2002. С. 332;
4 Бахтыбаев И.Ж. Конституционный надзор прокуратуры Республики Казахстан :
Автореферат диссертации на соискание ученой степени кандидата юридических наук.
Москва, 1997;
5 Бейбутов М.С. Институт конституционного контроля в Республике Казахстан:
эволюция и проблемы модернизации. Алматы: Комплекс, 2005. С 255;
6 Караев А. Конституционный контроль: Казахстан и зарубежный опыт: Учебное пособие.
– Алматы: КазГЮУ, 2002 – 248 с.
25
СЕКЦИЯ 1
SESSION 1
Қосымшылармен бірге АТ технологиялар және бағдарламаламарды өңдеудің
құрал-жабдықтары
IT technologies and software engineering with applications
Uygulamalı BT teknolojileri ve yazılım mühendisliği
ИТ технологии и программное обеспечение с приложениями
UDC 004.932
Abdinurova N.R.
1
, Tolebi G.A.
2
1
MSc in Computer Science, lect., Suleyman Demirel University, Kaskelen, Каzakhstan
e-mail: nazgul.abdinurova@sdu.edu.kz
2
MSc in Robotics, lect., Suleyman Demirel University, Kaskelen, Kazakhstan,
e-mail: gulnur.tolebi@sdu.edu.kz
WATERMARKING AND COMPARISON OF THE TWO SPECIES
Abstract. Watermark is the piece of information inserted into data for the copyright
protection. Two technologies of watermarking: LSB and DWT will be explained and compared
below .
Key words: LSB, DWT, robustness, spatial domain
1. Introduction
The immense popularity and expeditious widening of the Internet show the commercial
potential of proposing digital data through the networks. Since commercial interests mean a
chance to make a profit, the authors and creators are concerned about protecting their
ownership rights. Digital watermarking can be explained as one of the possible and effective
approach for defending intellectual property.
A digital watermark is a digital signal or pattern inserted in order to identify copyright
information. It is embedded information data within an insensible form for human visual
system but in a way that protects from attacks such as common image processing techniques.
As well as, watermarks can be inserted into papers (hard copies) by varying its
thickness when it is manufactured. There are many other special properties in use, such as
fluorescent threads. An extreme example is the Australian $10 note, which is printed on plastic
and has a see through window.[1,2] Since this essay is more concerned about digital
watermarking from a cryptographic perspective, some ways how the ownership of digital data
can be protected will be discussed below.
2.1 Basic on watermarking
Digital watermarking can be contrasted against publickey encryption, though it has
quite a lot of differences. Encryption is used mostly for messages that are involved in a
communication. The encryption procedure changes the messages completely and the original
message can only be retrieved by decryption. And once the message is decrypted, there are no
residues left on the message. But in case of Digital watermarking, a permanent signature is left
on the digital data (like music, movies and photos) so that the ownership can be verified later
on using special software. Also, the digital data can be perceived by anyone as there is no need
to decrypt the data to view/read it, i.e. it can be used or retransmitted. [3]
In Figure 1 you can see general scheme of watermark insertion and detection.
26
1.1 Watermarking Requirements
First of all, the requirements of watermarks are defined to provide maximum protection of
intellectual property.
Imperceptible
In terms of watermarking, imperceptible refers to the original data’s quality which must remain
intact or unaffected by the watermark. For a digital data to be imperceptible, the watermark
must be embedded in a transparent manner. For example, the hearing or viewing experiences
of a photo or music must not be affected by its watermarks.
Undeletable
Ideally watermark should be impossible to remove, at least to be difficult to delete without
obviously degrading the host signal.
Statistically Undetectable
A “pirate” should not be able to detect the watermark by comparing several watermarked
signals from one sender.
Robustness
The watermark should survive any compressions or other operations applied to it causing lot of
data or quality, like converting an image to JPEG.
Unambiguous
Retrieval of the watermark should be unambiguously identify the owner, and the accuracy of
identification should degrade gracefully in the face of attack
Capacity:
Watermark’s capacity refers to the amount of watermark information that can be applied to the
host/original data.
2.2 Types of Watermarks
Watermarks and watermarking techniques can be distinguished in various ways. According to
visibility we can classify digital watermarks into 2 types:
1. Visible
2. Invisible
A visible watermark is a viewable or noticeable watermark that is imprinted over the original
data. It is stronger and more robust in nature, and as a result is often preferred more to apply
strong copyright protection on digital data. On picture below you can see example of visible
watermark:
Figure 2
An invisible watermark is an embedded image which cannot be perceived with human’s eyes.
