Vous êtes sur la page 1sur 6

Distributed Fractal Image Compression on PVM

for Million-Pixel Images

Pou-Yah Wu
Department of Mathematics Education, National Tainan Teachers College, Tainan 700, Taiwan
E-mail :y??wu~,i),ipx.i~ti~tc.edu.
hv

Abstract measure the convergence, and Mandelbort had generated


images by the fractal theory [ I , 5, 12, 13, 231. Bamsley
In this paper, we propose the distributed fractal image developed a collage theorem and the iterated function
compression and decompression on the PVM system. We system (IFS), and was able to produce a method for fractal
apply the regional search for the fractal image coding, which greatly enhanced compression-ratios [ 1, 2,
compression to reduce the communication cost on the 3, 41. Jacquins improved on IFS by proposing the
distributed system PVM. The regional search is to search partitioned iterated hnction system (PIFS) which is able to
the partitioned iterated function system from a region of automatically determine the fractal code [8, 91. The main
the image instead of over the whole image. Because the disadvantage of previous techniques for fractal image
area surrounding of a partitioned block is similar to this compression is the large amount of computation that they
block possibly, finding the fractal codes by regional search require [6, 161.
has a higher compression ratio and less compression time. A number of more efficient methods have been
When implemented on the PVM, the fractal image proposed to achieve a higher compression-ratios and
compression using regional search reduces the smaller compression times. Efforts to increase the
compression time with less compression loss. When we compression ratio have been based on alternative methods
compress the image Lena with image size 1024 x 1024 of partitioning [ 171, altemative encoding methods [8, 9,
using the region size 512 x 512 on the PVM with 4 151, and heuristic search methods [18]: attempts to
Pentiuni 11-300 PCs, the compression time is 13.6 seconds. decrease compression time have also resorted to a variety
the compression ratio is 6.34 and the PSNR is 38.59. But it of techniques, including image classification [6], parallel
needs to take 176 seconds, have the compression ratio processing [ 19, 201, and processing using neural networks
6.30 and have PSNR 39.68 by the conventional fractal [IO, 1 I].
image compression. In addition, when the region size is Parallel processing, a method in which many small tasks
I 2 8 x I28, the compression time is 7.8 seconds, the are used to solve one large problem, has been facilitated by
compression ratio is 7.53 and the PSNR is 36.67. In the two major developments: the development of massively
future, we can apply this method to the fractal image parallel processors (MPPs) and the widespread use of
compression using neural networks. distributed computing. Distributed computing offers many
advantages; for instance, by using existing hardware, the
cost of this computing can be significantly reduced. In
1. Introduction addition, the virtual computer resource that is created by
distributed computing can grow in stages and can thus take
advantage of the latest computational and network
Today, distance diagnostic medical system, digital
technologies. As distributed computing and MPP have
library and Internet are applied popularly. One of the most
gained in popularity the notion of message passing has also
important problems in such applications is how to store become increasingly common. The Parallel Virtual Machine
and transmit images. Image compression reduces both
(PVM) system uses a message-passing model to allow
memory storage and transmission time [ 141. Previous
programmers to take advantage of distributed computing,
developed compression techniques include discrete cosine
using a wide variety of computer types, including MPPs. To
transform technique and pyramid coding. In this paper, we
be effective, distributed computing requires high
study a more recently developed method of compression
communication speeds [7].
technique, the fractal image compression.
As already mentioned by Fisher, the simplest way to
Fractal image compression is based on self-similarity
implement the fractal image compression with quadtree
property of an image, and performs image compression by
partition in parallel is to assign a piece of image to each
applying a series of transformations. These
processor element [6]. Uhl and Hammerle have previously
transformations are applied to perform image
studied some parallel methods for fractal image
decompression iteratively until the system converges.
compression using PVM [19, 201. We also studied that
Previous workers have used the Hutchinson metric to
fractal image compression using regional search had the

