You are on page 1of 6

What are fractals?

Features
Before attempting to explain what fractals are
Fractals it will be helpful to consider the differences
between the real world and the way we per-
ceive it. We live in an age where we can project
ourselves into a world of virtual reality and act
out our dreams and fantasies without regard
to normal physical constraints or risks. Most
Alan Flook
of us are aware that we exist in the real world
but, when we are designing structures or
investigating some physical phenomenon, we
often find it convenient to slip into a different
sort of virtual world where everything is mod-
elled on simple geometric principles. It has
become such a common practice that we
hardly recognize that we plan and theorize
The author using models to represent reality. Architects,
Alan Flook is a scientific computing consultant specializ- engineers and scientists draw models of build-
ing in image analysis and processing. He is based at 6, ings, engineering structures and atoms using
Widecombe Close, Bedford, Beds, UK. Tel/Fax: 01234 points, straight lines, planes and other con-
357572; E-mail: 100041.44@compuserve.com structs that are based on Euclidean geometry.
Euclidean geometry is based on the axioms
Abstract expounded by Euclid in the third century BC.
Since the term was first coined in 1977, fractals seem to It is the geometry that has been taught in
have pervaded every branch of science. Attempts to schools ever since. We have never needed to
explain what fractals are and what they are being used for. specify that the geometry was Euclidean
Are they a fad or are they really useful? Considers factors because it was tacitly assumed that it was the
including quantitative measurement, image compression only geometry; much in the same way that
and computer graphics. Concludes that the future will see nobody realized that they were wearing an
an increase in the use of fractal graphics. analogue watch until the advent of the digital
watch. The concepts of points, straight lines
and planes are so familiar to us that it may
come as a shock to realize that they do not
exist. What we are used to seeing are graphical
or physical representations of these conceptu-
al structures. For example, a point has zero
dimensionality yet a point drawn with the
sharpest pencil will, under magnification,
appear as a two-dimensional smudge. Under
a scanning electron microscope at a sufficient-
ly high magnification, the same point will be
seen as a three-dimensional heap of graphite
crystals and the paper will look like a tangled
bundle of fibres.
Author’s address Euclidean geometry has endured over the
Author’s name here centuries because it provides a good basis for
Address 1 modelling the world we live in, but there are
Address 2 many circumstances where it fails. In the late
Address 3 1950s the English mathematician, Lewis Fry
Address 4 Richardson, needed to estimate the length of
Address 5 the coastline of Great Britain. The coastline
e mail: insert here

Sensor Review The fractal fern shown in Figure 3 is the logo of


Volume 16 · Number 3 · 1996 · pp. 42–47 Iterated Systems and is reproduced with their kind
© MCB University Press · ISSN 0260-2288 permission.
42
Fractals Sensor Review
Alan Flook Volume 16 · Number 3 · 1996 · 42–47

