You are on page 1of 123

Master of Science Thesis

Fractal Dimension Analysis of Grain


Boundaries of 7XXX Aluminum Alloys and
Its Relationship to Fracture Toughness
Hamideh Khanbareh
22 December 2011
Faculty of Aerospace Engineering Delft University of Technology
Fractal Dimension Analysis of Grain
Boundaries of 7XXX Aluminum Alloys and
Its Relationship to Fracture Toughness
Master of Science Thesis
For obtaining the degree of Master of Science in Aerospace
Engineering at Delft University of Technology
Hamideh Khanbareh
22 December 2011
Faculty of Aerospace Engineering Delft University of Technology
Delft University of Technology
Delft University Of Technology
Department Of
Novel Aerospace Materials
The undersigned hereby certify that they have read and recommend to the Faculty of
Aerospace Engineering for acceptance a thesis entitled Fractal Dimension Analysis of
Grain Boundaries of 7XXX Aluminum Alloys and Its Relationship to Fracture
Toughness by Hamideh Khanbareh in partial fulllment of the requirements for the
degree of Master of Science.
Dated: 22 December 2011
Head of department:
prof.dr.ir. S. van der Zwaag
Supervisor:
dr. X. Wu
Reader:
dr.ir. R.C. Alderliesten
Reader:
dr. A.G. Miroux
Abstract
The main aim of this M.Sc. thesis is quantitative analysis of the grain boundaries in
partially recrystallized microstructures of hot-rolled 7050 Aluminum alloy. To this end
a MATLAB program was developed to automatically process both microstructures of
fully recrystallized and partially recrystallized materials in order to detect the desired
grain boundary, while removing all other features including sub-grains and precipitations.
The image processing performed on images of microstructures consists of several steps of
preparation, segmentation and postprocessing, that eventually lead to the extraction of
quantitative information. The output is a binarized image containing a one pixel wide
boundary suitable to be implemented in box counting calculations to estimate the fractal
dimension of the boundary.
To implement the box-counting method for fractal dimension calculations a MATLAB
program was developed based on the theoretical considerations. To examine the accuracy
of the algorithm, test images of Koch snowake curves with known theoretical fractal
dimension were generated at dierent iterations. The inuence of a number of iterations
and image resolution on the accuracy of estimated fractal dimension is evaluated using
the same Koch curves. It has been found out that at a constant resolution, increasing the
number of iterations results in a fractal dimension closer to the mathematical value, while
for each Koch curve higher image resolution results in less deviation from the theoretical
fractal dimension.
The veried box counting algorithm was used to quantify the irregularity of high angle
grain boundaries between recrystallized and un-recrystallized regions in the microstruc-
ture of heat-treated Aluminum alloys for three types of conventional-rolled, modied-
v
vi Abstract
rolled I ,and modied-rolled II materials each containing 0

and 90

groups of samples.
In order to investigate the eect of grain shape on fracture behavior of the material, box
counting calculations were carried out on high angle grain boundaries aligned in crack
growth direction in the microstructure of crack plane for tear-tested specimens. The eect
of grain boundary conguration on fracture toughness of the materials was investigated
using UPE data of tear test. The results show that there is a strong correlation between
UPE and the fractal dimension of the grain boundaries parallel to crack propagation di-
rection for both 0

and 90

group of samples. In other words by increasing the fractal


dimension of grain boundaries, the UPE of all studied material increases. For both 0

and 90

group of samples the modied type I with higher percentage of modied rolling
posses the highest fractal dimension and the highest fracture toughness value. Comparing
two dierent 0

and 90

orientations for each rolling process, it has been found out that
90

sample shows higher fractal dimension and higher UPE than 0

, this is in agreement
with the observation that the material tested in 0

mostly shows intergranular fracture


however, in 90

the fracture mode is more transgranular.


Quantitative analysis has been carried out on the Aluminum micrographs to estimate
the degree of recrystallization and the grain size in crack growth direction, in order to
compare the inuence of fractal dimension of grain boundaries on UPE to that of other
microstructural parameters. It has been found out that there is no clear correlation
between the fraction of recrystallized grains and the UPE. Studying the grain size in
crack propagation direction shows that increasing the grain size results in decreasing the
unit propagation energy. However comparing the general trends of data for both eects
of the grain size and the fractal dimension on the fracture toughness reveals that the
fractal dimension of the grain boundaries has the most signicant role in determining the
fracture toughness of the studied materials.
Contents
Abstract v
List of Figures xi
List of Tables xiii
1 Introduction 1
1.1 Fractals in materials engineering . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Fractals in the microstructure of metallic materials . . . . . . . . . 2
1.1.2 Serrated grain boundaries in the microstructure of alloys . . . . . . 3
1.2 Digital image processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Preprocessing or basic image enhancement . . . . . . . . . . . . . . 6
1.2.2 Segmentation or discrimination of features in microstructures . . . 11
1.2.3 Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 The concept of fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.1 The Koch snowake curve . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.2 Self-similarity and similarity dimension . . . . . . . . . . . . . . . 17
1.3.3 Statistical self-similarity . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.4 Characterization of serrated grain boundaries by fractal analysis . 20
2 Image processing 29
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 MATLAB program for boundary extraction . . . . . . . . . . . . . . . . . 29
2.2.1 Boundary extraction in partially recrystallized microstructure . . . 30
2.2.2 Boundary extraction in fully recrystallized microstructure . . . . . 39
2.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
vii
viii Contents
3 Box counting method for fractal dimension calculation 45
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Box-counting dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 MATLAB code for box-counting calculations . . . . . . . . . . . . . . . . 47
3.4 Generation of Koch snowakes . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Implementation of the box-counting program . . . . . . . . . . . . . . . . 49
3.5.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.5.2 The origin of possible errors . . . . . . . . . . . . . . . . . . . . . . 55
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4 Calculating fractal dimension of the grain boundaries 57
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.1.1 Fracture toughness testing . . . . . . . . . . . . . . . . . . . . . . . 58
4.2 Experimental work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3.1 Image processing on the Aluminum microstructure . . . . . . . . . 63
4.3.2 Box counting measurements on the grain boundaries of Aluminum
microstructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.3 Relationship between the fractal dimension of the grain boundaries
and the fracture toughness . . . . . . . . . . . . . . . . . . . . . . 78
4.3.4 Relationship between the fracture toughness and other microstruc-
tural features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5 Conclusions 87
A Matlab codes 99
A.1 MATLAB Code for Generating Koch snowake curves . . . . . . . . . . . 99
A.2 MATLAB Code for image processing followed by box-counting . . . . . . 101
A.3 Detect code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
A.4 MATLAB Code for box-counting calculation . . . . . . . . . . . . . . . . 105
A.5 MATLAB Code for calculating degree of recrystallization . . . . . . . . . 106
List of Figures
1.1 Flowchart of image-based description of microstructure. . . . . . . . . . . 6
1.2 An image and its histogram . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Examples of logic operations . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Background subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Sequence of application of a neighborhood operation . . . . . . . . . . . . 10
1.6 An example of SE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Properties of transformations based on erosion and dilation . . . . . . . . 14
1.8 Skeleton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.9 classical concept of the dimension versus fractal dimension . . . . . . . . . 16
1.10 Procedure for generating Koch snowake curve . . . . . . . . . . . . . . . 17
1.11 Self-similarity concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.12 Exact and statistically self-similar fractal curves and their fractal dimensions 21
1.13 Mandelbrot-Richardson diagram . . . . . . . . . . . . . . . . . . . . . . . 22
1.14 Divider method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.15 Aria-perimeter method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.16 Box-counting method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1 Microstructure of partially recrystallized Al7050. . . . . . . . . . . . . . . 30
2.2 Microstructure after binarization using T = 0.80. . . . . . . . . . . . . . . 32
2.3 Microstructure after hole-lling. . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Concept of connectedness. . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Disk-shaped structuring element. . . . . . . . . . . . . . . . . . . . . . . . 34
2.6 Microstructure after contour thickening followed by labeling. . . . . . . . 35
2.7 Detected grain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.8 Detected grain after closing. . . . . . . . . . . . . . . . . . . . . . . . . . . 36
ix
x List of Figures
2.9 Extracted grain boundary. . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.10 Initial image vs nal image. . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.11 Microstructure of fully recrystallized Al7050. . . . . . . . . . . . . . . . . 39
2.12 Microstructure after binarization. . . . . . . . . . . . . . . . . . . . . . . . 40
2.13 Microstructure after hole-lling. . . . . . . . . . . . . . . . . . . . . . . . . 41
2.14 Microstructure after contour thickening followed by labeling. . . . . . . . 41
2.15 Detected grain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.16 Detected grain after closing. . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.17 Extracted grain boundary. . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.18 Extracted grain boundary after opening. . . . . . . . . . . . . . . . . . . . 43
2.19 Initial image vs nal image. . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1 Surrounding the Koch curve with boxes . . . . . . . . . . . . . . . . . . . 46
3.2 Construction of the Koch curve. . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Koch snowakes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Results of box-counting calculation for Koch2 at two resolutions. . . . . . 52
3.5 Results of box-counting calculation for Koch3 at two resolutions. . . . . . 52
3.6 Results of box-counting calculation for Koch4 at two resolutions. . . . . . 53
3.7 Results of box-counting calculation for Koch5 at two resolutions. . . . . . 53
4.1 The sketches of an untested and tested Kahn-tear test specimens. . . . . . 59
4.2 Schematic result of tear test to measure UPE. . . . . . . . . . . . . . . . . 59
4.3 Sample orientations used for fracture toughness tests . . . . . . . . . . . . 60
4.4 Specimens for microstructure analysis . . . . . . . . . . . . . . . . . . . . 61
4.5 Microstructure of partially recrystallized 7050 alloy . . . . . . . . . . . . . 62
4.5 Microstructure of partially recrystallized 7050 alloy . . . . . . . . . . . . . 63
4.6 Procedure of sampling the grain boundary aligned in L direction on L-T
plane of CoR 0

sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.7 Procedure of sampling the grain boundary aligned in L direction on L-T
plane of MoR
I,1
0

sample. . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.8 Procedure of sampling the grain boundary aligned in L direction on L-T
plane of MoR
I,2
0

sample. . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.9 Procedure of sampling the grain boundary aligned in L direction on L-T
plane of MoR
II,1
0

sample. . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.10 Procedure of sampling the grain boundary aligned in L direction on L-T
plane of MoR
II,2
0

sample. . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.11 Procedure of sampling the grain boundary aligned in S direction on S-T
plane of CoR 90

sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.12 Procedure of sampling the grain boundary aligned in S direction on S-T
plane of MoR
I,1
90

sample. . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.13 Procedure of sampling the grain boundary aligned in S direction on S-T
plane of MoR
I,2
90

sample. . . . . . . . . . . . . . . . . . . . . . . . . . 72
List of Figures xi
4.14 Procedure of sampling the grain boundary aligned in S direction on S-T
plane of MoR
II,1
90

sample. . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.15 Procedure of sampling the grain boundary aligned in S direction on S-T
plane of MoR
II,2
90

