World Academy of Science, Engineering and Technology

International Journal of Computer and Information Engineering

Vol:2, No:11, 2008

Hybrid Genetic-Simulated Annealing Approach

for Fractal Image Compression
Y.Chakrapani, K.Soundera Rajan

 usually applied to optimize controlled parameters and

Abstract—In this paper a hybrid technique of Genetic Algorithm constrained functions.
and Simulated Annealing (HGASA) is applied for Fractal Image Simulated Annealing (SA) is a method for obtaining good
Compression (FIC). With the help of this hybrid evolutionary solutions to difficult optimization problems which has
algorithm effort is made to reduce the search complexity of matching received much attention over the last few years. The recent
between range block and domain block. The concept of Simulated interest began with the work of Kirkpatrick and Cerny. They
International Science Index, Computer and Information Engineering Vol:2, No:11, 2008 waset.org/Publication/6811

Annealing (SA) is incorporated into Genetic Algorithm (GA) in order

showed how a model for simulating the annealing of solids, as
to avoid pre-mature convergence of the strings. One of the image
compression techniques in the spatial domain is Fractal Image
proposed by Metropolis could be used for optimization
Compression but the main drawback of FIC is that it involves more problems, where the objective function to be minimized
computational time due to global search. In order to improve the corresponds to the energy of the states of the solid. Since then,
computational time along with acceptable quality of the decoded SA has been applied to many optimization problems occurring
image, HGASA technique has been proposed. Experimental results in areas such as computer design, image processing and job
show that the proposed HGASA is a better method than GA in terms scheduling. There has also been progress on theoretical results
of PSNR for Fractal image Compression. from a mathematical analysis of the method, as well as many
computational experiments comparing the performance of SA
Keywords—Fractal Image Compression, Genetic Algorithm, with other methods for a range of problems.
HGASA, Simulated Annealing. Even though few investigations have been carried out
involving GA and SA for fractal image compression [9], but
I. INTRODUCTION till now a hybrid technique involving both GA and SA has not

C OMPRESSION and decompression technology of digital

image has become an important aspect in the storing and
transferring of digital image in information society. Most
been proposed for FIC. Hence this paper proposes a hybrid
technique which involves incorporation of SA into GA in
order to avoid pre-mature convergence of strings.
of the methods in use can be classified under the head of lossy The remainder of the paper is organized in detail as follows:
compression. This implies that the reconstructed image is Section (II) focuses on the theory of transformations and
always an approximation of the original image. Fractal image fractal image compression technique. In Section (III) the
coding introduced by Barnsley and Jacquin [1-4] is the concept of GA is explained. Section (IV) explains the
outcome of the study of the iterated function system necessity of incorporating SA into GA. Section (V) explains
developed in the last decade. Because of its high compression the concept of SA and its incorporation into GA based FIC.
ratio and simple decompression method, many researchers The experimental results and discussions are furnished in
have done a lot of research on it. But the main drawback of Section (VI). In Section (VII) some conclusions are drawn.
their work can be related to large computational time for
image compression. At present, researchers focus mainly on II. FRACTAL IMAGE COMPRESSION
how to select and optimize the classification of the range The fractal image compression algorithm is based on the
blocks, balance the speed of compression, increase the fractal theory of self-similar and self-affine transformations.
compression ratio and improve the quality of image after
decompression [5]. Especially in the field of reducing the A. Self-affine and Self-similar Transformations
complexity of search, many outstanding algorithms based on In this section we present the basic theory involved in Fractal
classified search have been proposed. Image Compression. It is basically based on fractal theory of
GA is a search and optimisation method developed by self-affine transformations and self-similar transformations. A
mimicking the evolutionary principles and chromosomal self-affine transformation W : R n o R n is a transformation
processing in natural genetics. Especially GA is efficient to
of the form W ( x) T ( x)  b , where T is a linear
solve nonlinear multiple-extrema problems [6-8] and is
transformation on R n and b  R n is a vector.
Y.Chakrapani is with the Department of ECE, G.Pulla Reddy Engineering A mapping W : D o D, D Ž R n is called a contraction on D
College, Kurnool-518002, Andhra Pradesh, India (phone: 09849595014;
email: yerramsettychakri@ gmail.com). if there is a real number c, 0  c  1 such that
K.Soundera Rajan is with the Department of ECE, J.N.T.U College of
Engineering, Anantapur, Andhra Pradesh, India.