Only electronic devices (or specialized software) can extract the hidden information to identify
27
the copyright owner. Invisible watermarks are used to mark a specialized digital content (text,
images or even audio content) to prove its authenticity [4].
2.3 Watermarking Techniques
2.3.1 Spatial Domain
Watermarks also can be divided according the techniques used to embed them. For example,
there are number of ways that enable watermarking in the spatial domain. The easiest way for
many programs is to use least significant bit method.
Least significant bit
Now let’s describe technique of applying this method on images. Since image is the
two dimensional array of pixels, watermark will be embedded on that pixels. First of all, we
have to convert pixels (because in rgb format there 3 values for every pixel) into greyscale then
perform our operation on least significant bit (LSB) of each pixel. Knowing obviously
limitations of HVS we can conclude that processing of small difference in the LSB will not be
noticeable.
The steps to embed watermark image are given below.
A. Steps of Least Significant bit
1) The image is first converted into greyscale from RGB format
2) The double precision is then applied on the image
3) The most significant bits are shifted to the least significant bits of the image
4) Make least significant bits of host image to zero
5) Add shifted version (step 3) of watermarked image to modified (step 4) host image.
B. Limitations of Spatial Domain Watermarking
This method seems simple and effective, also it can survive transformations such as
cropping, any addition of noise. Disadvantage is that knowing about algorithm, everyone can
change least significant bit, thereby delete the watermark. For example, if you set all LSBs of
watermarked image to ‘1’s, image loss its watermark, but like in case of embedding it this will
not be noticeable for human eyes. [5]
2.3.2 Frequency Domain
Another method of watermarking image and do it with high quality is applying
watermarking in the frequency domain (and other transform domains). It is performed by
transforming with the Fourier, Discrete Cosine Transform (DCT) or Discrete Wavelet
transform (DWT) methods. Such as in spatial domain watermarking, the values of chosen
frequencies will be changed in the original image. Since compression or scaling can lead to
losing high frequency values, the watermark signal should be applied to frequencies with lower
value, or of it’s possible to perform, applied a to that frequencies which contain substantial
information such as edges of the original picture. Watermarks applied to important values of
image will be scattered over whole image, which means technique is not amenable to beating
as spatial domain method.[6]Now, let’s look at one of methods of this technique, explained
above.
Discrete wavelet transform watermarking
In the DWT the main idea of watermarking for image is to decompound the image into
subimage of sundry spatial domain and independently on value of frequency. Then alter the
coefficient of subimage. After DWT transforming the original image, we will decompose it
into 4 frequency fractions where one lowfrequency (LL) and three highfrequency parts (LH,
HL, HH). If the information of LL district is DWT transformed, the sublevel frequency region
information will be obtained. The number of how many times we decomposed LL district will
define level of DWT. Figure 3 shows us example of twodimensional image after DWT
decomposing 3 times. (It is called applying 3
rd
level DWT)
28
Figure 3
The information of low frequency part is an image nigh to the host image and most signal
information of original image is in LL. According to the character of Human Visual System,
our eyes are sensitive to the see minuscule change of edge, outline and strip. Thus, it’s difficult
to conscious that putting the watermarking signal into the big amplitude coefficient of high
frequency band of the image DWT transformed. Then
it can carry more watermarking signal and has good concealing effect
3.1 Watermark Embedding using DWT:
The Procedure of watermark embedding is shown in fig.4
Figure 4
Steps of watermark embedding using DWT method:
1. Convert it from RGB format to YCbCr
2. Apply 2nd level DWT
3. Embed the watermark components in to the frequency subcomponents.
4. Apply IDWT.
5. Convert YCbCr to RGB.
6. Get watermarked image
7. Check Authentication. [7]
3. Conclusion
In this paper we first try to explain what watermarks are and the aim of using them:
watermarks are some signal embedded into original data in order to show and prove authority.
Watermarks have some requirements such as robustness, imperceptibility and so on. We
determine types of watermarks and techniques of watermarking in spatial and frequency
domains. In order to compare them we elucidate LSB and DWT methods and list steps of
performing them. Now we can conclude that in case of robustness DWT watermarking is
comparatively much better than the LSB, since in DWT watermark is embedded into sub
image, which makes watermark stronger for altering or removing. But in case of cost, LSB is
computationally cheaper, since there is less number operations should be performed.
