quantitative analysis of fiber composite...
TRANSCRIPT
Quantitative analysis of fiber composite
microstructure
by: Jan Gajdošík
Thesis advisor: Doc. Ing. Michal Šejnoha Ph.D.
Ing. Jan Zeman Ph.D.
CTU in Prague
2004
2
Table of contents
QUANTITATIVE ANALYSIS OF FIBER COMPOSITE MICROSTRUCTURE 1
TABLE OF CONTENTS 2
ABSTRACT 4
ABSTRAKT 4
1. INTRODUCTION TO PROBLEM OF MICROSTRUCTURE ANALYSIS 5
2. DESCRIPTION OF MICROSTRUCTURE 7
2.1. Quantification of microstructure morphology 7
2.2. Basic concepts and hypotheses 8 2.2.1. Concept of an ensemble 9 2.2.2. Ergodic hypothesis 9 2.2.3. Statistical homogenity 10 2.2.4. Statistical isotropy 10
2.3. Microstructure description 11 2.3.1. n-point probability functions 11 2.3.2. Lineal path function 13
2.4. Numerical evaluation of microstructural statistics 15 2.4.1. n-point probability functions 15 2.4.2. Lineal path function 17
3. TAKING AND MODIFYING IMAGES 19
3.1. Application of image analysis in composite material research 19
3.2. C/E composites description 21
3.3. Tresholding 22
3.4. Image analysis of graphite fiber laminate 23 3.4.1. Preparation of samples 23
3
3.4.2. Taking images with camera 23 3.4.3. Taking images with scanner 24 3.4.4. Modification of input using program PHOTOSHOP 25 3.4.5. Modification using program LUCIA G 26
3.5. Conclusion of image analysis 29
4. STATISTICAL DESCRIPTION WITH VARIOUS BOUNDARY
CONDITIONS 30
4.1. Boundary conditions 30
4.2. Evaluation of the two-point probability function using Fast Fourier Transform method 32 4.2.1. Graphs of measures 36
4.3. Evaluation of the two-point probability function using the Monte Carlo Method 40 4.3.1. Description of used algorithm 40 4.3.2. Periodic boundary conditions 42 4.3.3. Plain boundary conditions 44 4.3.4. Mirror boundary conditions 46
4.4. Comparison of periodic boundary conditions with others 47
4.5. Conclusions 50
BIBLIOGRAPHY 52
5. APPENDIX: DESIGN OF OBJECT ORIENTED CODE SERVING FOR
STATISTICAL RESEARCH OF MICROSTRUCTURE 56
5.1. Introduction 56
5.2. General structure 57
5.3. The two-point probability function 58
5.4. Output possibilities 58
4
Abstract
This thesis is concerned with the description of the process of retrieving color images
from real composite systems together with their transformation into binary images. The fiber
composites were used as a real composite, because of its future perspective use. The main part
is dedicated to the comparison of various boundary conditions and processes for obtaining the
two-point probability function with various boundary conditions. The dominant purpose of
this thesis is to determine the differences in three boundary conditions. In particular, the plain
(no condition), mirror and periodic boundary conditions are considered. The possible
computational methods are taken in consideration, too. The speed of evaluation is one of the
most important issues and as such is emphasized.
Abstrakt
V p�edkládané diplomové práci se v�nuji popisu postupu získávání snímk� skute�ného
kompozitu a následného p�evodu snímk� na idealizované bitmapy. Pro tento ú�el byl použit
kompozitní materiál jednosm�rn� vyztužený vlákny a to s ohledem na jeho velmi perspektivní
využití ve stavebnictví. Hlavní �ást mé práce se zabývá srovnáváním r�zných okrajových
podmínek použitých p�i výpo�tu dvoubodové charakteristické funkce a také popisu postupu
získávání dvoubodové pravd�podobnostní funkce s r�znými okrajovými podmínkami.
D�ležitý úkol je ur�it rozdíly mezi r�znými okrajovými podmínkami. Prosté (žádné),
zrcadlové a periodické okrajové podmínky jsou vzaty v úvahu. Možné výpo�etní metody jsou
zmín�ny a popsány. Rychlost výpo�tu je jedním z nejd�ležit�jších faktor� a proto je
zd�razn�na.
5
1. Introduction to problem of microstructure analysis
The goal of this work is an extension of the knowledge in the field of evaluation and
modeling of a microstructure, so-called microanalysis. The microstructure generally consists
of two or more phases with fundamentally different properties. The microstructure description
has already been an objective of many works (see later references). Scientists can usually use
only a part of the section taken through the sample of a given microstructure and then
evaluate the measurable properties of a material (volume fractions, two and three point
probability functions, lineal path function etc.). This process is repeated on many sections in
order to determine average values or other characteristics.
One of the basic tasks is to decide which approach is the most suitable one for
examination of sections close to the sample boundary. In the neighborhood of edges the
images are inaccurate or rather incomplete (“cut” fibers of composite). The method of
investigation in this area that is rather unsure, which leads to further problems. For example,
the two point probability function is defined as a probability of finding two specified points in
specified phases. It is important to determine what to do in the case when one of the selected
points is found outside of the image. Such a pair can be either disposed or the distribution of
microstructure outside of the image can be predicted in some way. It is also very important to
determine the best size of the examined microstructure. In other words to decide, whether it is
comparable to compute two point probability functions of a large ensemble of small samples
and then averaged over the number of samples or rather use a big sample.
The very important factor of a reliable description of the real microstructure is the
preparation of an image representing such a microstructure. Note that samples of
microstructure are usually taken using microstructure sampling and conversion into electronic
form. This is described in the first part of this thesis.
It is worth mentioning that real microstructures usually fall into the category of
random composites. Literature offers many statistical descriptors of random composite
material. One of the most suitable descriptors is n-point probability function Sn defined in
[47], which gives the probability of simultaneous finding of n points randomly thrown into
the composite medium, which are in the same phase. Unfortunately, the determination of
higher order descriptors is computationally very demanding.
The first work dedicated to obtaining the function Sn and others for random
microstructure was presented in Corson’s work [6]. Although the suggested approach was
6
very slow and demanding (it was based on manual evaluation of photographs of samples of
real microstructures) the specified principle (evaluation of Sn values in lattice and subsequent
averaging) was used and expanded in following works. In [3] Corson’s technique was
automatized by image processing technology and finally [41] a simple simulation method for
setting Sn function suitable even for nondigitalized samples was proposed.
The thesis is separated into four main sections. The first one (Description of
microstructure) is devoted to the description of geometrical properties of the microstructure.
The basic descriptors and correlation between them are described in this section. The second
part is concerned with the description of microstructure images processing. The technique for
obtaining quality images of real woven composite (graphite fibers in epoxy matrix) and
following conversion into binary images is also presented. The third section (Statistic
descriptions with various boundary conditions) provides statistical processing of obtained
images. The different boundary conditions are described therein. The conclusion follows.
Appendix (Design of object oriented code serving for statistical research of microstructure) is
dedicated to the description of the developed program.
7
2. Description of microstructure
2.1. Quantification of microstructure morphology
Traditional micromechanical analysis of composite media with disordered
microstructure is typically based on very limited microstructural information such as volume
fraction of individual phases1. However, when an additional knowledge of the real
microstructure is available, the estimates of local fields can be improved by treating random
composites (see, e.g., [21], [22], [23], [45], [46], [50], [51] and references herein). Such a
modeling framework is considered throughout this text.
This opening chapter outlines evaluation of various statistical descriptors, which arise
in the analysis of binary microstructures with random arrangement of individual phases. With
regard to specific applications discussed in the following chapters (analysis of the graphite
fiber tow embedded in the polymer matrix), the background introduced in this chapter is quite
general and can be applied to any two-phase random heterogeneous medium of arbitrary
phase geometry2.
Section 2.2 reviews basic concepts and hypotheses associated with quantification of
microstructure morphology. Individual statistical descriptors used in the present work are
introduced in Section 2.3. The methods of their numerical evaluation are presented in
Section 2.4.
1 Namely, stiffness and compliance averaging method [14], [37], [48], dilute approximation [8],
selfconsistent method [4], [15], differential scheme [34] and Mori-Tanaka method [1], [33] fall into this
category. 2 In the case of particulate composites, a variety of specialized microstructural descriptors can be used
for microstructure characterization, see, e.g, [36], [38], [42], [45], [46], [49] and improved estimates of local
fields [30], [35], [45]. Moreover, we refer a more theoretically oriented reader to [42] for mathematically
rigorous discussion related to subjects of this chapter.
8
2.2. Basic concepts and hypotheses
Motivation. To introduce the subject, imagine a collection of a large number of micrographs
describing the geometry of a two phase fibrous composite. An example of such a micrograph
is displayed in Figure 2.1. Figure 2.1a represents a portion of a graphite–fiber tow containing
approximately twelve thousand fibers. A random cut consisting of about three hundred fibers
is shown in Figure 2.1b. Although having a large number of fibers, one can hardly assume
that such a representative sample can completely describe the morphology of the whole
composite. Simply taking similar micrographs from other parts of the fiber tow indicates
visual difference in the microstructure from sample to sample. At this point, we should
perhaps ask ourselves whether there is a reliable approach in modeling of composite
materials, which permits in some way incorporating elements of real microstructure into the
analysis. The answer is affirmative once we recognize the random nature of geometrical
arrangements of phases and treat random composites – it means that the particular
microstructure of a given part of a fiber tow yields only one possible arrangement of phases.