International Scholarly and Scientific Research & Innovation 2(11) 2008 3780 scholar.waset.org/1307-6892/6811
 World Academy of Science, Engineering and Technology
International Journal of Computer and Information Engineering
Vol:2, No:11, 2008

d (W ( x ), W ( y )) d cd ( x, y ) for x, y  D and for a metric d on Where F is the invariant set for the wi and d is the Hausdorff
R n .The real number c is called the contractivity of W . metric. As a consequence of this theorem, any subset R n can
d (W ( x),W ( y )) cd ( x, y ) then W is called a similarity. be approximated within an arbitrary tolerance by a self-similar
A family {w1 ,....wm } of contractions is known as Local set; i.e., given G ! 0 there exist contacting similarities
{w1 ,....wm } with invariant set F satisfying d ( E , F )  G .
Iterated function scheme (LIFS). If there is a subset F Ž D
Therefore the problem of finding a LIFS {w1 ,....wm } whose
such that for a LIFS {w1 ,....wm }
m attractor F is arbitrary close to a given image I is equivalent
F  wi ( F ) (1) § m ·
i 1 to minimizing the distance d ¨¨ I ,  wi ( I ) ¸¸ .
Then F is said to be invariant for that LIFS. If F is © i 1 ¹
invariant under a collection of similarities, F is known as a C. Fractal Image Coding
self-similar set.
The main theory of fractal image coding is based on iterated
Let S denote the class of all non-empty compact subsets of
function system, attractor theorem and Collage theorem.
D . The G -parallel body of A  S is the set of points within
Fractal Image coding makes good use of Image self-similarity
distance G of A, i.e. in space by ablating image geometric redundant. Fractal
International Science Index, Computer and Information Engineering Vol:2, No:11, 2008 waset.org/Publication/6811

AG ^x  D : x  a d G , a  A` (2) coding process is quite complicated but decoding process is

Let us define the distance d ( A, B ) between two sets A, B to very simple, which makes use of potentials in high
compression ratio.
be d ( A, B ) inf ^G : A  BG š B  AG `
The distance function is known as the Hausdorff metric on S .
We can also use other distance measures.
Given a LIFS {w1 ,....wm } , there exists an unique compact Wi ( D j )
invariant set F , such that
m Dj
F  wi ( F ) , this F is known as attractor of the system.
i 1
If E is compact non-empty subset such that wi ( E )  E and Range Blocks Domain Blocks
m
W (E)  wi ( F ) (3) Fig. 1 Domain – Range block transformations
i 1
We define the k-th iteration of W , W k (E ) to be Fractal Image coding attempts to find a set of contractive
W 0 (E ) E, W k (E) W (W ( k 1) ( E )) . For K t 1 then we transformations that map (possibly overlapping) domain cells
onto a set of range cells that tile the image. The basic
have that
f
algorithm for fractal encoding is as follows
k 1. The image is partitioned into non overlapping range cells
F  W (E) (4)
i 1 ^Ri ` which may be rectangular or any other shape such as
The sequence of iteration W k (E ) converges to the attractor triangles. In this paper rectangular range cells are used.
of the system for any set E . This means that we can have a 2. The image is covered with a sequence of possibly
family of contractions that approximate complex images and, overlapping domain cells. The domain cells occur in variety
using the family of contractions, the images can be stored and of sizes and they may be in large number.
transmitted in a very efficient way. Once we have a LIFS it is 3. For each range cell the domain cell and corresponding
easy to obtain the encoded image. transformation that best covers the range cell is identified.
If we want to encode an arbitrary image in this way, we will The transformations are generally the affined
have to find a family of contractions so that its attractor is an transformations. For the best match the transformation
approximation to the given image. Barnsley’s Collage parameters such as contrast and brightness are adjusted as
Theorem states how well the attractor of a LIFS can shown in Fig 1
approximate the given image. 4. The code for fractal encoded image is a list consisting of
information for each range cell which includes the location
B. Collage Theorm of range cell, the domain that maps onto that range cell and
If Let {w1 ,....wm } be contractions on R n so that parameters that describe the transformation mapping the
domain onto the range
wi ( x)  wi ( y ) d c x  y , x, y  R n š i , One attractive feature of fractal image compression is that it is
resolution independent in the sense that when decompressing,
Where c  1 . Let E  R n be any non-empty compact set.
it is not necessary that the dimensions of the decompressed
Then
image be the same as that of original image.
m 1
d ( E , F ) d d ( E ,  wi ( E )) (5)
i 1 (1  c)