sample. . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.16 Box counting results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.16 Box counting results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.17 Relationship between the fractal dimension of the grain boundaries and UPE. 80
4.18 The relationship between the fractal dimension of the grain boundaries and
UPE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.18 The procedure to calculate the degree of recrystallization of partially re-
crystallized microstrucures. . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.19 Relationship between degree of recrystallization and UPE. . . . . . . . . . 83
4.20 Schematic procedure of measuring the recrystallized grain size in crack
propagation direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.21 Relationship between grain size and UPE. . . . . . . . . . . . . . . . . . . 85
xii List of Figures
List of Tables
2.1 Input values for the MATLAB program . . . . . . . . . . . . . . . . . . . 39
3.1 Size of generated Koch images. . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Results of box-counting for dierent Koch islands at the resolution of 400
dpi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3 Results of box-counting for dierent Koch islands at the resolution of 600
dpi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4 Results of box-counting for Koch5 at 600 dpi. . . . . . . . . . . . . . . . . 54
4.1 Results of box-counting on Aluminum microstructures. . . . . . . . . . . . 79
4.2 UPE data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Measurement results of the degree of recrystallization. . . . . . . . . . . . 82
4.4 Measurement results of average recrystallized grain size in crack growth
direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
xiii
xiv List of Tables
Chapter 1
Introduction
The microstructure of metals and alloys is composed of grains, separated by grain bound-
aries. The mechanical properties of polycrystalline materials are strongly related to the
shape, distribution, and character of these grain boundaries. In classical stereology grains
are normally described by size and shape factors, for instance average grain size and as-
pect ratio. However these conventional descriptors are not capable to quantify an irregular
morphology of the grains. Recently a new family of hot-rolled Aluminum alloys have been
produced with very good toughness values. The alloys have a standard composition, and
a seemingly normal microstructure with the exception of one special feature: serrated
grain boundaries. Fractal geometry is found to be an ecient tool to describe shapes
having a rugged appearance. The tortuosity, fragmentation or roughness of such features
can be described via their fractal dimension. In general the fractal dimension is an exten-
sion of the Euclidean dimension and allows describing complex boundaries. For a straight
line, Euclidean and fractal dimensions are equal to 1. For a line inscribed in a plane, the
fractal dimension varies from 1 if the line is straight to 2 if the line is so tortuous that it
fully covers the whole plane. Introduction of the fractal concept and methods to calculate
fractal dimension are presented in section 1.3.
1.1 Fractals in materials engineering
The concept of fractal geometry has been applied in dierent elds of materials engineering
where shapes of objects have to be characterized. The main reported topics can be
1
2 Introduction
categorized as follows:
Metallic materials: will be treated in detail in 1.1.1.
Porous materials: surface fractal dimension, an important parameter reecting
the roughness of pore surfaces is the aim of characterization of dierent porous
systems [1]. For instance Al
2
O
3
-SiO
2
membranes with application in ltration [2],
mesoporous carbons with application as electrode materials in electric double-layer
capacitors [3, 4], SiC ceramics [5], palladium-alumina ceramic membrane [6], and
soil structure [7].
Thin-lm materials: surface roughness characterization by means of fractal ge-
ometry has been reported for gold lms deposited on quartz crystals [8], copper
tungsten thin lm deposited on silicon wafers for application in electronic devices
[9], thin lms of BaTiO
3
for electro-ceramic applications [10], and silver oxide lm
[11, 12] used as a photocathode.
Crack paths: analysis of the crack geometry using fractal dimensions is reported
for aluminazirconia composites [13]
Fracture surfaces: the fractal dimension as a measure of fracture surface rough-
ness [14, 15, 16, 17, 18, 19, 20] has been reported for ceramics[21], composites[22, 23,
24], glass [25], and metallic materials like Steels and Titanium alloys [26, 27, 28, 29].
Fine particles: irregularity of the structure of metal powder grains [30, 31], pig-
ments [31], wear particles [32, 33], colloidal particles in solid-liquid suspension [31],
and aerosol particles like carbon black and diesel soot [34, 35, 31] has been quantied
by fractal geometry.
Silicate minerals: serrated grain boundaries in the recrystallized microstructure of
quartz have been characterized by their fractal geometry. The relationship between
deformation parameters, temperature and strain rate, and fractal dimension of the
boundaries has been investigated. It has been reported that the fractal dimension
increases with increasing strain rate and decreasing temperature [36, 37, 38, 39, 40,
41, 42, 43].
1.1.1 Fractals in the microstructure of metallic materials
Fractal dimensions have been extensively used to characterize various features of metallic
microstructures, for instance:
1.1 Fractals in materials engineering 3
Dislocation patterns [20, 44, 45]
Grain boundary morphology [20, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58,
59, 60, 61, 62, 63, 64, 65, 66, 67]
Metallic powder particles [20]
Dendritic structures [20]
Slip lines [20]
Precipitates [20, 68, 69, 70]
Martensite structure [20]
Graphite structure in cast iron [71, 20]
Void structures [70, 20]
1.1.2 Serrated grain boundaries in the microstructure of alloys
Fractal analysis for quantifying the degree of the grain boundary serration as one of the
important indexes to characterize the microstructure and properties of alloys has been
attempted for dierent metallic materials. It has been shown that the grain boundary
conguration can be estimated by the fractal dimension of grain boundary surface prole
in two-dimensional section (the fractal dimension of the grain boundaries, D, 1 < D
< 2). When the fractal dimension of grain shapes is almost one, the grain boundaries
display linear shapes, and when the fractal dimension shows a higher value, towards 2,
the grain boundaries present a serrated gure. The evolution of rugged or serrated grain
boundaries in alloys, e.g. by cold work, hot work, or heat treatment, is one of the most
eective methods to improve the high-temperature strength of alloys, especially the creep
rupture properties [46, 48, 72, 47]. There is also a possibility to investigate a numerical
correlations between the fractal dimension of the grain boundaries and strength properties
such as fracture toughness, creep strength, and high temperature fatigue [47, 72].
The summary of the reported investigations and related mechanical properties for dierent
metallic systems is discussed below.
Hot-worked Al-Mg alloys [50, 46] has been reported to show serrated grain bound-
aries after hot deformation.
4 Introduction
Hot-worked Ti-15-3 alloy after hot deformation and solution treatment [49]. The
fractal dimension of recrystallized grains were studied and the inuence of processing
parameters on fractal dimension was investigated. With increasing deformation
degree and strain rate or decreasing deformation temperature, the fractal dimension
of recrystallized grains becomes larger.
Solution-treated AISI 316 stainless steel [51, 52] were studied to investigate the
formation of rugged grain boundaries. It has been reported that grain boundary
serration forms by interaction between grain boundary and second phase particles
at the boundary. It has been shown that creep properties improve when grain
boundary serration occurs, since serrations disturb grain boundary sliding.
Solution-treated Ni-based and Co-based superalloys. The development of serrated
grain boundary has been reported to be associated with the heterogeneous pre-
cipitations at grain boundaries and the migration of grain boundary segments in
between the particles. It has been demonstrated in several commercial superalloys
that a material with irregular, serrated grain boundaries has improved resistance to
creep crack growth over one with smooth grain boundaries [48, 54, 55, 58, 56].
Cold worked pure iron. The fractal geometry has been applied to describe the
morphologies of deformed and recrystallized structures of pure iron. It has been
found out that the fractal dimension increases with the amount of plastic defor-
mation. The fractal dimension of the deformed structures reaches a peak near the
recrystallization temperature and decreases with the grain growth [73].
Pure Zn poly-crystals [67] and austenitic steel [65] during creep deformation were
studied. The eect of creep deformation on grain morphology was investigated
using fractal geometry. The fractal dimension of the grain boundaries increases
with increasing the creep strain which can be correlated with the increase in the
density of slip lines in the grains that formed the serrations on grain boundaries.
Cu-Au alloy rened by platinum [61] has been reported to show higher fractal di-
mension than non-rened alloy.
Tungsten wire for light bulbs [62, 63]. Fractal dimension of tungsten grains has been
demonstrated to highly aect the creep behavior of the lamp lament wire which
must be very resistant to high-temperature creep. It has been observed that grain
boundary sliding is suppressed for grains with higher fractal dimensions.
1.2 Digital image processing 5
In order to quantitatively describe the fractal characteristics of grain boundaries in the
microstructures, the rst step is to abstract the grain boundary. It is necessary to conduct
several image processing operations on the images of the microstructure. This study
attempts to introduce the fundamentals of digital image processing and the basis of fractal
geometry required to quantitatively characterize the conguration of the grain boundaries
in the microstructure of partially recrystallized Aluminum alloys.
1.2 Digital image processing
In materials science images play very important role in the area of the quantitative de-
scription of the microstructure. In order to perform measurements on images of mi-
crostructures image processing is the rst step to be taken. Image processing, by de-
nition, include those methods that start with an image (an array of pixels each with a
gray-scale value) and end with an image. Various kinds of methods intend to produce a
modied image that emphasizes some aspects of the original image, for instance ltering
technique selects certain kinds of image data to be analyzed. More often the features
that are desired are particles, boundaries, or some other dened structures. Depending
on what is considered to be of interest, it is possible to construct a suitable processing
operation to make the feature ready to be measured. It is basically possible to measure
dierent stereologocal parameters which describe the geometrical properties of either a
group of features or one individual microstructural element from the imaged micrograph.
Point numbers, length and area of system of elements can be simply calculated using the
pixel-related information of the images. There are also various parameters that are not
directly measured, but can be derived from primitive measurements. Having described
the microstructural elements of interest it would be possible to evaluate the materials
properties. The procedure of image analysis starting from feature recognition to relation-
ship analysis is illustrated in gure 1.1.
The operations performed on images of microstructures in order to enhance them or to
make them more accessible for quantitative analysis, are described as image processing.
The process of image processing can be viewed as consisting of three steps:
1. Preprocessing or basic image enhancement
2. Segmentation or discrimination of features in microstructure
3. Postprocessing
6 Introduction
Figure 1.1: Flowchart of image-based description of microstructure.
After these steps have been performed, measurements and data analysis are undertaken
[74]. The above mentioned steps are described below.
1.2.1 Preprocessing or basic image enhancement
Enhancement is the rst step used to correct basic image defects, and it is considered
as a preparation for the segmentation step that eventually leads to the extraction of
quantitative information. Due to the digital nature of the computer images they can be
modied using usual mathematical functions. The simplest functions can be applied for
basic image enhancement, usually known as brightness and contrast control. Histograms,
look-up tables (LUT), and point operations are preprocessing mathematical manipula-
tions of pixel intensities which are widely used in image enhancement. Shading correction
is applied in this phase in case of uneven illumination, and noise reduction can be done
using suitable lters prior to any quantitative analysis.
The image histogram
It is noteworthy to make a brief review about the basis of digital images here. A gray-
scale digital image is composed of discrete points of gray tones, or brightness, rather than
1.2 Digital image processing 7
Figure 1.2: An image and its histogram. (a) Light micrograph, hypereutectic cast iron,
200, BF, CCD 1300 1030. (b) Intensity histogram [75].
continuously varying tones. A natural image is divided into a number of individual points
of brightness, and in addition, each of those points is described by a digital data value.
Each brightness point is a pixel of the digital image. A pixel is the most basic element
of any digital image. The pixels of an image form a rectangular array. Each pixel has a
coordinate (x, y) that corresponds to its location within the image, x being the vertical
component, and y the horizontal component. In general (0, 0) is the upper left corner of
the image. For 256 gray-tone images, a pixel can have one of the 256 brightness values,
ranging from 0 to 255. Black is represented by 0, and white is represented by 255.
The image histogram is actually the distribution of the pixel intensities in the image.
Figure 1.2 shows an image obtained through a light microscope and a CCD camera, and
its respective histogram. The horizontal axis represents the pixel intensities, in this case
between 0 (black) and 255 (white). The vertical axis measures the number of pixels in
each intensity value. For a gray scale image it corresponds to the gray-level distribution
[75].
Point Operations
Point operations are a class of image enhancements that do not alter the relationship
of pixels to their neighbors. This class of algorithms uses a type of transfer function to
translate original gray levels into new gray levels. In general terms, all LUT operations
are classied as point operations, where the intensity I
0
(x, y) of a pixel with coordinates
(x, y) in the output image is a function only of the intensity I
i
(x, y)of a pixel with the
same coordinates in the input image. That is:
8 Introduction
Figure 1.3: Examples of logic operations between binary images. (a) Image A. (b) Image
B. (c) A AND B. (d) A OR B. (e) A XOR B. (f) NOT A [75].
I
0
(x, y) = F[I
i
(x, y)] (1.1)
Where F is the function that relates input and output intensities. It can be represented
by a numerical table. Applying an LUT to an image is a very fast operation. Image-
processing programs use this fact to allow dynamic, interactive modication of parameters
such as brightness, contrast, and other display characteristics [75, 76]. Point operations
are the basis of algebraic and logic operations.
Algebraic and Logic Operations
Two or more images of the same size can be combined with dierent kinds of operations
in such a way that:
I
0
(x, y) = F[I
i1
(x, y), I
i2
(x, y)] (1.2)
Where I
i1
(x, y) and I
i2
(x, y) represent the intensities of pixels in the same positions of
two input images, F is a function that relates the two images, and I
0
(x, y)is the out-
put intensity. The simplest functions are algebraic operations like addition, subtraction,
multiplication, and division [75]. These functions that use more than one image and
combine them in mathematical way are very useful. For example, adding images is used
to increase the brightness in an image, averaging images is used to reduce noise, and
subtracting images is used to correct for background shading [77].
Similar to algebraic operations, logic operations relate binary images. The fundamental
operations, are negation (NOT), intersection (AND), union (OR), and dierence (ex-
clusive or, XOR). Figure 1.3 shows examples of these operations. Logic operations are
particularly useful in the postprocessing step that follows image segmentation [75].
1.2 Digital image processing 9
Figure 1.4: Background subtraction. (a) An image with uneven illumination (hypereutectic
cast iron, 200, BF, CCD 1300 1030). (b) Estimated background. (c)
Background subtracted image. ASM, Digital imaging [75].
Shading correction
Image defects that are caused by uneven illumination or artifacts in the imaging path,
typical of light microscopy, must be taken into account during image processing. Shading
correction is used when a large portion of an image is darker or lighter than the rest of
the image. The relative dierences between features of interest and the background are
usually the same, but features in one area of the image have a dierent gray level range
than the same type of feature in another proportion of the image. The main methods
of shading correction use a background reference image, either actual or articial. A
featureless reference image requires the acquisition of an image using the same lightening
conditions but without the features of interest. The reference image is then subtracted
from the shaded image to level the background as shown in Figure 1.4 [77].
Filters
Digital images are often polluted with noise produced, for example, by video cameras
in the case of insucient illumination or by SEM detectors. Obviously, noise should be
removed from such images prior to any quantitative analysis. This can be done using
suitable lters. Filtering is one of the neighborhood kernel operations, which translates
individual pixels based on surrounding pixels. In these transformations the output in-
tensity of any image pixel at position (x, y) depends not only on the initial gray value of
this pixel but also on the gray levels of its neighbors. The concept of using a kernel or
two-dimensional array of numeric operators which is briey discussed bellow, provides a
wide range of image enhancements.
The basic principle of operation is shown in Figure 1.5. A certain neighborhood size is
chosen in the input image, in this example, 3 3 pixels. The intensity of each pixel
10 Introduction
Figure 1.5: The sequence of application of a neighborhood operation [75].
in this neighborhood is multiplied by a certain weight; the results are summed together
and divided by the total weight for all pixels. The resulting value is written in the
output image in a coordinate that corresponds to the neighborhood center. The process
is repeated for a new neighborhood, one column to the right, and so on, until the right
edge of the image is reached. Then the neighborhood moves to the next line, and the
column scan is repeated. Eventually, the last line of the image is scanned.
The eect of the operation depends on the neighborhood size and on the weights that
multiply the intensities of the pixels. The neighborhood is always odd-sided, so that a
central pixel can be chosen. The larger the neighborhood the stronger is the eect of
the operation. The weights are normally represented by a matrix with the size of the
neighborhood, the kernel. Kernel operations are fundamental for noise ltering, back-
ground subtraction, edge detection, and other applications. Some lters will be described
in detail below.
Smoothing lter
Smoothing lter is probably the simplest possible lter that provides an image with
reduced noise and a somewhat out-of-focus appearance. The simplest way to smooth
noise in an image would be to replace the value of each pixel by the average of it and its
neighbors. The positive kernels are smoothing lters [75].
1.2 Digital image processing 11
Median lter
Median lter would provide better results than smoothing lter. For each neighborhood,
this lter sorts pixel intensities in ascending order and takes the median value of the
sequence, which is then written in the central pixel of the neighborhood in the output
image. A median lter can be eectively applied for treating heavily noisy images and in
most cases is the best solution available. This lter keeps the image sharp [75].
Maximum and the minimum lters
The maximum lter returns the value which is equal to the maximum of all the pixels
surrounding the pixel being analyzed. As a consequence, one obtains a new image which
is brighter than the original, with removed noise. The ltered image contains less details
than the initial one. The minimum lter returns the value which is equal to the minimum
of all the pixels surrounding the pixel being analyzed. The result of minimum ltering
is darker than the original and contains less details. The minimum and maximum lters
are widely used in practice and simultaneously are equivalent to some morphological
operations, which are sometimes decisive for the nal result of image analysis [78].
1.2.2 Segmentation or discrimination of features in microstructures
Segmentation is the technical term used for the discrimination of objects in an image.
When one looks at an image one uses many dierent inputs to distinguish the objects:
brightness, boundaries, specic shapes, or textures. The brain processes this information
in parallel at high speed, using previous experience. Computers, on the other hand, do
not have the same associative power. The recognition of objects in an image is made
through the classication of each pixel of the image as pertaining or not to an object.
Here the binarization method is discussed.
Binarization
The process of transformation of gray-scale images into binary ones is called binarization
or thresholding, in which 256 gray values are reduced to 2 gray values. It is accomplished
by selecting the gray level range of the features of interest. Pixels within the selected gray
level range are assigned as foreground, or detected features, and everything else as back-
ground, or undetected features. In other terms thresholding simply converts the image to
a series of 0s and 1s, which represent undetected and detected features respectively [77].
12 Introduction
The simplest and most commonly used parameter for thresholding is the intensity of the
pixel. A pixel is considered part of an object if it is bright enough. The segmentation
then proceeds through the choice of a certain threshold level T and the application of the
simple decision rule [75]:
If I(x, y) Tthen the pixel at (x, y) belongs to the object class.
If I(x, y) < Tthen the pixel at (x, y) belongs to the background class.
The proper choice of threshold level is decisive for the results of analysis. Threshold
selection methods can require user intervention or be fully automatic. In the interactive
thresholding the operator chooses the threshold T manually and adjusts it until a rea-
sonable segmentation is obtained. The image histogram is always used as a reference.
Interactive thresholding can be very ecient and provide fast, accurate results [78].
1.2.3 Postprocessing
Even with the best conditions, segmentation is seldom a single-step procedure. Most of
the sophisticated methods can leave behind spurious objects and other defects that must
be dealt with in the postprocessing step. There is one main group of methods based on
morphological operations to improve the segmentation results. These operations are used
to correct segmentation defects.
Morphological operators
Morphological operators are part of mathematical morphology [79], a powerful approach
to process images. Application of morphological operators enables detection of features
not available with other analysis methods. These operators are similar to the neighbor-
hood operators; they look at a given pixel and its neighbors. The main dierence is that
morphological operators are more commonly applied to the binary images created by
segmentation. The analysis of a pixel neighborhood denes if the pixel keeps its original
black or white color, or if it is inverted. Morphological operators depend on structuring
element, which denes the shape of the neighborhood analyzed around each pixel and,
will aect the nal shape of objects submitted to the operations. The central point of
mathematical morphology is the concept of structuring element. It can be understood as
a model of local pixel conguration. Usually, structuring elements are dened using the
following notation:
1 - for pixels belonging to the set of points under analysis
1.2 Digital image processing 13
Figure 1.6: An example of SE.
0 - for pixels belonging to the matrix
X - for pixels not taken into account (i.e., this point can have any value and has no eect
in the transformation)
An exemplary structuring element is shown in gure 1.6.
Morphological transformations can be dened as advanced lters, applied not for all the
pixels in the image, but only for pixels that t congurations dened by the structuring
element. From the very long list of morphological transformations available, the following
groups are more important:
Erosion, dilation, opening and closing
Skeletonization
Erosion and Dilation
Common operations that use neighborhood relationships between pixels include erosion
and dilation. These operations simply remove or add pixels to the periphery (both exter-
nally and internally, if it exists) of a feature based on the shape and location of neighbor-
hood pixels. Erosion often is used to remove extraneous pixels, which may result when
overdetection during thresholding occurs, because some noise has the same gray level
range as the features of interest. When used in combination with dilation (referred to as
opening), it is possible to separate touching particles. Dilation often is used to connect
features by rst dilating the features followed by erosion to return the features to their
approximate original size and shape (referred to as closing). Most image analysis systems
allow the option of using several neighborhood kernel patterns. However, great care must
be exercised because the feature shape can be signicantly dierent from the original
shape [77]. The basic properties of the whole family are shown graphically in Figure 1.7.
Skeletonization
A specialized use of erosion that prevents the separation of features while eroding away
pixels is called skeletonization, or thinning. This operation is useful when thinning thick,
14 Introduction
Figure 1.7: Illustration of the properties of transformations based on erosion and dilation.
(a) Initial image, (b) the same image after erosion, (c) opening, (d) dilation,
(e) closing, and (f) closing followed by opening [78].
Figure 1.8: An example of geometrical gure and its skeleton (thick black line) [78].
uneven feature boundaries. One can intuitively understand the idea of skeleton from
Figure 1.8. However, a more formal denition can be helpful for further analysis. We can
treat a skeleton as a central line, i.e., a line whose points are equally distant from two
closest points of the gure edges [77].
1.3 The concept of fractals
A traditional Euclidean classication of shape which is in harmony with the sense de-
veloped in the process of everyday observations is based on the analysis of the features
dimensions. From this point of view elements can be broadly divided into:
Point: 0-dimensional
Linear: 1-dimensional
1.3 The concept of fractals 15
Surface: 2-dimensional
Volume: 3-dimensional
Examples can easily be found of point, linear, surface, and volumetric shapes. However,
relatively new ndings from mathematics show that this categorization is incomplete
and some additional features may be dened that have their dimensionality expressed in
fractions. These features are called fractals. In other words, fractals are the geometrical
features whose non-integer dimensions are intermediate between 0 and 3.
The theory of fractals which is one of the most important developments in natural science
was rstly introduced by Mandelbrot [80]. Fractal geometry was derived mainly because
of the ineciency of classical geometry in describing complex shapes and phenomena.
This revolutionary theory has been applied in many elds ranging from molecular physics
to the large scale structure of universe, and provides new conceptual tools and insights.
In this theory fractals are considered as curves, surfaces, or volumes generated by some
repeated process involving successive subdivisions.
The fractal objects normally possess too irregular structure to be described by tradi-
tional topological or Euclidean approaches. Mandelbrot suggested that complex shapes
or boundaries can be described in useful terms by extending the classical concept of the
dimensionality of the system to add a fractional index to the whole number dimension to
generate a measure of the space-lling nature of a rugged structure. Mandelbrot proposed
to dierentiate between the topological dimension of a system and its fractal dimension
in the sense that topology concerns itself with the properties of a system which remain
invariant when their containing space is distorted. Thus, if one were to look at the set of
gure 1.9 they all have a topological dimension of one as any one of them can be drawn
on elastic graph stretched to t over a traditional straight line [80]. Mandelbrot points
out that in the non-topological sense the lines all represent systems intermediate between
one and two dimensions.
Here we summarize some of the major dierences between fractals and traditional Eu-
clidean shapes. First fractals are relatively modern inventions. Second, whereas Euclidean
shapes have one, or at most a few, characteristic sizes or length scales (typical length of
a shape, for example the radius of a sphere, the side of a cube), fractals, like a coastline
possess no characteristic sizes. Fractal shapes are said to be self-similar or scale-invariance
which will be discussed in detail later. Third, Euclidean geometry provides concise accu-
rate descriptions of man-made objects but is inappropriate for natural shapes. Euclidean
geometry with its well-dened and mathematically tractable planes and surfaces is usu-
16 Introduction
Figure 1.9: The classical concept of the dimension of a physical quantity can be extended
by adding fractional quantities related to the ruggedness of a system to the
topological dimension [81].
ally only found as an approximation over a narrow range of dimensions where mankind
has imposed it. For instance the surface of paper is at, straight, and Euclidean, but
only in the dimension or scale where humans perceive and control. Magnify the paper
surface and it becomes rough. Therefore the use of fractal dimensions is a new tool for
describing such roughness. Finally, whereas Euclidean shapes are usually described by a
simple algebraic formula (e.g. r
2
= x
2
+y
2
denes a circle of radius r), fractals, in general,
are the result of a construction procedure or algorithm that is often recursive (repeated
over and over) and ideally suited to computers [82]. The dierences between fractals and
traditional Euclidean shapes will be illustrated with the Koch curve.
1.3.1 The Koch snowake curve
The Koch snowake curve, one of the early mathematical shapes, (rst proposed around
1904) as shown in Figure 1.10, illustrates an iterative or recursive procedure for construct-
ing a fractal curve. A simple line segment is divided into thirds and the middle segment is
replaced by two equal segments forming part of an equilateral triangle. At the next stage
in the construction each of these 4 segments is replaced by 4 new segments with length
1.3 The concept of fractals 17
Figure 1.10: Recursive replacement procedure for generating Koch snowake curve and vari-
ations with dierent fractal dimensions [82].
1/3 of their parent according to the original pattern. This procedure, repeated over and
over, yields the Koch curve shown at the top right of Figure 1.10.
It demonstrates that the iteration of a very simple rule can produce seemingly complex
shapes with some highly unusual properties. Unlike Euclidean shapes, this curve has
detail on all length scales. Indeed, the closer one looks, the more detail one nds. More
important, the curve possesses an exact self-similarity. Each small portion can reproduce
exactly a larger portion. The curve is also invariant under changes of scale. At each stage
in its construction the length of the curve increases by a factor of 4/3. Thus, the limiting
curve crams an innite length into a nite area of the plane without intersecting itself.
At successive iterations corresponding to successive magnications, one nds new detail
and increasing length [80, 82].
1.3.2 Self-similarity and similarity dimension
The property of self-similarity or scaling, as exemplied by the Koch curve is one of the
central concepts of fractal geometry. It is closely connected with the intuitive notion of
dimension as illustrated in gure 1.11. An object considered as one-dimensional possesses
a similar scaling property. It can be divided into N identical parts each of which is scaled
down by the ratio r = 1/N from the whole. Similarly, a two-dimensional object, such as
a square area in the plane, can be divided into N self-similar parts each of which is scaled
18 Introduction
down by a factorr =
1