Therefore, instead of determining the exact value of some quantity at a given point (which is
sample dependent), attention is given to its expected or averaged or macroscopic value, which
incorporates information from all samples taken from a material.
(a) (b)
Figure 2.1: A real micrograph of a transverse plane section of the fiber tow
9
2.2.1. Concept of an ensemble
To reflect a random character of a heterogeneous medium it is convenient to introduce
the concept of an ensemble – a set of a large number of systems which are different in their
microscopic details but they are entirely identical within a point of view of macroscopic scale
(see, e.g., [1], [13], [21], [22], [25], [51]). Random nature of such material systems further
suggests that individual members of the ensemble, to be statistically representative of the
composite, should be sufficiently large compared to the microscopic length scale (e.g., fiber
diameter).
To begin, consider a sample space S defined here as a collection of material samples
similar to one of Figure 2.1b. Formation of S then opens a way to provide an estimate for
effective or expected value of some quantity, say stress or strain field, through the process of
its averaging over all systems in the ensemble. To proceed, identify individual members of
this space by � and define p(�) as the probability density of � in S (see [19], [21], [23], [51]
for further reference). Then, the ensemble average of function F(x, �) at a point x is provided
by
( ) ( ) ( ) .d,, �=S
pFF αααα xx (2.1)
Following the above definition would lead to experimental determination of the
ensemble average of function F(x, �) for a given point x through the cumbersome procedure
of manufacturing a large number of samples (which form the ensemble space S), measuring
F(x, �) for every sample and then its averaging for all samples. Therefore, it appears
meaningful to introduce certain hypotheses regarding the ensemble average, which
substantially simplify this task.
2.2.2. Ergodic hypothesis
This hypothesis demands all states available to an ensemble of the systems to be
available to every member of the system in the ensemble as well [1], [21], [25], [40], [42].
Once this hypothesis is adopted, spatial or volume average of function F(x,�) given by
10
( ) ( ) ,d,1
, � +=V
FV
F yyxx αα (2.2)
which is independent of and identical to the ensemble average for |V| � �, i.e.,
( ) ( ),,, αα xx FF = (2.3)
for all � ∈ S. This hypothesis allows an examination of only one arbitrary member of the
sample space, provided that the sample is “sufficiently large”. A possible way to fulfill this
condition is to assume a periodic composite described by a unit cell Y. Then [40],
( ) ( ) ,d,1
d,1
lim �� +=+∞→ YVV
FY
FV
yyxyyx αα (2.4)
so for the ergodic periodic composite medium, the ensemble average of F(x,�) is equal to the
volume average taken over the unit cell.
2.2.3. Statistical homogenity
Suppose that function F depends on n vectors x1, . . . , xn. If the material is statistically
homogeneous the ensemble average of F is invariant with respect to translation [1], [42], [47],
so the relation
( ) ( ),,...,,..., 11 yxyxxx −−= nn FF (2.5)
holds for an arbitrary value of y. The most common choice is to set y = x1, so
( ) ( ) ( ),,...,,...,,0,..., 1121121 nnn FFF xxxxxxxx =−−= (2.6)
where xij = xj − xi.
2.2.4. Statistical isotropy
Further simplification arises when assuming the material to be statistically isotropic
[1], [42], [47]. In such a case, the ensemble average is not only independent of the position of
the coordinate system origin but also of the coordinate system rotation. Under this hypothesis,
the ensemble average depends on the absolute value of vectors x12, . . . , x1n only:
( ) ( ),,..., 112 ijn rFF =xx (2.7)
where rij = xij, i = 1, . . . , n, j = (i + 1), . . . , n.
11
2.3. Microstructure description
A number of statistical descriptors are available to characterize the microstructure of a
two-phase random medium. This section describes two specific sets of descriptors which
proved to be useful in the present work. First, a set of general n-point probability functions,
applicable to an arbitrary two-phase composite, is introduced. A different statistical function
deserves attention when phase connectivity information is to be captured in more detail. The
lineal path function is discussed as an example of such a descriptor.
Both types of functions are introduced through a fundamental random function
relevant to the microstructure configuration. Then, statistical moments of this function are
identified as descriptors of the microstructure morphology. Finally, similarities and
differences between individual types of statistical descriptors are discussed.
2.3.1. n-point probability functions
Fundamental function and statistical moments. Consider an ensemble of a two-phase
random medium. To provide a general statistical description of such a systems it proves useful
to characterize each member of the ensemble by a stochastic function – characteristic function
�r(x,�), which is equal to one when point x lies in the phase r of the sample and equals to
zero otherwise [1], [42], [47],
( )��
��� ∈
=��
�
� ,if,1,otherwise,0,
ααχ r
r
Dxx (2.8)
where Dr(�) denotes the domain occupied by the r-th phase. Except where noted, composites
consisting of clearly distinguishable continuous matrix phase are considered. Therefore, r =
m, f is further assumed to take values m for the matrix phase while symbol f is reserved for the
second phase. For such a system the characteristic functions �f (x,�) and �m(x,�) are related
by
( ) ( ) 1,, =+ αχαχ xx fm (2.9)
Following [18, 238, 258, 286], we write the ensemble average of the product of
characteristic functions
( ) ( ) ( ),,,,..., 11,..., 11αχαχ nrrnrr nn
S xxxx �= (2.10)
12
where function Sr1,...,rn referred to as the general n-point probability gives the probability of
finding n points x1, . . . , xn randomly thrown into a medium located in the phases r1, . . . , rn.
Functions of the first and second order. Hereafter, we limit our attention to functions of the
order of one and two, since higher-order functions are quite difficult to determine in practice3.
Therefore, description of a random medium will be provided by the one-point probability
function Sr(x)
( ) ( ),,αχ xx rrS = (2.11)
which simply gives the probability of finding the phase r at x and by the two-point probability
function Srs(x1, x2)
( ) ( ) ( ),,,, 2121 αχαχ xxxx srrsS = (2.12)
which denotes the probability of finding simultaneously the phase r at x1 and the phase s at x2.
In general, evaluation of these characteristics may prove to be prohibitively difficult.
Fortunately, a simple method of attack can be adopted when accepting an assumption
regarding the material as statistically homogeneous, so that (compare with Eq. (2.5))
( ) ,rr SS =x (2.13)
( ) ( )., 2121 xxxx −= rsrs SS (2.14)
Further simplification arises when assuming the medium to be statistically isotropic.
Then Srs(x1,x2) reduces to (see also Eq. (2.7))
( ) ( ).2121 xxxx −=− rsrs SS (2.15)
Finally, making an ergodic assumption allows a substitution of the one-point
correlation function by its volume average, i.e., volume concentration or volume fraction of
the r-th phase cr,
.rr cS = (2.16)
Limiting values. In addition, the two-point probability function Srs incorporates the one-point
probability function Sr for certain values of its arguments such that
( ) ( ),,:for 12121 xxxxx rrsrs SS δ== (2.17)
( ) ( ) ( ),,lim:for 21212121
xxxxxxxx srrs SSS =∞→−
→∞− (2.18)
where symbol �rs stands for Kronecker’s delta. Relation (2.17) states that the probability of
finding two different phases at a single point is equal to 0 (see also Eq. (2.9)) or is given by
3 Note, however, that relatively efficient procedures for approximation of higher-order probability
functions for ergodic and statistically isotropic media were recently proposed in [45] and [46].
13
the one-point probability function if phases are identical. Equation (2.18) manifests that for
large distances points x1 and x2 are statistically independent. This relation is often denoted as
the no-long range orders hypothesis (see e.g. [30], [50]).
Finally, according to Eq. (2.9), we may determine one and two-point probability
functions for all phases provided that these functions are given for one arbitrary phase. For
one-point probability function of statistically homogeneous and ergodic medium, this relation
assumes a trivial form
.1 fm cc −= (2.19)
Relations for the two-point probability functions of statistically uniform and ergodic
medium are summarized in Table 2.14.
Known Function
Smm(x) Smf(x) Sff(x)
Smm(x) Smm(x) cm - Smf(x) cm - cf + Sff(x)
Smf(x) cm - Smm(x) Smf(x) cf - Sff(x)
Sff(x) cf - cm + Smm(x) cf - Smf(x) Sff(x)
Table 2.1: Relations among two-point probability functions
2.3.2. Lineal path function
As already noted in the previous section, the determination of probability functions of
order higher than two, encounters serious difficulties, both analytical and numerical5.
However, the importance of these functions for the characterization of morphology and
overall properties of heterogeneous materials is substantial (see, e.g., [8], [31], [32], [45] and
references therein). To overcome this difficulty, one can study low-order microstructural
descriptors based on a more complex fundamental function which contains more detailed
information about phase connectedness and hence certain information about long-range
orders. The lineal path function [29] described in this section is a representative of such
indicators.
