Computers & Geosciences 25 (1999) 1059±1071

High precision boundary fractal analysis for shape

characterization
p

Dominique BeÂrube a,*, Michel JeÂbrak b

a
Environmental Sciences, University of Quebec in Montreal, C.P. 8888, Succursale Centreville, Montreal, Que., Canada H3C 3P8
b
Department of Earth Sciences, University of Quebec in Montreal, C.P. 8888, Succursale Centreville, Montreal, Que., Canada H3C
3P8

Received 29 October 1998; received in revised form 28 January 1999; accepted 5 February 1999

Abstract

Numerous kinds of particles in geological and environmental sciences may be characterized by their boundary
fractal dimension. Several methods are available: structured walk, box-counting, dilation and euclidean distance
mapping (EDM). The precision and stability of these techniques is variable and usually low precision fractal
dimensions are obtained (20.1). Validation on mathematical fractals and tests of the e€ects of pixelization, size,
resolution and topology were performed with three computer-derived methods (box-counting, dilation and EDM),
using mathematical objects and fragments coming from impact and ore deposits breccias. Tests demonstrate that
high precision results can be yielded with the right technique and caution. EDM showed the highest precision
(20.01) and strongest reliability with less sensitivity to size and resolution, with reproducible results for fragments as
small as 10,000 pixels of area. It was also the most accurate for mathematical fractals. # 1999 Elsevier Science Ltd.
All rights reserved.

Keywords: Methodology; Fractal; Morphology; Outline; Breccia

1. Introduction and surfaces (Kaye, 1978; Allen et al., 1995; Russ,

1995). Fractal geometry is now applied to pore distri-
Fractal geometry represents a major advance over bution (Crawford and Matsui, 1996), networks
previous methods for quantifying and simulating the (Gillepsie et al., 1993), aggregation (Neimark et al.,
complex patterns encountered in nature that were pre- 1996), size distribution (Logdson, 1995) and many
viously described only qualitatively (Roach and other ®elds. The rapid growth of this research area has
Fowler, 1993). Fractals have been used for more than opened the way for developments in properties and
two decades in geology for the description of outline processes modeling such as permeability (Rawls and
Brakensiek, 1995) and brecciation (BeÂrube and JeÂbrak,
1996).
Boundary fractal analysis (BFA) is used to describe
p
Code available at http://www.iamg.org/CGEditor/ the ruggedness or complexity of a particle. Richardson
index.htm (1961) ®rst introduced the idea that an irregular out-
* Corresponding author. Tel.: +1-987-3000-3986; Fax: +1- line can be characterized by a dimension related to its
514-987-7749. ruggedness. Mandelbrot (1967) linked this dimension
E-mail addresses: [email protected] (D. to the fractal concept.
BeÂrubeÂ), [email protected] (M. JeÂbrak)

0098-3004/99/$ - see front matter # 1999 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 3 0 0 4 ( 9 9 ) 0 0 0 6 7 - 9
1060 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071

Fig. 1. Illustration of boundary fractal analysis techniques applied on same feature and derived Richardson plot for (A) Structured
walk (B) Box-counting (C) Dilation and (d) EDM.
 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071 1061