can be estimated by “walking” round a map grasped the idea behind recursion). The
with a pair of dividers with the points opened algorithm used to compute the graphic can be
to represent the scaled equivalent of five miles written in the same recursive fashion as the
for example. The estimate is obtained by last sentence. Recursive programming leads
multiplying the number of strides taken to to very efficient code and the whole program
return to the starting point by the stride will consist of very few lines. (Caution: com-
length. Richardson soon realized that the puters never tire and will continue the recur-
length of the coastline was indeterminate sion ad infinitum. You must instruct the
because it depended on the resolution (i.e. the computer when to break out of the loop, i.e.
stride length) with which the measurements limit the resolution of the graphic.)
were made. The smaller the stride length the The fractal dimension of a Koch island is
more closely the “walk” can follow all the readily calculated. If the sides of the square
small embayments and peninsulas on the are of unit length, it has a perimeter of 4 units.
coast and the longer will be the estimated The perimeter increases to 8 units after the
length. He found that a plot of the logarithm first generation and will continue to double at
of the estimated coastline length versus the
logarithm of the stride length used generated
a straight line. Furthermore the magnitude of Figure 1 Generation of Koch quadric island
the negative slope of this line depended on the
ruggedness of the coastline. These graphs are
now known as Richardson plots.
Benoit Mandelbrot’s book, Fractals, Form,
Chance and Dimension[1,2], was published in
1977. In it he developed the concept of a
dimension that could have non-integer values.
The term fractal was coined by Mandelbrot to
describe such dimensions. The fractal dimen-
sion of a curve is computed as 1 – the slope of
the Richardson plot (NB the slope is always
negative therefore the dimension will be 1 + Initiator
the absolute value of the slope). Curves, such
as coastlines, that have space filling properties
have fractal dimensions between 1 and 2 while
fractal surfaces have dimensions between 2
and 3. Mandelbrot’s work sparked an interest
in fractal geometry as a practical tool to be
used in circumstances where Euclidean
geometry had been found wanting.
Generator
Generating fractal structures
To gain an understanding of the fractal prop-
erties of curves it is helpful to look at comput-
er generated graphics with known fractal
dimensions. Fractal graphics can be produced
by using an initiating shape and a fractal
generator. Figure 1 shows the generation of a
Koch quadric island. Here the square is the
initiator and the generator is as shown. The
algorithm takes each straight line forming the
square and replaces it with the generator. The
generator consists * of a sequence of straight
lines so, at the next level of generation, all
these lines are replaced with a suitably scaled
copy of the generator which also consists
Quadric island
(Note to reader: loop back to * until you have
43
Fractals Sensor Review
Alan Flook Volume 16 · Number 3 · 1996 · 42–47

each subsequent level of generation. The Figure 3 Fractal fern


shape generated has a fractal dimension of
log(8)/log(4) =1.5. The shape shown in Fig-
ure 2 is a Koch triadic island. The fractal
generator is less complex than that of the
quadric island and the resultant fractal
dimension is log(4)/log(3) ≈1.26. The lower
fractal dimension of the boundary of this
island reflects that it is less rugged or convo-
luted than that of the quadric island.

Fractals in nature
It is astonishing how often fractal behaviour
can be observed in the natural world. One of
the most striking features about structures
produced by fractal generators is that many of
them look like familiar objects. Computer
graphics closely resembling mountain ranges,
ferns (see Figure 3), trees, fungal growth,
nerve fibres and electrical discharge patterns

Figure 2 Generation of Koch triadic island

can easily be generated with very simple


algorithms. Mandelbrot emphasized this
point in the title of his book, The Fractal
Geometry of Nature[2], published in 1982.
Natural fractals differ from computer
generated fractal shapes such as the triadic,
Initiator
and quadric islands. The method of generat-
ing these graphics causes them to exhibit self-
similarity over all scales of scrutiny. This
means that if you take a small part of the
structure it will always look exactly the same
no matter how much it is magnified. Naturally
Generator occurring fractal materials exhibit self-affinity
rather than self-similarity and this is usually
only maintained over a limited range of scales
of scrutiny.
That fractal-like structures abound in
nature begs the question: is this a coincidence
or something more fundamental? We have
seen that very complex structures can be
generated by simple computer algorithms.
Thus such structures necessarily contain very
little structural information. In biology, the
whole blueprint of any organism is contained
in a remarkably small number of genes. It is
tempting to speculate that the morphology of
Triadic island
organisms which have a fractal-like appear-
ance have at least some of their shape
44
Fractals Sensor Review
Alan Flook Volume 16 · Number 3 · 1996 · 42–47

characteristics encoded as fractal generators. Figure 4 Coastline dilation


In the world of physical phenomena, most
fractal-like structures are produced by mecha- a) Original b) Dilated 1 step c) Dilated 3 steps
nisms which dissipate energy, i.e. the energy
content of the propagating force is used up in
the process. In crack propagation, for exam-
ple, some energy will be used in generating
the primary crack. Secondary and higher
order cracks will have less propagation energy
and will result in progressively shorter path
lengths. The same is true for the tracking
behaviour of electrical discharge processes.

Are fractals useful?