International Scholarly and Scientific Research & Innovation 2(11) 2008 3781 scholar.waset.org/1307-6892/6811
 World Academy of Science, Engineering and Technology
International Journal of Computer and Information Engineering
Vol:2, No:11, 2008

III. GENETIC ALGORITHM Step 4: Next the process of reproduction is carried out on the
Genetic algorithms are procedures based on the principles strings to create a mating pool
of natural selection and natural genetics that have proved to be Step 5: The process of crossover and mutation is also carried
very efficient in searching for approximations to global out on the strings with probabilities of 0.8 and 0.05
optima in large and complex spaces in relatively short time. respectively.
The basic components of GA are: Step6: After the termination criteria is met with, the value of
x Representation of problem to be solved string with minimum fitness function value is
considered as optimum value.
x Genetic operators ( selection, crossover, mutation)
x Fitness function
IV. NECESSITY OF INCORPORATION OF SA
x Initialization procedure
GA starts by using the initialization procedure to generate the The performance of GA can be improved by introducing
first population. The members of the population are usually more diversity among the strings so that pre-mature
strings of symbols (chromosomes) that represent possible convergence can be eliminated. This can be achieved by
solutions to the problem to be solved as shown in Fig 2 replacing weaker strings i.e. the strings having low fitness
11010 1001001 value with better strings i.e. strings having higher fitness
0 1 0 ------ 0 0 1 0
value. Simulated Annealing may be used for this purpose.
International Science Index, Computer and Information Engineering Vol:2, No:11, 2008 waset.org/Publication/6811

x1 x2 x3 xn
V. HYBRID GENETIC-SIMULATED ANNEALING TECHNIQUE
Fig. 2 String representing ‘N’ variables
Simulated Annealing method follows the cooling process of
Each of the members of the population for the given molten metals through annealing. At high temperature, the
generation is evaluated and according to its fitness value, it is atoms can move freely and these movements get restricted to
assigned a probability to be selected for reproduction. Using shape gradually a structure of crystal as the temperature is
this probability distribution, the genetic operators select some reduced slowly. If the temperature is not reduced properly, the
of the individuals. By applying the operators to them, new system may end up in a polycrystalline state, which may have
individuals are obtained. The mating operator selects two higher energy state than the crystalline state. Based on this
members of the population and combines their respective annealing process, this algorithm begins with an initial point
chromosomes to create offspring as shown in Fig 3. The and a high temperature T. A second point is created at random
mutation operator selects a member of the population and .in the vicinity of the initial point and the difference in the
changes part of the chromosome as shown in Fig 4. function values 'E at these two points is calculated. If the
second point has a smaller function value, the point is
Parent A 0 0 0 0 0 1 0 1 accepted; otherwise the acceptance of that point is measured
Parent B 1 1 1 0 1 0 0 1 by Boltzmann probability distribution
In order to incorporate SA into GA, the strings resulted
Child A 0 0 1 0 1 0 0 1 from GA after performing the mutation operation in every
Child B 1 10 0 0 1 0 1 generation are modified by SA. New strings are generated
from current strings via a perturbation mechanism. The
Fig 3: Binary crossover similarity of the range block with the domain blocks computed
from new string and current string are calculated. If the
Child A 0 1 0 1 1 domain block constructed from the new string matches closely
with the range block than the current string domain block,
then the new string is carried to the next generation. If the new
New Child A 0 11 1 1 string is a deteriorated string, the probability of acceptance
P (T ) is calculated as, P (T ) exp  'E / T where 'E is the
Fig. 4 Binary mutation difference between mean squared error of current string
domain block and new string domain block when compared
The algorithm for the genetic algorithm can be represented as with the range block. T is the current temperature at which
follows the new solution is generated. A random number between 0
Step 1:As the genetic algorithm takes pairs of strings, we and 1 is then generated. If the value of P (T ) is greater than or
create a random number of strings depending upon our equal to the random number, the new string is passed to the
necessity and also note down their decoded values next iteration; otherwise, if P (T ) is less than random number,
along with setting a maximum allowable generation
number tmax the move is discarded and the current string is carried to the
Step 2:Using the mapping rule we next find out the next iteration. This completes one iteration and in the next
corresponding values of above created strings. iteration, the temperature is lowered and the acceptance or
Step 3: Using these values the fitness function values are rejection of the new strings selected by perturbation
found out. mechanism is decided by the same process.
The flowchart of HGASA based FIC is as shown in Fig 5.