4 Note that, by definition (2.12) and assumption of statistical homogeneity, Srs(x) = Ssr(x). 5 See, e.g., [3], [6], [8] for discussion of procedures for determination of third-order probability
functions for statistically isotropic ergodic media.
14
Fundamental function and statistical moments. To maintain a formal similarity with the
discussion of the n-point probability functions, we introduce a random function �r(x1, x2, �)
as
( )��
��� ⊂
=��
�
� ,,if,1,otherwise,0,, 21
21
ααλ r
r
Dxxxx (2.20)
i.e., a function which equals to 1 when the segment x1x2 is contained in the phase r for the
sample and zero otherwise. The lineal path function6, denoting the probability that the x1x2
segment lies in the phase r, then follows directly from the ensemble averaging of this function
[29]
( ) ( ).,,, 2121 αλ xxxx rrL = (2.21)
Under the assumptions of statistical homogeneity and isotropy, the function simplifies
equivalently to relations (2.5) and (2.7)
( ) ( ),, 2121 xxxx −= rr LL (2.22)
( ) ( ).2121 xxxx −=− rr LL (2.23)
Limiting values. Obviously, if the points x1 and x2 coincide, the lineal path function is
nothing else but the one-point probability function; for points x1 and x2 that are far apart the
lineal path function vanishes,
( ) ( ),,:for 12121 xxxxx rr SL == (2.24)
( ) ,0,lim:for 212121
=∞→−∞→−
xxxxxx rL (2.25)
The substantial difference between the lineal path function and n-point probability
function is that the functions related to different phases cannot be, in general, uniquely
determined by relations similar to Table 2.1. This is just another confirmation of the fact that
this function contains additional information which needn’t be captured by low-order
probability functions7.
6 The lineal path function can be related to the lineal contact distribution function Hr
l(u) introduced in
[42]. Indeed, the lineal contact distribution function for a line l starting at the origin and the r-th phase is defined
by relation Hrl(u) = 1 − P({Dr ∩ ul} = ∅)/(1 − cr). Then, e.g., for r = m, we get Hf
l(u) = 1 − Lm(ul)/(1 − cf ) and
finally Lm(ul) = cm(1 − Hfl). See also [152, 247].
7 For various deterministic periodic microstructures with smooth boundaries between phases, however,
the extensive numerical studies reported in [39] led the authors to the conjecture that the two-point probability
functions are sufficient to uniquely reconstruct the given microstructure (up to the translation and possible
inversion of the image).
15
2.4. Numerical evaluation of microstructural statistics
Numerical evaluation of microstructural statistics introduced in the previous sections
proceeds as follows. We begin with the n-point probability functions assuming an ergodic
medium. An approach suitable for digitized media is explored. Then, a procedure for the
determination of lineal path function is proposed. Both descriptors are evaluated for a selected
representative of theoretical microstructural models8.
2.4.1. n-point probability functions
To determine Sr1,...,rn we recall that the general n-point probability gives the probability
of finding n points x1, . . . , xn randomly thrown into a medium located in the phases r1, . . . ,
rn. Among all functions, the one–point and two–point probability functions deserve a special
attention. In view of Table 2.1 we further consider only the matrix probability functions.
To follow the above definition, the one–point matrix probability function Sm gives the
chance of finding a randomly placed point located in the matrix phase. To determine this
quantity, a simple Monte-Carlo like simulation can be utilized – we throw randomly a point
into the microstructure and count successful “hits“ into the matrix phase. Then, the value of
function Sm can be estimated as
,'
nn
Sm ≈ (2.26)
where n’ is the number of successful hits and n denotes the total number of throws. Entirely
similar procedure can be employed to determine values of Smm(x)9.
8 An interested reader may consult the overview article [45] or the books [42], [46] for exhaustive and
systematic discussion of various microstructural models. 9 For statistically isotropic microstructure, Smith and Torquato [41] proposed more efficient procedure for the
determination of Smm(||x12||). Instead of tossing a line corresponding to x into a medium, a sampling template is
used for the determination of two-point probability function. See also [49] for comparison of this method with
approaches presented hereafter.
16
Figure 2.2: Idealized binary image of Figure 2.1b. Bitmap
resolution is 976 × 716 pixels
Another, more attractive, approach is available when the real microstructure is
replaced by its binary image. A binary version of Figure 2.1b is shown in Figure 2.2. Such a
digitized micrograph can be imagined as a discretization of the characteristic function
�r(x,�), usually presented in terms of a W × H bitmap. Denoting the value of �r for the pixel
located in the i-th row and j-th column as �r(i, j) allows writing the first two moments of
function �r for an ergodic and statistically homogeneous medium in the form10
( ),,1 1
0
1
0��
−
=
−
=
=W
i
H
jrr ji
WHS χ (2.27)
( ) ( )( ) ( ) ( ),,,1
,1 1
��−
=
−
=
++−−
=M
m
N
n
i
ii
j
jjsr
nNmMrs njmiji
jjiinmS χχ (2.28)
where im = max(0,−m), iM = min(W,W−m) and jn = max(0,−n), jN = min(H,H−n). Observe that
to compute function Sr requires O(WH) operations, while O((WH)2) operations are needed for
function Srs. This might be computationally demanding, particularly for a large micrograph,
and does not seem to bring any advantages over simulation techniques.
10 Throughout the text, the C-language type of array indexing is consistently used, i.e., we denote the first
element of an array � as �0 and the last element of the array as �L−1, where L is the array length.
17
The required number of operations, however, can be reduced when writing the two–
point probability function Srs for a periodic ergodic medium as a correlation of functions �r
and �s, recall Eq. (2.4),
( ) ( ) ( ) .d1 yyxyx += � sY rrs Y
S χχ (2.29)
Then the Fourier transform of Srs is provided by [53]
( ) ( ) ( ),~~1~��� srrs Y
S χχ= (2.30)
where � now stands for the complex conjugate. Taking advantage of the periodicity of
function �r one may implement the Discrete Fourier Transform (DFT) [5] when evaluating
Eq. (2.30). To shed a light on this subject we first write the discrete version of Eq. (2.29) in
the form
( ) ( ) ( ) ( )( ),%,%,1
,1
0
1
0��
−
=
−
=
++=W
i
H
jsrrs HnjWmiji
WHnmS χχ (2.31)
where symbol “%” stands for modulo. The above equation, usually termed the cyclic
correlation [5], readily implies periodicity of function Srs. Note that the correlation property of
DFT holds for cyclic correlation. Referring to Eq. (2.30) it is given by the following relation
( ){ } ( ){ } ( ){ }.,DFT,DFT,DFT nmnmnmS srrs χχ= (2.32)
The inverse DFT denoted as IDFT then serves to derive function Srs at the final set of
discrete points as [3]
( ) ( ){ } ( ){ }{ }.,DFT,DFTIDFT1
, nmnmWH
nmS srrs χχ= (2.33)
This method is very economical and its accuracy depends only on the selected
resolution of the digitized medium. Usually, the Fast Fourier Transform, which needs only
O(WH log(WH) + WH) operations, is called to carry out the numerical computation11.
2.4.2. Lineal path function
Following the definition of the lineal path function as a probability of finding a
segment randomly thrown into a medium contained in a given phase, Eq. (2.21), an
11 The public-domain package FFTW version 2.1.3 [10] was used for the evaluation of (2.33).
18
elementary Monte Carlo-based procedure can be again used for its evaluation, i.e., we
randomly throw segments into a medium and count the cases when the segment meets the
given condition. Computationally more intensive approach, however, can be employed
following the idea of sampling template introduced in [41].
To that end, we form a sampling template with dimensions TW × TH pixels. Then, we
draw a set of segments from the center of a template to the points on the template boundary
separated by given discrete steps �W and �H. If the DDA algorithm (see, e.g., [16]) is used for
a construction of segments, the template can be rapidly assembled using only integer
operations. Moreover, this algorithm can be effectively combined with the bitmap
representation of the microstructure. Once a template is formed, the values of the lineal path
function for a given direction starts with placing the template center at a given point found,
say, in phase r and then marking the pixel at which the segment corresponding to the selected
direction meets the other phase, say s12. Then, counters corresponding to pixels of a given
segment which are closer to the center than the marked pixel are increased by one while
remaining counters are left unchanged. The value of the lineal path function can be then
obtained either by stochastic sampling (randomly throwing template center into a medium) or
deterministic sampling (template center is successively placed in all pixels of a bitmap) and
averaging the obtained results. Moreover, the latter method allows us to actually use only a
half of the sampling template, provided that the analyzed microstructure is statistically
homogeneous. Note that even though this procedure basically needs only integer-based
operations, it is still substantially slower than the FFT-based approach. Hence, a relatively
sparse sampling template is unavoidable if one wishes to keep the efficiency of this procedure
comparable to the determination of the two-point probability function.
12 In particular, the matrix phase is checked when Lf function is determined while the fiber phase represents
the “stop condition” for the Lm function.