There are three main applications of fractal
geometry. These are outlined in the following
sections.
good results. In general the choice of method
Quantitative measurement
of measurement will depend on the type of
As soon as it was realized how well fractal
fractal object under investigation.
graphics mimicked nature, researchers in
Fractal dimensions have been used as a
many, widely different fields started to explore
method of quantifying the complexity of a
how to measure and apply fractal dimensions
wide variety of materials and objects. For
in their investigations. example: a root and branch investigation into
The stepping divider procedure was the the growth of trees; nerve cell growth and
first practical method of estimating the fractal degeneration; osteoporosity; metal fatigue
dimensions of real objects. It is too labour fractures; diffusion limited aggregation and
intensive to be used as a general method. electrical discharge phenomena. One of the
Many computer assisted methods have been first uses of fractal dimensions was for the
devised. One of the simplest of these methods measurement of texture where it may
is based on the Minkowski sausage[3]. The replace/augment Fourier methods in provid-
image of the object is stored in an image ing an index for quantifying surface rough-
analysing computer where it is thresholded to ness. Although the word fractal had not then
produce a binary image. The binary image is been coined, it would be wrong to omit their
processed so that only the edge pixels forming use in studying the role of the complexity of
the perimeter are retained. These edge pixels borders between countries on the potential for
are then progressively dilated and the number armed conflict. This was the main aim of
of pixels (i.e. the area) is counted at each Richardson’s work.
dilation step. (Dilation is the process where The main difficulty in employing fractal
each pixel is replaced by a circle, or the near- dimensions has already been implied by the
est pixelated equivalent. This thickens the statement that natural objects exhibit self-
perimeter into a ribbon.) The diameter of the affinity over a limited range of measurement
circle used for each dilation step is equivalent scales. This leads to Richardson plots that are
to the stride length and the perimeter estimate either a series of straight line segments or are
at that resolution is given by the area of the curved. The transition points between the
dilated perimeter divided by the stride length. straight sections of the plots can yield extra
The dilation process is illustrated in Figure 4. insight into the structure of the material but
There are many other methods, including also lead to a greater degree of complexity in
one that uses the, so-called, Euclidean dis- their interpretation. A further limitation is
tance transform. There is a double irony here: that the fractal dimension provides a good
not only is a Euclidean distance being used to index for the complexity of a boundary but
measure a non-Euclidean dimension but also contains no information on the structure of
the term Euclidean has no meaning at all in that complexity. Many very different types of
the discrete geometry of pixel space. Never- structure can have the same fractal dimen-
theless the method is very practical and yields sion.
45
Fractals Sensor Review
Alan Flook Volume 16 · Number 3 · 1996 · 42–47