393
$10.00 0 2001 IEEE
0-7695-0951-7/01
minimum communication cost on a PVM system, but in Theorem 1.
these methods a part of an image in one node of the PVM
was considered as a region for searching the fractal code
[21]. In dealing with images consisting of more than one
million pixels, we applied the regional search to reduce the
compression time [22].
In this study, we parallelize fractal image compression
using regional search on a PVM. In Section 2, we present a
technique for sequential fractal image compression using
regional search. The distributed fractal image compression
and experimental results are discussed in Section 3 and Figure 1. Nonoverlapped search regions of regional search.
Section 4. Finally, Section 5 gives conclusions. Theorem 1: Assume an image G E P(A) be partitioned
into sufficiently large nonoverlapped search regions GIE
2. Fractal Image Compression with Regional P(AJ where l l i l s and W,={CDIJ=(pL,, T,)}J=l.2..m be a
Search partitioned iterated function system of GI. Then U w,is a
,=I

In this section, we develop a new search strategy for partitioned iterated function system of the image G.
fractal image compression. Searching for fractal codes is The encoding procedure for fractal image compression
one of the major tasks of fractal image compression and using regional search partitions the original image into
there are several directions that we may pursue as we seek search regions. For each search region, find a partitioned
to improve the compression process. First, we can design a iterated function system to represent the search region.
search method which will reduce the execution time of The union of these partitioned iterated function systems is
image compression. Secondly, we can seek to reduce the a partitioned iterated function system of the image. The
loss ratio. Finally, we can aim to improve the compression algorithm for fractal image compression is shown in
ratio. Figure 2.
In an image, the area which surrounds a partitioned Similarly, the algorithm for fractal image
block is naturally similar to the block. Therefore, a search decompression using regional search is shown in Figure 3.
which is prioritized by neamess may be quicker than a
search which begins from the upper left corner of the Input : An original image G
image. Prioritizing the search by neamess also improves Output : The compressed file of the image
the compression ratio at the same time. Although the Method :
compression method used in this study classifies reference Choose a partition number n and a tolerance level e,;
regions according to archetype, it searches the fractal codes Partition G into nonoverlapped search regions Go, G,, .. .,
from the same archetype class that is permuted from the G-1;
upper left comer to the right-down comer of the whole For each GJ do {
image. Set D , = GJand mark it uncovered;
However, assuming that we wish to compare the While there are uncovered destination regions D, do {
nearest reference region with the destination region each Find the reference region RI and the corresponding w,
time, we find that this requires a length execution time. In that best cover D,;
the regional search method, the original image is If d(D,, w,(R,)) <e, or size(D,) < rmlnthen
partitioned into sufficiently large nonoverlapped regions, Mark D, as covered, and write out the
as shown in Figure 1. In this method of fractal image transformation w,;
compression, the search is confined to similar reference Else
regions from the same search region of the destination Partition D, into smaller destination regions that are
region. Simultaneously, we also use the archetype marked as uncovered, and remove D, from the
classification in the regional search. list of uncovered destination range;
Assume an original image GE P(A) be partitioned into 1
sufficiently large nonoverlapped search regions G i e P(A,), 1
where 1Gls. Since Gi is sufficiently large, there is a Figure 2. Algorithm for the fractal image compression
partitioned iterated function system Wi={CDij=(p,J, using regional search.
Tij)}j=,,2.,,,,m
of G,, for any 1 l i S According to the regional
The encoding algorithm is appropriate to compress
search method, we need show that U wi is a partitioned images with million pixels. Because the size of search
i=
I
region is constant, the time complexity to find a partitioned
iterated function system of the image G, as shown in