19
3. Taking and modifying images
3.1. Application of image analysis in composite material research1
A rapid development of image analysis device enabling us to obtain digital record of
image information has occurred in the last decade [54]. Computer based image acquisition,
processing and analysis system allows scientists, engineers, researchers, etc., capturing video
information into the digital world of computers, where images can be viewed, enhanced,
analyzed, measured, annotated, archived, transmitted and more [17].
Image data acquisition is processed by various image input devices according to the
type of scanned specimen:
• The scanners are mainly used for square stationary patterns (prints, slides, negatives).
• Individual images of micro- and macro-objects are scanned by digital photocamera
either directly or by the camera connected to microscope.
• Sequence of images is recorded by the digital CCD camera (usually connected to
other optical equipment).
Imaging system consists of the following components in addition to the computer
(Figure 3.1):
• Image input device
• Frame grabber, or image capture computer add-in board, and any associated interface
cables.
• Software to acquire, analyze, measure, process and store images.
1 This section is based o work[44]
20
Figure 3.1: Imaging system diagram [17]
Computer added image processing may consist of:
• Detection of optical system, conversion and digitizing influence on the distortion of
image data and its subsequent correction
• Image adjustment before own analysis (e.g. decomposition into color components,
conversion to other color spaces, filtration)
• Basic image data analysis (e.g. measurement of lengths and areas, determination of
color component histograms in individual color spaces)
• Complex image data analysis (e.g. harmonic, wavelet, fractal)
The development of image analysis device has qualified its use in material
engineering. The opportunity of direct structure viewing increases the chances of difficult
structure investigation (e.g. various composite structures). The outcome of material image
analysis is basic information of image structure and its attributes, which are used in the proper
description of recorded objects or processes.
21
3.2. C/E composites description
Carbon epoxy (C/E) composites are materials which possess similar structure of the
reinforcement and matrix built on carbon base. In this thesis, structural research of plain
weave reinforced C/E composites is further described. These materials are prepared by
molding together a number of plies, each ply consisting of preimpregnated woven carbon-
fibre fabric (Figure 3.2).
Figure 3.2: a) Plain weave fabric b) Plain weave reinforced composite
Plain weave reinforced composite structure has been the topic of many works arisen in
last few years. These works have been oriented on the description of woven reinforcement,
relation between the structure and mechanical properties of a composite, influence of carbon
tows deformation caused by woven technology on effectiveness of the composite
reinforcement [7], [18], [20], [52].
Many researchers employ themselves in the creation of geometrical models of woven
reinforcement and their application in the prediction of composite properties. These models
have been usually based on idealized description of woven structure [24], [27], [28]. In the
current thesis, however, we restrict our attention to analysis of a cross-section of a tow, which
is unidirectional fibrous composite.
Expansion of image analysis and its use in composite research have allowed detailed
examination of real woven reinforcement structure and specification of its geometry
description. Also other structural characteristics (e.g. shape and distribution of individual
structural components, its size and volume fraction, etc.) can be measured and evaluated
directly [26].
22
However, image analysis of composite structures requires careful composite specimen
preparation and skilled worker for image acquisition and analysis to obtain valuable data set
for effective description of a real structure.
3.3. Tresholding
In this part, one of the terms used in the description of deriving binary images from the
real microstructure is explained.
The thresholding [43] is defined as „an image operation which produces a binary
image from a gray scale image“ [12]. Furthermore, the thresholding can produce a binary one
on the output digital image whenever a pixel value on the input digital image is below a
specified threshold level. A binary zero is produced otherwise. Although the composite
material is not always composed of two phases, it is considered as a two-phase composite for
simplicity. Nonetheless, the definition can be easily extended by increasing the number of
thresholding values.
The threshold values may be determined interactively by the operator. While the video
display affords direct comparison of the thresholded image with the working image, the
software enables the operator to modify or fix the thresholded images. Since it is assumed that
the phases of the composite could be distinct in the original image, the thresholding is done by
referring to the histogram on the video screen. If the original image has enough resolution and
little noise, then the histogram can provide most of the information needed to choose the
threshold value required for generating an approximated geometry.
Furthermore, if the volume fraction of every phase in the two dimensional image is
given, the software may provide a function to calculate the ratio between pixel values, by
which the threshold value is determined. On the other hand, if there is some noise, e.g., due to
bad resolution of the captured image, it has to be eliminated manually by the drawing function
in the software so that the desired geometry can be obtained. This additional operation is
called adjusting. Once a threshold value has been chosen, it is easy to convert the selected
image into a binary image. If the obtained image does not seem to be a satisfactory
representation of the original one, new threshold values may be chosen until the satisfaction
with the binary image is achieved.
23
3.4. Image analysis of graphite fiber laminate2
The laminated composite material in the shape of tube is used in the present study. In
particular, the fiber tow with the assumption of unidirectional fibers embedded in the matrix is
considered.
In this study LUCIA G software is used. It is an image editing software, enabling the
processing of color images – like, e.g., ADOBE PHOTOSHOP, and supporting conversion
into binary image. Binary image is a color image reduced just to two colors (black and white),
which is suitable when dealing with two-phase composites. Approach of converting images
into binary images is very efficient for the present material pattern. The program is designed
for scientific work – it allows programming simple macros leading to a great improvement of
the work efficiency.
3.4.1. Preparation of samples
A part of the laminated tube was put into the form of blunted cone and bonded by
dentacryle. The blunted cone was very steep and almost cylindrical. After solidification of
dentacryle a piece of material perpendicular to the longitudinal axis of the tube was cut. Both
surfaces perpendicular to the longitudinal axis of the tube were parallel. The sample was
grinded with metallographic grinder with decreasing abrasivity with 280, 400, 600 particles
per mm2. The perfectly plane parallel surfaces were achieved this way, which was necessary
for the quality of taken images. At the end, the sample was polished with diamond substance
on felt pad. The whole procedure was repeated many times to obtain images of the desired
quality.
3.4.2. Taking images with camera
Images of low quality were gained with the use of video camera. However, it was
impossible to convert these images with the help of the program LUCIA G. Brightness of
2 This section is based on work [11]
24
images was uneven, that was the reason, why it was not possible to use the function threshold
or to increase contrast. For illustration, we present a preview (Figure 3.3) of these images.
Each image was gained as an average of two hundred images made using microscope. It was
tried to get useable output from these inputs, but it was just proved, that the only way to arrive
at suitable results is a complicated increase of intensity, which could not often be done
sufficiently. The subsequent tresholding was again unsatisfactory, owing to the low contrast
so that the final manual drawing of missing parts of fibers with the use of original image was
necessary.
In overall it took about 3 hours to obtain one binary output. The output was
comparable with outputs gained with improved technique, but lasted too long. That is the
main reason, why for getting useable images we switched to scanner with higher resolution
and better illumination.
Figure 3.3: Input image made by camera
3.4.3. Taking images with scanner
Taking images by scanner was done using the microscope with lenses 12,5 with lateral
output. First, a few color images were taken, in order to determine the effectiveness of this
technique. An example is shown in Figure 3.4. This image served as a starting point in the
process of getting final simplified diagrams. Nevertheless, even this technique was too long
25
and inefficient in order to get larger sets of images. Therefore, the direct modification in input
of scanner images was used.
Figure 3.4: Color output made by scanner3
3.4.4. Modification of input using program PHOTOSHOP
First, the sample was scanned with full color, but it was found impossible to
determine, using only visual determination, whether the image had sufficient quality. Balance
of brightness is important factor of quality of the image, but it is hard to decide, whether the
image is balanced enough for subsequent modification.
That was the reason, why the first part of the image modification in the analytical
graphical program LUCIA G (increasing of contrast) was changed and left to scanning of
sample. During scanning the contrast was adjusted (increased) and saved in the grayscale.
Note that because of increased contrast the images seem to be monochromatic, but they are
not. As suggested by previous works, images with about 100 fibers are sufficient for the
subsequent work. The resolution of scanner was 4 400x3 600 pixels and the saved images had
3 The dark color stripe in the left part of the image is caused by defect of the scanner. The
example shown here is just for illustration, images without defects were used for modification.
Unluckily the undefected ones were lost by computer collapse.
26
resolution 1 500x1 500 pixels. The image was cut out from the large scanned image. Owing to
the sizes of images it was theoretically possible to receive up to four independent and non
overlapping images from one prescan. In reality, it was almost impossible to get all four
samples, especially because of unbalanced exposure. Thanks to increased contract it was,
despite of absence of color in images, possible to determine, whether the image was reliable
or not. If edges were very sharp and evidently artificially made it was obvious that the sample
was in bad quality. Such images were automatically disposed from the set of inputs.
Figure 3.5: Input grayscale image made by scanner.
3.4.5. Modification using program LUCIA G
There is only passive version of LUCIA G available on our faculty. It allows just
modification of images, but not grabbing or direct editing camera or scanner input.
The aim of modification in program LUCIA G was to obtain binary (monochromatic)
images. In these images the fiber should be idealized by circles with the equal radii. The
centers of these circles are placed in gravity centre of original imperfect fibers. The total area
of imperfect fibers equals to the total area of idealized fibers. The number of fibers is
identical, too. These conditions fulfill expectation of the same material fraction in real and
idealized image.
