ARTICLE IN PRESS

Signal Processing: Image Communication 18 (2003) 925945

Multiple description image and video coding

for wireless channels
Manuela Pereira*,1, Marc Antonini, Michel Barlaud
I3S Laboratory of CNRS, University of Nice-Sophia Antipolis, Algorithmes/Euclide B, 2000 route des Lucioles BP-121 06903
Sophia Antipolis Cedex, France

Abstract
We consider the problem of efcient image/video transmission over wireless channels. Such a problem involves good
compression rates and effectiveness in presence of channel failures. In this work, we use the multiple description coding
(MDC) techniques, based on Wavelet Transforms, that have been shown to be powerful against channel failures. We
propose a bit allocation procedure that dispatches the source redundancy between the different channels when
compressing to a target bit rate with a bounded side distortion. In this way, we develop an MDC scheme well adapted
to channel noise. Then, we extend this bit allocation to video. This extension uses the 3D scan-based DWT of Parisot
et al. (MMSP, IEEE, New York, 2001); that involves scan-based MDC with rate or quality control. In this work we
propose to adapt redundancy between the descriptors in function of channel model and state (BER). Furthermore, the
channels can have time-varying states. We consider BSC channels, Gaussian channels, and UMTS channels. The
method can be efciently extended to other channels when the channel model that matches the channel behavior is
known.
We evaluate the performance of the proposed MDC image/video coder for two descriptions. In a rst group of
simulations, we compare our application with some different MDC techniques presented by Vaishampayan (IEEE
Trans. Inform. Theory 39 (3) (1993) 821) and Servetto et al. (ICIP, IEEE, Chicago, USA, 1998; IEEE Trans. Image
Process. 9 (5) (2000) 813826). In a second group of simulations, we evaluate the robustness of the proposed MDC
when submitted to different kinds of channels noise.
r 2003 Elsevier B.V. All rights reserved.
Keywords: Joint source channel coding; Multiple description coding; Bit allocation; Adaptability to channel model and state; Error
resilience video coding; Video over 3G wireless systems

1. Introduction
*Corresponding author.
E-mail addresses: pereira@i3s.unice.fr (M. Pereira), am@
i3s.unice.fr (M. Antonini), barlaud@i3s.unice.fr (M. Barlaud).
URLs: http://www/i3s.unice.fr/pereira, http://www.i3s.uni
ce.fr/barlaud.
1
On leave from Universidade da Beira Interior, Portugal.
Research partially supported by PRAXIS XXI Grant SFRH/
BD/1234/2000.

The use of mobile communication and multimedia communication knew an enormous increase
in the latest decade, being the wireless channels
considered as a transport medium for various
types of multimedia information. However, the
scarcity of wireless bandwidth, the time-varying
characteristics of the channel, and the power

0923-5965/$ - see front matter r 2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.image.2003.08.009

ARTICLE IN PRESS
926

M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

limitations of wireless devices, imposes tremendous challenges on wireless multimedia communications. The present work involves image and
video transmission over wireless channels.
Given the bandwidth limitations, compression
will be required for image and video transmission
over wireless channels. Wireless video transmission is specially difcult because the huge volume
of data required to describe a video greatly slows
down transmission and then, involves the use of
lossy compression at low bit rates. The transmission delay is very important since information that
arrives too late at decoder is considered as lost.
Moreover, data are corrupted by channel noise.
The channel noise can occur in the form of
random bit errors, bursty bit errors or packet
losses. Both problems result on losses of perceptual quality at decoder. This made robust compression schemes very useful especially for
transmission at low bit rates.
1.1. Error control techniques or multiple
description coding (MDC)?
The classic technique to combat transmission
errors is forward error correction (FEC)
[5,6,8,27,34,35,38]. FEC involves the addition of
redundant data to the compressed signal, which
allows the decoder to correct errors up to a certain
level. This redundancy increases the total number
of bits required and thus reduces compression.
Moreover, FEC code must be designed with a
worst case channel scenario in mind. For channels
that have a highly variable quality, this worst case
may imply the need for a very powerful code, and
hence highly or even prohibitive amount of
redundancy, which will severally reduce the
compression performance. In the case of burst
errors, the error correction capability is often
exceeded or the block is error-free in which case
additional redundancy is wasted. To overcome this
limitation, FEC is often enhanced by a technique
known as interleaving. For burst errors, this
effectively reduces concentration of errors in single
code words, more precisely, a burst of b consecutive symbol errors causes a maximum of b=M
symbol errors in each code word. Thought
interleaving can be implemented with low com-