394
iterated function system in a search region is O(c). Assume a decompressed file. The algorithm is shown in Fig. S .
that the size of the original image be wxh, the time
complexity of the encoding algorithm is O(wxh), as shown Algorithm of master program
in Theorem 2. Input : An original image G and a node number n
Output : The compressed file of the image G
Theorem 2: Assume that the size of the original image be Method :
wxh, then the time complexity of the encoding algorithm is Choose a tolerance level e,;
O(wx h). Partition G into s nonoverlapped search regions, then
decompose these regions into Goo,Go,,. .., Go,,-,,. ..,
Input : The fractal image compression code Gn-l.o,.. ., G,.,,,.,, where s = n x t;
Output : The image decompression file Broadcast the slave program to each child node;
Method : Broadcast the tolerance level and the size of search region
Choose any initial image Gowith search regions having the to each child node;
same size as the search regions in the original image as For each child j do {
the region of the reference image and an iteration send G,,,,, ,G,+,,,,..., G,+,,,.,to the child node j; }
number i,; For each Go, do {
For each search region do { Set D, = GOkand mark it uncovered;
Apply the fractal image code to the search region of the While there are uncovered destination regions Di do {
domain image iteratively until the iteration number is Find the reference region R,and the corresponding wi
1,;
that best cover D,;
1 If d(D,, wi(R,))<e, or size(Di) < rminthen
Output the transformed image as the image decompression Mark Di as covered, and record the transformation
file; w,;
Figure 3. Algorithm for fractal image decompression using Else
regional search. Partition Di into smaller reference regions that are
marked as uncovered, and remove Di from the
3. Distributed Fractal Image Compression on list of uncovered destination regions;
PVM }
1
For each child j do {
Next, we parallelize the regional search method of Receive the transformations from the child node j; }
fractal image compression by using PVM. Since the issue Write out the transformations;
of performance on PVM is primarily concerned with Exit PVM;
communication overhead, the main objective of our design Algorithm for the slave program at node j
is to reduce communication costs. Method :
Fractal image compression partitions the original image
Receive the tolerance level e, and the size of the search
into s search regions and sends search regions to nodes by
region from the master;
region decomposition. The region decomposition divides
the work based on different search regions. For example, Receive Gj+,o, . .., G,,,,., from the master;
consider the fractal image compression of s search regions, For each Gj+l,kdo {
GI, G2, ..., G,. If we assume that n processes are working Set D , = G,+l,kand mark it uncovered;
on this compression, region decomposition involves the While there are uncovered destination regions Di do {
allocation of s/n search regions to each process. Each node Find the reference regions Ri and the corresponding
performs conventional fractal image compression on s/n withat best cover Di;
search regions. Then, the fractal codes for each search If d(D,, w,(Ri))<e, or size(Di) < rminthen
region are sent to the master node and are merged to a Mark Di as covered, and record the transformation w,:
compressed file. The algorithm is shown in Fig. 4. Else
Similarly, the master program for the regional search Partition Di into smaller destination regions that are
method of the fractal image decompression sends the marked as uncovered, and remove Ri from the list
fractal codes for search regions to child nodes. Then, the of uncovered destination regions;
master and the slave node apply the fractal codes for a }
given search region to a region of the same size in any }
initial image iteratively. After the decompression of all the Send the recorded transformations to the master;
search regions is complete, the slave program sends the Exit PVM;
decompressed regions to the master and they are merged to Figure 4. Algorithm for fractal image compression using
regional search on PVM.

