Escolar Documentos
Profissional Documentos
Cultura Documentos
.
Applications in Electrical Engineenng
by ARNAUD JACQUIN
ABSTRACT : Wepresent a very brief history offractals, describe their generation, their charac-
teristics, and their relation with chaos. We point to where and why,fractals and chaotic s_wtems
are commonly found in nature, which implies that they should be good candidates for modeling
dijyerent types of real worldsignals. We then list a number ofdioerse research areas in electrical
engineering where ,fractals and,fractal-based techniques have ,found applications. Finally, we
present in some detail applications of ,fractals in signal processing, more spec~jically in the
areas qf digital ima~qe modeling, synthesis, and compression.
I. Introduction
The applications of fractals in signal and image processing have been losing their
boundaries more and more, as one starts to witness the merging of various fields :
image synthesis and computer animation (which traditionally derive from computer
graphics), with digital image and video compression (which derive from digital
si~gnalprocessing, specifically information theory). To a large extent, this merging is
due to the need to find ever more powerful models for real world digital signals in
order to represent, store, and transmit these signals efficiently. We will give exam-
ples involving different types of signals, and correspondingly, different types of
fractal models, and will try to describe the advantages and shortcomings of each
of these models. The reader should be warned that our choice was to present several
frameworks and examples which prevented mathematical rigor. This tutorial is
mostly aimed at engineers who are interested in an overview of fractals and chaos,
rather than in mathematical detail. However, references to articles and books
which give the details of the various frameworks are indicated in each section.
The general problem statement which underlies the use of fractals for signal
modeling and compression-the main purpose of the last and largest section of
this paper-is the following. Given any original object or signal, for example a
discrete monochrome image specified by an array of S-bit pixel values, how can a
computer construct a fractal object-the encoded object-which is both visually
close to the original one, and has a digital representation which requires fewer bits
than the original. Note that this latter requirement is not part of the notion of
modeling peu se but is essential to the efficiency of the modeling procedure for
compression or coding. For each type of object/signal to model, this problem will
659
A. Jucquin
be presented in the following format. First, we describe the objects under study,
the class of iterated systems which can generate them, and the generation process.
Second, we describe several approaches to solving what is commonly referred to
as the inverse problem, which consists of devising a procedure for controlling the
generation process in such a way that it produces fractal models of original real
world objects. While the first part is mathematically straightforward and merely
represents several instances of a very general theory of iterated contractive trans-
formations in a metric space, the inverse problem can only be dealt with on a per-
framework basis and does require a significant amount of creativity on the part of
the designer of a fractal modeling system.
A = ,liir ?(A,,),
&,A,,Az,..., (3)
where A, = z(A,_,). From Eq. (l), it is clear that fractals satisfy the invariance
equation :
A = z(A), (4)
which confers to them a property which we generically refer to as self-trans-
,formability, and which leads to the well-known properties of fractals to be rugged
objects with an infinite amount of detail; objects which can be found again and
again in magnified pieces of themselves, however small.
This very simple formulation can lead to objects which have pathological
mathematical propertiesI, as well as a tremendous visual complexity. It is sufficient
tNote that where mathematicians see iteration, engineers might be more likely to see
a feedback process. The idea is the same.
ISuch as nowhere differentiability for some continuous fractal curves and surfaces, or
infinite length or area. Pathological is meant here as opposed to the regularity of the
typical curves and surfaces of Cartesian geometry and differential calculus.
660
Fractals and their Applications in Electrical Engineering
1. Take as initial object a line segment of unit length S,, = [0, 11.
2. To produce the first iterate S, distort the segment by introducing a bump
in the middle of S, in the form of two sides of an equilateral triangle with
side of length f of the length of So.
3. Repeat this introduction of triangular bumps to each line segment until
convergence.
One can see that this construction can be equivalently formulated in the framework
of Eq. (l), with a transformation r consisting of the union of four similarities
which map S, to S,. The construction of the snowflake is very similar and is left
as a simple exercise to the reader. The iterative construction of both objects is
illustrated in Fig. 1.
?A series of such measurements was studied by L. F. Richardson, who showed that the
approximate length L(q) of the west coast of Great Britain, as measured with a yardstick
of size q, satisfies an empirical law of the form L(q) cc q- (9). In (1, 8), Mandelbrot
interprets the quantity D = 1+cc as the ,fractaldimension of a curve. According to the
measurements, the west coast of Great Britain has a fractal dimension approximately equal
to 1.25. It is easy to see that a curve with a fractal dimension strictly greater than one has
infinite length. This feature is characteristic of fractal curves.