plexity it suffers from increased delay, depending

on the number of interleaved blocks M: Therefore
interleaving is a frequently used technique for
bursty channels if additional delay is acceptable.
These problems can be solved if used unequal loss
protection [20].
Closed-loop error control techniques like Automatic Repeat reQuest (ARQ) have been shown to
be more effective than FEC and successfully to
wireless video transmission [17]. Retransmission of
corrupted data frames, however, introduces additional delay. Moreover, it is possible to combine
FECARQ to be successfully to wireless video
transmission [4,12,18].
The frameworks for wireless communication
should join optimization of source coding and
channel coding, should present great robustness
and adaptability to adverse transmission condition
and should make efcient use of limited wireless
network resources. In this way, a particular joint
source and channel coding method, known as
MDC, has proven to be an effective way to
provide error resilience with a relatively small
reduction in compression ratio. This coding
scheme assumes that there are several parallel
channels between the source and destination, and
that each channel may be temporarily down or
suffering from long burst errors. Furthermore, the
error events of different channels are independent
so that the probability that all channels simultaneously experience losses is small. These channels
could be physically distinct paths between the
source and destination in, for example, a wireless
multipath network or a packet switched network.
Even when only a single physical path exists
between the source and destination, the path can
be divided into several virtual channels by using
time interleaving frequency division, etc.
In the MD problem (reduced to the simplest
case of two descriptions), a source is described by
two descriptors with side rates R1 and R2 : These
two descriptions individually lead to reconstructions with side distortions D1 and D2 ; respectively;
the two descriptions together yield a reconstruction with central distortion D0 pD1 (and D2 ).
Almost all multiple description codes to date
assume the existence of multiple independent
onoff channels between the transmitter and the

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

927

receiver (e.g. Internet). When a link is broken, all

of the symbols or packets passing through that
channel are lost; when it is functioning properly,
the symbols are transmitted error free. Some
exceptions are presented in the following.
In [14] the authors replace the onoff channel
model with wireless channel, where they assume
that the system will employ multiple transmit and
multiple receive antennas. In this work, they
observe that it is possible to improve average
transmission error probability by a proper choice
of the correlating transform. In [3] they introduce
fading models, such as Rayleigh, Rician, or
Nakagami channel, within the context of MDC.
Their simulations show the efciency of MDC for
fading channels with multiple antennas.
Vaishampayan in [46] showed that for transmitting information from memoryless Gaussian
source over a Rayleigh fading channel, the multiple description approach results in good performances at low interleaving delays as compared to
standard channel coding approaches. This conclusion was extended to sources with memory in [15]
where, on an equal interleaving delay basis,
signicant performance improvements are obtained over channels codes for speech transmission
on Rayleigh fading channels.
The great amount of works dedicated to MDC
for onoff channels vs. the very reduced amount of
work dedicated to MDC for wireless channels is
explained by the relatively large overhead associated with MDC that implies that when channel
loss rate is small, the reconstruction performance
in the error-free case dominates and the single
description coding (SDC) perform best. So, as
cited in [45] A challenging task is how to design
the MDC coder that can automatically adapt the
amount of added redundancy according to underlying channel error characteristics. It is that
challenge we propose to respond in the present
work.

descriptions, using a generalized Lloyd-like clustering algorithm that minimizes the Lagrangian of the
rates and expected distortions R1 ; R2 ; D1 ; D2 ; D0
[31,32,4042]: In the second approach, pioneered by
Wang, Orchard, and Reibman [44], MD quantizers
are constructed by separately describing (i.e.,
quantizing and coding) the N coefcients of an N