There still exist major uncertainties in shape charac- reported include the generation of insucient data
terization, due to the lack of precision in the de®nition points to calculate the fractal dimension, the e€ects of
of the quantitative morphology descriptors (circularity, very small and large step measurements, the e€ect of
shape factor or ruggedness). As a shape parameter, size and resolution of images or the identi®cation of
boundary fractal dimensions (BFD) o€er (1) a visually multiple fractal dimensions for a single outline.
acceptable model of total line characterization, (2) a Questions on the real scaling of measured features
good description and generation of natural outline and were also asked (Dutch, 1993; Gillepsie et al., 1993).
ultimately (3) an enhanced understanding of the var- These problems have lowered the reliability of the
ious processes which unite to produce a ®nal outline BFA methods and the best method for providing re-
(Longley and Batty, 1989). BFD have therefore been liable fractal dimensions is still a matter of discussion.
considered a useful addition to shape factors, improv- This study will analyze the validity of the di€erent
ing the quantitative description of morphology, es- matrix-based BFA methods ideally suited for compu-
pecially for very complex objects (Orford and Whalley, terization like box-counting, dilation and euclidean dis-
1987). tance mapping (EDM). EDM was proposed as a fast
Complex outlines can be de®ned in 2D using a frac- and e€ective algorithm to derive fractal dimensions,
tal dimension ranging from 1 to 2, that is related to with high precision and low variability. Until now, the
the outline perimeter. Depending upon the scale of ob- EDM method has rarely been used in BFA (Russ and
servation, the measured perimeter length will vary. On Russ, 1989; Adler and Hancock, 1994) and box-count-
a log±log plot, a linear relationship exists between the ing or structured walk have been generally preferred.
perimeter of an object and the scale of measurement.
This relationship represents the main property of frac-
tals, called scaling (or physical self-similarity) which is
2. Methods for boundary fractal analysis
the property of a ®gure to look the same at all scales
(Mandelbrot, 1967). In other words, the amount of
Vector-based methods such as the structured walk
ruggedness would appear the same at all scales and
and matrix-based such as box-counting or Minkowski
therefore, if no reference scale were present, looking at
logic are equivalent in the continuous domain. When
an image of a fractal object would give no indication
applied to digitized data, they lead to di€erent results
of its magni®cation. In reality, this principle is not
(Dubuc et al., 1989). Methodologies and di€erences of
always applicable to natural outlines for which con-
the most widely used are presented here.
cepts of statistical self-similarity or self-anity are bet-
ter suited (Kaye, 1993). Self-anity is de®ned by
2.1. Structured walk method (Richardson) and the
anisotropic scaling, where axes are scaled di€erently to
Hausdorf dimension
preserve shape (LeMeÂhauteÂ, 1990), and scaling is
usually limited to two or three orders of magnitude.
The structured walk, as named by Kaye (1990) (also
Despite studies which addressed the validation of
called caliper, divider, equipaced random walk or yard-
BFA methods and related problems (Longley and
stick), is a vector-based method based on the initial
Batty, 1989; Carr and Benzer, 1991; Roach and
study of Richardson (1961). It consists of walking
Fowler, 1993; Baumann et al., 1994; Klinkenberg,
around the perimeter of an object with a pair of com-
1994; Allen, 1995), there still exists some confusion on
passes (or divider) of a ®nite stride length. The per-
the use and utility of boundary fractal dimensions
imeter of the outline is estimated from the number of
(BFD). Di€erent BFA methods have been developed:
steps needed to span the outline multiplied by the
physical measurements such as adsorption rate, turbid-
stride length (l ). Changing the stride length, which is
ity or settling velocity, stochastic analysis of the noise
equivalent to changing the scale of observation, will
of an image (e.g. Brownian noise) and analytical
produce another estimate of the perimeter. Plotting the
measurements applied to binary image features.
log of the perimeter against the log of the step length
Confusion often arises when the object of analysis
(Richardson plot) yields a linear relationship from
(outline, surface, porosity, volume, distribution) is not
which the fractal dimension (Fd) is derived by the re-
speci®ed, when fractal dimensions derived from di€er-
lationship:
ent methods are compared, or when methods of analy-
sis are inadequately described. Only analytical Fd ˆ 1 ÿ s
measurements are considered in this study.
BFA techniques have limitations that should be where s is the slope of the plot and 1 represents the
respected to ensure valid results. The accuracy, re- topological dimension of a line (Fig. 1(a)).
liability and the associated error margin of the various Several implementations of the structured walk
methods have been questioned (Dubuc et al., 1989; method have been developed (Schwarz and Exner,
Klinkenberg, 1994; Allen et al., 1995). Problems 1980; Batty and Longley, 1986; Clark, 1986; Kennedy
1062 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071

and Lin, 1986; Orford and Whalley, 1987; Hayward et developed (Liebovitch and Tibor, 1989; Block et al.,
al., 1989). Use of this method can result in varying 1990; Buczkowski et al., 1998; Gonzato, 1998).
degrees of scatter, depending on the algorithm chosen. Box-counting must be performed on a one pixel out-
Many of these artifacts have already been reviewed in line that can be generated by image processing algor-
the literature (Kaye, 1990; Klinkenberg, 1994; Allen et ithm. The diculty of the generation of the outline
al., 1995): varies with the chosen algorithm, depending upon the
desired precision and speed and is dependent on the
1. The structured walk does not lend itself well to
original image. Box-counting is then applied to the
computerization because of the discretization of the
entire image containing only one feature. In this study,
data coming from the digitization of images.
box-counting was performed with di€erent box sizes,
Walking the divider around the outline produces
from 2 to 100 pixels with a geometric progression of
polygons of di€erent side lengths, most of which
box sizes. This is necessary to make possible a linear
will have intercepts between pixels. This results in
regression if the data points are to be weighted equally
data scatter on the Richardson plot, particularly for
(Klinkenberg, 1994). All results involving less than 15
small-scale strides. At large step measurements, the
boxes (large step measurements) were eliminated to
object is too coarse and the data is not representa-
reduce bias. Computing time is relatively long com-
tive of the feature because of truncation.
pared to the other methods (approximately 2,5 times
2. Inbound and outbound swings result in di€erent
as much as the dilation and 7 times as much as the
intercepts with the outline.
EDM method, with the available algorithms, on NIH-
3. Perimeter estimates vary with the starting point of
Image software by the National Institute of Health1).
the walk.
4. Structured walk is orientation sensitive. Some problems have been reported, such as low pre-
cision and less sensitivity to ®ne details that can result
Structured walk is still used because interpolation, in an underestimation of the fractal dimension (Roach
multiple starting points and averaging are now possible and Fowler, 1993; Baumann et al., 1994). The rela-
with powerful computers. More recent studies tively narrow range of possible box sizes on a digitized
(Hamblin and Stachowiak, 1993) tested the method image of a particle outline gives a low statistical pre-
against curves of known fractality with results closer cision and a least squares ®t of the data is dicult to
to theoretical values. Still, this method is not utilized obtain. The arbitrary placing of the grid on the image
here since fundamental problems remain and its com- will result in biases, which can be reduced by averaging
puter implementation is far more dicult than for over multiple starting points for the grid. Box-counting
other methods presented here. is also highly dependent upon image resolution
(Baumann et al., 1994).
2.2. Box-counting method and the Kolmogorov
dimension 2.3. `Sausage' method and the Minkowski dimension