27
Input images (Figure 3.5) were modified in program LUCIA G. The program creates
binary image based on tresholding of input color images. The tresholding works as setting of
range of color scales or intensity and saturation (depends, whether the RGB or HIS
representation of image is used), for which the output binary image should be black (or
white). More detailed explanation of this procedure can be found in section 3.3. The function
could be used only providing the quality (balanced) illuminated image was
sufficientpresented. In case of gradual edges it was better to first increase the contrast. This
was fulfilled by input modification from the scanner in program PHOTOSHOP.
Figure 3.6: Binary image created in LUCIA G by tresholding.
The binary image in Figure 3.6 was not good enough for the analysis. Thanks to
existing scratches on the sample surface the image was locally incomplete – parts of fibers
were missing (Figure 3.6). It was necessary to complete the shape of fibers to the expected
shape. Original shape was usually evident from fragments of a single fiber. To automatize this
step the function convex envelope was used. It was accepted that the shape of fibers is circular
or oval, but nevertheless still non concave.
It was necessary to erase too small “cut” pieces of fibers at edges of the image. If the
major part of shape was preserved at the edge, it was completed to expected original shape
(based on similarity with nearby fibers).
28
Next, the fibers were separated from each other. The separation was done with the use
of function automatic separation of objects in LUCIA G. The function is very efficient,
provided that sufficient setup of number of steps and size of neighborhood is done, but just
with the use of powerful computer.
Figure 3.7: Binary image after first step of modification in LUCIA G.
The area of fibers, number of objects and positions of their gravity centers of the
image in Figure 3.7 was measured. After that the binary image with the same area, number of
objects and positions of their gravity centers was generated. The only difference was that in
the idealized image all objects were circles with the same radius. Whole technique was done
with the help of macro in LUCIA G. There is a tool in the program, which allows getting
history of used commands and setting of various functions. This tool was used for creating
macro in programming language (sort of C++). Resulting image, which is suitable for
statistical simulations, is shown in Figure 3.8.
29
Figure 3.8: Example of final output.
3.5. Conclusion of image analysis
This section presented the most effective approach for obtaining binary images of real
composite material. This approach was tested and used for retrieval of c. 25 images of real
microstructure of the composite material. The next chapter discusses the main objective of
this work; process of statistical description of the microstructure.
30
4. Statistical description with various boundary conditions
4.1. Boundary conditions
There are many possibilities how to statistically describe microstructure. Some of
them are summarized in chapter 2 (Description of microstructure). In this thesis the attention
is focused on the two-point probability function. This function is quite simple and easy to
compute when compared with other descriptors. The two point probability function can be
determined applying various boundary conditions. The literature offers the three main
possibilities. Recall that the two point probability function can be obtained by throwing a
needle of certain length in the image and counting how many times the ends fall in the
selected color. The question is what to do, when one end of the needle falls out of the image.
The first possibility is to expect that the image of microstructure is periodically
repeated in each direction. This condition is the simplest one, as it allows using very efficient
computational method described below. This first possibility is shown in Figure 4.1.
Figure 4.1: Example of binary image of microstructure with periodical boundary conditions.
The second possibility is to expect that the microstructure is mirrored. This approach
is more time consuming, but is presented for the sake of comparison with other results. In
case of unidirectional fibers it is probably not very efficient, but research of microstructure
consists of many others fields of research, where technique can be very useful(Figure 4.2)
31
Figure 4.2: Example of binary image of microstructure with mirror boundary conditions.
The third possibility, here termed as plain boundary conditions, is to disregard the
throws, where at least one of the ends of the needle falls out of the image. This approach is
probably the most accurate one, but significantly reduces the amount of obtained information.
It is evident that long needles can be thrown just a few times to be totally contained in the
image. If the needle with length of the diagonal of the rectangle is used, it can be thrown just
one time. That is the reason, why the data of two point probability function have to be reduced
of data with low reliability. (The needle is too long and therefore can be thrown just few
times.) (Figure 4.3)
Figure 4.3: Example of binary image with no mirroring or repeating.
32
Based on the previous work it was possible to use 25 images, but only 18 images were
eventually selected for the analysis. The visual evaluation consists of elimination samples
which seem to be unuseful. If about one fifth of the image was fiber free it was clear, that the
image experienced some error.
4.2. Evaluation of the two-point probability function using Fast
Fourier Transform method
The fast Fourier transform method already explained in section 2.4.1 can be used only
with the periodic boundary conditions. This method is much faster than the classical Monte-
Carlo method and was used to as a basis to compare effectiveness of both methods.
After selecting suitable inputs the two point probability function was calculated. In the
first step the edge of the input image was cut to obtain a square bitmap. The reason was that
the edge could be influenced by inaccuracy when taking images. That means that the largest
image had size of 1 148 x 1 148 pixels. In the next step the size was reduced by 70 pixels in
both directions to 1 078 x 1 078 pixels. The smaller image was selected 10 times from random
positions of the largest image. The step 70 pixels was chosen, because the average size of
fibers in images is c. 70 pixels. In each subsequent step the size was reduced by 70 pixels in
both directions and the random selection was increased 10 times compared to the previous
case. Sizes of binary images are summarized in Table 4.1.
33
Height of bitmap Width of bitmap Size of bitmap Number of bitmaps
1 148 1 148 1 317 904 1
1 078 1 078 1 162 084 10
1 008 1 008 1 016 064 20
938 938 879 844 30
868 868 753 424 40
798 798 636 804 50
728 728 529 984 60
658 658 432 964 70
588 588 345 744 80
518 518 268 324 90
448 448 200 704 100
378 378 142 884 110
308 308 94 864 120
238 238 56 644 130
168 168 28 224 140
98 98 9 604 150
Table 4.1: Sizes of binary images used in FFTM. Units in table are pixels.
After preparation and sorting of bitmaps the two point probability function was
calculated for each bitmap. The results of this calculation were stored in the matrix of the
same dimensions as the bitmap. Next step was to create the average result for each size. The
average was calculated not only on the whole set, but also on partial sets constructed from 2,
3, 4, … , Number of bitmaps of the specified size bitmaps. The value of average two-point
probability function at the position ij was obtained as average value of independently
computed two-point probability functions at the position ij.
The last step was to determine a suitable measure of a difference among the two-point
probability functions. As “an exact” two point probability function the function calculated on
the largest image was chosen. To calculate the measure M, the following formula was used
��−
=
−
=
−=
1
0
1
0
..1 W
i
H
j ij
ijij
S
PS
HWM
. (4.1)
In equation (4.1) W means the number of used points in the direction of width of the bitmap,
H the number of used points in the direction of height of the bitmap. If every point of bitmap
34
is used for evaluation of the measure W means the width of bitmap and H is the height of
bitmap. Sij is the value of the two point probability function of an image taken in the point
with coordinates i and j. Pij is the value of the two point probability function of the largest
bitmap. This enables comparison of measures taken from bitmaps with variable sizes.
Another measure MM was also considered and compared to measure M. The used
formula follows
��−
=
−
=
−=
1
0
1
0 ),(..
.1 W
i
H
j ij
ijij
jidistS
PS
HWM (4.2)
in which
��
��
� =∧=
+= 00,1
otherwise,22),( ji
jijidist (4.3)
The variable dist should guarantee that the values of the two point probability function, which
are more distant from the origin, have smaller significance than those which are closer. Quite
surprisingly, this correction results in rather negligible difference. Both kinds of measures are
compared in Table 4.2.
35
Number in the first column is the number of the two point probability functions from which the
average two point probability function was computed. The dimensions of the bitmap size over
which the averaging was carried out, are 1 008 x 1 008 pixels
Number Measure M Measure MM M/MM abs((M/MM)i-(M/MM)avg)
1 0,0000141996 4,640461E-11 305994,60 523,93
2 0,0000144160 4,708017E-11 306201,10 317,43
3 0,0000150730 4,915416E-11 306648,10 129,58
4 0,0000151216 4,931404E-11 306639,04 120,51
5 0,0000153525 5,005355E-11 306721,30 202,77
6 0,0000152362 4,970160E-11 306553,11 34,58
7 0,0000152719 4,981294E-11 306585,20 66,67
8 0,0000153377 5,001717E-11 306648,50 129,97
9 0,0000153923 5,019365E-11 306658,11 139,58
10 0,0000154202 5,028348E-11 306665,33 146,80
11 0,0000153869 5,018206E-11 306620,93 102,40
12 0,0000153068 4,992734E-11 306580,52 61,99
13 0,0000152420 4,972088E-11 306550,69 32,16
14 0,0000151973 4,958246E-11 306504,76 13,77
15 0,0000151477 4,942139E-11 306501,29 17,23
16 0,0000151423 4,940891E-11 306469,82 48,70
17 0,0000151201 4,934129E-11 306439,09 79,44
18 0,0000151799 4,953054E-11 306474,55 43,98
19 0,0000151807 4,953722E-11 306451,19 67,34
20 0,0000151436 4,941407E-11 306463,32 55,21
Average M/MM: Max difference:
Max difference/Average = 0,17% 306518,53 523,93
Table 4.2:Example of comparison of measure M and MM.