N
. A three-dimensional object like a solid cube may be divided
into N little cubes each of which is scaled down by a ratio r =
1
3

N
. With self-similarity
the generalization to fractal dimension is straightforward. A D-dimensional self-similar
object can be divided into N smaller copies of it self each of which is scaled down by a
factor v where r =
1
D

N
so;
N(r) =
1
r
D
(1.3)
Where N(r) is the number of self-similar parts, each scaled by a factor of r from the whole
that comprises the object. Conversely, given a self-similar object of N parts scaled by a
ratio r from the whole, its fractal or similarity dimension is given by
D =
log(N)
log(
1
r
)
(1.4)
The fractal dimension, unlike the more familiar notion of Euclidean dimension, need not be
an integer. Any segment of the Koch curve is composed of 4 sub-segments each of which is
scaled down by a factor of 1/3 from its parent. Its fractal dimension is D = log(4)/log(3)
or about 1.26. This non integer dimension, greater than one but less than two, reects
the unusual properties of the curve. It somehow lls more of space than a simple line (D
= 1), but less than a Euclidean area of the plane (D = 2). Hence the similarity dimension
may be regarded as an index of complexity [83, 82].
Mandelbrot gives some variations of the Koch construction and one is presented in gure
1.10. At the bottom a segment is replaced by 8 new segments each of the initial one to
yield:
D =
log(8)
log(4)
= 1.5 (1.5)
As D increases from 1 toward 2 the resulting curves progress from being line-like to
lling much of the plane. Indeed, the limit D 2 gives a space-lling curve. The fractal
dimension D, thus, provides a quantitative measure of ruggedness of the curves. Although
these Koch curves have fractal dimensions between 1 and 2, they all remain a curve with
a topological dimension of one [82].
Because fractal scaling is demonstrated easily to non-mathematicians using self-similar
fractals, the concepts of fractal and self-similarity may be confusingly mixed. Clearly,
not all self-similar objects are fractals; for example a straight line is self-similar but not
fractal, because its fractal and topological dimensions are both equal to 1 [84].
1.3 The concept of fractals 19
Figure 1.11: Interpretation of standard integer dimension gures in terms of exact self-
similarity and extension to non-integer dimensioned fractals [82].
20 Introduction
1.3.3 Statistical self-similarity
Fractal objects can be divided in two groups: linear (or ideal or exact) fractals and non-
linear (or natural) fractals. Linear (or ideal) fractals result from an absolute generating
process and are mathematical objects such as those of gure 1.10 which are constructed
by the innite iteration of a construction algorithm. In this case the determination of the
fractal dimension is analytical and it can be exactly calculated over a semi-innite domain.
It is a property of an ideal fractal curve that it appears the same at any magnication.
Therefore, if one photographs a portion of the boundary of an ideal fractal, one would
not know from the photograph what scale of magnication was being used to inspect
the boundary, since increasing resolution would only reveal increased detail similar in
structure to the boundary visible at low resolution [85]. Although mathematically exact
fractals are excellent for demonstrating fractal scaling, their mathematical exactness and
inferred innite scaling make them unlike most natural fractals. Most fractal-looking
objects in nature are nonlinear and do not display quite this precise form, for instance, a
magnied view of one part of the coastline will not precisely reproduce the full picture, but
it will have the same qualitative appearance if a sucient number of high magnication
images are taken. A coastline displays the kind of fractal behavior that is called statistical
self-similarity. Natural-appearing fractal curves are demonstrated in gure 1.12 b and c.
Note that these patterns are neither innitely self-similar nor exact. Instead, they possess
a statistical self-similarity which can be estimated within a nite range of scales. The
property that objects can look statistically similar while at the same time dierent in
detail at dierent length scales, is the central feature of fractals in nature. A coastline
for instance is random in the sense that (unlike the Koch curve) a large scale view is
insucient to predict the exact details of a magnied view [84, 82].
1.3.4 Characterization of serrated grain boundaries by fractal analysis
As soon as the grain boundaries are extracted from the microstructure, fractal analysis
to quantify the degree of grain boundary serration can be carried out by means of several
methods as described below.
1.2.4.1. Mandelbrot-Richardson method
Mandelbrot [80] discusses various approaches for the evaluation of fractal dimensions. The
most widely used method is to estimate the length of the prole with varying resolution
and to calculate the fractal dimension from a diagram called Mandelbrot-Richardson
plot in which the logarithm of the length estimate is plotted versus the logarithm of the
1.3 The concept of fractals 21
Figure 1.12: Exact and statistically self-similar fractal curves and their fractal dimensions.
(a) Koch curve, D
f
= 1.26, (b) and (c) fractional curves of fractal dimension
1.2 and 1.8, respectively, produced by midpoint displacement and successive
random additions algorithm [86].
resolution as shown in gure 1.13. Briey, in this case fractal dimension would be the
rate at which the perimeter (or surface area) of an object increases as the measurement
scale reduces. The measurement scale varies from ne resolution on the left to coarse
resolution on the right side of horizontal axes, as shown in gure 1.13. In practice, the
resolution is chosen to range from very ne (depending on the method used) to the order
of the total prole length. The data points are substituted by a straight line of best t
which can easily be obtained by a linear regression analysis. The slope equals 1 - D where
D is the fractal dimension of the curve [80].
The practical problem is to measure the length of the prole with varying resolution. The
following three methods have been shown to be feasible:
Divider method or yard-stick technique
In this method [80], the irregular curve of the original prole is approximated by polygons
each with constant side length (= resolution) which increases in consecutive steps. In
practice, a pair of dividers can be set to a prescribed length L (gure 1.14). Walking
along the curve with this step size, the length of the original curve is approximated by
the polygon. The number of steps multiplied by the stride length produces a perimeter
measurement.
Considering the self-similar fractal curve in gure 1.14, the distance between the two end
points isL
max
. The perimeter can be measured using dividers of length (L), where
22 Introduction
Figure 1.13: Mandelbrot-Richardson diagram for evaluating fractal dimension [80].
Figure 1.14: Fractal boundary spanned by three segments of ruler length L. Using this
size ruler measured perimeter is P = L N, where N is number of ruler
steps. Note how details smaller than L are spanned by ruler and excluded from
perimeter calculation. Using smaller rulers, more detail is included in perimeter
calculation, hence measured perimeter increases [84].
1.3 The concept of fractals 23
L = r L
max
, for r < 1 (1.6)
The measured perimeter (P) is determined by
P(L) = L N (1.7)
where N is the number of ruler steps of length (L) needed to measure the curve. Variations
along the curve that are smaller than L are spanned by the ruler. Therefore, as the ruler
size decreases, more details of the curve are accounted for in the perimeter measurement.
Recalling Equation 1(N(r) =
1
r
D
) for self-similar fractals,and combining with Equations
1.6 and 1.7 yields
P(L) = L (
L
max
L
)
D
(1.8)
or
P(L)
1
L
D1
(1.9)
P(L) = L
1D
(1.10)
also N as the number of sides of the polygon would be
N(L) = L
D
(1.11)
where is a constant [31]
Taking logarithm gives:
log(P) (1 D) log(L) (1.12)
log(N) (D) log(L) (1.13)
If a range of divider sizes is used, then the fractal dimension can be estimated from the
slope of the linear least-square t in a log (P) versus log (L) or log (N) versus log(L) plots
[31], where in the rst case
24 Introduction
D = 1 slope (1.14)
And in the second case
D = slope (1.15)
In principle, a smaller yardstick can take into account details of the (complicated) shape
which are invisible with larger strides. Therefore, the smaller the yard-stick length L, the
larger the perimeter P measured with it [31].
Minkowski method
The second method is based on an algorithm due to the mathematician G. Cantor [81].
Each point of the curve is the center of a circle of radius R, which results in a tape of
width 2R. Measuring the area of this tape and dividing it by 2R gives an estimate of
the prole length. The smaller R (= resolution), the better is the approximation to the
original curve. Plotting the area swept out by the circle versus the radius on log-log
axes, produces a line whose slope gives the fractal dimension. This method is easily
implemented in automatic image analyzers with a dilation unit (available in most of the
higher-priced commercial instruments). Dilation produces tapes of dierent width the
area of which is readily measured by these instruments [74].
Smoothing method
Another method to construct Mandelbrot-Richardson plot can be employed by applying
smoothing kernels to blur the serrated boundary, so small irregularities are moved and the
perimeter decreases. Plotting the measured perimeter vs. the standard deviation of the
smoothing kernel, on log-log axes, produces a straight line whose slope gives the fractal
dimension [74].
1.2.4.2. Area-perimeter
This method [40, 39, 84] is to compare grain boundary length P (perimeter) and grain
size d (diameter) (see gure 1.15) which is calculated by a circle having the same area A
of the grain. For fractal grain boundaries, the perimeter P is related to the diameter d or
the area A as is:
P d
D
A
D/2
The areas of a series of self-similar grains are measured to calculate the diameters, d,
of equivalent circles. By log-log plotting the actual perimeters, versus d, the fractal
1.3 The concept of fractals 25
Figure 1.15: The concept of the aria-perimeter analysis shown schematically. (a) A having
the area, A, and serrated perimeter, P. (b) Circle has the same area. A, and
diameter, d.
dimension, D, is dened as the slope of a least squares tted line. Where 1 D 2
because the measurement is in two-dimensional Euclidean space [87, 80]. This schematic
explanation is illustrated in gure 1.15. Obviously it is also possible to use grain area
in order to obtain the fractal dimension. The areaperimeter method obtains the fractal
dimension as a collective property for a set of objects with various sizes. However other
methods provide the fractal dimension that expresses the complexity of an individual
grain boundary as the individual fractal dimension [39, 80, 87]. It has been reported
that the evaluation of the fractal dimension of the grain boundaries by means of the
perimeter-area relation requires the verication of the self-similarity of the structure rst.
For self-similar random fractals this requires the application of a quantitative method
based on the dening equation (P(S) = S
1D
), for instance the yardstick method, to
examine the fractal dimension at dierent magnications [46, 37].
1.2.4.3. Box counting
This technique is one of the most popular methodologies for fractal analysis of grain
boundary shapes [40, 42, 57, 58, 60, 66, 92, 93]. Box counting method consists in the
superposition of a series of boxes with specic edge on top of the feature whose fractal
dimension is to be determined. Then count the number of cells needed to cover the
shape to be characterized. The size of the boxes, in pixels, is usually varied following
an exponent 2 progression, i.e. 1 x 1, 2 x 2, 4 x 4, etc. In this method, two values
26 Introduction
Figure 1.16: Schematic illustration of box-counting method to obtain the fractal dimension
of a grain boundary D [39].
are recorded: the number of square boxes containing at least one pixel of the projection
contour, N(r), and the size of the squares, r. The number N of boxes of size r needed to
cover a fractal set follows a power-law:
N(r) = N
0
r
D
(1.16)
where N
0
is a constant and D D
space
(D
space
is the dimension of the space, usually
D
space
= 1, 2, 3). The regression slope D of the straight line formed by plotting log(N(r))
against log(r) indicates the degree of complexity, or fractal dimension, between 1 and 2
(1 < D < 2) (see equation 1.17). D is known as box-counting dimension [80, 83]. Figure
1.16 shows schematically the box-counting procedure to obtain the fractal dimension of a
grain boundary.
Log(N(r)) = DLog(r) +Log(N0) (1.17)
The ledges and steps on serrated grain boundaries have both the lower and the upper
bounds of characteristic size in scale. The grain boundary length or the grain boundary
diameter may be the upper bound, and the atomic radius or the inter-atomic spacing may
be the lower bound. The degree of serration of interest lies between these two bounds.
Since grain boundaries do not have fractal nature beyond these two bounds, the maximum
size of grids are normally of the order of the grain boundary length [48]. If the measured
object is an individual grain boundary, the obtained fractal dimension is the individual
fractal dimension. However, the box-counting method can also give collective fractal
1.3 The concept of fractals 27
dimension when the measured object is not an individual or partial grain boundary but
an entire network of grain boundaries [82, 41, 88].
The structural complexity of an object can be characterized using the dierence between
fractal dimension and topological dimension (D
0
= D d) which is called the fractal
dimensional increment [47]. A growth in value of (D
0
) would correspond to an increase
in complexity. Since fractal dimension of grain boundaries is measured in a plane, the
topological dimension for a curve in a plane is one. (D
0
= D 1) can be used to rep-
resent the complexity of grain boundaries, that is to say, the degree of serration of grain
boundaries increase with the increase in D [49, 40].
In the next chapter of this study we will focus on details of practical image processing to
extract grain boundaries in the microstructure of Aluminum alloys. A MATLAB program
will be introduced to automatically detect the grain boundaries of interest with reasonable
preferences for using box counting analysis. The third chapter will be mainly dealing with
box counting calculations of articially generated Koch curves using a written MATLAB
program to verify the accuracy of the implemented method. Then the same procedure
was carried out to detect high angle grain boundaries in the partially recrystallized mi-
crostructure of hot-rolled Aluminum alloy, and calculate the fractal dimensions. Having
UPE data of the materials, it has been attempted to investigate the relation between frac-
ture toughness and fractal dimension of the grain boundaries. The results are reported
in the forth chapter.
28 Introduction
Chapter 2
Image processing
2.1 Introduction
The use of computer algorithms to enhance the digital images of the microstructures in
order to prepare them for fractal analysis forms the main aim of this chapter. Thus
MATLAB as a powerful programming language for technical computing is used. The
basic MATLAB distribution can be expanded by adding a range of toolboxes for specic
applications. The particular toolbox of interest to us is the image processing toolbox.
This software provides a comprehensive set of reference-standard algorithms and graph-
ical tools for image processing, analysis, visualization, and algorithm development. In
the present work a MATLAB program was developed to process two types of microstruc-
tures for Aluminum alloy, fully recrystallized, and partially recrystallized to extract the
grain boundary of interest. The MATLAB function and the resulting microstructures are
discussed in detail below.
2.2 MATLAB program for boundary extraction
The MATLAB function developed in this M.Sc. project is able to automatically pro-
cess both microstructures of fully recrystallized and partially recrystallized materials in
order to detect the desired grain boundary, while removing all other features including
sub-grains and precipitations. The output is a one pixel wide boundary which will then
29
30 Image processing
Figure 2.1: Microstructure of partially recrystallized Al7050.
be implemented in box counting calculations to get the fractal dimension of the bound-
ary. The operation of the program is described in the following sections for a partially
recrystallized (section 1.2.1) and fully recrystallized (section 1.2.2) microstructures. The
program code is given in the Apendixes.
2.2.1 Boundary extraction in partially recrystallized microstructure
The microstructure of partially recrystallized Aluminium alloy sample shown in gure 2.1,
is a tif type image of size 1461 pixels by 1222 pixels at a resolution of 96,000 pixels/inch.
It was taken by the Neophot 30 optical microscope using two lenses of 12.5X and 8X.
This raw image is the input for the MATLAB program.
The program basically works in three main phases;
reading the raw image
processing the raw image
returning the extracted grain boundary
2.2 MATLAB program for boundary extraction 31
Reading the image
In MATLAB images are considered as matrices whose elements are the pixel values of
the image. If the input is a true-color image or any other types rather than gray-scale, it
has to be converted to gray-scale before running the code. I = imread(filename.fmt)
, reads the gray-scale image from the graphics le into a matrix I. If the le contains a
gray-scale image, I is an M-by-N array. M refers to number of rows and N is number of
columns respectively [89].
Thresholding the image
The next step is the image binarization after choosing the appropriate threshold value.
MATLAB has the im2bw function, which thresholds an image of any data type, using
the general syntax I = im2bw(image, Tlevel). where T level is a value between 0 and
1, indicating the fraction of gray values to be converted white. The im2bw function
automatically scales the T level to a gray value, and then performs a thresholding in the
way that it will return 1 (white) for all those pixels for which the gray values are greater
than T level, and 0 (black) for all those pixels for which the gray values are less than
or equal to T level. The result would be a matrix of 0s and 1s, which shows up in a
new window. In image processing, the pixels in a binary image having logical value 1 are
called the image foreground pixels, whilst those pixels having logical value 0 are called the
image background pixels. The program keeps performing a loop thresholding in which
the optimized T level should be dened by the user. As soon as the grain boundary of
interest is well-dened as a continoius boundary, next step appears [90]. The optimized
threshold value for gure 2.1 was obtained to be 0.80. The resulting binary image is
shown in gure 2.2.
Removing particles
Impurities and particles within a grain can strongly aect the quantitative analysis on
grain boundaries and thus should be removed. To this aim the region lling operation
is used. The imfill function, I = imfill(I,