One of the ®rst methods to be developed speci®cally Minkowski logic follows the same principle as the
for computerized fractal analysis is the box-counting box-counting method in that it coarsens the initial
method. Averaging adjoining pixels to form a new image but in a more isotropic way. Circles of ®nite di-
image with coarser base elements progressively coar- ameters are drawn around each point of the outline to
sens the representation of the object. The fractal form a ribbon, usually called covering or `sausage'
dimension derived from this method is called the (Dubuc et al., 1989; Russ, 1995). The Minkowski
Kolmogorov dimension and is close to the Hausdor€ dimension derived from this method is not a true
dimension for ideally isotropic conditions (Dubuc et Hausdorf dimension, in which circles with touching
al., 1989). The easiest method to compute the dimen- edges (like the divider method) are used to cover the
sion consists of placing a grid of ®nite box sizes on an outline, and is usually of equal or greater value. (Russ,
outline, and counting the intersects between the bound- 1995).
ary with the grid (Fig. 1(B)). Varying the box sizes in The sausage method was mainly developed to ease
the grid produces a di€erent number of intersections. the computerization of fractal analysis, thus avoiding
Box-counting has been widely used, mainly due to its the diculties encountered with the structured walk
versatility in the type of measurement (boundary, sur- method. Two implementations are presented here, the
face or lacunarity) and its ease of implementation. dilation and the euclidean distance mapping methods.
Faster and more reliable algorithms have since been
2.3.1. Dilation method
Flook (1978) introduced dilation (or erosion-di-
1
NIH-Image software download site. http://rsb.info.nih.- lation) as a computerized method to calculate the
gov/nih-image/download.html Minkowski dimension. Dilation is an image-processing
 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071 1063

Fig. 2. Fragments from Sudbury breccias (1±20), Epidote Abitibi breccia (21±24), Cirotan breccia (25±31) and carbonblack particle
(32) (from Kaye, 1990).

algorithm that adds a background pixel for every pixel volution of kernels of di€erent diameters with the out-
in contact with the feature and erosion retrieves a fea- line. The resultant area for each kernel is divided by its
ture pixel for every pixel that touches the background. diameter. The log of that result is then plotted against
Successive erosion and dilation operations on a one- the log of the kernel diameter and the fractal dimen-
pixel width outline produce a ribbon of a ®nite width sion (Fd) is related to the slope of the plot(s) through
(Fig. 1(C)). Dilation can also be obtained by the con- the relationship:
1064 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071