FIG. 1. Construction of the von Koch curve (left) and snowflake (right).
4, = ~(&,)+&I, (5)
662
Fractals and their Applications in Electrical Engineering
2x if O<x<i
s(x) = (8)
i 2x-1 if i6x61,
to denote that
x0 =a,*2-'+a2*2-2+.". (10)
One can easily see that the transformation s is equivalent to the left-shif operator
S which acts on the binary expansions of points in I in the following way :
remains a challenge at bit rates under 64 kbpst. It seems that classical image
compression techniques will not be sufficient, and different types of model-based
approaches are being considered (3942).
A = w(A), (14)
which means that A is exactly made of the union of k affinely transformed copies
of itself. It is said to be a glohull~ self-gffine fructul$. The famous 2D Black
Spleenwort fern of Barnsley (5, 44) is an example of such an object; it is the self-
affine attractor of a set of four two-dimensional affine transformations. It is easy
to show that any iterated sequence of the type
tEven though many approaches have been proposed and many systems have been
developed, the image quality of these systems is well below that of analog television signals.
$The smallest .s, for which Eq. (13) holds is called the contractivity factor of the trans-
formation IV,.
If the transformations M,are contractive sirdaritie.s, i.e. each W,contracts the p dimen-
sions of RP with the same factor, A is said to be a (globally) self-similar fractal.
11Fractal plants and trees, for example, are objects which arise as attractors of such
systems.
666
Fractals and their Applications in Electrical Engineering
The union of the transformed copies in Eq. (18) is called a collage of Aor,y (5, 44).
The goal of the modeling is to find the best possible-in the sense of Eq. (18)-
collage of Aorrywith as few affine transformations as possible, in the shortest amount
of time. Note that although the notion of closeness should be in the sense of the
Hausdorff metric, it was interpreted in the sense of visual closeness as defined
by a human operator in the examples of natural scenes described at the end of this
section.
The collage theorem of Barnsley et al. (5,44) states that the fixed point A of the
transformation
1. Real world color images were segmented by a human operator who would
strive to identify various self-affine objects.
2. Each object was modeled as the attractor of an IFS, using an interactive
software package?.
3. The final images were synthesized by successively reconstructing, rendering,
and overlaying the pieces in a specific order-from background to fore-
ground-using another software package described in (30).
The fact that this procedure, when viewed as an image coding and decoding
schemet, can be seen as compressing images with ratios of 1000: I and higher
provoked great, and partially unfulfilled, excitement in the signal processing com-
munity. The major hurdles with the approach described above are clear : (i) natural
scenes are in general not globally self-affine (the way a fern is), and scene seg-
mentation-a difficult problem by itself-is one essential feature, and (ii) the
scheme fundamentally involves a human operator whose function cannot easily be
automated.
Piece-Iv&e selfltran.~forn?ahility, recurrent IFS. Foreseeing that the above pro-
cedure could only be automated with great difficulty, if it could be automated at
all, lead the author to try and bypass the segmentation step by capturing the notion
of piececvise self-tran:formahility in the transformations themselves, albeit by con-
sidering slightly more general transformations. Instead of segmenting a priori an
object Aorjg into several pieces and modeling each of the pieces separately, one
could look for a set of contractive transformations which operate on an object
restricted to pieces of its support, i.e. of the general form
w(B) = u
I <izsk
w,(B,.,). (21)
and which would leave the object approximately invariant, i.e. such that
668
Fractals and their Applications in Electrical Engineering
Note that Z(t) is a closedcurve if go = +. Let Z ,rl,,J(t) denote the restriction of the
parametrized curve Z(f) to the segment [ti, t,].
To each ordered pair of consecutive range nodes (z,, zi+ I), is associated an ordered
pair of domain nodes (z,,,, Q+ , ), also among the set of interpolation points, and a
contractive two-dimensional affine transformation We,such that
(25)
The action of this mapping is illustrated in Fig. 2. It is easy to see that the
transformed curve T(Z) is also a continuous interpolating curve. It can also be
tThe term recurrent was coined by Barnsley et ul. (5, 48). Its selection will probably
not seem obvious to the reader, whom we refer to (48) for a justification of that choice.