holes

), identies the pixels constituting


holes within connected components in the binary image and lls them. A hole is a set
of connected background pixels that cannot be reached by lling in the background from
the edge of the image [90]. The resulting output image is shown in gure 2.3.
32 Image processing
Figure 2.2: Microstructure after binarization using T = 0.80.
Figure 2.3: Microstructure after hole-lling.
2.2 MATLAB program for boundary extraction 33
(a) 4-connection. (b) 8-connection.
Figure 2.4: Concept of connectedness.
Extracting the connected objects
An object in a binary image consists of any group of connected pixels. Two denitions
of connection between pixels are commonly used. If a given foreground pixel has at
least one neighboring foreground pixel to the north, south, east or west of itself to be
considered as part of the same object, then the 4-connection concept is applied. If,
however, a neighboring foreground pixel to the north-east, north-west, south-east or south-
west is sucient for it to be considered as part of the same object, then the 8-connection
defenition is used. The basic concept is shown in gure 2.4.
For any foreground pixel, the set of all foreground pixels connected to it is called the
connected component of that pixel. A group of pixels which are all connected to each
other in this way is dierentiated from others by giving it a unique label. Extracting
connected components can be achieved with the function bwlabel. The function I =
bwlabel(I), analyses an input binary image so as to identify all the connected objects in
the image. It returns a so called labeled image in which all the pixels belonging to the
rst connected component are assigned a value 1, all the pixels belonging to the second
connected component are assigned a value 2 and so on. In this process the default value
for connection is 8 [90, 91].
Thickening the contour of the object
To apply the labeling method, the user needs to make sure that the object of interest(in
this case the largest grain in the image) is well isolated from the rest of the microstrusture,
while preserving the details, otherwise the features would not be labeled correctly. For
this type of microstructure to be analized, the optimized approach is a one step of line
34 Image processing
Figure 2.5: Disk-shaped structuring element.
thickening, followed by a morphological opening operation.
I = bwmorph(I,

thin

, n) removes pixels from the exterior of the foreground object,


resulting in a thicker boundary. It applies the operation n times. In this case it is set to
1, which means only one step thickening is performed.
I = imopen(I, se) performs morphological opening on the binary image I with the struc-
turing element se. The general eect of opening is to remove small, isolated objects from
the foreground of the image, placing them in the background. So it tends to smooth the
contour of the binary object. The structuring element is the key factor which determines
exactly which image pixels surrounding the given foreground/background pixel must be
considered in order to make the decision to change its value or not. Using the MATLAB
function se = strel(shape, parameters) creates a morphological structuring element of
the type specied by shape. Depending on shape, strel can take additional parameters.
se = strel(

disk

, R, N) creates a at, disk-shaped structuring element (see gure 2.5),


where R species the radius [89]. In the present case R is dened as 2, which determines a
small structuring element. After this operation, the small interconnections such as broken
grain boundaries are removed and the image is ready for correct labeling. The resulting
image after labeling is demonstrated in gure 2.6.
Detecting the desired component
Imtool(I) opens a new Image Tool displaying the image. To access the Pixel Region
Tool, user should select Tools > Pixel region in the Image Tool menu. Then the Pixel
Region Tool opens a separate gure window containing an extreme close-up view of a small
region of pixels in the target image, then the user can drag the pointer through the object
of interest and read the pixel value which is constant all over the connected component.
This is the input matrix for the Detect function. Inew = Detect(I, pixelvalue) is a
MATLAB function which gets the labeled matrix and creates a matrix the same size as
I consisting of all zeros, then replaces an array including the elements with the indicated
2.2 MATLAB program for boundary extraction 35
Figure 2.6: Microstructure after contour thickening followed by labeling.
pixel value in the zeros matrix while all other elements are set to zero. Having dened
the pixel value of the largest grain in the current microstructure as 612, Detect function
returns an output as shown in gure 2.7.
Closing
Morphological closing tends to remove small holes in the foreground, changing small
regions of background into foreground. The main reason to apply closing in this stage
is to eliminate small unavoidable articial traces on the grain boundary caused by the
previous thresholding operation. If second phase particles or subgrain-boundaries are
located very close to the main grain boundary to be extracted (see gure 2.1), they
appear to be connected to the grain boundary after thresholding operation (see gure 2.2),
consequently the detected grain as shown in gure 2.7 has some tiny articial intrusions
on the grain boundary. Here a disk of radius 7 was chosen as the structuring element
which returns the image illustrated in gure 2.8. The radius of 7 was found to be the
optimized value for the group of microstructures being processed in this study, in other
words it might not be valid for other types of grain strustures or other materials.
36 Image processing
Figure 2.7: Detected grain.
Figure 2.8: Detected grain after closing.
2.2 MATLAB program for boundary extraction 37
Figure 2.9: Extracted grain boundary.
Edge detection
An edge is dened as a signicant, local change in image intensity. There are a large
number of edge detection algorithms available. Here we focus on a particular one de-
veloped by John F. Canny ([92]) to produce a one pixel wide edge, while preserving the
structural properties of the desired features. The Cannys Edge Detector is optimal with
regards to localization, in other words the detected edges are as close as possible to the
real edges. I = edge(I,