Image compression blocks on to the domain blocks. At each stage


In recent years there has been a rapid growth of the compression, the range block positions
in the number of applications for computers and the transformation coefficients are
which require the mass storage of images. recorded. These are the data that are used to
Stored on a pixel by pixel basis, a medium reconstitute the image. Because the data are
resolution, monochrome image (512 × 512 stored as mathematical formulae, very high
pixels) will need 0.25 Mbytes of memory. compression ratios can be achieved. Further-
This storage requirement is trebled if the more, the formulae are not size related and
image is in real colour. Most images exhibit a can be used to decompress files to give larger
high degree of redundancy in their informa- images than the original without loss of reso-
tion content: for example most adjacent pixels lution. When using this “resolution indepen-
will be the same or of similar colour or intensi- dence” properly to expand images, it should
ty. This redundancy can be used to repack the be remembered that it is impossible to
image so that it takes up less storage space. increase the information content beyond that
There are two main types of image compres- of the original image and that the additional
sion: lossless and lossy. detail in the image will be an artefact of the
Lossless compression methods pack the transform. However, examples of images that
image in a way that allows the image to be have been expanded beyond the original file
reconstituted as a faithful replication of the size show that the formulae can predict the
original image. These methods usually pro- form of the missing data to give images with a
duce relatively low compression ratios. The much more natural appearance than the
actual value will depend on the content of the blocky effect produced in images zoomed by
original image but will generally be of the pixel replication, which is the alternative
order of 4:1. Lossy image compression meth- method. As with JPEG transforms, the fractal
ods only preserve the general appearance of compression ratio will depend on the image
the image and some information content, content and the required degree of fidelity of
mostly the fine detail, is lost. For many IT the reproduction but typical ratios will be in
applications this loss of information is not as the range of 20:1 to 70:1. Usable compression
important as the need for efficient storage and ratios of over 100:1 can be achieved for some
rapid retrieval of images and for their fast images.
transfer over networks. Until recently the Fractal compression transforms are asym-
JPEG standard was used for most lossy image metric: they can take a lot of processing time
compression. The JPEG method is based on during the compression process but the
the discrete cosine transform which is related images can be very quickly extracted. Many
to the Fourier transform. JPEG compression images exhibit a high degree of affine redun-
ratios depend both on the original image dancy and can withstand high compression
content and the resolution required in the ratios without a significant loss of detail and,
retrieval process. Typical values are in the therefore, this method of compression offers
range 10:1 to 20:1. many advantages over the alternatives. A
Fractal image compression methods were further advantage of this method is that inter-
developed by Michael Barnsley in 1987. He mediate resolution images can be displayed as
used affine transforms of regions within the decompression progresses. This allows the
image. An affine transform of an image pro- user to determine the image content before it
duces an output which is a rotated, translated is fully resolved and abort the process in the
(shifted), magnified, shrunk or sheared ver- event of an erroneous choice.
sion of the original image. These distortions
of the image can be combined into more Computer graphics
complex affine transforms. Fractal compres- Early attempts at using computer graphics to
sion uses affine transforms that are contrac- produce realistic scenery failed mainly
tive (i.e. they make things smaller). The image because natural textures were difficult to
to be compressed is divided into domain reproduce without resorting to storing and
blocks. These are non-overlapping tiles that processing masses of textural data. It is rela-
cover the whole image. The image is also tively easy to generate fractal graphics which
divided into larger, and therefore overlapping, closely resemble natural textures and struc-
range blocks. The strategy is to try to find tures. It is no surprise therefore that they are
affine transforms that best map the range heavily used for this purpose.
46
Fractals Sensor Review
Alan Flook Volume 16 · Number 3 · 1996 · 42–47

The future There is a growing use of computers for all


sorts of animation. There is no doubt that we
Although it had been recognized that Euclid-
shall see an increase in the use of fractal
ean geometry posed some difficulties in the
graphics to create structures and textures to
mensuration of real objects, very little
produce realistic scenery for cartoons, com-
progress had been made until the pioneering
puter games and for more serious applications
work of Mandelbrot. Since then, many scien-
such as flight and battlefield simulators. In
tists have explored the use of fractal dimen-
other words, the world of virtual reality. This
sions in their fields of interest. The trickle of
is where this contribution started and it seems
research papers in the late 1970s has now
entirely appropriate that an article on fractals
become a flood. People who were initially
should have a recursive feel about it.
sceptical about non-integer dimensions now
recognize that they have an important role in
quantifying many complex structures. It is Note and references
certain that this trend will continue as fractal
1 This is now out of print. The Fractal Geometry of
dimensions gain acceptance in more research
Nature[2] is an updated version.
disciplines.
2 Mandelbrot, B.B., The Fractal Geometry of Nature,
Fractal image compression is finding ever
W.H. Freeman, San Francisco, CA, 1982.
more applications since Barnsley set up his
3 Flook, A.G., “The use of dilation logic on the quan-
company, Iterated Systems, in 1987. It is timet to achieve fractal dimension characterisation of
widely used in CD-ROM based applications textured and structured profiles”, Powder
such as encyclopaedias. Microsoft has used it Technology, Vol. 21, 1978, pp. 295-8.
to compress the 7,000 images in Encarta.
Hutchinson and Grolier also have followed
Further reading
suite. The increase in the speed of modern
processors and the improvements in decom- Kaye, B., A Random Walk through Fractal Dimensions,
pression algorithms have rendered fractal VCH ISBN 0-89573-888-0.
methods fast enough to be used for video
applications on compact discs. This, together
Contact address
with the emergence of digital video discs and
the ever increasing need for images to be Iterated Systems, Wyvols Court, Swallow-
archived, will ensure the future of fractals as field, Reading, Berkshire RG7 1PY. Internet –
an efficient image compression tool. http://www.iterated.com

47

You might also like