Fd ˆ 1 ÿ s, of 2 to limit the approximation of diagonal distances

(square root of 2). This has the disadvantage that the
as per the Richardson plot.
size of the measurable features is considerably lowered
The dilation process is straightforward and requires
unless the software can manage more than 256 levels
the simple application of ®lters available in image pro-
of gray. Otherwise, the euclidean distance map is lim-
cessing software. Its main restriction comes from the
ited to 128 levels of gray and the data points to 128.
fact that erosion and dilation are not isotropic oper-
The chosen algorithm for generating the map is of
ations on a square grid (pixels). After a few steps, the
major importance. The intermediate results of the pro-
original image will show a directional bias and the de-
cessing must be stored with enough precision (real
rived dimension will be closer to the Kolmogorov
mode) to maintain isotropy. The algorithm requires
dimension (box-counting method) than to the
the computation of two partial maps for the feature
Minkowski dimension (Dubuc et al., 1989). The bias
and the background (an inverted image of the original
will tend to increase with more erosion-dilation oper-
feature), which are then added to produce the ®nal
ations (Fig. 1(C)). As with structured walk or box-
euclidean distance map.
counting methods, the small measuring elements are
As for all other techniques, log plotting of the data
biased, which leads to deviations in the slope of the
results in a signi®cant increase of data points in the
plot.
region of large steps (fewer points in the region of slen-
The process used in this study allowed 100 appli-
der sausages than in the region of the wider sausages,
cations of dilation and erosion ®lters, creating a maxi-
cf. Fig.1(D)) and biases the linear regression.
mum ribbon width of 200 pixels. The number of
Measuring at an exponentially increasing gray number
erosion and dilation steps can be increased for bigger
of levels prevents this problem.
fragments although biases due to the nonisotropic
More than one feature can be present on the image,
treatment will increase. The dilation procedure must be
providing there is enough space between them (at least
applied to one feature at a time to compute its dimen-
the width of the feature) for the future purpose of
sion.
analysis. The EDM method allows the measurement of
the fractal dimension of the entire image or individual
2.3.2. Euclidean distance mapping (EDM) method
features.
The euclidean distance mapping (EDM) method is
an image processing algorithm applied to a black and
white image, that results in a gray-scale image where
each pixel is given a brightness value proportional to 3. Methodology of boundary fractal analysis
its distance to the nearest pixel of the outline of a fea-
ture. The feature then has a dark backbone that Tests of the various methods were limited to box-
whitens up to the outline and gets darker again toward counting, dilation and EDM. BFA was performed on
the background (Fig. 1(D)). The fast computing algor- euclidean and fractalized ®gures (circle, square and
ithm proposed by Danielsson (1980) consists of calcu- stick), on mathematical curves (Koch triadic and quad-
lating the distance to the nearest border. Thresholding ratic islands) and on breccia fragments of various ori-
gins (Sudbury breccias, Ont., Canada; Cirotan Silver
the image at varying levels of gray produce an isotro-
pic (i.e. of equal width for every point of the outline) Gold Mine breccia, West Java, Indonesia; and Epidote
ribbon around the outline. The fractal dimension can breccia, Cartwright Mountains, Abitibi, Que.,
Canada). Clasts were chosen for their structural and
be computed by plotting the log of the perimeter (cal-
culated by dividing the area by the estimated sausage textural aspects as to cover the widest possible range
width) against the log of the width of the covering cir- of morphologies (Fig. 2). The Cirotan and Epidote
cles (approximated by the number of selected gray clasts were photographed at high resolution and the
levels). However, it is much more practical to plot the negatives scanned on a Polaroid Sprintscan with a
log of the area of the sausage against the number of 1200 dpi resolution. Clasts from the Sudbury breccias
selected gray levels. The fractal dimension (Fd) is de- were taken from Muir and Peredery (1984) and
scanned on an Epson Scantastik scanner (ES-1200C)
rived from the slope of the plot(s) using the relation-
at 300 dpi resolution on a Macintosh 7500/100, with
ship:
the Adobe Photoshop v 3.0 program. Resulting frag-
Fd ˆ 2 ÿ s ments have a total area of more than 50,000 pixels.
Euclidean and fractalized ®gures were drawn with the
where 2 is the embedding space of the outline (since a Macintosh Canvas software (v. 3.5.2). Image analysis
surface is being measured instead of a line). was performed on a Macintosh PPC8100/100 computer
Russ (1995) proposed a modi®cation to this pro- using the public domain NIH-Image program devel-
cedure that reduces the bias due to anisotropy in pixe- oped at the US National Institute of Health and avail-
lated images. The image can be multiplied by a factor able at their internet web site1. Box-counting was
 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071 1065

Fig. 3. Validation of box-counting, dilation and EDM methods on triadic and quadratic Koch islands. Closer calculated fractal
dimensions are computed with EDM methods.

achieved with a NIH-Image macro (Smith et al., 1996). ical values of 1.262 and 1.500 respectively). The pro-
Dilation was also performed using a NIH-Image cess for generating these ®gures is well known and the
macro. Euclidean distance maps were created with the theoretical background can be readily found in the lit-
latest released version of NIH-Image (v. 1.61). A new erature (Kaye, 1990). A wide range of measured values
EDM macro was developed (DB) to automate multiple for these mathematical fractals have been reported
image-processing steps, including (1) production of (Kaye, 1990; Adler and Hancock, 1994) with as much
euclidean distance maps, (2) tresholding at various as 10 to 20% error on the results (Dubuc et al., 1989).
levels of gray, (3) measuring the resulting sausage area For the triadic curve, the results are fairly uniform
for each threshold and (4) exporting data for proces- but the dilation and box-counting estimates are lower
sing. Listing of the macro can be found in Appendix 1 than the EDM results (Fig. 3). For the more complex
or the code can be downloaded from the user-contribu- curve, the quadratic Koch, the analytical limitations
ted macro directory at the NIH Internet site, under the are clearly shown. Only the EDM approaches the
name BFD.hqx. theoretical value of 1.5 with a value of 1.492
(20.53%). Dilation and box-counting are limited to
1.452 (23.13%) and 1.447 (23.53%) respectively. The
4. Results two curves were scanned at high resolution to increase
the number of data points. Small steps measurements
In this section, box-counting, dilation and EDM are more biased for such highly complex curves.
methods are compared using several criteria that can Linear regressions on the Richardson plots re¯ect such
cause variations in fractal dimensions: precision, pixeli- problems and ®rst data points are ignored (Fig. 4).
zation, size and resolution of feature and geometry.
4.2. E€ect of pixelization of objects
4.1. Testing of the methods against mathematical
fractals BFA is a€ected by the discretization of images. As
previously discussed, pixelization in the structured
Box-counting, dilation and EDM were tested on two walk method causes the strides to be of unequal length
complex structures: the triadic and quadratic Koch unless interpolation between pixels is performed, which
curves (Figs. 3 and 4) of known fractality (mathemat- considerably lengthens the computing time. All
1066 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071

Table 1
Variation of fractal dimension with size of feature (circle and
carbonblack)