canny

) nds edges in the gure 2.8 and returns the extracted


boundary as shown in gure 2.9. Please note that as a result of the procedure followed,
the boundary is always continious. This is a suitable input image for box counting and
fractal dimension calculations.
Final particle removing step
If there is any small particle still remained inadvertently after the edge detection opera-
tion, bwareaopen function helps to elliminate it. IB = bwareaopen(I, P), removes from
a binary image all connected components (objects) that have fewer than P pixels, pro-
ducing another binary image, IB. The default connectivity is 8 for two dimensions [89].
38 Image processing
(a) Original microstructure. (b) Extracted grain boundary.
Figure 2.10: Initial image vs nal image.
The basic steps are:
Determining the connected components
Computing the area of each component
Removing small objects
This operation has no eect on the current case, however for fully recrystallized mi-
crostructure the remnants of small objects are removed by this nal operation. In gure
2.10 the original microstructure and the extracted boundary are compared, showing that
the grain boundary features are well preserved. Please note that the procedure used aims
to nd the perfect topology of one grain boundary, rather than a less accurate represen-
tation of all grain boundaries. As pointed out earlierly in this chapter the input value
for thresholding and pixel value of labled feature are chosen based on the grain to be ex-
tracted. This increases the accurcy of preserving real prole of the boundary of interest.
Therefore it is not possible to detect dierent recrystallized grains in one microstruc-
ture (see 2.10) at the same time, since every single recrystallized grain is labled as one
connected component having dierent pixel value from the other objects. So the whole
procedure should be redone if more than one grain is aimed to be extracted in the same
image.
2.2 MATLAB program for boundary extraction 39
Figure 2.11: Microstructure of fully recrystallized Al7050.
Table 2.1: Input values for the MATLAB program
Microstructure Threshold value Pixel region
Partially recrystallized 0.80 612
Fully recrystallized 0.62 127
2.2.2 Boundary extraction in fully recrystallized microstructure
Figure 2.11 shows a gray-scale tif image of a fully recrystallized Aluminum alloy as the
second case study for boundary extraction using the presented MATLAB program. The
image size is 1461 pixels by 1222 pixels and the resolution is 96,000 pixels/inch.
Suppose the nal goal is to detect the grain boundary on the bottom-right side of the
image. Having read the input, the function operates the same procedure as for the
rst case of partially recrystallized microstructure. The input values should be adjusted
depending on the image and the desired object to be detected (see table 2.1). The
consequent processed images are illustrated in the following section.
40 Image processing
Figure 2.12: Microstructure after binarization.
Since the processed image (gure 2.17) still contains a few traces of particles, obviously
it is not a suitable input for the fractal analysis. Thus one step bwareaopen operation
enhances the image by removing small particles containing fewer than 100 pixels. The
output is shown in gure 2.18. The original fully recrystallized microstructure and the
extracted boundary are compared In gure 2.19, showing that the desired grain boundary
feature is detected without missing details.
2.2 MATLAB program for boundary extraction 41
Figure 2.13: Microstructure after hole-lling.
Figure 2.14: Microstructure after contour thickening followed by labeling.
42 Image processing
Figure 2.15: Detected grain.
Figure 2.16: Detected grain after closing.
2.2 MATLAB program for boundary extraction 43
Figure 2.17: Extracted grain boundary.
Figure 2.18: Extracted grain boundary after opening.
44 Image processing
(a) Original microstructure. (b) Extracted grain boundary.
Figure 2.19: Initial image vs nal image.
2.3 Conclusions
The presented MATLAB program has been shown to be capable of extracting the grain
boundaries of interest in both fully and partially recrystallized microstructures of Alu-
minum alloys, while preserving the details needed for the fractal dimension calculations.
Chapter 3
Box counting method for fractal
dimension calculation
3.1 Introduction
Fractal concept is a new approach for quantifying the irregularity of complex shapes
that may not be treated by Euclidean geometry. In general, a fractal is dened as a set
having non-integer dimension. Consequently the fractal dimension (FD) is introduced as
a factor highly correlated with the human perception of objects roughness. FD lls the
gap between one- and two-dimensional objects. The more complex the contour of the
curve, the more it lls the plane and the more its fractal dimension will be closer to 2
[93]. The FD can be calculated in dierent ways. The present study concentrates on the
box-counting method. To implement this method a MATLAB program was developed
based on the theoretical considerations. To validate the code, test images of Koch curves
having known fractal dimension were generated and the estimated fractal dimensions are
compared to the known fractal dimension of the input gure.
3.2 Box-counting dimension
The box-counting dimension is probably the most frequently used method in characteri-
zation of irregular sets in science. The reason is that it is easy, automatically computable,
and applicable for patterns with or without self-similarity [94, 95]. The denition goes
45
46 Box counting method for fractal dimension calculation
Figure 3.1: Surrounding the Koch curve with boxes [96].
back at least to the 1930s and it has been variously termed Kolmogorov entropy, entropy
dimension, capacity dimension, and metric dimension [95]. In this method, each image
is covered by a sequence of grids of increasing sizes and for each of the grids, two values
are recorded: the number of square boxes containing at least one pixel of the projection
contour, N(r), and the size of the squares, r. The number N of boxes of size r needed to
cover a fractal set follows a power-law:
N(r) = N
0
r
D
(3.1)
where N
0
is a constant and D D
space
(D
space
is the dimension of the space, usually
D
space
= 1, 2, 3). The regression slope D of the straight line formed by plotting log(N(r))
against log(r) indicates the degree of complexity, or fractal dimension, between 1 and 2
(1 < D < 2) (see equation 3.2). D is known as box-counting dimension [80, 83].
Log(N(r)) = DLog(r) +Log(N0) (3.2)
The concept of box covering procedure for a Koch curve is illustrated in gure 3.1.
3.3 MATLAB code for box-counting calculations 47
3.3 MATLAB code for box-counting calculations
A quantitative analysis of perimeter roughness is carried out to illustrate the degree of
complexity of a binary image. In the input image all foreground pixels having a value of
1 are considered as the feature the dimension of which is to be calculated. BoxCount(I)
counts the number of boxes (boxnum) of size res needed to cover the nonzero elements
of a 2D array I. For practical purposes it is often convenient to consider a sequence of
grids where the mesh size is increased by a factor of 2 from one grid to the next. So
in this approach the box sizes are powers of two, i.e., res = 1, 2, 4...2
n
, where n is the
smallest integer such that max(size(I)) 2
n
. Noting that a 2D image is represented by a
N M matrix in MATLAB environment, n is dened based on the maximum dimension
of the input image. If the sizes of I over each dimension are smaller than 2
n
, I is padded
with zeros to size 2
n
over each dimension. The output vectors boxnum and res are of
size n + 1. The algorithm superimposes a grid of squares over the edge, while counting
the occupied squares that the edge passes through. This is continued for an increasing
number of squares. So plotting Log((boxnum)) versus Log(res), the points should lie
approximately on a straight line. It should be noted that in the MATLAB environment
Y = log(X) function operates the natural logarithm. This program ts a least-square
straight line to all the points of log-log data allocating equal weight to all data points.
Then the box-counting dimension is given as the negative gradient of this line. This
is a close approximation of the fractal dimension of the pattern. In order to check the
applicability of this method for characterizing grain boundaries in the microstructures,
a set of triadic Koch curves with known theoretical fractal dimension were articially
generated and the results are shown in the next section.
3.4 Generation of Koch snowakes
Since our goal was to examine the accuracy of box-counting algorithm for fractal di-
mension analysis of boundaries, the Koch snowake(also called Koch island) with known
theoretical fractal dimension of 1.2619 was generated at dierent iterations using a MAT-
LAB function [97]. The Koch snowake is constructed by starting with an equilateral
triangle, then recursively altering each line segment as follows:
1. divide the line segment into three segments of equal length.
2. draw an equilateral triangle that has the middle segment from step 1 as its base
and points outward.
48 Box counting method for fractal dimension calculation
3. remove the line segment that is the base of the triangle from step 2.
Construction steps are well shown in gure 3.2.
The actual Koch snowake is the limiting curve obtained by applying this construction
an innite number of times. It should also be noted that the precise relationship between
an estimator of fractal dimension and a mathematical concept of dimension, such as box-
counting dimension is known only under innite iteration. Obviously, it is not possible to
generate and save innite Koch snowakes because of the limited physical memory avail-
able on the computers. What has been done in this study was constructing Koch islands
in 2, 3, 4, and 5 iterations, and analyze them at two dierent resolutions of 400 dpi and
600 dpi in order to explore the eect of resolution and iteration number on the accuracy
of calculated fractal dimension. Resolutions less than 400 dpi are inadequate to show all
the small features in the graphic image of mostly detailed Koch 5 (Koch curve after 5
iterations), while images with resolutions higher than 600 dpi are too large to be analyzed
with regard to the available computer memory of 2,00 GB for the Intel(R)Core(TM)2 Duo
CPU processor. Generated Koch islands are shown in gure 3.3. Canny edge detection
[92] has been applied to the generated images in order to produce uniform thickness of
one-pixel for all lines. This step of image analysis is necessary to improve the accuracy
of box-counting calculations. In the next section the results of box-counting analysis on
the Koch curves illustrated in gure 3.3 are reported.
3.5 Implementation of the box-counting program 49
Figure 3.2: Construction of the Koch curve.
Table 3.1: Size of generated Koch images.
All Koch images size(pixel pixel) n-value new dimension (2
n
2
n
)
400 dpi 1867 2144 12 4096 4096
600 dpi 2800 3217 12 4096 4096
3.5 Implementation of the box-counting program
The presented box-counting procedure was adopted to determine the FD of articially
generated Koch curves after 2, 3, 4, and 5 iterations. The binary images of one-pixel wide
Koch islands at two resolutions of 400 dpi and 600 dpi as illustrated in gure 3.3 are the
inputs for this program. Due to the limited memory of the computer for calculations,
600 dpi was the highest limit for resolution, and 5 was the highest available number of
iterations. The estimate of FD was based on the measurements made over the whole
range of box sizes, starting from 1 pixel, to 2
n
, where n is the smallest integer such that
max(size(I)) 2
n
, as explained earlier in this chapter. Based on the original image sizes
generated earlier, n is calculated as shown in table 3.1.
The following plots (see gures 3.4, 3.5, 3.6, and 3.7) are the results of box-counting
experiments on four articially generated Koch islands, each at two dierent resolutions.
For each case the best tted line and the fractal dimension are illustrated. The results of
these calculations are summarized in tables 3.2 and 3.3.
50 Box counting method for fractal dimension calculation
Table 3.2: Results of box-counting for dierent Koch islands at the resolution of 400 dpi.
Koch images calculated fractal dimension deviation from theoretical value of 1.2619
Koch2 1.1151 0.1468
Koch3 1.1603 0.1016
Koch4 1.2036 0.0583
Koch5 1.2305 0.0314
Table 3.3: Results of box-counting for dierent Koch islands at the resolution of 600 dpi.
Koch images calculated fractal dimension deviation from theoretical value of 1.2619
Koch2 1.1416 0.1203
Koch3 1.1869 0.0750
Koch4 1.2298 0.0321
Koch5 1.2605 0.0014
3.5.1 Discussion
In this section the calculation of data points and the resulting fractal dimension for Koch5
at 600 dpi is discussed in detail as an example. Then the general trends related to the box-
counting results of all cases are explored while discussing how the results are aected by
the number of iterations and the resolution. Finally the origin of some possible calculation
errors will be explained.
Koch5 as an example
Upon implementing the box-counting analysis on the binary image of Koch5, grids of
increasing size starting from one pixel cover the digital image of Koch5. The rst box
is put over the rst element of the 2-D matrix, in other words I(1,1), the most top-left
element in the array. Number of boxes containing at least one pixel of foreground are
counted and saved. The procedure is repeated for increasing box sizes as 2
n
, until the
largest box ts the whole image and returns 1 for the box number. Results are shown
in table 3.4. Plotting the results on a graph with log(boxnum) on the vertical axis and
log(res) on the horizontal axis (see 3.7), illustrates that the data points are well aligned
on a straight line. The best line is tted using the least-square method, while the slope
of this line gives an approximation of the fractal dimension of Koch5.
3.5 Implementation of the box-counting program 51
(a) koch2. (b) koch3.
(c) koch4. (d) koch5.
Figure 3.3: Koch snowakes.
52 Box counting method for fractal dimension calculation
Figure 3.4: Results of box-counting calculation for Koch2 at two resolutions.
Figure 3.5: Results of box-counting calculation for Koch3 at two resolutions.
3.5 Implementation of the box-counting program 53
Figure 3.6: Results of box-counting calculation for Koch4 at two resolutions.
Figure 3.7: Results of box-counting calculation for Koch5 at two resolutions.
54 Box counting method for fractal dimension calculation
Table 3.4: Results of box-counting for Koch5 at 600 dpi.
res boxnum
1 155822
2 48217
4 16532
8 6175
16 2371
32 925
64 374
128 145
256 54
512 18
1024 8
2048 4
4096 1
Eect of iteration number and resolution on the calculated fractal dimension
As shown in tables 3.2 and 3.3, at a constant resolution, increasing the number of iterations
results in a FD closer to the theoretical value of 1.2619. In other words, the more the
generated Koch island tends to innity, the less deviation is reported in the calculated
box-counting dimension. While Koch5 well represents a fractal Koch curve, its calculated
box-counting dimension shows only a very small deviation from the mathematical FD of
innitely reproduced Koch snowake.
By increasing the resolution from 400 dpi to 600 dpi, for all Koch curves the calculated FD
tends more to the mathematical value of 1.2619. In other words the higher the resolution,
the higher is the calculated fractal dimension. So it can be concluded that at the higher
resolution the greater number of boxes are counted for the smaller box sizes, thus the best
tted line is higher in gradient. This can be clearly related to the image size dierence at
dierent resolutions. The same Koch curve is spread broader at higher resolution, thus
takes up more pixels. This eect is not profound for large box sizes since the images are
padded with zeros to 4096pixels over each dimension as shown in table 3.1. It has been
also reported in literature for two types of mathematically generated fractal surfaces, that
the estimated FD tend consistently towards the theoretical FD with increasing the image
resolution [98].
3.6 Conclusions 55
3.5.2 The origin of possible errors
The thickness of the lines in a digital image of real microstructures is an important
parameter aecting the range of box sizes in the box-counting method. If the image
has lines or curves with a thickness greater than the smallest box size (one pixel for the
present case) over-counting problems may occur. Over-counting means that the number
of counted boxes for the range of box sizes smaller than the line thickness increases
compared to the number of counted boxes for boxes greater than or equal to the line
thickness. This results in a higher gradient of the best tted line, thus higher fractal
dimension. In other words the calculated fractal value would reect the dimension of
a 2-dimensional area for the boxes smaller than the line thickness. In order to prevent
such a problem caused by the thickness of the lines, the Koch curves were processed to
generate one-pixel-wide edges all over the images. It is also of great importance to avoid
lack of uniformity in thickness for lines over the images. However all of the images in this
chapter were computer-generated with uniform thickness, and the Canny edge detector
also keeps uniformity of the lines in reducing them into one-pixel-wide boundaries.
3.6 Conclusions
In the present chapter the box-counting program was implemented to calculate the fractal
dimension of the articially generated Koch islands of 2, 3, 4, and 5 iterations at two
dierent resolutions of 400 dpi and 600 dpi. The conclusions can be summarized as
following:
The calculated values of fractal dimension for the sequence of Koch curves converges
to the mathematical value of 1.2619, as the number of iterations increases.
At each constant iteration number, increasing the resolution results in less deviation
of calculated fractal dimension from the theoretical value.
It is possible to apply the present box-counting method to non-mathematically com-
plex objects such as a serrated grain boundary and calculate the fractal dimension.
Since the increase in resolution for Koch curves has been shown to improve the
calculated fractal dimensions, the use of the microstructure images with the highest
available resolution is a requirement for accurate estimation of the fractal dimension
of grain boundaries in Aluminium alloys.
56 Box counting method for fractal dimension calculation
Chapter 4
Calculating fractal dimension of the
grain boundaries
4.1 Introduction
The high strength 7050 Aluminum plate alloy has been widely used for manufacturing
of the thick section products in structural aircraft applications, where fracture toughness
and damage tolerance properties are important [99, 100]. The industrial production of
these structural components involves various processing steps, including casting, homog-
enization, preheat, hot rolling, solution heat treatment and aging. The microstructure of
the material which determines the mechanical properties, is controlled by dierent factors
during all processing steps. Many investigations have been carried out on the fracture
toughness of the these alloys showing that the fracture toughness is inuenced by a range
of microstructural features, such as the coarse particles [101, 102], the aging precipitates
[103, 104], the conguration of the precipitation free zone (PFZ) [105, 106], and the grain
structure [107]. These features can contribute to the initiation and propagation of frac-
ture. Consequently, the overall fracture of partially recrystallized AA7050 alloy involves
three main mechanisms, intergranular fracture, fracture and decohesion of coarse par-
ticles, and microvoid-induced transgranular fracture. Hence, the fracture toughness of
these alloys is determined by the actual fraction of these micromechanisms in the overall
fracture, which are in turn governed by the microstructure [108, 109, 107].
57
58 Calculating fractal dimension of the grain boundaries
The current research focuses on understanding the microstructure/toughness relation-
ships by investigating the eect of grain boundary structure on the fracture toughness of
hot-rolled AA7050. In order to obtain improved mechanical properties special rolling pro-
cedures have been developed by our material supplier, introducing additional deformation
into the plate. In this report they are referred to as the modied hot-rolling processes
compared to the conventional rolling procedure. It has been shown that depending on
how rolling is performed in the so-called modied rolling processes, it is possible to ma-
nipulate shape of the grain boundaries, and generate dierent grain boundary structures
for the same overall thickness reduction. Since signicant improvement in fracture tough-
ness has been reported for a modied-rolled materials compared to conventionally-rolled
plates it might be expected that grain boundary shape plays an important role in gov-
erning the fracture toughness of the material. To quantitatively characterize the grain
boundary structures, the fractal dimension analysis of high-angled grain boundaries in
the partially-recrystalized microstructure of unconventionally and conventionally rolled
AA7050 samples is performed following the procedure presented in previous chapters.
4.1.1 Fracture toughness testing
Fracture toughness is an indication of the amount of stress required to propagate a preex-
isting crack in the material. The Kahn tear test is one of the most widely used methods
to characterize the notch resistance of precipitation hardening Aluminum alloys in terms
of the resistance of the material to tearing [103, 110, 111]. Despite the fact that the de-
formation mode is quite dierent from the classical plane-strain fracture toughness test,
this method has been proven to be remarkably representative of toughness [112, 113].
Tear tests are performed on thin (2 mm thick), notched samples loaded in tension at con-
stant speed. The geometry of specimen is shown in gure 4.1. The load and deformation
during the test are measured and used to characterize the notch resistance by calculating
the energy required to initiate and propagate crack as demonstrated in gure 4.2. The
crack initiation region and the crack propagation region are illustrated with yellow and
blue respectively. The unit propagation energy (UPE) is calculated by dividing the area
under the load-deformation graph by the specimen cross sectional area for crack initiation
[114].
4.1 Introduction 59
(a) un-tested
tear speci-
men.
(b) Tear-tested
specimen.
Figure 4.1: The sketches of an untested and tested Kahn-tear test specimens.
Figure 4.2: Schematic result of tear test to measure UPE.
60 Calculating fractal dimension of the grain boundaries
Figure 4.3: Schematics of sample orientations used for fracture toughness tests of the plates.
4.2 Experimental work
The microstructure characterization was carried out on the tear tested specimens of
three types of heat-treated AA7050 plates of standard commercial composition, including
conventional-rolled (CoR), modied-rolled-type I (MoR
I,1
and MoR
I,2
), and modied-
rolled type II (MoR
II,1
and MoR
II,2
), for which the UPE data were available. All sam-
ples received 50% thickness reduction, while for each modied-rolled case, two dierent
percentage of modied rolling has been applied, in type I category MoR
I,2
had higher
percentage of modied rolling than MoR
I,1
, while in type II group MoR
II,2
had higher
percentage of modication than MoR
II,1
. The heat treatment was T74 condition which
is solution treatment and articial ageing, including 2.5 hours at 475

C plus 5 hours at
120

C and 18 hours at 165

C. The heat treatment has yielded specimens that are partially


recrystallized. The test specimen were cut o at 0

and 90

orientation from the rolled


slab as shown in gure 4.3. The three orthogonal characteristic directions in the rolled
structures are commonly labelled as L, T, and S, for the longitudinal (rolling direction),
transverse and short transverse directions, respectively. The schematic illustration of tear
test specimen used for fracture toughness measurement is shown in gure 4.1. The rect-
angular labled with A (10mm15mm) shows the region for microstructure analysis which
is chosen from non-deformed area of the sample far from the fracture surface. The reason
to study this plane is that during the fracture toughness test the main crack grows on
this plane, thus the microstructure of the material on this plane is expected to govern
the crack propagation behaviour. So quantitative microscopy is carried on on S-T plane
for 90

orientation samples, and L-T for 0

specimens (see gure 4.4). Arrow shows the


crack propagating direction on each plane.
4.3 Results and discussion 61
(a) S-T plane for 90

samples.
(b) L-T plane for 0

sam-
ples.
Figure 4.4: The specimens for microstructure analysis labeled as A in 4.1. Arrows show the
crack growth direction.
The samples for microstructure observation were cut as indicated, then mounted in an
epoxy resin and ground to p4000 with sandpaper and polished with diamond suspension
to 1/4 microns. For etching, a 50

C 10% phosphoric acid was used. Samples were soaked


in the solution for 4 minutes and rinsed and dried. The microstructure of the sample
were examined and analyzed at 20X magnication with an Olympus LEXT confocal
microscopy, then saved as images of size 1024 pixels 768 pixels, at the resolution of 96
dpi.
4.3 Results and discussion
Figure 4.5 shows the optical micrographs of partially recrystallized conventional-rolled,
modied-rolled type I and modied-rolled type II samples, on the metallographic plane
of L-T (0