International Scholarly and Scientific Research & Innovation 2(11) 2008 3782 scholar.waset.org/1307-6892/6811
 World Academy of Science, Engineering and Technology
International Journal of Computer and Information Engineering
Vol:2, No:11, 2008

Start

This range
No block covered
More range cells ?
within tolerance
Done
Yes No
Fit domain to
range(rotate,shri Fit Search
Yes
Select range block nk,scale);Keep within for best
from partition list track of best fit error domain
so far Tol ? ?

No Yes
International Science Index, Computer and Information Engineering Vol:2, No:11, 2008 waset.org/Publication/6811

Get domain block from

domain pool using End of domain Yes
HGASA No list ?

No
Yes
End of domain list ?

Fig. 5 Flow Chart of HGASA based Fractal Image Encoding

VI. RESULTS AND DISCUSSIONS This work is carried out in MATLAB 7.0 version and the
In this paper a gray level image of 256 u 256 size with 256 original image is classical 256 u 256 Lena and Barbara face
gray levels is considered. A Range block of size 4 u 4 and image coded with 8 bits per pixel. An optimal bit allocation
strategy for GA is as follows: 14 bits for the location of
Domain blocks of size 8 u 8 are considered. The domain
matched domain block (horizontal and vertical coordinate), 3
blocks are mapped to the range block by affine
bits for isomorphic types. For each of the range block fractal
transformations and the best domain block is selected using coding includes 17 bits allocation. During the iteration process
the above proposed technique. In the above technique the of the above proposed methods the gray level values beyond 0
individual chromosome is coded as shown in Fig 6 and 255 are replaced by average of its four neighbours to
11010101 10010011 010 avoid block diverging. Figs 7 and 8 show the reconstructed
images using FIC with GA and HGASA as search algorithms
x y along with the original images of Barbara and Lena. The
Domain block coordinates Isometry type various parameters of HGASA are shown in Table 1.The
Coding scheme of FIC using GA and SA is given in Table 2.
Fig 6: Binary code for searching domain block As seen from Table 2 the PSNR of the reconstructed image
using HGASA has a better value than the image reconstructed
The mean squared error (MSE) and PSNR considered in this through normal GA. Since the proposed technique involves
work are given by two evolutionary algorithms joined together, the
N Rows N Cols 2 computational time of HGASA is slightly more than that of
1
MSE= ¦ ¦ f i, j  d i, j (6) normal GA. But this can be compensated through the
N Rows N Cols i 1 j 1
improvement in the reconstructed image quality. So it can be
§ 255 2 · concluded that the proposed method can be used in all
PSNR = 10 log10 ¨ ¸ (7) applications of image compression where the quality of
¨ MSE ¸
© ¹ reconstructed image is demanding.
Where N Rows N Cols 256

International Scholarly and Scientific Research & Innovation 2(11) 2008 3783 scholar.waset.org/1307-6892/6811
 World Academy of Science, Engineering and Technology
International Journal of Computer and Information Engineering
Vol:2, No:11, 2008

(a) (b) (c)

Fig. 7 (a) original image (b) reconstructed image using GA (c) reconstructed image using HGASA
International Science Index, Computer and Information Engineering Vol:2, No:11, 2008 waset.org/Publication/6811

(a) (b) (c)

Fig. 8 (a) original image (b) reconstructed image using GA (c) reconstructed image using HGASA

TABLE I HGASA PARAMETERS

Range block Size 4*4 Population size 50
Crossover Probability Pc 0.8 Fitness f 1/(1+MSE)
Mutation Probability 0.05 Temperature 5000