Form Size (pixels)EDMDilationBox-counting

Circle 32 1.099 0.960 1.709

137 1.096 1.016 1.000
690 1.025 1.018 1.105
2644 1.029 1.018 1.083
3794 1.020 1.013 1.033
5734 1.018 1.008 1.076
7705 1.017 1.006 1.004
10,302 1.012 1.004 1.044
15,856 1.012 1.003 1.034
23,092 1.012 1.002 1.063
31,280 1.012 1.001 1.011
40,860 1.011 1.001 1.032
51,889 1.011 1.001 0.999
63,560 1.010 1.001 1.033
Variation range 0.008 0.007 0.077
Fragment `The Poodle' 1027 1.374 1.352 1.306
4176 1.348 1.363 1.365
9343 1.342 1.360 1.278
16,662 1.331 1.356 1.331
37,504 1.335 1.349 1.178
66,737 1.337 1.348 1.208
84,455 1.338 1.340 1.176
104,379 1.336 1.347 1.140
126,148 1.337 1.363 1.173
150,154 1.339 1.367 1.129
203,915 1.335 1.373 1.132
Variation range 0.008 0.033 0.202

methods will show biases for small increments of

measure. A simple way to illustrate this bias is to com-
pute the fractal dimension of an euclidean shape such
as a circle or square. Using the EDM method, it can
be shown that boundary fractal analysis on a circle
Fig. 4. Boundary fractal analysis on triadic and quadratic
results in a higher fractal dimension (1.0107) than that
Koch curves with box-counting, dilation and EDM methods. for a square (1.0068) of the same area (approximately
Filled data points in regions of small and large steps were not 32,000 square pixels). However, when the square is
used in linear regression. rotated, pixelization appears and the sides of the
pivoted square show the same stair-like appearance as
that of the circle (Fig. 5). The fractal dimension for
the square is now 1.016, even higher than that for the
circle.

4.3. E€ect of object size

Size and resolution are important factors when com-

puting fractal dimensions since the computing algor-
ithms are in¯uenced by the anisotropic treatment of
Fig. 5. E€ect of discretization of outline on boundary fractal the images (Graf, 1991). Table 1 and Fig. 6 show the
dimensions. Circle and tilted square display higher fractal e€ect of size (or resolution) on the fractal dimension of
dimensions than square. a circle. The same analysis was performed on a well-
 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071 1067

Fig. 6. Variation of boundary fractal dimension of circle and carbonblack fragment (Poodle) with size (pixels). Size limit for EDM
analysis is 5000 pixels (area). Recommended size is 10,000 pixels.

known carbonblack particle (herein referred as `The EDM method but consistent results are produced for
Poodle') reproduced from Kaye (1990, Fig. 2). This features with area size greater than 5500 pixels. For
particle is often used in fractal outline studies to illus- circles over that size, results vary from 1.018 (small cir-
trate the complexity of features. cle) to 1.010 (larger circles). For the same size range,
For the euclidean circle, dilation and EDM methods dilation shows a better accuracy and a regular pattern,
exhibit an increase of the fractal dimension for very close to the theoretical value, with dimensions varying
small sizes. This increase is more pronounced for the between 1.001 to 1.008. The more pronounced e€ect of

Fig. 7. E€ect of overall structure on boundary fractal dimension for box-counting, dilation and EDM methods. Graph shows larger
range of fractal dimensions for EDM method.
1068 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071