) and S-T (90

). In the micrographs long elongated bright areas are the recrys-


tallized regions, and the network of small grains are the subgrain structure shown in dark
gray on the images. Constituent particles appear black which are found mostly in con-
junction with the recrystallized areas. The boundaries between the recrystallized regions
and non-recrystallized regions are the high angle grain boundaries on which quantitative
analysis is performed.
62 Calculating fractal dimension of the grain boundaries
(a) Conventional-rolled (CoR), 0

sample (b) Conventional-rolled (CoR), 90

sample
(c) Modied-rolled type I (MoRI,1), 0

sample
(d) Modied-rolled type I (MoRI,1), 90

sample
(e) Modied-rolled type I (MoRI,2), 0

sample
(f) Modied-rolled type I (MoRI,2), 90

sample
Figure 4.5: Microstructure of partially recrystallized 7050 alloy observed on the planes of
S-T and L-T.
4.3 Results and discussion 63
(g) Modied-rolled type II (MoRII,1), 0

sample
(h) Modied-rolled type II (MoRII,1), 90

sample
(i) Modied-rolled type II (MoRII,2), 0

sample
(j) Modied-rolled type II (MoRII,2), 90

sample
Figure 4.5: Microstructure of partially recrystallized 7050 alloy observed on the planes of
S-T and L-T.
4.3.1 Image processing on the Aluminum microstructure
Since the main crack grows parallel to S and L directions in 90

and 0

specimens respec-
tively (see gure 4.4), the fractal dimension of grain boundaries between recrystallized
and un-recrystallized regions are characterized along S and L directions. For this pur-
pose Fiji software [115] was used for image analysis to generate rectangular masks having
aspect ratio of 3 (300 pixels100 pixels), while the length of the box is always located
in the crack propagation direction. Then the part of high angle grain boundary tted
in the mask were cut for fractal analysis. The procedure is illustrated for example mi-
crostructures of 0

and 90

of all samples in gures 4.6 to 4.15. In this work 10 partial


grain boundaries for each sample were chosen, cut, and processed following the procedure
64 Calculating fractal dimension of the grain boundaries
introduced in chapter 2. Box counting measurements were then performed on the binary
images of the one pixel wide grain boundaries as follows.
4.3 Results and discussion 65
(a) Image of the L-T plane of CoR, the blue rectangular box shows the
selected part of the grain boundary aligned in L direction chosen for
fractal analysis.
(b) Selected boundary, 300 pixels X 100 pixels
(c) Detected grain boundary
Figure 4.6: Procedure of sampling the grain boundary aligned in L direction on L-T plane
of CoR 0

sample.
66 Calculating fractal dimension of the grain boundaries
(a) Image of the L-T plane of MoRI,1, the blue rectangular box shows the
selected part of the grain boundary aligned in L direction chosen for
fractal analysis.
(b) Selected boundary, 300 pixels X 100 pixels
(c) Detected grain boundary
Figure 4.7: Procedure of sampling the grain boundary aligned in L direction on L-T plane
of MoR
I,1
0

sample.
4.3 Results and discussion 67
(a) Image of the L-T plane MoRI,2, the blue rectangular box shows the
selected part of the grain boundary aligned in L direction chosen for
fractal analysis.
(b) Selected
boundary,
100 pixels X
300 pixels
(c) Detected
grain bound-
ary
Figure 4.8: Procedure of sampling the grain boundary aligned in L direction on L-T plane
of MoR
I,2
0

sample.
68 Calculating fractal dimension of the grain boundaries
(a) Image of the L-T plane of MoRII,1, the blue rectangular box shows
the selected part of the grain boundary aligned in L direction chosen
for fractal analysis.
(b) Selected boundary, 300 pixels X 100 pixels
(c) Detected grain boundary
Figure 4.9: Procedure of sampling the grain boundary aligned in L direction on L-T plane
of MoR
II,1
0

sample.
4.3 Results and discussion 69
(a) Image of the L-T plane of MoRII,2, higher modication percentage,
the blue rectangular box shows the selected part of the grain boundary
aligned in L direction chosen for fractal analysis.
(b) Selected
boundary,
100 pixels X
300 pixels
(c) Detected
grain bound-
ary
Figure 4.10: Procedure of sampling the grain boundary aligned in L direction on L-T plane
of MoR
II,2
0

sample.
70 Calculating fractal dimension of the grain boundaries
(a) Image of the S-T plane CoR, the blue rectangular box shows the se-
lected part of the grain boundary aligned in S direction chosen for
fractal analysis.
(b) Selected boundary, 300 pixels X 100 pixels
(c) Detected grain boundary
Figure 4.11: Procedure of sampling the grain boundary aligned in S direction on S-T plane
of CoR 90

sample.
4.3 Results and discussion 71
(a) Image of the S-T plane of MoRI,1, the blue rectangular box shows the
selected part of the grain boundary aligned in S direction chosen for
fractal analysis.
(b) Selected boundary, 300 pixels X 100 pixels
(c) Detected grain boundary
Figure 4.12: Procedure of sampling the grain boundary aligned in S direction on S-T plane
of MoR
I,1
90

sample.
72 Calculating fractal dimension of the grain boundaries
(a) Image of the S-T plane of MoRI,2, the blue rectangular box shows the
selected part of the grain boundary aligned in S direction chosen for
fractal analysis.
(b) Selected
boundary,
100 pixels X
300 pixels
(c) Detected
grain bound-
ary
Figure 4.13: Procedure of sampling the grain boundary aligned in S direction on S-T plane
of MoR
I,2
90

sample.
4.3 Results and discussion 73
(a) Image of the S-T plane of MoRII,1, the blue rectangular box shows the
selected part of the grain boundary aligned in S direction chosen for
fractal analysis.
(b) Selected boundary, 300 pixels X 100 pixels
(c) Detected grain boundary
Figure 4.14: Procedure of sampling the grain boundary aligned in S direction on S-T plane
of MoR
II,1
90

sample.
74 Calculating fractal dimension of the grain boundaries
(a) Image of the S-T plane of MoRII,2, the blue rectangular box shows the
selected part of the grain boundary aligned in S direction chosen for
fractal analysis.
(b) Selected
boundary,
100 pixels X
300 pixels
(c) Detected
grain bound-
ary
Figure 4.15: Procedure of sampling the grain boundary aligned in S direction on S-T plane
of MoR
II,2
90

sample.
4.3 Results and discussion 75
4.3.2 Box counting measurements on the grain boundaries of Aluminum
microstructure
Box counting dimension as an index of irregularity of grain boundaries between recrys-
tallized and un-recrystallized regions of the microstrudture is calculated following the
procedure described in chapter 3. Noting that an image is a 2D array I in the MATLAB
environment, the automatic box-counting program requires binary images of grain struc-
tures in which the grain boundary of interest is indicated with white pixels (foreground)
and the rest pixels are assigned as black (background). The image processing is carried
out on 10 randomly selected part of the grain boundaries for each sample. The original
size of the images which are cut from the whole micrograph according to the mask size is
300 pixels100 pixels for the horizontal masks, and 100 pixels300 pixels for the vertical
masks. Since grids of increasing size in the order of res = 1, 2, 4...2
n
cover the image, n is
dened as the smallest integer such that max(size(I)) 2
n
, where max(size(I)) is the
maximum dimension of the input image. In order to fulll max(size(I)) 2
n
, images
are padded with zeros to size 2
9
over each dimension. Meshes of increasing size starting
from one pixel cover the digital images and number of boxes containing at least one pixel
of the grain boundary are counted. The procedure is repeated for increasing box sizes
until the largest box ts the whole image and returns 1 for the box number, in this case
the largest box size is 512 pixels. The fractal dimension corresponding to the slope of the
log(boxnum) versus log(res) plot is calculated by least-squares linear tting. Examples
of box counting results of the microstructures presented earlier in this chapter are shown
in gure 4.16, the estimated fractal dimension for each case is reported. The correlation
coecients show a value of more than 0.99 for all of the results. The overall fractal dimen-
sion for each sample is averaged over 10 randomly selected grain boundaries and shown
in table 4.1. The results reported in table 4.1 shows that the grain boundary congura-
tions were successfully evaluated on the basis of fractal geometry in the microstructure
of heat-treated AA7050 alloy.
76 Calculating fractal dimension of the grain boundaries
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(a) Conventional-rolled (CoR), 0

sample,
FD = 1.0308
0 1 2 3 4 5 6 7
0
1
2
3
4
5
6
7
log(res)
l
o
g
(
b
o
x
n
u
m
)
(b) Conventional-rolled (CoR), 90

sam-
ple, FD = 1.0079
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(c) Modied-rolled type I (MoRI,1), 0

sample, FD = 1.0086
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(d) Modied-rolled type I (MoRI,2), 0

sample, FD = 1.0188
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(e) Modied-rolled type II (MoRII,1), 0

sample, FD = 1.0129
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(f) Modied-rolled type II (MoRII,2), 0

sample, FD = 1.0277
Figure 4.16: Box counting results for processed images shown in gures 4.6 to 4.15.
4.3 Results and discussion 77
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(g) Modied-rolled type I (MoRI,1), 90

sample, FD = 1.0385
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(h) Modied-rolled type I (MoRI,2), 90

sample, FD = 1.07053
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(i) Modied-rolled type II (MoRII,1), 90

sample, FD = 1.0086
0 1 2 3 4 5 6 7
1
0
1
2
3
4
5
6
7
log(res)
lo
g
(
b
o
x
n
u
m
)
(j) Modied-rolled type II (MoRII,2), 90

sample, FD = 1.0526
Figure 4.16: Box counting results for processed images shown in gures 4.6 to 4.15.
78 Calculating fractal dimension of the grain boundaries
4.3.3 Relationship between the fractal dimension of the grain bound-
aries and the fracture toughness
It has been reported that the modication of grain boundary congurations by making
them serrated has an important inuence on strengthening the grain boundaries [47,
116, 53]. Since the grain boundaries are the preferred path for crack propagation during
fracture, the crack growth behavior may be aected by the shape of the boundaries. So the
quantitative data of fractal dimension of grain boundaries can be used to investigate the
relationship between fracture toughness and ruggedness of high angle grain boundaries.
The UPE data for each sample were averaged over three available test results and the
average value is shown in table 4.2. Figure 4.17 shows the plot of UPE versus fractal
dimension in which blue data points are the results of 90

samples, while the results


of 0

samples are illustrated in red. Each data point is also labeled with the name of
related sample. The standard deviation of UPE and FD data are shown using error bars.
Concerning all data points, increasing the fractal dimension of the grain boundary clearly
increases the fracture toughness of the material for both 0

and 90

samples. For 0

samples there is a linear relationship between UPE and FD of the boundaries. In both
0

and 90

data group the high Mod. I material shows the highest FD and the highest
value of UPE. Comparing 0

and 90

samples for each type of rolling process reveals that


both fracture toughness and FD are higher for 90

than 0

. This observation supports


the idea that the crack growth in L direction on L-T plane of 0

sample consumes less


energy compared to crack propagation on S-T plane of 0

sample when the crack grows in S


direction. This idea is strengthened by the fact that 0

samples mostly show intergranular


mode of fracture. So less serrated grain boundaries provide easier path for cracks to grow
along, which requires less energy for crack propagation. For highly rugged grain boundary,
the fracture crack either grows along the serrated path of the boundary or through the
matrix material, both of which result in high propagation energy. So the higher the
serration of the grain boundaries along the crack propagation direction, the higher the
UPE data are obtained. However the rate of increasing is not constant and this holds
true until a certain point at which the material having highly serrated boundaries is very
tough that a small increase in FD of the boundary does not make very much dierence
for the crack to grow along the boundaries, which means that the fracture mode is mostly
transgranular and controlled by the matrix material. For 0

data series if the last data


for MoR
I,2
sample is excluded, the rest of data points can be considered as a straight
line as shown in gure 4.18.
4.3 Results and discussion 79
Table 4.1: Results of box-counting on Aluminum microstructures.
sample Average fractal dimension over 10 microstructure samples standard deviation
CoR90 1.032 0.012
MoR
I,1
90 1.053 0.025
MoR
I,2
90 1.094 0.027
MoR
II,1
90 1.023 0.012
MoR
II,2
90 1.049 0.023
CoR0 1.019 0.010
MoR
I,1
0 1.023 0.016
MoR
I,2
0 1.036 0.021
MoR
II,1
0 1.018 0.015
MoR
II,2
0 1.027 0.012
4.3.4 Relationship between the fracture toughness and other microstruc-
tural features
In order to investigate any possible relation between the fracture toughness and the other
attributes of the partially recrystallized microstructure, the degree of recrystallization and
the recrystallized grain size in crack growth direction are measured for microstructures
at the magnication of 20.
To estimate the degree of recrystallization or the fraction of recrystallized regions in the
microstructure rstly all of the recrystallized grains in one image are detected performing
the program developed previously. In the binary images of detected grains the whole
recrystallized region is converted into white color using Fiji software. Then another
MATLAB program (see the Appendix) is implemented to scan the whole binary image,
calculating the number of white pixels and the fraction of white pixels to the total number
of the pixels (see gure 4.18). This gives a ne indication of degree of recrystallization.
The procedure was performed on 5 microstructures and the average results are reported
in table 4.3.
Using the reported data for UPE as shown in table 4.2, the relationship between fracture
toughness and degree of recrystallization is plotted as shown in gure 4.19. Standard
deviations for each group of microstructures including ve measurements are illustrated
on the plot by error bars.
80 Calculating fractal dimension of the grain boundaries
Table 4.2: UPE data.
sample Average UPE (KJ/m2) over 3 reported data standard deviation
CoR 90 76 11.26
MoR
I,1
90 130 22.51
MoR
I,2
90 151.67 12.01
MoR
II,1
90 36.67 7.09
MoR
II,2
90 119.33 17.03
CoR 0 23.67 7.23
MoR
I,1
0 60 7
MoR
I,2
0 118 13.74
MoR
II,1
0 21.67 5.03
MoR
II,2
0 65.67 26.15
Figure 4.17: Relationship between the fractal dimension of the grain boundaries and UPE.
4.3 Results and discussion 81
Figure 4.18: The relationship between the fractal dimension of the grain boundaries and
UPE.
(k) An example microstructure of
MoRII,1 0 sample.
(l) Detected recrystallized grains
Figure 4.18: The procedure to calculate the degree of recrystallization of partially recrystal-
lized microstrucures.
82 Calculating fractal dimension of the grain boundaries
Table 4.3: Measurement results of the degree of recrystallization.
sample Degree of recrystallization averaged over 5 samples standard deviation
CoR90 0.412 0.034
MoR
I,1
90 0.340 0.038
MoR
I,2
90 0.387 0.072
MoR
II,1
90 0.260 0.016
MoR
II,2
90 0.254 0.051
CoR0 0.433 0.056
MoR
I,1
0 0.329 0.037
MoR
I,2
0 0.381 0.029
MoR
II,1
0 0.275 0.041
MoR
II,2
0 0.279 0.051
The plot shows that for each type of rolling process the fraction of recrystallized grain is
almost the same for 0

and 90

samples, and the conventional rolled samples show the


highest values for degree of recrystallization of all data. Data points are widely spread
and no clear trend observed. As for the same sample the degree of recrystallization should
be always the same, but clearly 0

and 90

samples have distinctively dierent UPE.