$For the sake of simplicity, we use the same notation for a curve 2(*)-a parametric
vector function, and its graph-the set of points {Z(t), t E [0, I]}.
shown (50) that the mapping T is contractive in the space of interpolating curves
endowed with an appropriately defined Hausdorff metric and therefore has a
unique fixed point A-the attractor of the RIFS {{z,};=,, T}-which is a con-
tinuous fractal curve that interpolates the set of pomts {zi};=,. The invariance
equation
shows that the curve A is the union of N affinely distorted copies of pieces of itself.
This type of fractal curve is therefore said to be piecew-ise selfhffine. It can be
generated by plotting the iterates of the sequence
FIG. 4. Modeling contours extracted from Cloud image with recurrent IFSs. (a) Cloud
image, (b) Extracted contours, (c) Interpolation points, (d) Collage, @1988, SPlE Vol. 1001
Visual Communications and Image Processing 88, p. 128.
FIG. 5. Six iterations converging to the attractor of a RIFS model for Cloud contours,
$31988, SPIE Vol. 1001 Visual Communications and Image Processing 88, p. 129.
(29)
where
9 = {R,)S,<,V (30)
p= 1 i+R,, (31)
O<,<N
simply to indicate that an image is the union of its restrictions to the cells of the
partition. The symbol zi denotes an elemental block transformation from a domain
cell D,t to the range cell Ri. For clarity, 7i is written as the composition of two
transformations Si and T, :
T, = T,oS,, (32)
where Si and T, are the so-called geometric and massic parts of T,, respectively. In
the case of square blocks, S, contracts domain blocks the size of range block R,,
while T, processes an image block without altering its square support.
Designing a fractal block coding system consists in defining a priori the fol-
lowing: (i) a procedure for constructing a good image dependent partition,
which can consist for example of a quadtree structure or any other type of tiling,
and (ii) classes of discrete block transformations and a procedure for searching
these classes. The construction of a fractal code 5 for p,,, consists in finding, for
each value of the index i, the best pair of domain and block transformation
(D,, 70, i.e. the pair such that the distortion
The fractal code consists of the description of the image partition along with the
quantized parameters of the block transformations. Its structure is illustrated in
j-These cells have to be larger than the range cells so as to ensure contractivity. A common
choice is to consider square blocks of twice the size of the range blocks.
Fig. 6, in which the arrows indicate block transformations from domain block to
range block.
Decoding/reconstruction. As always with fractal techniques, the reconstruction
of a decoded image is easily achieved through iteration. Here, the fractal block
code z is iterated on an initial image, such as a black square (or any other image)
until convergence to a stable image-in practice after about ten iterations. This
reconstruction procedure is illustrated in Fig. 7 for a fractal code of the familiar
digital image Lena, with the initial image peppers.
Extensions. Over the past five years, a large number of extensions of fractal
image block coding have been proposed by various researchers in the signal pro-
cessing research community (5%79), which shows that the field remains a very
active research area. These extensions broadly address the following issues: (i)
the influence of the type of image partition, pool of block transformations, and
optimization of the parameters defining these transformations, (ii) the reduction
of the complexity of the encoding process-the search for the best matching blocks
in the image to encode, (iii) the comparison and possible merging of fractal block
coding with more traditional block-based image coding techniques, and (iv) the
extension of the theory to the three-dimensional case for fractal block coding of
sequences of images. A summary of these extensions can be found in (80).
Unfortunately, the compression ratios achieved with block fractal coding tech-
niques turn out to be quite similar to those obtainable with other more classical
image coding techniques, in the order of 30: 17, which seems modest compared
with the 1000 : 1 ratios claimed for the pictures described in Section 4.1.2. This is
because, in order to make it tractable, the inverse problem had to be broken up
into elementary subproblems which can be efficiently dealt with automatically. In
particular, the image segmentation with human operator described in Section 4.1.2
?A rather severe breakdown seems to occur fairly systematically at coding rates below
0.2 bpp.
has been replaced by an image partitioning into fairly small blocks, which is of
course reminiscent of other block-based image coding techniques such as block-
based DCT (81L83), or vector quantization (VQ) (84,SS). However, by doing so,
the possibility to model very large image segments by only very few trans-
formations-as was done in the images of Section 4. l-has been lost.
V. Conclusion
References
(1) B. Mandelbrot, The Fractal Geometry of Nature, Freeman, San Francisco, CA,
1982.
(2) K. Falconer, The Geometry of Fractal Sets, Cambridge University Press,
Cambridge, 1985.
(3) K. Falconer, Fractal Geometry, Mathematical Foundations and Applications,
J. Wiley & Sons, NY, 1990.
(4) G. Cherbit, Fractals, Dimensions non Entieres et Applications, Masson, Paris, 1987.