pixelization for smaller features and the lack of data dimension (0.1) between the fractalized circle and the
points to compute a reliable linear regression explain stick that re¯ects the complexity of the features.
larger dimensions for smaller circles. Box-counting
yields erratic results, even if results from smaller circles
(<5000 pixels) were eliminated, producing fractal 5. Discussion
dimensions of up to 1.076 for an euclidean shape.
For the `Poodle', the EDM method yields stable Various tests revealed that box-counting cannot be
fractal dimensions for sizes greater than 10,000 pixels, used for the precise characterization of outlines.
with a variation in the third decimal place only (error Because of the large scatter of results, it is shown that
range of 0.008). No signi®cant increase is observed for it is dicult to compute high precision reproducible
very large sizes (e.g. more than 50,000 pixels). Dilation results with this method.
is less stable with a variation of 0.033 for sizes over Dilation and EDM are two methods based on the
10,000 pixels and an increase for larger fragments is same type of measurements (sausage covering) and
observed. Results are generally higher overall than nonsurprisingly, yield similar results in terms of repro-
those obtained with the EDM method. Box-counting ducibility. Still, precision and accuracy are di€erent
results decline from small features to larger ones and overall. Dilation yielded more precise results (closer to
in our tests, this method yielded underestimated fractal theoretical values) for euclidean shapes such as the
dimensions when compared to EDM and dilation square, stick and circle. EDM was more precise for the
results (e.g. `The Poodle' and the quadratic Koch complex shapes such as the quadratic Koch curve.
curve). Box-counting fractal dimensions vary over a Fractal dimensions are used in the purpose of describ-
range of 0.2. ing complex shape, not euclidean ones, which gives
EDM method displays a low variability of the more importance to the behavior of the method with
results (20.01 or <1%) for fragments larger than complex curves. As for the accuracy, the two methods
10,000 pixels, that represents the basic uncertainty for showed small variations with size and resolution, pro-
the EDM technique. Below that size limit, the fractal ducing constant fractal dimensions. However, dilation
dimension tends to be slightly high, especially for more seemed slightly less accurate for complex curves (cf.
euclidean shapes. Dilation showed better accuracy for The Poodle and the Koch curves), showing variations
the euclidean circle (closer to 1) but more scatter for with size or underestimation of the fractal dimension.
larger irregular features and the dilation margin of The lower sensitivity of the dilation technique is
error is estimated at 0.03. more evident in the study of the three fractalized
®gures (circle, square and stick). The fractal dimen-
4.4. E€ect of geometry sions for these three ®gures range over only 0.04,
although they present an overall very di€erent form.
To study the e€ect of topography, BFA was per- The margin of error for the dilation technique (20.03)
formed for three di€erent structures: a circle, a square is too close from the range of variation to use this
and a stick (elongated structure). Each object was dimension as a shape descriptor.
modi®ed using the Fractalize algorithm in the Canvas EDM method yielded results ranging over 0.1 for
software v. 3.5.2. The resulting six objects (Fig. 7) the three ®gures with a margin of error of 20.01. This
were analyzed with the three BFA methods. smaller margin of error is due to the fact that the cov-
For the euclidean forms, box-counting produce high ering of the outline is more isotropic with this algor-
fractal dimensions for the circle and the stick and the ithm than with the erosion-dilation algorithm. Fractal
highest variability (Fig. 7). Dilation gives excellent dimensions computed with the EDM method for the
results for the square and the stick with perfect values quadratic Koch curve were closer to the mathematical
of one (no pixelization e€ect), but calculates a value value than the other methods. These results indicate a
less than one for the circle, which may be explained by greater sensitivity to complexity. Since fractal analysis
the anisotropy of the method. EDM shows a good was developed for the description of complex outlines,
stability, with a maximum variation of 0.007 between the methods should be judged on their ability to
the euclidean dimensions and a maximum value of describe complexity, not euclidean geometry. This sen-
1.011 for the circle (pixelization e€ect). For the fracta- sitivity can be used to classify features and is therefore
lized forms, box-counting behaves erratically with a the main advantage of the EDM method.
too low fractal dimension for the stick. Dilation The lower size limit for a valid analysis is an import-
dimensions increase from the circle to the stick, but ant parameter to determine. Minimum perimeters of
values remain low (<1.09) and the di€erence between 500 to 1500 pixels (equivalent of circles of 18,000 to
the three dimensions is only in the second decimal 20,000 pixels of total area) have been suggested for the
place, to a maximum of 0.04. Results obtained with use of the structured walk method (Orford and
the EDM method show a large increase in fractal Whalley, 1987; Hamblin and Stachowiak, 1993). Other
 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071 1069

workers prefer the ferret diameter (Kaye, 1990) or the portant than accuracy in BFA. Choice of the method
fraction of the maximum diameter (Schwarz and should be based upon these considerations.
Exner, 1980) as a lower limit. . Dilation and EDM methods produce the more re-
In the examples of the dilation and EDM methods, liable results with a low sensitivity to object size and
the reference must be the total area of the feature since resolution. The precision of the EDM method
it is the area of the sausage that is measured and not (20.01) is higher than for the dilation method
the perimeter. Both methods produced reliable results (20.03). Box-counting behaves erratically when used
for areas down to 5000 pixels. Larger dimensions for to analyze outlines.
smaller features are explained by the in¯uence of the . The recommended minimum object area is 10,000
pixelization of the border that becomes more import- pixels. For smaller features, the margin of error
ant for smaller outlines and the lack of data points. would increase. Narrow objects (stick form) should
For a circle, a total area of 5000 pixels corresponds to be analyzed at greater resolution to increase the
a perimeter of only 125 pixels, which illustrate the measurements in the narrow regions.
large range of size that can be analyzed reliably. More . EDM method is more sensitive to complexity and
complex features, such as `The Poodle' will be more
yields dimensions over wider ranger of values. This
sensitive to the lower size limit. A decrease in resol- advantage enables the use of the fractal dimensions
ution would result in a greater loss of details than for as a shape descriptor.
an euclidean shape. Objects smaller than 5000 pixels . Implementation of dilation and EDM techniques is
should therefore be rejected and the minimum size of fast and easy, using processing algorithms readily
10,000 pixels is recommended. available in most imaging systems.
No e€ects were observed for very large size features
with the EDM method. Results showed variations only
in the third decimal for very large fragments
(area > 200,000 pixels). The largest feature that can be
analyzed is de®ned by the number of grays available in Acknowledgements
the imaging software. For example, NIH-Image can
manage 256 levels of gray, which limits the sausage We would like to thank Mr. John Russ for the use-
width to 512 pixels (256 in the background and 256 in ful conversations and Mr. Tony Fowler and an anon-
the feature). Objects with a diameter larger than 1024 ymous reviewer for their comments that helped
pixels would be dicult to analyze. Dilation and box- to improve this paper. We would also like to thank
counting are not limited in the number of data points the FCAR for providing funds for one of the author
but size test showed that analysis at large scale result (DB).
in less reliability of the data. The increase in the di-
lation results for larger fragments such as `The Poodle'
can be explained by the anisotropy of the algorithm. Appendix A. EDM macro
Very narrow features, although they might present a
sucient total area, could yield biased results if not Macros are available on server at http:\\www.iam-
analyzed at high enough resolution with the EDM and g.org/CGEditor/index./htm. Macros are loaded directly
dilation methods. The sausage forming around the out- in NIH-Image 1.61, available on the NIH web site.
line will close up rapidly in the narrow portions, not
allowing enough data points to proceed with the linear {Production of euclidean distance map and bound-
regression. The minimum size of 10,000 pixels would ary fractal dimension computing
therefore not be sucient.
Although computing time is less of a factor with the by Dominique BeÂrubeÂ
use of more powerful computers, dilation and box- Version 2.1
counting are time-consuming algorithms. Each involves 28/11/98
an operation of amalgamation (coarsening of the
®gure) and subsequently, the measure of the new fea- The macro must be applied to a binary image
ture, with further similar steps. EDM enables to com- only}
pute the entire image and then make the measurements var
by simply tresholding the image. edm1, temppic:integer;
i, sf, map: integer;
begin
6. Conclusion RequiresVersion (1.61);
{because of power function and euclidean dis-
Precision and reproducibility of results is more im- tance map functions}
1070 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071