N block linear transform, which has been designed
to introduce a controlled amount of correlation
between the transform coefcients [10,9]. In the
third approach, pioneered by Goyal, Kovacevic,
and Vetterli, MD quantizers are constructed by
separately describing the N coefcients of an overcomplete N
K tight frame expansion [10,11].
Other works were developed by Jiang et al. who
propose an MD extension to SPIHT coder
presented in [30], by separating Zerotrees into
polyphase components [13]. Rogers et al. propose
to rearrange bits at the output of one conguration
of the SPIHT coder, in such a way that the loss of
one packet results in an error that does not
propagate beyond the image region contained in
that packet [29]. Mohr et al. propose the use of
error correcting codes of different strengths applied
to different portions of a progressive bit stream
such as that generated by SPIHT coder [20].
Previous MD coding dedicated to video was
proposed by Vaishampayan in [43]. A predictive
MD system was applied along with transform
coding to construct an interframe balanced MD
video coder based on the H.263 standard. In [2],
Apostolopoulos and Wee show that MD coding
and path diversity provide improved reliability in
systems with multiple paths with equal or unequal bandwidths. In [28], Reibman et al. proposed MD video coders which use motion
compensated predictions. In [23], we propose an
extension of [25] for video that uses the 3D scanbased DWT.

1.2. Prior work on MDC

In this paper, we propose a joint source and

channel coding method that presents robustness
and adaptability to channel characteristics and
state. For its compression and synchronization
capabilities, it is suitable for real-time transmission. This is specially true for the video case.

We can nd three distinctly different approaches

of MD quantization. In the rst approach,
pioneered by Vaishampayan, MD scalar, vector,
or trellis quantizers are designed to produce two

1.3. Main contributions of the paper

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

928

The proposed method uses an MD scheme

based on the DWT and an efcient bit allocation
technique. Our goal is to nd an optimal trade-off
between efcient compression and robustness from
losses due to communications using unreliable
channels. The new bit allocation technique is
derived from [25] and adapted to the 3D scanbased DWT video coder presented in [23]. Here,
we propose to control automatically the amount of
redundancy dispatched on the different descriptions by taking into account the channel model
and state. The 3D scan-based DWT transform
allows us to develop a stripe-based MDC and so to
use different redundancies to take into account
changes in channels state while coding.
We show that the proposed MDC automatically
adapted to different channels and states, is efcient
for transmission over time varying channels. In
this perspective, the proposed method is an
alternative to methods that use error control
schemes, such as FEC or ARQ.
Our simulations are divided into two main
parts. The rst one serves to prove the effectiveness of our MDC scheme when compared to other
MDC schemes. The results presented show that
the proposed MDC overcomes previous MDC
schemes in the relation side PSNR vs. central
PSNR. The second group of simulations show the
robustness of the proposed MDC when submitted
to different kinds of channel noise. Simulations are
performed using the BSC channel, the Gaussian
channel, and the standard UMTS channel.

The paper is organized as follows: Section 2

presents the problem statement and is followed by
the proposed bit allocation for multiple description scheme in Sections 3 and 4. Section 5 presents
the proposed algorithm and Section 6 presents
concisely the channels models used in this work.
Finally, results are presented in Section 7. We
conclude and propose some future works in
Section 8.

2. Problem statement
Our MD scheme focus on the special case in
which there are two channels of equal capacity
between a transmitter and a receiver. In such a
scheme, a sequence of source symbols is given to
an encoder to produce two independent bitstreams
of equal importance. These bitstreams are transmitted to three decoders over two noisy channels.
One decoder (the central decoder) receives information sent over both channels while the remaining two decoders (the side decoders) receive
information only from their respective channel.
The amount of redundancy is dispatched on the
different descriptions by taking into account
the channel model and state (or bit error rate).
The general coder is presented in Fig. 1.
Given the side rate Rl and the side distortion
Dl ; the generation of the two descriptions is
constrained to some conditions that we detail in
the following.