395
Algorithm for the master program that obtained by the conventional compression. Moreover,
Input : The fractal image compression code and an iteration the vision quality may be poor and the encoding time is
number i, long. But if the area which surrounds a partitioned block is
Output :The image decompression file similar to the block, the destination region can find the
Method : similar reference region under the tolerance level without
Broadcast the slave program to each child node; more partitions. The compression ratio obtained by the
Broadcast the search region size and the iteration number regional search method will larger than that obtained by
to each child node; the conventional compression, the vision quality will be
For each child node do { tolerant and the encoding time is short. In this section, we
Send the fractal codes for search regions to the child use the experimental results to illustrate that the area
node;} which surrounds a partitioned block is similar to the block.
For each search region do { Our PVM system consists of 4 Pentium 11-300 PCs
Choose any initial image Go in which search regions are connected by 100Mbits/sec Ethemet. The original image is
the same size as search regions of the reference Lena's image with size 1024x1024. The sizes of search
image ; regions are 1024x1024, 512x512, 256x256, 128x128,
Apply the fractal image code for a search region of the 64x64 and 32x32. Due to 4 nodes in the PVM system,
initial image to the corresponding region of the search regions are partitioned into a 2x2 mesh by the
reference image iteratively until the iteration number region decomposition. But when the size is 1024x 1024, we
is i,; use the conventional fractal image compression on one
Record the transformed search region of the image; node. The decompressed images using different region
} sizes are shown in Figure 6. Encoding times, compression
Receive the transformed search regions of the image from ratios and PSNRs (Peak Signal Noise Ratio) are shown in
each child node; Table 1; where
Output the transformed image as the image decompression Compression ratio = (the file size of original
file; image)/(the file size of compressed image), and
Exit PVM; PSNR = 20xlog,,(B/rms),
Algorithm for the slave program at node j B is the largest gray value (255) and rms is the root mean
Method : square of gray values of original image and decompressed
Receive the search region size and the iteration number image.
from the master; First, we observe the vision quality of decompressed
Receive the fractal codes on the respective regions from the images using different region sizes in Figure 6. When sizes
master;
For each search region do { are not smaller than 128x128, the vision quality is equal to
that of the image decompressed by the conventional fractal
Choose any initial image Goin which search regions are
image compression. As the size of search region is 64x64,
the same size as search regions of the reference
if we compare the decompressed image with the original
image ;
image then we can find the edge of the hat has fuzz. In the
Apply the fractal image code for a search region of the
initial image to the corresponding region of the decompressed image using the search region size 32x32,
reference image iteratively until the iteration the edge of the hat has fuzz clearly. It means that the
number is i,; regional search at least can find the tolerant similar
reference region.
1 Secondly, the less the size of search region, the quicker
Send the transformed search regions of the image to the
master; the encoding and the larger the compression ratio, as
Exit PVM; shown in Figure 7 and Figure 8. Though reference regions
for search decrease, the destination region can find the
Figure 5. Algorithm for fractal image decompression using
regional search on PVM. similar reference region under the tolerant level without
more partitions. This says that the area which surrounds a
partitioned block is similar to the block.
4. Experimental Results Finally, when the size of search region decreases half
each side, the value of PSNR reduces 1, as shown in
Because the number of reference regions in a search Figure 9. The small compression loss can save many
region is smaller than that in the whole image, the execution times. Hence, images with more than million
destination regions may be partitioned into smaller regions pixels will need the regional search method to reduce the
for finding the most similar reference regions. It is encoding time and increase the compression ratio with the
possible that the compression ratio will be smaller than small compression loss.

396
myw 5M!2 EFZ mn Iwz

Figure 8. The relation between compression ratios and


region sizes

Figure 6. Using regional search


method to compress Lena image
with size 1024x1024, where the 41
region size is below the 40
decompressed image. 39
33
I 32x32 I
Table 1. Encoding times, compression ratios and PSNRs of
%
a
35
34
Lena's image using the regional search.
33
he search region size 1024 512 256 128 64 32 32-
31
/The encoding time (sec.) I 176 113.6(10.217.81 3.9 15.1 I
IThe compression ratio
I I I I I I II
6.30 6.34 6.85 7.5 8.59 10.00

lpsm 139.7 138.6137.7136.135.31 34.0


Figure 9. The relation between PSNRs and region sizes

5. Conclusions

This paper has presented the distributed fractal image


compression using regional search on the PVM. This
method is especially appropriate to images with more than
million pixels. We shall apply this method to fractal image
compression using other partitions. In addition, we will
research to parallelize the fractal image compression using
neural networks.

6. References

M. F. Bamsley, Fractal eveiydiere. New York,


[I]
Figure 7. The relation between encoding times and region Academic Press, 1992.
sizes. [2] M. F. Bamsley and A. D. Sloan, "A better way to
compress images," Byte, vo1.13, pp. 215-223,1988.