{Compute euclidean distance map} References

SelectAll;
temppic:=PidNumber; Adler, J., Hancock, D., 1994. Advantages of using a distance
Duplicate (`Inverted'); transform function in the measurement of fractal dimen-
Invert; sion by the dilation method. Powder Technology 78 (3),
191±196.
edm1:=PidNumber;
Allen, M., Brown, G.J., Miles, N.J., 1995. Measurement of
Binary (`edm'); boundary fractal dimensions: review of current techniques.
SelectPic (temppic); Powder Technology 84 (1), 1±14.
Binary (`edm'); Batty, M., Longley, P., 1986. The fractal simulation of urban
ImageMath (`add', temppic, edm1, 1, 0, `Sum'); structure. Environment and Planning A 18 (9), 1143±1179.
SelectPic (temppic); Dispose; SelectPic (edm1); Baumann, G., Barth, A., Nonnenmacher, T.F., 1994.
Dispose; Measuring fractal dimensions of cell contours: practical
SelectWindow (`Sum'); approaches and their limitations. In: Nonnenmacher, T.F.,
Losa, G.A., Weibel, E.R. (Eds.), Fractals in Biology and
{Measure the surface of the Minkowski sausage as Medicine. BirkhaÈuser Verlag, Bale, Switzerland, pp. 182±
determined by the number of levels of gray selected. 189.
An exponential suite of levels of gray is used to BeÂrubeÂ, D., JeÂbrak, M., 1996. Fractal descriptors for the
optimize the number of points on the graph} classi®cation of Sudbury breccias. Program with Abstracts.
Geological Association of Canada, Mineralogical
SetOptions (`Area, User1');
Association of Canada, Annual Meeting, Winnipeg,
SetUser1Label (`Width'); Canada 21, p. A-8.
SetScale (1, `pixel'); Block, A., Von Bloh, W., Schellnhuber, H.J., 1990. Ecient
ResetCounter; box-counting determination of generalized fractal dimen-
map:=pidNumber; sions. Physical Review A 42 (4), 1869±1874.
for i:=1 to 50 do begin Buczkowski, S., Hildgen, P., Cartilier, L., 1998. Measurement
y:=power (1.1, i); of fractal dimension by box-counting: a critical analysis of
{1.1 can be changed in the power function to data scatter. Physica A 252 (1), 23±34.
Carr, J.R., Benzer, W., 1991. On the practice of estimating
generate a satisfying number of data}
fractal dimension. Mathematical Geology 23 (7), 945±958.
y:=round(y); Clark, N.N., 1986. Three techniques for implementing digital
If lg<>y fractal analysis of particle shape. Powder Technology 46
then begin (1), 45±52.
Crawford, J.W., Matsui, N., 1996. Heterogeneity of the pore
lg:=y;
and solid volume of soil. distinguishing a fractal space
SetDensitySlice (0, lg); from its nonfractal complement. Geoderma 73 (3/4), 183±
Measure; 195.
rUser1[rcount]:=2 lg; Danielsson, P.E., 1980. Euclidean distance mapping.
Setprecision (0); Computer Graphics Image Processing 14, 227±248.
Dubuc, B., Quiniou, J.F., Roques-Carmes, C., Tricot, C.,
{necessary to compute results in french version of Zucker, S.W., 1989. Evaluating the fractal dimension of
EXCEL, can be removed} pro®les. Physical Review A 39 (3), 1500±1512.
Dutch, S.I., 1993. Linear Richardson plots from nonfractal
data sets. Mathematical Geology 25 (6), 737±751.
Flook, A.G., 1978. The use of dilation logic on the quantimet
UpdateResults; to achieve fractal dimension characterization of textural
and structural pro®les. Powder Technology 21 (2), 295±
end; 298.
SetDensitySlice (255, 255); Gillepsie, P.A., Howard, C.B., Walsh, J.J., Watterson, J.,
ShowResults; 1993. Measurement and characterization of spatial distri-
selectpic (map); butions of fractures. Technophysics 226 (1±4), 113±141.
end; Gonzato, G., 1998. A practical implementation of the box
counting algorithm. Computers & Geosciences 24 (1), 95±
Macro `Export results [e]'; 100.
begin Graf, J.C., 1991. The importance of resolution limits to the
SetExport (`Measurements'); interpretation of fractal descriptors of ®ne particles.
Export (`Fd-EDM'); Powder Technology 67 (1), 83±85.
DisposeAll; Hamblin, M.G., Stachowiak, G.W., 1993. Comparison of
boundary fractal dimensions from projected and sectioned
Close;
particle images. Part I. technique evaluation. Journal of
End; Computer-Assisted Microscopy 5 (4), 291±300.
 D. BeÂrubeÂ, M. JeÂbrak / Computers & Geosciences 25 (1999) 1059±1071 1071