29
References:
1 Ross Anderson, (2001). Security Engineering: A guide to Building Defendable Distributed
System. p247, USA: “Wiley Computer Publishing”.(0471389226)
2 http://en.wikipedia.org/wiki/Australian_tendollar_note
3 http://www.computerweekly.com/feature/WhitePaperDigitalwatermarking
4 Sk.Shamshad , K.L.Sailaja, P.Rameshkumar, Encryption of Watermarked Images using
Chakra Symmetric Key Approach, November 2013 International Journal of Advanced
Research
in
Computer
Science
and
Software
Engineering,
Available
at:
http://www.ijarcsse.com/docs/papers/Volume_3/11_November2013/V3I110109.pdf
5 Darshana Mistry (2010) Comparison of digital watermarking methods, September 2010,
(IJCSE) International Journal on Computer Science and Engineering,Available at:
http://www.researchgate.net/profile/Darshana_Mistry/publication/50235154_Comparison_of_
Digital_Water_Marking_methods/file/d912f50f510ef83de1.pdf
6 Mahmoud ElGayyar (2006)Watermarking Techniques Spatial Domain Digital Rights
Seminar, May 2006, Media Informatics University if Bonn Germany, Available at:
http://wob.iai.unibonn.de/Wob/images/55867298.pdf [Accessed on: 15 March 2014]
7 PravinM.Pithiya, H.L.Desai(2013) DWT Based Digital Image Watermarking, De-
Watermarking & Authentication, June 2013, International Journal of Engineering Research and
Development, Available at:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.415.8220&rep=rep1&type=pdf
УДК004.421.5 A28
Aitbayev Y.K.
1
, Kabulov B.M.
2
, Amirgaliyev Y.N.
3
1
MSc, International Information Technology University, Almaty, Kazakhstan
e-mail: mansure1991@gmail.com
2
MSc,International Information Technology University, Almaty, Kazakhstan
e-mail:kaboul777@gmail.com
3
Prof. Dr. Ing., Suleyman Demirel University, Kaskelen, Kazakhstan
e-mail:amir_ed@mail.ru
ENSEMBLE LEARNING ALGORITHMS IN PATTERN RECOGNITION TASKS
Аннотация. Статья посвящена теме использования моделей коллективного
принятия решений в автоматизированных интеллектуальных системах. Рассматривается
применение данных моделей для решения задач распознавания образов. Под
коллективным распознаванием подразумевается задача использования множества
классификаторов, каждый из которых принимает решение о классе одной сущности с
последующим согласованием решений с помощью некоторого алгоритма.
Ключевые слова: распознавание образов, групповые решения, коллективный
анализ, интеллектуальные системы
1. Introduction
The current degree of technological and scientific progress requires a focused
development of computer vision systems as an important mechanism of providing effective
interaction between machinery and humans. One of the most important areas of computer
vision is pattern recognition. Successful solution of pattern recognition tasks is necessary to
develop systems capable of intelligently evaluating the environment and doing certain actions.
There has been growing interest in pattern recognition tasks in the last decade. This is
determined by the prevalence of the problems that is being solved in recognizing images and
30
characters, scene analysis, technical and medical diagnostics, signal identification, analysis of
expert data, speech recognition, creation of expert and artificial intelligence systems.
Basic theoretical and practical issues of this area are reflected in scientific and practical
works of domestic and foreign experts, such as M.Z. Zgurovsky, G.S. Osipov, V.P. Gladun,
V.I. Donskoy, O.P. Kuznetsov, V.F. Khoroshevsky et al [1].
Fundamental work in the theory of pattern recognition and classification associated with
the names of such foreign scientists as J. von Neumann, K. Pearson, A. Wald, F. Rosenblatt. A
great contribution to the development of recognition and classification theory was made by
Soviet scientists Yzerman M.A., Braverman E.M., Rozonoer L.I. (the method of potential
functions), Vapnik V.N., Chervonenkis A.Y. (statistical pattern recognition theory,
"generalized portrait" approach), Ivakhnenko A.G. (group method of data handling), Zhuravlev
J.I., Galushkin A.I. [2]
An important requirement for the classification algorithms is resilience to changes in the
classified set of objects. Nowadays, among specialists, collective classifiers are becoming more
popular as a tool to improve the efficiency of pattern recognition [3]. Its essence consists in the
fact that the final decision is taken on the basis of individual classifiers’ partial decisions
"integration". In classification problems, the group method is the synthesis of the results
obtained from different algorithms applied to a given initial information, or selecting the
optimal algorithms of the given set [4]. When solving practical recognition problems, a user is
interested in algorithms, providing nearoptimal solution of applied problem. Given a set of
different recognition models and means for collective decisionmaking, certain guarantees of
success can be obtained [5].