4.3 Results and discussion 83
Figure 4.19: Relationship between degree of recrystallization and UPE.
The second microstructural parameter which might aect the fracture toughness of par-
tially recrystallized Aluminum alloys can be the size of recrystallized grains in crack
growth direction on crack plane. Crack growth direction as indicated earlier in gure
4.5 is L and S for 0

and 90

samples respectively. The practical procedure to measure


the grain size along L and S as demonstrated in gure 4.20 is simply using test lines
aligned in L or S direction and measuring the length of intercepts. This was carried out
on 10 random grains for each sample and the average result is shown in table 4.4. To
investigate the inuence of grain size on fracture toughness, UPE data is plotted versus
grain size as shown in gure 4.21. The observed general trend is that decreasing grain size
in crack growth direction increases the fracture toughness which is reasonable since the
recrystallized region of the microstructure is soft compared to the un-recrystallized area.
Compared to the strong inuence of fractal dimensions of the grain boundaries on UPE,
as shown in gure 4.17, the inuence of grain size appears to be of secondary importance.
84 Calculating fractal dimension of the grain boundaries
Figure 4.20: Schematic procedure of measuring the recrystallized grain size in crack propa-
gation direction.
Table 4.4: Measurement results of average recrystallized grain size in crack growth direction.
sample Average rain size (m) in crack growth direction standard deviation
CoR90 81.1 24.9
MoR
I,1
90 72 17.1
MoR
I,2
90 67.4 27.7
MoR
II,1
90 96 33.2
MoR
II,2
90 74.2 30.1
CoR0 141.7 57.8
MoR
I,1
0 113.1 37.5
MoR
I,2
0 133.7 46.3
MoR
II,1
0 130.2 65.5
MoR
II,2
0 144 72.9
4.3 Results and discussion 85
Figure 4.21: Relationship between grain size and UPE.
4.3.5 Conclusions
The aim of the present chapter was implementing the established quantication method
to measure the irregularity of high angle grain boundaries between the recrystallized and
un-recrystallized regions in the microstructure of heat-treated Aluminum alloys by means
of fractal geometry. The eect of grain boundary conguration on fracture toughness of
the materials was investigated using UPE data of tear test for three types of conventional-
rolled, modied-rolled I ,and modied-rolled II materials.
Shape characterization of high angle grain boundaries of partially recrystallized AA7050
carried out on tear tested specimens shows that there is a strong correlation between UPE
and the fractal dimension of the grain boundaries aligned in crack propagation direction
for both 0

and 90

group of samples. In other words by increasing the fractal dimension


of grain boundaries, the fracture toughness of all studied material increases. While for
both 0

and 90

type of samples the modied type I with higher percentage of modied


rolling posses the highest fractal dimension and the highest fracture toughness value.
In order to compare the inuence of fractal dimension of grain boundaries on the fracture
toughness to that of other microstructural features, two series of quantitative study were
carried out on the Aluminum micrographs to estimate the degree of recrystallization and
the grain size in crack growth direction. It has been shown that there is no clear correlation
86 Calculating fractal dimension of the grain boundaries
between the fraction of recrystallized grains and the UPE. Studying the eect of grain
size in crack propagation direction on the fracture toughness revealed that increasing the
grain size results in decreasing the unit propagation energy, which sounds reasonable since
the crack requires less energy to propagate through the soft recrystallized region of the
microstructure. However comparing the general trends of data for both eects of the grain
size and the fractal dimension on the fracture toughness shows that the measured data
points for grain size are widely spread compared to that of fractal dimension. So it can
be concluded that FD has the most signicant role in determining the fracture toughness
of the studied materials.
Chapter 5
Conclusions
In this work the morphology of high angle grain boundaries for the microstructure of
Aluminum alloy has been automatically processed to quantitatively describe the degree
of irregularity by means of fractal dimensions. The developed image processing program
has been shown to be capable of extracting the grain boundaries of interest in both
fully and partially recrystallized microstructures of Aluminum alloys, while preserving
the details needed for the fractal dimension calculations. The box-counting program
developed based on theoretical considerations has been validated using test images of
Koch curves at dierent iteration levels having known fractal dimensions. The estimated
values of fractal dimension for the sequence of Koch curves has been observed to converge
to the mathematical value of 1.2619, as the number of iterations increases. So it has been
concluded that implementing the present box-counting method to non-mathematically
complex objects such as a serrated grain boundaries is an easy, automatically computable
method to calculate the fractal dimension.
In order to obtain improved mechanical properties special rolling procedures have been
developed by our material supplier. To understand the microstructure/toughness re-
lationship in hot-rolled AA7050, the eect of grain boundary structure on the fracture
toughness was investigated. It has been shown that depending on how rolling is performed
in the modied rolling processes, the shape of the grain boundaries were changed, and dif-
ferent grain boundary structures were generated for the same overall thickness reduction.
A signicant improvement in fracture toughness has been reported for a modied-rolled
materials compared to conventionally-rolled plates.
87
88 Conclusions
Quantitative characterization of high angle grain boundaries of partially recrystallized
AA7050 carried out on Kahn-tear specimens has shown that there is a strong correlation
between UPE data of Kahn-tear test and the fractal dimension of the grain boundaries
aligned in crack propagation direction for both 0

and 90

group of samples. In other


words by increasing the fractal dimension of grain boundaries, the UPE of all studied
material increases. While for both 0