Because the measure M and measure MM were the same, just multiplied by almost a
constant value, in the subsequent research, only the measure M was used as it is more simple
to compute.
36
4.2.1. Graphs of measures
This part presents an overview of the measures M evaluated from the computed two
point probability functions. It is evident that bigger bitmaps allow getting results with better
reliability. Quite surprisingly, even for lower resolutions there is no common value, where all
measures stabilize. The measure of the two point probability function oscillates around
different value.
The resolution displayed above each graph represents the area over which the measure
M was evaluated. The values at horizontal axis are the numbers of used bitmaps for evaluation
of the average two point probability function. The accuracy does not gain zero value at
numbers of repeats 11, 21, 31, … . The line connecting the value of measure with zero is
displayed just for the lucidity of graphs. The described example follows.
938 ×××× 938pixels
37
1078 ×××× 1078pixels 1008 ×××× 1008pixels
1,05E-05
1,06E-05
1,07E-05
1,08E-05
1,09E-05
1,10E-05
1,11E-05
1,12E-05
1 2 3 4 5 6 7 8 9 10
Number of repeats
Mea
sure
0,00E+00
2,00E-06
4,00E-06
6,00E-06
8,00E-06
1,00E-05
1,20E-05
1,40E-05
1,60E-05
1,80E-05
1 3 5 7 9 11 13 15 17 19
Number of repeats
Mea
sure
938 ×××× 938pixels 868 ×××× 868pixels
0,00E+00
2,00E-06
4,00E-06
6,00E-06
8,00E-06
1,00E-05
1,20E-05
1,40E-05
1,60E-05
1,80E-05
2,00E-05
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Number of repeats
Mea
sure
0,00E+00
5,00E-06
1,00E-05
1,50E-05
2,00E-05
2,50E-05
1 4 7 10 13 16 19 22 25 28 31 34 37 40
Number of repeats
Mea
sure
38
798 ×××× 798pixels 728 ×××× 728pixels
0,00E+00
5,00E-06
1,00E-05
1,50E-05
2,00E-05
2,50E-05
3,00E-05
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Number of repeats
Mea
sure
0,00E+00
5,00E-06
1,00E-05
1,50E-05
2,00E-05
2,50E-05
3,00E-05
3,50E-05
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57
Number of repeats
Mea
sure
658 ×××× 658pixels 588 ×××× 588pixels
0,00E+00
5,00E-06
1,00E-05
1,50E-05
2,00E-05
2,50E-05
3,00E-05
3,50E-05
4,00E-05
4,50E-05
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69
Number of repeats
Mea
sure
0,00E+00
1,00E-05
2,00E-05
3,00E-05
4,00E-05
5,00E-05
6,00E-05
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76
Number of repeats
Mea
sure
39
518 ×××× 518pixels 448 ×××× 448pixels
0,00E+00
1,00E-05
2,00E-05
3,00E-05
4,00E-05
5,00E-05
6,00E-05
7,00E-05
8,00E-05
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86
Number of repeats
Mea
sure
0,00E+00
2,00E-05
4,00E-05
6,00E-05
8,00E-05
1,00E-04
1,20E-04
1 9 17 25 33 41 49 57 65 73 81 89 97
Number of repeats
Mea
sure
378 ×××× 378pixels 308 ×××× 308pixels
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
1 10 19 28 37 46 55 64 73 82 91 100 109
Number of repeats
Mea
sure
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
3,00E-04
3,50E-04
1 10 19 28 37 46 55 64 73 82 91 100 109 118
Number of repeats
Mea
sure
40
238 ×××× 238pixels 168 ×××× 168pixels
0,00E+00
1,00E-04
2,00E-04
3,00E-04
4,00E-04
5,00E-04
6,00E-04
7,00E-04
1 11 21 31 41 51 61 71 81 91 101 111 121
Number of repeats
Mea
sure
0,00E+00
2,00E-04
4,00E-04
6,00E-04
8,00E-04
1,00E-03
1,20E-03
1 12 23 34 45 56 67 78 89 100 111 122 133
Number of repeats
Mea
sure
98 ×××× 98pixels
0,00E+00
1,00E-03
2,00E-03
3,00E-03
4,00E-03
5,00E-03
6,00E-03
7,00E-03
1 12 23 34 45 56 67 78 89 100 111 122 133 144
Number of repeats
Mea
sure
4.3. Evaluation of the two-point probability function using the
Monte Carlo Method
4.3.1. Description of used algorithm
Due to computational demands of the Monte-Carlo method the simplifying conditions
had to be applied to the evaluation. The detail description of the used algorithm is presented in
this section. Two possibilities of simplification were considered. First, the value of the two
41
point probability function could be computed for all available vectors of function, but not all
data contained by the graph are used for the function evaluation. In other words the needles of
every available length are used but just limited numbers of throws are done. Other possibility
is to evaluate the function just for a few specific vectors, but all data stored in the bitmap are
used for evaluation. In other words to throw the needles of just few chosen lengths but the
needles are thrown into every point of bitmap.
In the present work the second possibility was employed. The reason was that this
approach allows comparison with results obtained by Fast Fourier transform method, which
evaluates the function with use of all data stored in the image. This property is important as
the goal of this research is to compare different boundary conditions.
The step used through the bitmap was based upon the average size of one fiber. The
average diameter of fiber was c. 70 pixels and one quarter of the diameter is approximately 17
pixels. The used step was 15 pixels. The quarter of diameter of fiber is value determined from
results presented in [53]. For obtaining the step in horizontal direction the same relation was
used and since the fibers are circles, both steps were the same.
Another difference compare to the Fast-Fourier transform method is that the step
reducing the bitmaps sizes was bigger. Recall that in Fast Fourier transformation method the
step 70 pixels was used which in the Monte Carlo method was set to 140 pixels and just one
half of bitmap size reductions was done (Compare with Table 4.1). The resolutions and
number of reductions are summarized in Table 4.3.
Height of bitmap Width of bitmap Size of bitmap Number of bitmaps
1 148 1 148 1 317 904 1
1 008 1 008 1 016 064 10
868 868 753 424 20
728 728 529 984 30
588 588 345 744 40
448 448 200 704 50
308 308 94 864 60
168 168 28 224 70
Table 4.3: Sizes of binary images used in MCM. Units in table are pixels.
42
4.3.2. Periodic boundary conditions
1008 ×××× 1008pixels 868 ×××× 868pixels
1,12E-04
1,14E-04
1,16E-04
1,18E-04
1,20E-04
1,22E-04
1,24E-04
1,26E-04
1,28E-04
1,30E-04
1 2 3 4 5 6 7 8 9 10
Number of repeats
Mea
sure
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
1 3 5 7 9 11 13 15 17 19
Number of repeatsM
easu
re
728 ×××× 728pixels 588 ×××× 588pixels
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
3,00E-04
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Number of repeats
Mea
sure
0,00E+00
1,00E-04
2,00E-04
3,00E-04
4,00E-04
5,00E-04
6,00E-04
1 4 7 10 13 16 19 22 25 28 31 34 37 40
Number of repeats
Mea
sure
43
448 ×××× 448pixels 308 ×××× 308pixels
0,00E+00
2,00E-04
4,00E-04
6,00E-04
8,00E-04
1,00E-03
1,20E-03
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Number of repeats
Mea
sure
0,00E+00
5,00E-04
1,00E-03
1,50E-03
2,00E-03
2,50E-03
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57
Number of repeats
Mea
sure
168 ×××× 168pixels
0,00E+00
1,00E-03
2,00E-03
3,00E-03
4,00E-03
5,00E-03
6,00E-03
7,00E-03
8,00E-03
9,00E-03
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69
Number of repeats
Mea
sure
44
4.3.3. Plain boundary conditions
1008 ×××× 1008pixels 868 ×××× 868pixels
1,08E-04
1,09E-04
1,10E-04
1,11E-04
1,12E-04
1,13E-04
1,14E-04
1,15E-04
1,16E-04
1,17E-04
1,18E-04
1,19E-04
1 2 3 4 5 6 7 8 9 10
Number of repeats
Mea
sure
0,00E+00
2,00E-05
4,00E-05
6,00E-05
8,00E-05
1,00E-04
1,20E-04
1,40E-04
1,60E-04
1 3 5 7 9 11 13 15 17 19
Number of repeatsM
easu
re
728 ×××× 728pixels 588 ×××× 588pixels
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
3,00E-04
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Number of repeats
Mea
sure
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
3,00E-04
3,50E-04
4,00E-04
1 4 7 10 13 16 19 22 25 28 31 34 37 40
Number of repeats
Mea
sure
45
448 ×××× 448pixels 308 ×××× 308pixels
0,00E+00
1,00E-04
2,00E-04
3,00E-04
4,00E-04
5,00E-04
6,00E-04
7,00E-04
8,00E-04
9,00E-04
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Number of repeats
Mea
sure
0,00E+00
5,00E-04
1,00E-03
1,50E-03
2,00E-03
2,50E-03
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57
Number of repeats
Mea
sure
168 ×××× 168pixels
0,00E+00
2,00E-03
4,00E-03
6,00E-03
8,00E-03
1,00E-02
1,20E-02
1,40E-02
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69
Number of repeats
Mea
sure
46
4.3.4. Mirror boundary conditions
1008 ×××× 1008pixels 868 ×××× 868pixels
1,34E-04
1,35E-04
1,36E-04
1,37E-04
1,38E-04
1,39E-04
1,40E-04
1,41E-04
1,42E-04
1,43E-04
1 2 3 4 5 6 7 8 9 10
Number of repeats
Mea
sure
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
1 3 5 7 9 11 13 15 17 19
Number of repeatsM
easu
re
728 ×××× 728pixels 588 ×××× 588pixels
0,00E+00
5,00E-05
1,00E-04
1,50E-04
2,00E-04
2,50E-04
3,00E-04
3,50E-04
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Number of repeats
Mea
sure
0,00E+00
1,00E-04
2,00E-04
3,00E-04
4,00E-04
5,00E-04
6,00E-04
1 4 7 10 13 16 19 22 25 28 31 34 37 40
Number of repeats
Mea
sure
47
448 ×××× 448pixels 308 ×××× 308pixels
0,00E+00
2,00E-04
4,00E-04
6,00E-04
8,00E-04
1,00E-03
1,20E-03
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Number of repeats
Mea
sure
0,00E+00
5,00E-04
1,00E-03
1,50E-03
2,00E-03
2,50E-03
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57
Number of repeats
Mea
sure
168 ×××× 168pixels
0,00E+00
1,00E-03
2,00E-03
3,00E-03
4,00E-03
5,00E-03
6,00E-03
7,00E-03
8,00E-03
9,00E-03
1,00E-02
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69
Number of repeats
Mea
sure
4.4. Comparison of periodic boundary conditions with others
In previous section statistical characteristics for different boundary conditions were
compared. In this section the two-point probability functions with periodic, mirror and plain
boundary conditions derived from one bitmap are shown. The difference between different
boundary conditions presented, too. The difference is provided in a absolute value, since it is
confusing to show negative and positive values with just one color. The two point probability
48
functions are shown in first quadrant. The enormous computational demand associated with
an exact evaluation of the function in every point is the main reason for this simplification.