2. Collective recognition
What is meant by the term “collective recognition” is the task of using multiple
classifiers (committee, ensemble, etc.), each of which will decide on the class of one entity
with the subsequent coordination of their decisions with the help of a certain algorithm. An
important condition for the efficient formation of the committee is to comply with the
necessary balance between accuracy and diversity of committee members. Committee diversity
is the degree of errors noncorrelatedness between committee members, which demonstrated a
significant impact (including experimentally). In particular, the advantage of combining 3
classifiers, each of which had an accuracy at the rate of 67% and a low rate of errors
correlation, compared with the same association with the accuracy of members ≈ 95% had
been demonstrated.
An important factor in the efficiency of a committee is members' votes combining
scheme. There are various voting schemes, the choice of which depends on the feature space,
classifiers models, etc. In this study, the most universal schemes are shown, for which the
winner is the class:
1) the maximum – with a maximum response of the committee members;
2) averaging – with the highest average response of the committee members;
3) a majority – with the largest number of votes of the members [6].
The following algorithms for constructing collective decisions exist: Bayesian method,
competence areas, decisionmaking patterns, Woods’ dynamic method, complex committee
methods, logical correction, convex stabilizer, and a generalized polynomial and algebraic
corrector. Generally, using collective algorithms strategy can improve the prediction accuracy
due to mutual compensation of an algorithm’s disadvantages for the benefits of others.
There are different approaches of partial decisions integration. In some cases, it is
proposed to use the majority vote method or label ranking method. In others – use schemes
based on averaging or linear combination of the posterior probabilities that are estimated by
individual classifiers, or fuzzy rules algorithms can be used. It is also proposed to carry out
independent fitting of the combined classifier, considering the partial decisions as the new
complex features. Approaches based on allocation of local areas in observation space, in each
of which only one partial classifier is "competent" to make a decision, are also developing [3].
31
The essence of the collective decisionmaking task is to develop an agreed collective
decision on the order of preference of the observable objects based on individual assessments
of group members. The need to use multiple classifiers and then combining their decisions
explained in different ways, depending on the problem definition. The main reasons of using
multiple classifiers’ coordinated combination of decisions are the following two ideas:
reducing complexity of a problem being solved (increasing computational
efficiency of a procedure).
increasing the decisionmaking competence (increasing accuracy rate) [7].
Despite the fact that one of the classifiers have superior properties compared to other,
sets of misclassified objects from different classifiers would not necessarily overlap. For this
reason, different classifiers may provide different information about the classified object,
which may be essential for improving the system properties.
As different recognition algorithms manifest themselves in different ways on the same
sample of objects, then the question arises about the synthetic decision rule that adaptively uses
the strengths of these algorithms. This decision rule is based on twolevel recognition scheme.
On the 1
st
level, partial recognition algorithms work, the results of which are combined on the
2
nd
level in synthesis unit. The most common ways of such union based on assigning areas of
competence of a certain partial algorithm. The easiest way to find the areas of competence is to
partition the feature space. Then, for each of the selected areas its own recognition algorithm is
developed. Another method is based on the use of formal analysis to determine local regions of
feature spaces as a surrounding area of recognizable objects, for which successful functioning
of any partial recognition algorithm is proved.
The general approach to the construction of the synthesis unit considers the resulting
performance of partial algorithms as initial indications for the construction of a new
generalized decision rule. In this case, all of the above methods with intensional and
extensional trends in pattern recognition can be used.
Consider the collective decisionmaking block diagram (Fig. 1). The decision rules
collective is some finite subset {R} of all possible decision rules set C, {R}, where C, {R} =
{R
l
}; l = 1, 2, …, L, formed to develop collective decision where R
l
lth decision rule, Y
l
–
the decision on the output of lth rule, C – a collective decision. Type of collective decision
concretized by the type of a problem to be solved by the collective. Since this is a pattern
recognition problem, both collective and individual decisions made by members of this
collective, consist in classifying a certain situation or object X to one of the classes or sets K
k
,
k = 1, 2, ..., K.
The situation X is characterized by the vector of parameters or features:
P = {p
1
, p
2
, ..., p
m
, ..., p
M
}.