and 90

type of samples the modied type I with


higher percentage of modied rolling posses the highest fractal dimension and the highest
fracture toughness value.
In order to compare the inuence of fractal dimension of grain boundaries on the fracture
toughness to that of other microstructural features, quantitative study were carried out on
the Aluminum micrographs to estimate the degree of recrystallization and the grain size
in crack growth direction. It has been observed that there is no clear correlation between
the fraction of recrystallized grains and the UPE. Studying the eect of grain size in crack
propagation direction on UPE revealed that increasing the grain size decreases the unit
propagation energy. However comparing the general trends of data for both eects of the
grain size and the fractal dimension on the UPE shows that the measured data points for
grain size are widely spread compared to that of fractal dimension. So it can be concluded
that the fractal dimension of high angle grain boundaries has the most signicant role in
determining the UPE of the studied materials.
References
[1] P. Hildgen, F. Nekka, F. Hildgen, and J. McMullen, Macroporosity measurement
by fractal analysis, Physica A: Statistical Mechanics and its Applications, vol. 234,
no. 3-4, pp. 593603, 1997.
[2] Q. Wei and D. Wang, Pore surface fractal dimension of sol-gel-derived al2o3-sio2
membranes, Materials Letters, vol. 57, no. 13-14, pp. 20152020, 2003.
[3] S. Pyun and C. Rhee, An investigation of fractal characteristics of mesoporous
carbon electrodes with various pore structures, Electrochimica Acta, vol. 49, no. 24,
pp. 41714180, 2004.
[4] G. Lee and S. Pyun, The eect of pore structures on fractal characteristics of
meso/macroporous carbons synthesised using silica template [8], Carbon, vol. 43,
no. 8, pp. 18041808, 2005.
[5] C. Atzeni, G. Pia, and U. Sanna, Fractal modelling of medium-high porosity sic
ceramics, Journal of the European Ceramic Society, vol. 28, no. 14, pp. 28092814,
2008.
[6] A. Ahmad and N. Mustafa, Pore surface fractal analysis of palladium-alumina
ceramic membrane using frenkel-halsey-hill (fhh) model, Journal of colloid and
interface science, vol. 301, no. 2, pp. 575584, 2006.
[7] N. Bird, M. Daz, A. Saa, and A. Tarquis, Fractal and multifractal analysis of
pore-scale images of soil, Journal of Hydrology, vol. 322, no. 1-4, pp. 211219,
2006.
89
90 REFERENCES
[8] J. Krim and V. Panella, Characterization of The Surface Fractal Dimension of Evap-
orated Silver and Gold Films Through Adsorption Isotherm Measurements, vol. 62
of Studies in Surface Science and Catalysis. 1991.
[9] Y. Wang and K. Xu, Characterization of surface morphology of copper tungsten
thin lm by surface fractal geometry and resistivity, Thin Solid Films, vol. 468,
no. 1-2, pp. 310315, 2004.
[10] K. Malek, A study of the fractal dimension and percolative structure of lithium-
inserted batio3 lm, Thin Solid Films, vol. 408, no. 1-2, pp. 7378, 2002.
[11] S. Hou, M. Ouyang, and H. Chen, Fractal structure in the silver oxide thin lm,
Thin Solid Films, vol. 315, no. 1-2, pp. 322326, 1998.
[12] J. Yehoda and R. Messier, Are thin lm physical structures fractals?, Applications
of Surface Science, vol. 22-23, no. 2, pp. 590595, 1985.
[13] A. Celli, A. Tucci, L. Esposito, and C. Palmonari, Fractal analysis of cracks in
alumina-zirconia composites, Journal of the European Ceramic Society, vol. 23,
no. 3, pp. 469479, 2003.
[14] J. J. Mecholsky, J. K. West, and D. E. Passoja, Fractal dimension as a charac-
terization of free volume created during fracture in brittle materials, Philosophi-
cal Magazine A: Physics of Condensed Matter, Structure, Defects and Mechanical
Properties, vol. 82, no. 17-18, pp. 31633176, 2002.
[15] J. J. Mecholsky, Estimating theoretical strength of brittle materials using fractal
geometry, Materials Letters, vol. 60, no. 20, pp. 24852488, 2006.
[16] J. J. Mecholsky, Fractography and fractal geometry: What can we learn?, Ce-
ramic Transactions, vol. 199, pp. 5365, 2007.
[17] J. J. Mecholsky, Fractal analysis and fractography: What can we learn thats
new?, Key Engineering Materials, vol. 409, pp. 145153, 2009.
[18] J. J. Mecholskyand, D. E. Passoja, and K. S. Feinberg-Ringel, Quantitative anal-
ysis of brittle fracture surfaces using fractal geometry, Journal of the American
Ceramic Society, vol. 72, no. 1, pp. 6065, 1989.
[19] F. H. R.H. Dauskardt and R. Ritchie, On the interpretation of the fractal character
of fracture surfaces, Acta Metall. Mater., vol. 38, no. 2, p. 143159, 1986.
REFERENCES 91
[20] E. Hornbogen, Fractals in microstructure of metals, International Materials Re-
views, vol. 34, no. 6, pp. 277296, 1989.
[21] M.Tanaka, Y. Kimura, and et al., Fractal analysis of three-dimensional fracture
surfaces in metals and ceramics, ISIJ International, vol. 44, no. 7, pp. 12501257,
2004. Cited By (since 1996): 1.
[22] Q. Chang and D. L. e. a. Chen, Three-dimensional fractal analysis of fracture
surfaces in titanium-iron particulate reinforced hydroxyapatite composites: Rela-
tionship between fracture toughness and fractal dimension, Journal of Materials
Science, vol. 46, no. 18, pp. 61186123, 2011.
[23] V. V. Silberschmidt, Multifractal characteristics of matrix cracking in laminates
under t-fatigue, Computational Materials Science, vol. 13, no. 1-3, pp. 154159,
1998.
[24] L. Wang and L. W. W. Fei, Fractal analysis of fracture surfaces in aluminum borate
whisker-reinforced aluminum alloy 6061 composite, Transactions of Nonferrous
Metals Society of China (English Edition), vol. 21, no. 3, pp. 461466, 2011.
[25] J. J. Mecholsky and J. R. Plaia, Fractal analysis on fracture surfaces of glass
using replication techniques, Journal of Non-Crystalline Solids, vol. 146, no. C,
pp. 249255, 1992.
[26] B. Venkatesh, D. Chen, and S. Bhole, Three-dimensional fractal analysis of fracture
surfaces in a titanium alloy for biomedical applications, Scripta Materialia, vol. 59,
no. 4, pp. 391394, 2008. Cited By (since 1996): 8.
[27] O. Hilders and D. Pilo, On the development of a relation between fractal dimension
and impact toughness, Materials Characterization, vol. 38, no. 3, pp. 121127,
1997.
[28] B. Mandelbrot, D. Passoja, and A. J. Paullay, Fractal character of fracture surfaces
of metals, Nature, vol. 308, no. 5961, pp. 721722, 1984.
[29] E. E. Underwood and K. Banerji, Fractals in fractography, Materials Science and
Engineering, vol. 80, no. 1, pp. 1 14, 1986.
[30] B. Kaye, Applied fractal geometry and powder technology, Chaos, Solitons and
Fractals, vol. 6, no. C, pp. 245253, 1995.
[31] B. Kaye, A Random Walk Through Fractal Dimensions. VCH, 1989.
92 REFERENCES
[32] T. Kirk and G. Stachowiak, Fractal characterization of wear particles from synovial
joints, Journal of Computer-Assisted Microscopy, vol. 3, no. 4, pp. 157170, 1991.
[33] T. Kirk, G. Stachowiak, and A. Batchelor, Fractal parameters and computer image
analysis applied to wear particles isolated by ferrography, Wear, vol. 145, no. 2,
pp. 347365, 1991.
[34] C. Luo, W. Lee, and et al., Measuring the fractal dimension of diesel soot ag-
glomerates by fractional brownian motion processor, Atmospheric Environment,
vol. 39, no. 19, pp. 35653572, 2005.
[35] B. H. Schmid, B. Al-Zaitone and et al., Evolution of the fractal dimension for
simultaneous coagulation and sintering, Chemical Engineering Science, vol. 61,
no. 1, pp. 293305, 2006.
[36] T. Masuda and A. Fujimura, Microstructural development of ne-grained quartz
aggregates by syntectonic recrystallization, Tectonophysics, vol. 72, no. 1-2,
pp. 105128, 1981.
[37] J. Kruhl and M. Nega, The fractal shape of sutured quartz grain boundaries:
Application as a geothermometer, International Journal of Earth Sciences, vol. 85,
no. 1, pp. 3843, 1996.
[38] M. Toriumi, J. Teruya, M. Masui, and H. Kuwahara, Microstructures and ow
mechanisms in regional metamorphic rocks of japan, Contributions to Mineralogy
and Petrology, vol. 94, no. 1, pp. 5462, 1986.
[39] M. Takahashi, H. Nagahama, T. Masuda, and A. Fujimura, Fractal analysis of
experimentally, dynamically recrystallized quartz grains and its possible application
as a strain rate meter, Journal of Structural Geology, vol. 20, no. 2-3, pp. 269275,
1998.
[40] M. Takahashi and H. Nagahama, The sections fractal dimension of grain bound-
ary, Applied Surface Science, vol. 182, no. 3-4, pp. 297301, 2001.
[41] M. Takahashi and H. Nagahama, Fractal dimensions of recrystallized quartz grain
boundaries and grain fabrics, Arabian Journal for Science and Engineering, vol. 28,
no. 1 C, pp. 213221, 2003.
[42] M. Takahashi and H. Nagahama, Fractal properties of dynamic recrysatallized
grain boundaries, Materials Science Forum, vol. 426-432, no. 4, pp. 35633568,
2003.
REFERENCES 93
[43] M. Takahashi and H. Nagahama, Fractal grain boundary migration, Fractals,
vol. 8, no. 2, pp. 189194, 2000.
[44] P. Hahner and M. Zaiser, Dislocation dynamics and work hardening of fractal
dislocation cell structures, Materials Science and Engineering A, vol. 272, no. 2,
pp. 443454, 1999.
[45] M. Zaiser, K. Bay, and P. Hahner, Fractal analysis of deformation-induced dislo-
cation patterns, Acta Materialia, vol. 47, no. 8, pp. 24632472, 1999.
[46] E. Hornbogen, Fractal analysis ofgrain boundaries in hot-worked poly-crystals,
Z. Metallkd, vol. 78, pp. 622625, 1987.
[47] M. Tanaka, Characterization of grain-boundary conguration and fracture surface
roughness by fractal geometry and creep- rupture properties of metallic materials,
Journal of Materials Science, vol. 27, pp. 47174725, 1992.
[48] M. Tanaka and H. Iizuka, Characterization of grain boundaries by fractal geometry
and creep-rupture properties of heat-resistant alloys, Z. Metallkd, vol. 82, pp. 442
447, 1991.
[49] P. Li, Q. Zhang, and K. Xue, Fractal characteristics and prediction of ti-15-3 alloy
recrystallized microstructure, Journal of Materials Science and Technology, vol. 24,
no. 6, pp. 835839, 2008.
[50] E. Konopleva, H. McQueen, and E. Evangelista, Serrated grain boundaries in hot-
worked aluminum alloys at high strains, Materials Characterization, vol. 34, no. 4,
pp. 251264, 1995.
[51] J. Kim, H. Hong, and S. Nam, Investigation on the formation of serrated grain
boundaries with grain boundary characteristics in an aisi 316 stainless steel, Jour-
nal of Nuclear Materials, vol. 393, no. 2, pp. 249253, 2009.
[52] M. Yamazaki, The eect of two-step solution treatment on the creep rupture prop-
erties of a high carbon 18cr-12ni stainless steel, J. Japan Institute of Metals, vol. 30,
pp. 10321036, 1966.
[53] M. Tanaka, O. Miyagawa, T. Sakaki, H. Iizuka, F. Ashihara, and D. Fujishiro,
Creep rupture strength and grain-boundary sliding in austenitic 21 cr-4ni-9mn
steels with serrated grain boundaries, Journal of Materials Science, vol. 23, no. 2,
pp. 621628, 1988.
94 REFERENCES
[54] J. Larson and S. Floreen, Metallurgical factors aecting the crack growth resistance
of a superalloy, Metall. Trans. A, vol. 8A, pp. 5155, 1977.
[55] A. Koul and G. Gessinger, On the mechanism of serrated grain boundary formation
in ni-based superalloys, Acta Metall., vol. 31, p. 10611069, 1983.
[56] M. F. Henry, Y. S. Yoo, D. Y. Yoon, and J. Choi, The dendritic growth of pre-
cipitates and grain boundary serration in a model nickel-base superalloy, Metall.
Trans., vol. 24A, p. 17331743, 1993.
[57] A.-C. Yeh, K.-W. Lu, C.-M. Kuo, H.-Y. Bor, and C.-N. Wei, Eect of serrated
grain boundaries on the creep property of inconel 718 superalloy, Materials Science
and Engineering: A, vol. 530, no. 0, pp. 525 529, 2011.
[58] R. Mitchell, H. Li, and Z. Huang, On the formation of serrated grain bound-
aries and fan type structures in an advanced polycrystalline nickel-base superalloy,
Journal of Materials Processing Technology, vol. 209, no. 2, pp. 1011 1017, 2009.
[59] P. Streitenberger, D. Frster, G. Kolbe, and P. Veit, Temperature induced smooth-
ing of initially fractal grain boundaries, Scripta Materialia, vol. 34, no. 1, pp. 111
119, 1996.
[60] P. Streitenberger, D. Frster, and P. Veit, Thermal relaxation of initially fractal
grain boundaries, Fractals, vol. 5, no. SUPPL. 2, pp. 521, 1997.
[61] K. Hisatsune, Y. Takuma, Y. Tanaka, K. Udoh, and K. Kawasaki, Fractal dimen-
sion of grain boundary in cuau alloys rened by platinum addition, Journal of
Materials Science, vol. 33, no. 19, pp. 47834785, 1998.
[62] J. Pugh and W. Lasch, Image analysis for grain shape characterization in lamp
laments, Metall. Trans. A, vol. 21A, p. 22092214, 1990.
[63] K. Tanoue, Roles of potassium bubbles in high temperature creep resistance of
doped tungsten ne wires, Materials Science Forum, vol. 426-432, no. 5, pp. 4277
4282, 2003.
[64] M. Tanaka, A. Kayama, R. Kato, and Y. Ito, Estimation of the fractal dimension of
fracture surface patterns by box-counting method, Fractals, vol. 7, no. 3, pp. 335
340, 1999.
[65] M. Tanaka, A. Kayama, R. Kato, and Y. Ito, Eects of the creep deformation
on the fractal dimension of the grain boundaries in an austenite steel, Journal of
Materials Science, vol. 33, no. 13, pp. 33513359, 1998.
REFERENCES 95
[66] M. Tanaka, Kayama, Y. Sato, and Y. Ito, Fractal nature of creep-fracture surface
patterns in pure zn polycrystals, Journal of Materials Science Letters, vol. 17,
no. 20, pp. 17151717, 1998.
[67] M. Tanaka, A. Kayama, R. Kato, and Y. Ito, Change in the fractal dimension of
the grain boundaries in pure zn polycrystals during creep, Journal of Materials
Science, vol. 33, no. 24, pp. 57475757, 1998.
[68] P. Streitenberger and D. Frster, The eect of surface fractal dimension of particles
on ostwald ripening, phys. stat.sol, vol. 164, p. K65, 1991.
[69] G. Laird, J. Rawers, and A. Adams, Fractal analysis of caride morphology in high-
cr white cast irons, Metallurgical Transactions A, vol. 23, no. 10, pp. 29412945,
1992.
[70] M. Tanaka, Fractal nature and quantitative evaluation of microstructures in metal-
lic materials, Journal of Materials Science, vol. 30, no. 14, pp. 36683673, 1995.
[71] S. Lu and A. Hellawell, An application of fractal geometry to complex microstruc-
tures: Numerical characterization of graphite in cast irons, Acta Metallurgica Et
Materialia, vol. 42, no. 12, pp. 40354047, 1994.
[72] Z. Huda and B. Ralph, Kinetics of grain growth in powder-formed in-792: A nickel-
base superalloy, Materials Characterization, vol. 25, no. 2, pp. 211220, 1990.
[73] T. Nishihara, Fractal analysis of recrystallized structure, Contributions to Min-
eralogy and Petrology, vol. 57, p. 209, 1993.
[74] J. Russ, The Image Processing Handbook-third ed. CRC Press and IEEE Press,
1998.
[75] G. V. Voort, ed., ASM Handbook Vol. 9. Metallography and Microstructures. ASM
International.
[76] K. K. B. Ralph, The Quantitative Description of The Microstructure of Materials.
CRC Press, 1995.
[77] J. Friel, J. Grande, D. Hetzner, K. Kurzydlowski, and et al., Practical Guide to
Image Analysis. ASM International, 2000.
[78] L. Wojnar, Image Analysis, Applications in Materials Engineering. CRC Press,
1999.
96 REFERENCES
[79] R. Gonzalez and R. Woods, Digital Image Processing-Second Edition. Prentice-Hall,
2001.
[80] B. Mandelbrot, The Fractal Geometry of Nature. Freeman, 1983.
[81] B. Kaye, Applications of recent advances in ne-particle characterization to mineral
processing, Part. Charact., vol. 2, pp. 9197, 1985.
[82] M. Barnsley, R. Devaney, B. B. Mandelbrot, and et al., The Science of Fractal
Images. Springer-Verlag, 1988.
[83] H. Takayasu, Fractals in the Physical Sciences. Manchester Univ Pr, 1990.
[84] D. E. Roach and A. D. Fowler, Dimensionality analysis of patterns: fractal mea-
surements, Part. Charact., vol. 19, no. 6, pp. 849869, 1985.
[85] B. Kaye, Modern Methods in Fine Particle Characterization. CRC Press, 1983.
[86] X. Wang, L. Dong, and L. Xiong, The change of fractal dimensionality in the recov-
ery and recrystallisation process, J. Phys.: Condens. Matter, vol. 2, p. 38793884,
1990.
[87] S. Lovejoy, Area-perimeter relation for rain and cloud areas, Science, vol. 216,
no. 4542, pp. 185187, 1982.
[88] R. Colas, On the variation of grain size and fractal dimension in an austenitic
stainless steel, Materials Characterization, vol. 46, no. 5, pp. 353358, 2001.
[89] I. The MathWorks, Matlab help, image processing toolbox.
http://www.mathworks.nl/help/toolbox/images/.html, 2011. Visited on:
June 2011.
[90] Fundamentals of Digital Image Processing A Practical Approach with Examples in
Matlab. John Wiley and Sons, Ltd, 2011.
[91] Digital Image Processing Using Matlab. 2nd ed. Gatesmark Publishing, 2009.
[92] J. Canny, A computational approach to edge detection., vol. 8, no. 6, p. 679698,
1998.
[93] B. Mandelbrot, Fractal geometry: What is it, and what does it do? Princeton
University Press, 1989.
REFERENCES 97
[94] H. Peitgen, H. Jurgens, and D. Saupe, Chaos and Fractals New Frontiers of Science.
Springer, 1992.
[95] K. Falconer, Fractal geometry, Mathematical Foundations and Applications. Wiley,
1990.
[96] G. Elert, The chaos hypertextbook. about dimension.
http://hypertextbook.com/chaos/33.shtml, 2011. Visited on: September
2011.
[97] J. Briot, Source code for koch curve generation.
[98] Q. Huang, J. Lorch, and R. Dubes, Can the fractal dimension of images be mea-
sured?, Pattern Recognition, vol. 27, no. 3, pp. 339349, 1994.
[99] A. Heinz, A. Haszler, C. Keidel, S. Moldenhauer, R. Benedictus, and W. Miller,
Recent development in aluminium alloys for aerospace applications, Materials
Science and Engineering: A, vol. 280, no. 1, pp. 102107, 2000.
[100] E. S. Jr. and J. Staley, Application of modern aluminum alloys to aircraft,
Progress in Aerospace Sciences, vol. 32, no. 2-3, pp. 131172, 1996.
[101] Z. Cvijovi, M. Vratnica, and M. Rakin, Micromechanical modelling of fracture
toughness in overaged 7000 alloy forgings, Materials Science and Engineering A,
vol. 478, no. 1-2, pp. 163172, 2008.
[102] Z. Cvijovi, M. Rakin, and M. Vratnica, Microstructural dependence of fracture
toughness in high-strength 7000 forging alloys, Engineering Fracture Mechanics,
vol. 75, no. 8, pp. 21152129, 2008.
[103] G. Garrett and J. Knott, The inuence of compositional and microstructural varia-
tions on the mechanism of static fracture in aluminum alloys, Metallurgical Trans-
actions A, vol. 9, no. 9, pp. 11871201, 1978.
[104] D. Dumont, A. Deschamps, and Y. Brechet, A model for predicting fracture mode
and toughness in 7000 series aluminium alloys, Acta Materialia, vol. 52, no. 9,
pp. 25292540, 2004.
[105] T. Pardoen, D. Dumont, A. Deschamps, and Y. Brechet, Grain boundary ver-
sus transgranular ductile failure, Journal of the Mechanics and Physics of Solids,
vol. 51, no. 4, pp. 637665, 2003.
98 REFERENCES
[106] N. Ryum, The inuence of a precipitate-free zone on the mechanical properties of
an al-mg-zn alloy, Acta Metallurgica, vol. 16, no. 3, pp. 327332, 1968.
[107] N. Deshpande, A. Gokhale, D. Denzer, and J. Liu, Relationship between frac-
ture toughness, fracture path, and microstructure of 7050 aluminum alloy: Part i.
quantitative characterization, Metallurgical and Materials Transactions A: Physi-
cal Metallurgy and Materials Science, vol. 29, no. 4, pp. 11911201, 1998.
[108] D. Dumont, A. Deschamps, and Y. Brechet, On the relationship between mi-
crostructure, strength and toughness in aa7050 aluminum alloy, Materials Science
and Engineering A, vol. 356, no. 1-2, pp. 326336, 2003.
[109] N. Kamp, I. Sinclair, and M. Starink, Toughness-strength relations in the over-
aged 7449 al-based alloy, Metallurgical and Materials Transactions A: Physical
Metallurgy and Materials Science, vol. 33, no. 4, pp. 11251136, 2002.
[110] J. Kaufman and A. Knoll, Kahn-type tear tests and crack toughness of aluminium
alloy sheets, Materials Research and Standards, vol. 4, pp. 151155, 1964.
[111] H. Dudgeon, N. Parson, S. Court, and R. Ricks Proceedings of the 4th Int.Conf.on
Aluminum Alloys, vol. 1, p. 443, 1994.
[112] T. Kobayashi, M. Niinomi, and K. Ikedai, Evaluation of fracture toughness of alu-
minum alloys by tear test., Keikinzoku/Journal of Japan Institute of Light Metals,
vol. 38, no. 1, pp. 915, 1988.
[113] T. Kobayashi, M. Niinomi, Y. Takabayashi, and S. Kohmura, Finite element anal-
ysis aided fracture toughness evaluation in tear test of aluminum alloys, Keikin-
zoku/Journal of Japan Institute of Light Metals, vol. 38, no. 11, pp. 723730, 1988.
[114] Fracture toughness testing. http://aluminium.matter.org.uk, 2011. Visited
on: October 2011.
[115] Fiji open sourse software. http://rsb.info.nih.gov/ij/index.html, 2011.
Visited on: September 2011.
[116] M. Tanaka, H. Iizuka, and F. Ashihara, Eects of serrated grain boundaries on
the crack growth in austenitic heat-resisting steels during high-temperature creep,
Journal of Materials Science, vol. 23, no. 11, pp. 38273832, 1988.
Appendix A
Matlab codes
A.1 MATLAB Code for Generating Koch snowake curves
function [x,y] = Kochcurve(sx,sy,N)
if nargin == 0
N = 4;
end
X = [0;1/3;1/2;2/3;1];
Y = [0;0;sqrt(3)/6;0;0];
x{1} = sx;
y{1} = sy;
for n=1:N
for k =1:numel(x{n})-1
th = atan2((y{n}(k+1)-y{n}(k)),(x{n}(k+1)-x{n}(k)));
99
100 Matlab codes
d = sqrt((y{n}(k+1)-y{n}(k))^2+(x{n}(k+1)-x{n}(k))^2);
temp = [d*X(:) d*Y(:) ones(numel(X),1)] * ...
[cos(th) sin(th) 0 ; -sin(th) cos(th) 0 ; x{n}(k) y{n}(k) 1];
x{n+1}((k-1)*5+1:k*5) = temp(:,1);
y{n+1}((k-1)*5+1:k*5) = temp(:,2);
end
x{n+1}(6:5:end) = [];
y{n+1}(6:5:end) = [];
end
% sx et sy variables are vectors containing initial polygon coordinates.
% For the Koch curve :
sx = [0;1]
sy = [0;0];
% For Koch snowflake :
sx{2} = [0;.5;1;0];
sy{2} = [0;sqrt(3)/2;0;0];
% Draw function file
function p = drawKochcurve(x,y)
figure(color,w)
p = plot(nan,nan);
axis equal off
for k = 1:numel(x)
set(p,xdata,x{k} ,ydata,y{k})
pause(1)
end
% Demo file
A.2 MATLAB Code for image processing followed by box-counting 101
function demoKochcurve(n)
if nargin == 0
n = 1;
end
% Koch curve
sx{1} = [0;1];
sy{1} = [0;0];
% Koch snowflake
sx{2} = [0;.5;1;0];
sy{2} = [0;sqrt(3)/2;0;0];
N = 4;
[x,y] = Kochcurve(sx{n},sy{n},N);
drawKochcurve(x,y);
A.2 MATLAB Code for image processing followed by box-
counting
%function [s1 s2 boxnum res] = Main(RawImage)
clear all
close all
clc
warning off
%% reading the raw images
if ~exist(RawImage,var)
RawImage = input(Image name: ,s);
end
102 Matlab codes
cd ../RawImages
if ~exist(RawImage,file)
disp(Image does not exist);
else
I_raw = imread(RawImage);
end
cd ../Code
%% Analysisng
figure(1)
imshow(I_raw);
pic_check = n;
while strcmpi(n,pic_check)
ts_check = n;
while strcmpi(n,ts_check)
ts = input(Threshould: );
I = im2bw(I_raw,ts);
figure(1)
imshow(I);
ts_check = input(Is the image OK [y,n]? ,s);
end
A.2 MATLAB Code for image processing followed by box-counting 103
I = imfill(I,holes);
I=bwmorph(I,thin,1);
se=strel(disk,2);
I=imopen(I,se);
I = bwlabel(I);
imtool(I);
pic_check = input(Is the image OK [y,n]? ,s);
end
num_check = y;
pixel_count = 1;
while strcmpi(num_check,y)
t = input(pixel value: );
II = Detect(I,t);
se=strel(disk,7);
II=imclose(II,se);
IB=edge(II,canny);
IB=bwareaopen(IB,100);
figure(2);
imshow(IB);
104 Matlab codes
P = input(Do you want to save this picture? ,s);
if strcmpi(P,y)
cd ../ReducedImages
print(-r400,-deps,[RawImage _400.eps]);
print(-r600,-deps,[RawImage _600.eps]);
cd ../Code
end
BC = input(Start box counting [y,n]? ,s);
%%
if strcmpi(BC,y)
close all
[s1(pixel_count,:) s2(pixel_count,:) boxnum(pixel_count,:) res(pixel_count,:)] =
pixel_count = pixel_count+1;
end
num_check = input(Do you want to enter a new pixel value? ,s);
end
A.3 Detect code
function A_new = Detect(A,x)
[a b]=size(A);
A_new = zeros(a,b);
for i=1:a
for j=1:b
if A(i,j)==x
A_new(i,j)=x;
A.4 MATLAB Code for box-counting calculation 105
end
end
end
A.4 MATLAB Code for box-counting calculation
function [s1 s2 boxnum res] = BoxCount(I)
maxDim = max(size(I));
n = ceil(log2(maxDim));
newDim = 2^n;
newI = zeros(newDim);
for i = 1:size(I,1)
for j = 1:size(I,2)
newI(i,j) = I(i,j);
end
end
boxnum = zeros(1,n+1);
res = zeros(1,n+1);
for dim = 1:n+1
pix = 2^(dim-1);
res(1,dim) = pix;
for i = 1:newDim/pix
for j=1:newDim/pix
Icount = newI(i*pix-pix+1:i*pix,j*pix-pix+1:j*pix);
if sum(Icount(:))>=1
boxnum(1,dim) = boxnum(1,dim)+1;
end
clear Icount
106 Matlab codes
end
end
end
P = polyfit(log(res),log(boxnum),1);
D = polyval(P,log(res));
s1=-gradient(D)./gradient(log(res));
s2=-gradient(log(boxnum))./gradient(log(res));
figure(1)
hold on
plot(log(res),log(boxnum),or);
plot(log(res),D,b-);
xlabel(log(res))
ylabel(log(boxnum))
A.5 MATLAB Code for calculating degree of recrystalliza-
tion
function [ratio] = Ratio(A,x)
[a b]=size(A);
count = 0;
for i=1:a
for j=1:b
if A(i,j)==x
count = count +1;
end
end
end
A.5 MATLAB Code for calculating degree of recrystallization 107
total = a*b;
ratio = count/total;

You might also like