Figure 4.4: Two point probability function with periodic boundary conditions.
Figure 4.5: Two point probability function with plain boundary conditions.
49
Figure 4.6: Two point probability function with mirror boundary conditions
Figure 4.7: Comparison of periodic and mirror boundary conditions.
The difference is plotted in an absolute value.
50
Figure 4.8: Comparison of periodic and plain boundary conditions.
The difference is plotted in an absolute value.
4.5. Conclusions
It is obvious from graphs in section 4.3 that all boundary conditions can be considered
as equal. The measure used for comparison is almost the same and behaves in a similar
manner. The suggested measure cannot fully contain the difference between functions set at
different resolutions of bitmaps. The reason is explained in the next paragraph at parable.
The two point probability function behaves little bit like surface of water in pond after
being hit by stone in the middle of the pond. If stones of different shapes, but quite similar
size are thrown in pond, the waves close to place of hit are quite different but in a farther
distance from the place of hits the waves are small. If the wave is small, the difference is
small too. If the observer is monitoring just small area of the surface close to the place of hit,
he will see that the differences between shapes of waves made by stones of different shapes
are bigger than if he monitors larger area. The results from this research are same. The
average values of measure are always smaller if determined from functions computed for
large bitmaps.
In section 4.4 it is shown that the difference between periodic and plain or mirror
boundary conditions are random and do not follow any regular pattern. Recall that the more
important values of the two-point probability functions are close to the origin. The origin of
51
graphs displayed in this work is in the left upper corner. All graphs look to be very similar in
this area and even the differences are close to zero (are bright) in this part.
Providing no boundary condition is superior to the other the only objective then
remains the speed of evaluation. The periodic boundary conditions can be determined with
the use of Fast Fourier transform, which is significantly faster than the classic Monte-Carlo
method. The largest bitmaps in this work have resolution about 1 000x1 000pixels. FFTM is
approximately 125000 times faster than MCM for these dimensions. In particular, when
using FFTM the evaluation lasted two minutes, while with MCM it would have taken almost
half a year.
The speed of evaluation with use of Monte-Carlo method is heavily influenced by the
dimensions of input bitmap. Out of the measures M it follows that it is faster to use a set of
smaller sections of the largest bitmap. The results of one big bitmap and set of sections are
comparable but the result for one big bitmap is nevertheless more accurate. Due to the Fast
Fourier method small time consumption I generally recommend to use this method on the
largest bitmap, because it is still very simple.
52
Bibliography
[1] Y. Benveniste, A new approach to the application of Mori-Tanaka theory in composite
materials, Mechanics of Materials 6 (1987), 147–157.
[2] M. J. Beran, Statistical continuum theories, Monographs in Statistical Physics,
Interscience Publishers, 1968.
[3] J. G. Berryman, Measurement of spatial correlation functions using image processing
techniques, Journal of Applied Physics 57 (1984), no. 7, 2374–2384.
[4] B. Budiansky, On the elastic moduli of some heterogeneous materials, Journal of the
Mechanics and Physics of Solids 13 (1965), 223–227.
[5] C.S. Burrus and T. W. Parks, DFT/FFT and convolution algorithms: Theory and
implementation, Topics in Digital Signal Processing, A Wiley-Interscience
Publication, 1985.
[6] P. B. Corson, Correlation functions for predicting properties of heterogenous materials.I.
Eperimental measurement of spatial correlation functions in multiphase solids, Journal
of Applied Physics 45 (1974), no. 7, 3159–3164.
[7] B. N. Cox, G. Flanagan, Handbook of Analytical Methods for Textile Composites,
NASA Contractor Report 4750, Hampton, Virginia, 1997
[8] R. Derr, Statistical modeling of microstructure with applications to effective property
computation in materials science, Ph.D. thesis, University of North Carolina, 1999, Also
available as http://citeseer.nj.nec.com/derr99statistical.html.
[9] A. Einstein, Eine neue bestimmung der molek¨uldimensionen, Annalen der Physic 19,34
(1906,1911), 289–305,591–592.
[10] M. Erigo and S.G. Johnson, FFTW: An adaptive software architecture for the FFT,
Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and
Signal Processing, ICASSP98, vol. 3, IEEE, New York, 1998, (See also
http://www.fftw.org), pp. 1381–1384.
[11] J. Gajdošík, M. Šejnoha a J. Zeman, P�íprava geometrického modelu pro
mikromechanickou analýzu, Stavební obzor, 12, no. 3, (2003), 80-82
[12] R. M. Haralick and L. G. Shapiro, Glossary of computer vision terms, E.R.Dougherty
Digital image processing methods. Marcel Dekker Inc.
[13] Z. Hashin, Analysis of composite materials - A survey, Journal of Applied Mechanics -
Transactions of the ASME 50 (1983), no. 3, 481–505.
53
[14] R. Hill, The elastic behavior of a crystalline aggregate, Proceedings of the Physical
Society, London, A 65 (1952), 349–354.
[15] R. Hill, A self-consistent mechanics of composite materials, Journal of the Mechanics
and Physics of Solids 13 (1965), 213–222.
[16] B. Hudec, Základy po�íta�ové grafiky [Elements of computer graphics], CTU, 1993.
[17] I-CUBE, About Image Analysis/Processing, www.i-cubic.com/about.htm, Integrated
Imaging,Inc., 2002
[18] J. Jortner, Microstructure of cloth-reinforced carbon-carbon laminates, Carbon, Vol.30,
No.2, 1992, p.p.153-163
[19] W. Kohler and G. C. Papanicolau, Bound for the overall conductivity of random media,
Macroscopic properties of disordered media (R. Burridge, S. Childress, and G.
Papanicolau, eds.), Lecture Notes in Physics, vol. 24, Courant Institute, Springer-
Verlag, Berlin, 1982, pp. 111–130.
[20] M. Košek, B. Košková, Analysis of yarn wavy path periodicity of textile composites
using discrete Fourier transform, Proc. of International Conference on Composite
Engineering, ICCE/6, Orlando, Florida, 1999, p.p.427-428
[21] W. Kreher and W. Pompe, Internal stresses in heterogeneous solids, Physical
Research,vol. 9, Akademie-Verlag, Berlin, 1989.
[22] E. Kröner, Bounds and effective elastic moduli of disordered materials, Journal of the
Mechanics and Physics of Solids 25 (1977), 137–155.
[23] E. Kröner, Statistical modelling, Modelling of small deformation of polycrystals (J.
Gittus and J. Zarka, eds.), Elsevier Applied Science Publishers, 1986, pp. 229–291.
[24] J. L. Kuhn, P. G. Charalambides, Modeling of planar weave fabric composite geometry,
J. of Composite Materials, Vol.33, No.3, 1999, p.p.188-220
[25] J. Kvasnica, Statistická fyzika [Statistical physics], second ed., Academia, 1998.