( R l , Dl )

Sequence
Video

3D
Scan Based
DWT

( R 1 , D1)
MDC

( R 2 , D2)

Quality Control
or
Rate Control

bitstream

Side
Decoder

( R 1 , D1)

Central
Decoder

( R 0 , D0)

Side
Decoder

( R 2 , D2)

noisy channel

bitstream

BER

Fig. 1. 3D scan-based MDC scheme.

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

Condition 1. The central decoder has to reconstruct the original sequence from the two descriptors with minimal central distortion D0 :
Condition 2. An MDC coder must generate two
descriptors each with a side rate R1 R2 Rl :

This allocation problem is a constrained problem which can be solved by introducing the
Lagrange operators. The Lagrangian functional
for the constrained optimization problem, is
given by
Jfqi;1 ; qi;2 g D0

Condition 3. The side decoders must reconstruct

the original sequence from a single descriptor with
a side distortion D1 pDl and D2 pDl :

929

2
X

lj Rj pRl

j1

2
X

mj Dj pDl ;

j1

The problem is then to minimize the central

distortion D0 (Condition 1) when Conditions 2
and 3 are veried. That is, for # SB the number of
spatio-temporal subbands, we have to nd the
sets of bit rates fRi;1 g; fRi;2 g that minimizes the
central distortion D0 ; where Ri;j is the bit rate of
subband iAf1; y; #SBg for descriptor jAf1; 2g:
More precisely, we have to nd the set of
quantization steps fqi;1 g; fqi;2 g that produce the
sets of bit rates fRi;1 g; fRi;2 g: This problem is
known as the bit allocation problem we propose to
solve in Section 3.

where, for a source with generalized Gaussian

distribution, D0 can be written as [25]:


#SB
X
qi;1 qi;2
D0
Di s2i;0 Di;0
;
;
2
si;1 si;2
i1
where Di is an optional weight for frequency
selection.
The expected central distortion is estimated
based on the channels states and the a priori
channels models as we will see in the next sections.
3.2. Central distortion modeling

3. Proposed bit allocation for MDC

3.1. Proposed scheme for DWT
The problem is to nd, for a given redundancy
between the descriptors, which combination of
scalar quantizers across the various wavelet
coefcients subbands will produce the minimum
total central distortion while satisfying the side bit
rates, and side distortions constraints.
Then, the purpose of our bit allocation for MD
scheme is to determine the optimal sets of
quantization steps fqi;1 ; i 1; y; #SBg; fqi;2 ; i
1; y; #SBg; for descriptors 1 and 2. More
precisely, the bit allocation nds which combination of quantizers in the various subbands of the
two descriptors considered jointly minimizes the
central distortion D0 (Condition 1) for a bit rate
2Rl (Condition 2). This goal should be met while
the side distortion is kept below a given distortion
Dl (Condition 3). The parameters Rl and Dl are
given for the bit allocation (Fig. 1).

Recall that the central distortion is the distortion of the decoded image when using both
descriptors at decoding. When the decoder receives
both descriptors, each subband appears twice,
with different bit rates (different associated quantization steps). In this case, if the subbands are
noiseless, the central decoder chooses the subbands with the smaller quantization step and the
others, that we will call the redundant ones, are
only considered for side decoders. Hence, we can
calculate, the
central distortion of the decoded
P#SB
image as
i1 minDi;1 ; Di;2 ; where Di;1 ; Di;2
are the distortions of subband i for descriptors
1 and 2, respectively. In the general case we
have to take into account channel noise. So, we
cannot despise the redundant subbands as in
the noiseless case. Actually, the level of redundancy should increase when the BER increases,
such in the case of very noisy channels where the
redundant subbands are as important as the
others. In this case the central
P#SB distortion can
simply be written
as
i1 minDi;1 ; Di;2
P
maxDi;1 ; Di;2 #SB
D

