P1252109244ROZBA

1
April 30th, 2003
Parallel Design of JPEG2000 Image Compression
Xiuzhen Huang
CS Department UC Santa Barbara
2
Outline

• Introduction to image compression
• JPEG2000 compression scheme
• Parallel implementation of JPEG2000
• On distributed-memory multiprocessors
• On shared-memory multiprocessors
• Conclusion

3
Introduction to Image Compression
Why do we need image compression?
File size of a small digital photo without
compression
1280 ? 800 ? 3 (RGB) 3 M bytes
800 pixels
To speedup the image transmission over
Internet and reduce image storage space, we need
compression
1280 pixels
4
Introduction to Image Compression
Original Picture 3 M bytes
JPEG2000 Compression 19 K bytes

• Compression Ratio gt150 times !
• No noticeable difference in picture quality

5
JPEG2000 International Standard
JPEG2000 the new international standard for
image compression, is much more efficient than
the old JPEG international standard. For the same
compression ratio / bit rate / file size, the
JPEG2000 picture has much better quality.
JPEG
JPEG2000
Original Picture
Compression ratio 501
Strong blockiness
6
JPEG2000 International Standard
JPEG2000 has a much Higher computational
complexity than JPEG, especially for larger
pictures.
Need parallel implementation to reduce
compression time.
7
JPEG2000 Compression Scheme
Major steps of JPEG2000 image compression
Wavelet Transform
Blockwise Partition
Coding of each block
Input
Binary Compressed data

• Wavelet transform uses most of the image
  compression time (gt80)
• parallel implementation should focus on wavelet
  transform

8
JPEG2000 Compression Scheme
Brief Introduction to Wavelet Transform
Step 1 Horizontal wavelet transform of an image
for each row do 1-D wavelet transform end
What is 1-D wavelet transform ?
9
JPEG2000 Compression Scheme
A simple example 1-D Haar wavelet transform
Low- Frequency coefficients
Low-pass filter
Average of neighboring pixels
1, 1
2
First half of the output
One array of image data
Down-sample by 2
High- Frequency coefficients
Difference of neighboring pixels
1, -1
2
Second half of the output
high-pass filter
High
Low
Horizontal Wavelet Transform of Each Row
10
JPEG2000 Compression Scheme
Wavelet Transform
Step 2 Vertical transform of image
for each column of the new image do 1-D wavelet
transform end
11
JPEG2000 Compression Scheme
Low
High
Horizontal Wavelet Transform of Each Row
Vertical Wavelet Transform of Each Column
Low Low
High Low
High High
Low High
12
Parallel Design of JPEG2000 Compression
Two Parallel Computing Architectures
Distributed-Memory Multiprocessors

• Each processor has its own memory module
• Processors communicate to each other over a
  high-speed network
• Programming tool MPI (Message Passing Interface)

Shared-Memory Multiprocessors

• Has a single address space.
• Allow processors to communicate through variables
  stored in a shared address space
• Programming tool openMP

13
Parallel Implementation of JPEG2000
Compression on Distributed-Memory Multiprocessors
14
Parallel Design of JPEG2000 Compression-DMP
Traditional Approach

• The image is first divided into n regions on
  rows.
• Each processor performs 1-D horizontal wavelet
  transform
• Then, the new image is divided into n regions on
  columns.
• Each processor performs 1-D vertical wavelet
  transform.

This approach requires intensive data
transmission among processors, has very high
network communication cost.
15
Parallel Design of JPEG2000 Compression-DMP
Tiling Approach
P1
P2
P3

• JPEG2000 international standard supports
  tile-based image compression.
• A large image is divided into several tiles and
  each image tile is compressed independently.

P5
P4
P6
P8
P7
P9
16
Parallel Design of JPEG2000 Compression-DMP
Choose MPI for parallel implementation of
JPEG2000, because the JPEG2000 software is
written in C, which supported by MPI. Basic
framework is
17
Parallel Design of JPEG2000 Compression-DMP
Image 512x512
Size 32
Compression Time (Sec)
Size 256
Number of processors
The picture shows the compression time using
different tile size. For each tile size,processor
number increases,compression time is reduced.The
small tile need larger computation overhead.
18
Parallel Design of JPEG2000 Compression-DMP
Note

• There is a jump between one process and two
  processes.
• When there is only one process, JPEG2000
  compression is sequential
• If there are more than two processes involved in
  the program, Process 1 is responsible for
  collecting data, while the others are responsible
  for processing different tiles and sending
  processed data back to the Process 1.

19
Parallel Implementation of JPEG2000
Compression on Shared-Memory Multiprocessors
20
Parallel Design of JPEG2000 Compression-SMP
A problem with tile-based approach
Images compressed by JPEG, JPEG2000, and JPEG2000
with relatively small tiles.
Each tile is compressed independently, which
causes discontinuity across tile edges, also
called blockiness.
21
Parallel Design of JPEG2000 Compression-SMP

• Another parallel architecture is shared-memory
  multiprocessors.
• The excellent price-performance ratio of
  Intel-based SMPs make such systems very popular
  in many data processing applications.
• There are also many available programming tools
  for shared memory processor, such as openMP and
  Java Threads.

22
Parallel Design of JPEG2000 Compression-SMP

• In SMP, we do not need worry about data
  communication over network, because the data is
  in the shared memory. So there is no need for
  tile partitioning.
• Therefore, we can use the traditional data
  partitioning approach for horizontal and vertical
  wavelet transforms.

23
Parallel Design of JPEG2000 Compression-SMP

• JPEG2000 image compression is implemented on a
  4-processor SMP system using direct openMP.
• The speedup in wavelet transform is only about
  1.6 times, which is supposed to be near 4 times.
• Why?

24
Parallel Design of JPEG2000 Compression-SMP
It is found that the vertical wavelet transform
requires more than 10 times the horizontal
transform. But we know that both vertical and
horizontal transforms have the same number of
operations.
vertical
horizontal
25
Parallel Design of JPEG2000 Compression-SMP
Cache Miss Problem

• In computer memory, the image data is stored line
  by line in a raster-scan order (from left to
  right, from top to bottom).
• Each continuous block of image data is brought
  into the cache from memory for wavelet transform.
• In horizontal wavelet transform, as the filter
  window is moving, the data of next transform is
  often available, few cache miss.

26
Parallel Design of JPEG2000 Compression-SMP
Cache Miss Problem
data

• In vertical wavelet transform, the filtering is
  done in the vertical direction, however, the data
  is brought into cache in a horizontal way. So,
  there are very frequent cache miss.

filtering
Solution
Do vertical transform of several columns at the
same time to make full use of the existing data
in the cache. , instead of column by column
Significantly reduces cache miss.
27
Parallel Design of JPEG2000 Compression-SMP
The vertical transform is speed up by about 10
times.
Original Vertical transform
Improved Vertical transform
28
Parallel Design of JPEG2000 Compression-SMP
Using the improved vertical wavelet transform,
the overall speedup times of wavelet transform is
now close to the number of processors.
29
Conclusion

• Give a brief review JPEG2000 image compression.
• Discussed two approaches for parallel
  implementation of JPEG2000 image compression
  distributed memory multiprocessor and shared
  memory multiprocessor.

Question?