(1)
Formally, the task of making a collective decision is stated as follows: if the Y
l
, l = 1, 2,
…, L – the individual decisions made by members of the collective – by the decision rules R
n
=
1, 2, ..., n, then the collective decision is determined as a function of individual decisions:
C = F (Y
1
, Y
2
, ..., Y
L
, X),
(2)
where F – a collective decision making algorithm
Figure 1. Collective decisionmaking block diagram
Decision C in the recognition task consists in choosing the number of one of the classes
K
k
, k = 1, 2, ..., K, for each particular situation X, for which rules R
l
make different decisions:
32
R
l
: X
∈K
k
, then Y
l
(X); l = 1, 2, .., L; k = 1, 2, .., K.
A voting algorithm when the final decision is determined by the majority of algorithms
can serve as the most obvious approach. In practice, such methods of decisions associations do
not always show high quality results, because the collective majority error may occur. The
weights of individual algorithms are fixed, i.e. the peculiarities of some specific situation are
not taken into account.
There are decision combining algorithms based on probabilistic approaches, when
selecting among the decisions of different algorithms, the one that has the highest probability is
selected. There are also matching algorithms based on metaclassification, when generalization
of decisions is performed by special metaсlassifier. The input data for it is the decisions of base
classifiers, which are interpreted as a set of features of the new feature space.
Collective recognition is effective in these cases:
decision is to be made by different algorithms;
algorithms use different feature spaces or different data sources;
algorithms trained with different training data;
dimension of the feature space is too large and/or it comprises the features measured at
different scales;
feature space comprises the features of different levels of abstraction (aggregation);
specific requirements are set for the type I and type II errors (false alarm and signal
pass).
There are three collective recognition strategies:
1) selection of the classifier, whose result determines the solution of a recognition task
(assuming that each classifier is an expert in a certain area of feature space);
2) fusion of classifiers decisions (assuming that all classifiers are equally competent in all
feature space);
3) a combination of the above strategies.
Recognition methods using not one, but several concurrent decision rules have already
been mentioned. Each rule provides a partial decision. The final decision is taken on the basis
of these options with the help of a certain generalization procedure. It is expedient to extend
the group decision approach to the case when more than one group of decision rules used, i.e.
"collective" of groups. The hierarchy of groups or collectives can be arbitrarily large. At each
level, partial decisions produced, according to them – the generalized decisions of current
level, which play the role of partial for the next level, etc. On the basis of the foregoing, the
following general scheme of the class of efficient algorithms for solving pattern recognition
problems with the help of the collective decision rules is proposed. Algorithms of the class
consists in performing four consecutive steps [8]:
1) generating groups of decision rules;
2) obtaining partial decisions and evaluating competences of groups;
3) formation of a generalized decision;
4) expected error estimation.
For the rational use of the characteristics of different algorithms in solving recognition
problems, it is possible to combine different in nature recognition algorithms into groups that
make the classification decision on the basis of rules adopted in the collective decisionmaking
theory. Suppose that in some situation X the decision taken is S. Then S = R(X), where R –
decision making algorithm in situation X. Suppose that there are L different algorithms for
solving the problem, i.e., S
l
= R
l
(X), l = 1, 2, ..., L, where S
l
– solution obtained by the
algorithm R
l
. We define the set of algorithms {R} = {R
1
, R
2
, ..., R
i
} as collective of algorithms
for solving a problem (collective of decision rules), if on the set of decisions S
l
in any situation
X a decision rule F is determined, i.e. S = F (S
1
, S
2
, ..., S
L
, X). Algorithms R
l
are called group
members, S
l
the solution of lth member of the group, and S – collective decision. The
function F defines the method of generalization of individual decisions into the collective S
33
decisions. Therefore, the synthesis of F function, or a method of generalization, is the central
point in organization of a collective.
In the recognition tasks a situation X is a description of the object X, i.e. its image, and
the decision S – the pattern number that corresponds to the image. Individual and collective
decisions in the recognition task consist in assigning a certain picture to one of the patterns.
The most interesting groups of recognition algorithms are those in which there is a dependence
between weight (influence rate) of each decision rule R
l
and the recognizable image. For
example, the weight of the decision rule R
l
may be determined by the relation:
( ) =
1,
∈
0,
∉
(3)
where B
l
– competence area of R
l
.
The weights of decision rules are chosen so that:
( ) = 1
(4)
for all possible values of X. Equation (3) means that the collective decision determined
by the decision of the decision rule R
i
, whose areas of competence belong to the image of the
object X. This approach represents a twolevel recognition procedure. On the 1
st
level image
belonging to a particular area of competence is determined, and on the 2
nd
the decision rule, the
competence of which is maximum in the found area, comes into force. The decision of the rule
is identified with the decision of the whole group.
Достарыңызбен бөлісу: |