(5) M. F. Barnsley, Fractals Everywhere (Second edition), Academic Press, New York,
1988.
(6) M. SchrBder, Fractals, Chaos, Power Laws, W. H. Freeman and Co., New York,
1991.
(7) H.-O. Peitgen, H. Jiirgens and D. Saupe, Chaos and Fractals, New Frontiers of
Science, Springer-Verlag, Berlin, 1992.
(8) B. Mandelbrot, How long is the coast of Britain?-Statistical self-similarity and
fractional dimension, Science, Vol. 155, pp. 636-638, 1967.
(9) L. F. Richardson, in General Systems Yearbook, 6, 139, 1961.
(10) B. Mandelbrot, Les Objects Fractals (Deuxikme tdition), Flammarion, Paris, 1984.
(11) J. Gleick, Chaos, the Making of a New Science, Vicking, New York, 1987.
(12) K.-H. Becker, M. DBrfler, Dynamical Systems and Fractals, Cambridge University
Press, Cambridge, 1989.
(64) T. Laurencot and A. Jacquin, Hybrid image block coders incorporating fractal coding
and vector quantization, with a robust classification scheme, AT&T Technical
Memorandum, February 1992.
(65) S. Lepsoy, Attractor image compression-fast algorithms and comparisons to related
techniques, Ph.D. Thesis, The Norwegian Institute of Technology, June 1993.
(66) S. Lepsoy, G. E. 0ien and T. A. Ramstad, Attractor image compression with a fast
non-iterative decoding algorithm, Proc. ICASSP-93, pp. 331-340, 1993.
(67) H. Li, M. Novak and R. Forchheimer, A fractal-based image sequence compression
scheme, Preprint, 1993.
(68) L. Lundheim, An approach to fractal coding of one-dimensional signals, Proc.
KOMPRESJON-89 (Oslo), 1989.
(69) L. Lundheim, Fractal signal modelling for source coding, Ph.D. Thesis, The
Norwegian Institute of Technology, September 1992.
(70) D. M. Monro and F. Dudbridge, Fractal approximation of image blocks, Proc.
ICASSP-92, Vol. III, pp. 485488, 1992.
(71) D. M. Monro and F. Dudbridge, Fractal block coding of images, Elec. Letters, Vol.
28, pp. 1053-1055, May 1992.
(72) D. M. Monro, Generalized fractal transforms : Complexity issues, Proc. Duta Com-
pression Cor~fi, pp. 254261, IEEE Computer Society Press, March-April 1993.
(73) G. E. 0ien, S. Lepsoy and T. A. Ramstad, An inner product space approach to image
coding by contractive transformations, Proc. ICASSP-91, pp. 211332776, 1991.
(74) G. E. Oien, S. Lepsoy and T. A. Ramstad, Reducing the complexity of a fractal-
based image coder, EUSIPCO-92.
(75) G. E. Mien, A. Sollid and T. A. Ramstad, Hybrid image compression combining
transform coding and attractor coding, NOBIM-92.
(76) G. E. 0ien, Lz-Optimal attractor image coding with fast decoder convergence, Ph.D.
Thesis, The Norwegian Institute of Technology, 1993.
(77) G. E. 0ien, Z. Baharav, S. Lepssy and E. Karnin, A new improved collage theorem
with applications to multiresolution fractal image coding, Proc. ICASSP-94, 1994.
(78) D. van Schooneveld, Fractal coding of monochrome images, Delft University of
Technology, Internal Report, December 1991.
(79) G. Vines and M. H. Hayes, III, Adaptive IFS image coding with proximity maps,
Proc. ICASSP-93, pp. 3499352, 1993.
(SO) A. E. Jacquin, Fractal image coding : A review, Proc,. IEEE, Vol. 81, pp. 1451-1465,
October 1993.
(81) N. S. Jayant and P. Nell, Digital Coding of Waveforms : Principles and Applications
to Speech and Video, Prentice-Hall, Englewood Cliffs, NJ, 1984.
(82) R. J. Clarke, Transform coding of images, Academic Press, London, 1985.
(83) A. N. Netravali and B. G. Haskell, Digital Pictures : Representation and
Compression, Plenum Press, New York, 1989.
(84) R. M. Gray, Vector quantization, IEEE ASSP Mug., April 1984.
(85) A. Gersho and R. M. Gray, Vector Quantization and Signal Compression, Kluwer
Academic Publishers, Netherlands, 1992.