unequal protection of jpeg2000 code-streams in wireless channels ambarish natu & david taubman...
TRANSCRIPT
Unequal Protection of JPEG2000 Unequal Protection of JPEG2000 Code-Streams in Wireless ChannelsCode-Streams in Wireless Channels
Ambarish Natu & David Taubman
School of Electrical Engineering & Telecommunications
The University of New South Wales, Australia
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
2
Introduction• JPEG2000 suited for Wireless Image Transmission
– Better quality at lower bit rates compared to its predecessors.
– Error Resilience tools provided within the standard.• Option to include RESYNC Markers
• Error concealment and error localizing tools.
• Partition compressed data into independently decodable elements.
• Objective– Development of unequal error protection schemes for
JPEG2000 compressed imagery. – Optimize JPEG2000 coding parameters.– Maximize image quality in the presence of random bit errors.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
3
Previous Work• Hamming Codes used to provide unequal error
protection to JPEG2000 code-stream (1999)
• Turbo codes also proposed. (2000 & 2002)– Both approaches do not consider the problem of
optimizing JPEG2000 coding parameters.– Do not consider application of different levels of
protection to different quality layers in the code-stream.
• RCPC codes proposed by Z.Wu, A.Bilgin & M.Marcellin (ICIP’02).
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
4
Wireless Channels
• Wireless Models– Bit Level
• Each bit may be corrupted.
– Packet Level• Data is partitioned into packets
– Each packet received or lost
• We restrict our work to bit level errors, assuming a memoryless error process.– Characterized only by BER.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
5
• RS codes over Galois Field GF( ) are of particular interest to us. – For the present study we work with GF(16).
• Each symbol is a nibble
• Investigate the use of (15,7), (15,9), (15,11) and (15,13) RS Codes
– Simpler to decode than turbo or convolutional codes– Loss of multiple consecutive bits ( ) is rarely
worse than loss of single bit in J2K.
m2
Reed-Solomon (RS) Codes
km
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
6
Error Resilience in JPEG2000
• Code-Blocks,Precincts and Packets:
LLLL22
LHLH22HHHH22
HLHL22
HLHL11
HHHH11LHLH11
Precinct in the highest resolution
embedded code-block bit-streams
Packets
Precinct in the next lowest resolution
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
7
Error Resilience in JPEG2000 (ctd.)
• JPEG2000 Packets (not network packets)– Each packet consists of a packet head and a packet
body• Incremental contributions from code-block bit-streams
belonging to the relevant precinct.
• To extract code-block bit-stream contributions from packet body
– Must correctly decode the header of that packet and all preceding packets from the same precinct.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
8
Error Resilience in JPEG2000 (ctd.)• Quality Layer Contribution
– Layers • The first layer is a collection of all first packets from each precinct in
the image. The second layer consists of the second packet from each precinct and so forth.
• Effect of Layering– Layer bit-rates may be set by J2K compressor
• More smaller layers. – More packets per precinct.
– Less information in each packet header
– Less likely to lose whole precinct if corrupted.
– More significantly, can assign different levels of protection to layers
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
9
Error Resilience in JPEG2000 (ctd.)• Error Concealment
– ERTERM• Decoder can exploit predictable termination to detect and
conceal errors in code-block bit-streams.
• Additional ER Tools– SEGMARK
• Four-symbol code inserted immediately before the first new coding pass in each magnitude bit-plane.
– If an error occurs in the preceding 3 coding passes there is 1 in 16 chance that it will go undetected.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
10
Impact of Existing Error Resilience Tools
•Resync Markers and Error Concealment are very useful tools•Multiple quality layers of little benefit when multiple precincts are employed.
Resync Markers & Error Concealment
0
5
10
15
20
25
30
35
40
45
0 0.0001 0.001
BER
PS
NR
(dB
)
Precinct Sizes and Number of Layers
15
20
25
30
35
40
0 0.0001 0.001
BER
PSNR
(dB)
Concealment & Resync
No Concealment & Resync
No Concealment & No Resync
No of Layers=1, Single Precinct
No. of layers=1, Precinct Size
{256,128,64}
No. of layers=6, Precinct Size
{256,128,64}
No of Layers=6, Single Precinct
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
11
Uniform Error Protection
15
20
25
30
35
40
0 0.0001 0.001
BER
PS
NR
(d
B)
• Almost 9 dB improvement in image quality with (15,9) code at both BER compared to the existing ER tools
• 4 dB loss in image quality under noiseless condition for (15,9) code
• Disadvantage:– All elements protected
equally
(15,9) code
(15,13) code
Existing error resilience tools
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
12
Unequal Error Protection• Code-stream organized into 6 quality layers
with cumulative bit-rates of 0.03125, 0.0625, 0.125, 0.25, 0.5 and 1.0 bits per sample– Error sensitivity increases from lower to higher
quality layers.
• Key factor is spacing between layers– 2 layers for each factor of 2 change in cumulative
bit-rate i.e.11 layers in all. – Finer layer spacing gives little further improvement.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
13
Unequal Error Protection (ctd.)
Scheme-A: Layer (0,1 & 2) protected with (15,9) code
Layer (3 & 4) protected with (15,11) code
Layer (5) protected with (15,13) code.
Scheme-B: Layer (0) protected with (15,7) code
Layer (1,2 & 3) protected with (15,9) code
Layer (4 & 5 ) protected with (15,11) code
26
28
30
32
34
36
38
0 0.0001 0.001
BER
PS
NR
(d
B)
Scheme-A
Scheme-B
Uniform FEC
(15,9) code
2 layers for each factor of
two changes in bit-rate
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
14
Result Interpretation
• Unequal Protection across quality layers is of significant benefit– Improvement in noiseless compression performance
• Strongest codes used to protect only initial quality layers , which contain many fewer data bytes than later layers.
– Simple codes robust to BER conditions. – Little impact on error-free conditions.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
15
Unprotected Vs Protected JPEG200 Code-stream
Unprotected JPEG2000 code-stream (using only existing ER tools) for BER of
Protected JPEG2000 code-stream (using (15,9) RS code) for BER of
a) PSNR: 17.64 dB
BER:
c) PSNR: 26.24 dB
BER:
b) PSNR: 23.50 dB
BER:
d) PSNR: 34.34 dB
BER:
410
410
310
310
410 310and
and 310
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
16
Other Questions
• Precincts are of significant benefit when FEC codes are not used to protect JPEG2000 code-stream.
• Interestingly, multiple precincts do not help when combined with equal or unequal error protection. – Additional cost for independently coding each
packet header and aligning packet on a whole codeword boundary.
Natu, TaubmanUNSW
GLOBECOM’02, November 17-21Taipei, Taiwan
17
Summary• Resync markers, error concealment, layering &
precincts improve error resilience when FEC codes are not used to protect JPEG2000 code-stream. – Layering largely irrelevant unless unequal error protection
employed.
• Unequal Protection of quality layers definitely beneficial – Use Octave bit-rate spacing.
• 2 layers per octave offer some help at lower bit error rates.
• Multiple precincts of little benefit when RS codes are used to protect compressed data.