Hayward, J., Orford, J.D., Whalley, W.B., 1989. Three im- In: Pye, E.G., Naldrett, A.J., Giblin, P.E. (Eds.), The
plementations of fractal analysis of particles outlines. Geology and Ore Deposits of the Sudbury Structure.
Computers & Geosciences 5 (2), 199±207. Ontario Geological Survey, Ontario, Canada, pp. 139±210.
Kaye, B.H., 1978. Speci®cation of the ruggedness and/or tex- Neimark, A.V., KoÈyluÈ, UÈ.OÈ., Rosner, D.A., 1996. Extended
ture of a ®ne particle pro®le by its fractal dimension. characterization of combustion-generated aggregates: self-
Powder Technology 21 (1), 1±16. anity and lacunarities. Journal of Colloid and Interface
Kaye, B.H., 1990. A Random Walk through Fractal Science 180 (2), 590±597.
Dimensions, 2nd ed. VCH Publishers, New York. Orford, J.D., Whalley, W.B., 1987. The quantitative descrip-
Kaye, B.H., 1993. Fractal dimensions in data space: new tion of highly irregular sedimentary particles: the use of
descriptors for ®ne particle systems. Part. Syst. Charact. the fractal dimension. In: Marshall, J.R. (Ed.), Clastic
10, 191±200. Particles. Van Nostrand Reinhold Company, New York,
Kennedy, S.K., Lin, W.-H., 1986. FRACT-A fortran subrou- pp. 267±289.
tine to calculate the variables necessary to determine the Rawls, W.J., Brakensiek, D.L., 1995. Utilizing fractal prin-
fractal dimension of closed forms. Computers & ciples for predicting soil hydraulic properties. Journal of
Geosciences 12 (5), 705±712. Soil and Water Conservation 50 (5), 463±465.
Klinkenberg, B., 1994. A review of methods used to determine Richardson, L.F., 1961. The problem of contiguity: an appen-
the fractal dimension of linear features. Mathematical dix of statistics of deadly quarrels. In: General Systems
Geology 26 (1), 23±46. Yearbook, 6, pp. 39±187.
LeMeÂhauteÂ, A., 1990. Fractal Geometries. HermeÁs, Paris. Roach, D.E., Fowler, A.D., 1993. Dimensionality analysis of
Liebovitch, L.S., Tibor, T., 1989. A fast algorithm to deter- patterns: fractal measurements. Computers & Geosciences
mine fractal dimensions by box-counting. Physics Letters 19 (6), 849±869.
A 141 (8/9), 386±390. Russ, J.C., 1995. The Image Processing Handbook, 2nd ed.
Logdson, S.D., 1995. Analysis of aggregate fractal dimensions CRC Press, Boca Raton, FL.
and aggregate densities back-calculated from hydraulic Russ, J.C., Russ, J.C., 1989. Uses of the euclidean distance
conductivity. Soil Science Society of America Journal 59 map for the measurement of features in images. Journal of
(5), 1216±1221. Computer Assisted Microscopy 1 (4), 343.
Longley, P.A., Batty, M., 1989. Fractal measurement and line Schwarz, H., Exner, H.E., 1980. The implementation of the
generalization. Computers & Geosciences 15 (2), 167±183. concept of fractal dimension on a semi-automatic image
Mandelbrot, B.B., 1967. How long is the coast of Britain? analyzer. Powder Technology 27 (2), 207±213.
Statistical self-similarity and fractional dimension. Science Smith Jr., T.G., Lange, G.D., Marks, W.B., 1996. Fractal
156, 636±638. methods and results in cellular morphology. Journal of
Muir, T.L., Peredery, W.V., 1984. The Onaping Formation. Neuroscience Methods 69, 1123±1126.