397
M. F. Bamsley, A. Jacquin, L. Renter and A. D. Sloan, model for fractal image coding," IEEE Trans. on
"Harnessing chaos for image synthesis," Computer Image Processing, vo1.6, no.3, pp.373-382, 1997.
Graphics, v01.22,n0.4, pp.131-141, 1988. [16] S. F. Reddaway, "Fractal graphics and image
S . Demko, L. Hodges and B. Nayloy, "Construction compression on a DAP," International Specialist
of fractal objects with iterated function systems," Seminar on the Design and Application of Parallel
Computer Graphics, vo1.19, no.3, pp.271-278, 1985. Digital Processors ( IEEE Conference Publication,
K. J. Falconer, The geometiy of fractal sets, 298 ) , London: Institute of Electrical Engineers 201,
Cambridge University Press, 1985. 1988.
Y. Fisher, Fractal image compression: Theorem and [17] E. Shusterman and M. Feder, "Image Compression
Application, New York, Spring-Verlag ,1994. via improved quadtree decomposition algorithm,"
A. Geist, A. Beguelin, J. Dongarra, W. Jiang, R. IEEE Trans. on Image Processing, vo1.4, no.6,
Manchek and V. Sunderam, PVM: Parallel Virtual pp.207-215, 1994.
Machine-A Users' Guide and Tutorial for Networked [ 181 L. Thomas and F. Derrai, "Region-based fractal
Parallel Computing, The MIT Press, Cambridge, image compression using heuristic search," IEEE
Massachusetts London, England, 1994. Trans. on Image Processing, ~01.4,no.6, pp.832-838,
A. E. Jacquin, "Image coding based on a fractal 1995.
theory of iterated constructive image [19] A. Uhl and J.Hammerle, "Image compression on a
transformations," IEEE Trans. on Image Processing, workstation cluster using PVM," Proceedings of the
vol. 1, no. 1, pp. 18-30, 1992. third European PVM Users Meeting, Euro PVM'96,
A. E. Jacquin, "Fractal image coding: a review," Lecture Notes in Computer Science, Springer-Verlag,
Proceeding ofthe IEEE, vo1.81, no.10, pp.1451-1465, 1996.
1993. [20] A. Uhl and J.Hammerle, "Fractal Image Compression
[lO]S. L. Lee, P. Y. Wu and K. T. Sun, "Fractal Image on MlMD Architectures I: Basic Algorithms,"
Compression Using Neural Networks," IEEE Parallel Algorithms and Applications, vol. 11, pp.187-
International Joint Conference on Neural Networks, 204, 1997.
Anchorage, Alaska, May 5-9, pp.613-618, 1998. [21]P. Y. Wu, "Minimum Communication Cost Fractal
[ I l l S . J. Lee, K. T. Sun and P. Y. Wu, "Fractal Image Image Compression on PVM," Proceedings of the
Compression and Decompression by Neural Network sixth European PVM Users Meeting, Euro PVM'99,
Approaches," Proceedings of I998 International Lecture Notes in Computer Science, Springer-Verlag,
Computer Symposium, Workshop on Artifical 1999.
Intelligence, pp.54-61, 1998. [22]P. Y. Wu, "Fast Fractal Image Compression,"
[ 121Ning Lu, Fractal Imaging, Academic Press, 1997. International Conference on Information Technology:
[13] B. Mandelbort, The fractal geometry of nature, San Coding and Computing ZTCC2000, Las Vegas,
Francis0 CA: freeman, 1982. Nevada, USA, 2000.
[ 141 M. Nelson, The data compression book, 2"ded.,M&T [23] G. Zorpette, "Fractal: Not just another pretty picture,"
books, New York, 1996. IEEE Spectrum, vo1.25, pp.29-3 1, 1988.
[15] D. C. Propescn, A. Dimca and H. Yang, "A nonlinear

398

Vous aimerez peut-être aussi