TABLE II FIC CODING SCHEME COMPARISON USING HGASA AND GA

Image FIC with HGASA FIC with GA
Compression Ratio Barbara 6.73:1 6.73:1
Lena 6.73:1 6.73:1
PSNR (db) Barbara 28.89 28.34
Lena 28.86 26.22
Encoding Time (sec) Barbara 5580 4470
Lena 5390 4230

it can be seen that the PSNR is better in the case of decoded

VII. CONCLUSION image using HGASA over the one obtained by GA.
The concept of HGASA has been successfully applied to
fractal image compression of a gray level image. Instead of REFERENCES
traditional exhaustive search which takes large computational [1] A.E.Jacquin, ‘Image coding Based on a Fractal theory of Iterated
contractive Image Transformations,’ IEEE Transactions on Image
time in FIC the evolutionary computational technique like GA
Processing. vol 1, Jan 92, pp 18-30
is implemented which goes for random search thus reducing [2] A.E.Jacquin, “Fractal Image coding: A Review,” Proc. IEEE, vol 81, pp.
the computational time. In order to improve the performance 1451-1465, 1993.
of GA in terms of PSNR, a new technique which combines the [3] M.F.Barnsley, Fractals Everywhere. New York: Academic 1988.
concept of genetic algorithm and simulated annealing has [4] M.F Barnsley and A.E Jacquin, “Application of recurrent iterative
function systems to images”, Proc SPIE, 1001(3),122-131 (1998)
been proposed. Experimental results show that the proposed [5] Mingshui Li, Shanhu Ou and Heng Zhang, “The new progress in
HGASA scores over normal GA in the case of fractal image research approach of fractal image compression” journal of engineering
compression. Normally the PSNR ratio for a decoded image graphics, 4(3), 2004, pp 143-152 .
should be very high to have a better image. Based on Table 3 [6] Xiaoping Wang, Liming Cao, “Genetic Algorithms-theory, application
and software reliability Xi” An Jiao Tong University Press (2002).

International Scholarly and Scientific Research & Innovation 2(11) 2008 3784 scholar.waset.org/1307-6892/6811
 World Academy of Science, Engineering and Technology
International Journal of Computer and Information Engineering
Vol:2, No:11, 2008

[7] David Goldberg, “Genetic algorithms in search, optimization and

machine learning” Addison-Wesley Publishing Company: MA (1989).
[8] Kalyanmoy Deb, “Optimization for engineering design” Prentice Hall
of India,2000.
[9] Y.Chakrapani, K.Soundera Rajan, “A comparative approach to fractal
image compression using genetic algorithm and simulated annealing
technique” Asian Journal of Information technology 7(7), pp 285-289.
[10] M.Hassaballah, M.M.Makky, Youssuf B. Mahdy, “A fast fractal image
compression method based entropy” Electronic letters on computer
vision and image analysis, 5(I) 2005, pp 30-40.
[11] Zhang Chao, Zhou Yiming, Zhang Zengke, “Fast fractal image encoding
based on special image features” Journal of Tsinghua science and
technology, Vol12, No 1, 2007,pp 58-62

Y.Chakrapani is presently with the Department of

Electronics and Communications, G.Pulla Reddy
Engineering college, Kurnool, A.P, India. Currently
he is pursuing Ph.D at JNT University under the
International Science Index, Computer and Information Engineering Vol:2, No:11, 2008 waset.org/Publication/6811

guidance of Dr K.Soundera Rajan. His research

areas are image compression, image retrieval and
evolutionary algorithms.

Dr K.Soundara Rajan , born on April 1953, obtained

Post graduation from JNTU, Hyderabad and Ph.D from
Indian Institute of Technology, Roorkee. He has been
with JNT University for the past 22 years. Prof
K.Soundara Rajan is presently the Principal of JNT
College of Engg, Anantapur. He has published 45
papers in various reputed International and National
Journals. His areas of Interest are Microcomputers,
Instrumentation, Computer Operating systems and
Digital Image Compression. He received Best Teacher award in 2006 by the
State government of Andhra Pradesh. He is also member of many expert
committees.

International Scholarly and Scientific Research & Innovation 2(11) 2008 3785 scholar.waset.org/1307-6892/6811