[26] LIM, System LUCIA G, User guide, Laboratory Imaging s.r.o., Praha, 2001
[27] S. V. Lomov, Textile composites modelling strategies, Composites: part A32, 2001,
p.p.1379-1394
[28] S. V. Lomov, I. Verpoest, T. Peeters, Nesting in textile laminates: Geometrical
modelling of the laminate, Composites Science and Technology, Vol.63, No.7, 2003,
p.p.993-1007
[29] B. Lu and S. Torquato, Lineal-path function for random heterogeneous materials,
Physical Review E 45 (1992), no. 2, 922–929.
54
[30] K. Z. Markov, On the cluster bounds for the effective properties of microcraked solids,
Journal of the Mechanics and Physics of Solids 46 (1998), no. 2, 357–388.
[31] G. W. Milton, Bounds on the elastic and transport properties of two-component
composites, Journal of the Mechanics and Physics of Solids 30 (1982), no. 3, 177–191.
[32] G. W. Milton and N. Phan-Thien, New bounds on effective elastic moduli of 2-
component materials, Proceedings of the Royal Society of London A 380 (1982),no.
1779, 305–331.
[33] T. Mori and K. Tanaka, Average stress in matrix and average elastic energy of elastic
materials with misfitting inclusions, Acta Metallurgica 21 (1973), 571.
[34] A.N. Norris, A differential scheme for the effective elastic moduli of composites,
Mechanics of Materials 4 (1985), no. 1, 1–16.
[35] P. Ponte Casteñada and J.R. Willis, The effect of spatial distribution on the effective
behaviour of composite materials and cracked media, Journal of the Mechanics and
Physics of Solids 43 (1995), no. 12, 1919–1951.
[36] R. Pyrz, Correlation of microstructure variability and local stress field in two-phase
materials, Materials Science and Engineering A177 (1994), 253–259.
[37] A. Reuss, Berechnung der Fliessgrenze von Mischkristallen auf Grund der
Plastizatätsbedingung für eirkristalle, Zietschirft für Angewandte Matematik und
Mechanik 9 (1929), 49.
[38] B.D. Ripley, Modelling of spatial patterns, Journal of the Royal Statistical Society -
Series B (methodological) 39 (1977), no. 2, 172–192.
[39] M. G. Rozman and M. Utz, Uniqueness of reconstruction of multiphase morphologies
from two-point correlation function, Physical Review Letters 89 (2002), no. 13,
135501–1–135501–4.
[40] K. Sab, On the homogenization and the simulation of random materials, European
Journal of Mechanics A-Solids 11 (1992), no. 5, 585–607.
[41] P. Smith and S. Torquato, Computer simulation results for the two-point probability
function of composite media, Journal of Computational Physics 76 (1988), 176–191.
[42] D. Stoyan, W.S. Kendall, and W.S. Mecke, Stochastic geometry and its applications,
Akademie-Verlag Berlin, 1987.
[43] K. Terada, T. Miura and N. Kikuchi, Digital image based modeling applied to the
homogenization analysis of composite materials, Computational Mechanics 20 (1997)
331—346 Sprigler Verlag 1997
55
[44] B. Tomková, Applycation of image analysis in composite research. A survey Technical
University of Liberec, Department of Textile Materials
[45] S. Torquato, Random heterogenous media: Microstructure and improved bounds on
effective properties, Applied Mechanics Review 44 (1991), no. 2, 37–76.
[46] S. Torquato, Random heterogeneous materials: Microstructure and macroscopic
properties, Springer-Verlag, 2001.
[47] S. Torquato and G. Stell, Microstructure of two-phase random media.I. The n-point
probability functions, Journal of Chemical Physics 77 (1982), no. 4, 2071–2077.
[48] W. Voight, Über die Beziehung zwischen den beiden Elastizitätskonstanten isotroper
Köper, Wiedemanns Annalen der Physics und Chemie 34 (1889), 537–587.
[49] M. Šejnoha and J. Zeman, Micromechanical analysis of random composites, vol. 6,
CTU Reports, no. 1, Czech Technical University in Prague, 2002.
[50] J. R. Willis, Bounds and self-consistent estimates for the overall properties of
anisotropic composites, Journal of the Mechanics and Physics of Solids 25 (1977), 185–
202.
[51] J. R. Willis, Variational and related methods for the overall properties of composites,
Advances in Applied Mechanics, vol. 21, 1981, pp. 2–74.
[52] S. W. Yurgartis, K. Morey, J. Jortner, Measurement of yarn shape and nesting in plain
weave composites, Composites Science and Technology, Vol.46, No.1, 1993, p.p.39-50
[53] J. Zeman, Analysis of Composite Materials with Random Microstructure, vol. 7, CTU
Reports, no. 1, Czech Technical University in Prague, 2003.
[54] O. Zmeškal, O. Sedlák, M. Nežádal, Image analysis methods, Digital imaging in biology
and medicine, Czech Academy of Science, �eské Bud�jovice,2002, p.p.34-43
56
5. Appendix: Design of object oriented code serving for
statistical research of microstructure
5.1. Introduction
The scientific community concerned with the research of microstructure needs to
explore microstructure statistically. This statistical research is predetermined by using
computer programs. There are many statistical functions we need to measure. Some of them
are simple while other need a lot of computer time. One of the basic approaches is to make an
image of some microscopic material and after some adjustment it is possible to use routines
for statistical evaluation of a sample.
Because the original few steps are generally the same, it is not necessary for each
scientific team dealing with statistical evaluation of samples to create their own code. It was
impossible to find any application, which could be extended just by adding new statistical
functions. It takes a lot of time to write service programs, which load a work with bitmap
instead of writing just the new part. Up to that it is hard to simulate the same conditions for
repeating the measurement in other scientific team.
Due to these facts the necessity of a suitable bitmap processing computational
environment is evident. Such computational environment must be able to reflect the
continuing development in terms of its wide and unrestricted extensibility, teamwork support
and portability.
The main goals can be summarized as follows:
Ability to work with various types of microstructures. (not just a two-phase material
system). That is why we need to be able to work with color bitmaps. The program should be
able to read and create generally used and multiplatform supported bitmaps.
Easy readability and extensibility of new functions without influencing the existing
code. The best way is if every researcher creates his own code file and does not have to
interfere with the rest of the code.
The code must provide output, which can be used both for the next research and for
the presentation of results.
57
Creating idealized models of a microstructure. Some functions have known graph for
specified model. This provides determination whether the used function is correct or not.
Even if the best organization of microstructure is searched the efficient modeling tool should
be prepared for use.
Computational performance. Bitmaps of microstructure samples often take many
MegaBytes and the run of program should be repeated many times.
The created application fulfils all mentioned requirements. It is written in C++ with
the use of object oriented features. This programming language was used, because it can be
used in various OS types and it is quite simple and powerful. Object organization of the code
provides easy addition of new functions.
The input and output graphical interface is standard windows bitmap (BMP). The 1-bit
(2 colors), 4-bit (16 colors) and 16-bit (256 colors) bitmap can be loaded and output is just in
the 16-bit mode. The function of program is tested in Windows XP and Linux. The code is
open and anybody can change any part of it.
5.2. General structure
The code is separated into independent files. Each file contains the part of code, which
performs specified tasks. The main part contains general functions. One part is aimed for the
bitmap analysis. Loading, saving and basic operation with bitmaps as is loading or storing
selected pixel, invert colors of the whole bitmap, add or substract two bitmaps etc. The bitmap
is loaded into an object bitmap. Since object bitmap stores also information about color table,
number of colors etc., it is better to work with a matrix. The matrix is a discrete description of
microstructure, where the value at any coordinates is the number of phases present at these
coordinates. The possibilities of work with matrix or with bitmap are the same. There are
some more operations due to mathematical essence of matrix. The last part contains the user
defined functions. So far, volume fraction function and the two-point probability function are
implemented.
58
5.3. The two-point probability function
One of the basic statistical functions describing microstructure is the two point
probability function. It describes many aspects of a microstructure. It is necessary to find the
influence of boundary conditions on the shape of the two-point probability function. There
exist two computational models for this function. The simple direct Monte Carlo method or
Fast Fourier transform of matrix is available. The direct way is quite slow for big parts of
bitmap, but it provides many possibilities for changing behavior of bitmap boundary
conditions. The Fast Fourier transform is faster, but builds up on the bitmap periodicity.
5.4. Output possibilities
The output is available in the form of text file with function values stored in the matrix
or in the graphical format. Some sample outputs are displayed in the following part. It is
possible to change properties of the output depending on the shape of graph.
Figure 5.1: Plan view of the graph. The darker places have the higher value than the lighter
colors representing the smaller values. The tag on the right gives information about the values
in the scheme.
59
Figure 5.2: Isometric view of the graph. The darker places have the higher value than the
lighter colors representing the smaller values. The tag on the right gives information about the
values in the scheme.
Figure 5.3: Isometric view of the wire frame graph. In some cases this image can supply the
better evidence value than the previous two images.