Di;2 :
i;1
i1

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

930

Then, introducing a weighting parameter rN to

the redundant subbands, called redundancy parameter, we propose to write the central distortion
for a subband as




 
qi;1 qi;2
1
1
qi;1
2
Di;0
min si;1 Di;1
;
2
;
si;1 si;2
si;1
si;0 rN 1
 
qi;2
s2i;2 Di;2
si;2

 
 
qi;1
qi;2
2
2
rN
max si;1 Di;1
; si;2 Di;2
;
si;1
si;2
3
s2i;j Di;j qi;j =si;j

is the mean square error for

where
the ith subband, in the case of a generalized
Gaussian distribution.
Eq. (3) can be simplied as
8 2
 
si;1 1
qi;1
>
>
D
>
>
2 r 1 i;1 s
>
s
N
i;1
>
>
> i;0
 
>
2
>
>
si;2 rN
qi;2
>
>
Di;2
2
if mins2i;1 Di;1 ; s2i;2 Di;2
>
>
>
r

1
s
s
N
i;2
>
i;0
<
s2i;1 Di;1 ;
>
>
 
>
2
>
1
qi;2
> si;2
>
D
>
i;2
>
2
>
si;2
si;0 rN 1
>
>
>
 
>
2
>
>
si;1 rN
qi;1
>
>
otherwise:

D
:
i;1
si;1
s2i;0 rN 1

Then, we propose to use (5) to compute the

redundancy parameter:
rN

Hy x
;
maxpx Hx

where px stands for the distribution of the input

channel symbols.
Hy x is unknown at encoding; however, we can
nd some bounds for equivocation by stating the
following proposition.
Proposition 1.
minHy xp maxHx  Cp maxHy x;
px

px

px

where C is the channel capacity defined by C

maxpx Hx  Hy x:
Proof. Being C maxpx Hx  Hy x we can
infer that Cpmaxpx Hx  minpx Hy x:
Thus, minpx Hy xpmaxpx Hx  C: For the
right bound, we start also from the channel
capacity denition, C maxpx Hx  Hy x
and
we
infer
that
CXmaxpx Hx 
maxpx Hy x maxpx Hx  maxpx Hy x:
Thus, we can conclude that maxpx Hy xX
maxpx Hx  C:
&

4
The amount of redundancy, i.e., the importance of
the redundant subbands, depends on the channel
BER and we propose in the following some
strategies for the choice of rN :

Therefore, instead of (5), we use (7) to compute the

redundancy parameter.

3.3. Redundancy parameter

We
know
that
0pHy xpHx:
Thus,
0pCpmaxpx Hx: Using (7) we can conclude
that 0prN p1 as pretended.
We will show (in Section 6) that the channel
capacity is known for several interesting channel
models. Thus, more details about the calculi of rN
parameters for different channel models will be
done in that section.

Taking into consideration what we said above, it

is easy to conclude that the redundancy parameter
domain is [0, 1]. rN 0 when the channel is
noiseless and rN 1 when is expected a very noisy
channel. The problem is how to choose intermediate redundancies, and implicitly intermediate
values of rN : Taking into account the Shannon
theorem, [33, Theorem 10], we propose to solve
this problem using the equivocation Hy x: Indeed,
in this theorem, Shannon states that the equivocation Hy x is the amount of redundancy that the
decoder needs to correct the received message.

rN

maxpx HX  C
:
maxpx Hx

3.4. Constraints
Conditions 2 Rj1;2 pRl and 3 Dj1;2 pDl
have to be dened for each descriptor. For the
different descriptors j 1; 2; we write Condition 2

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

as a constraint F1 such that

!
 
#SB
X
qi;j
F1
ai Ri;j
 Rl ;
si;j
i1

where ai is the quotient of the size of the subband

divided by the size of the whole image (e.g., ai
1=22i in the dyadic case) and Ri;j qi;j ; is the bit rate
in bits per sample for the ith subband.
Condition 3 is forced using a penalty. The side
distortions D1 ; D2 are dened by
 
#SB
X
qi;j
Dj
Di s2i;j Di;j
; for all jAf1; 2g:
9
s
i;j
i1
The penalty method is simple and efcient.
Consider a constraint x > 0 the penalty is written
as Px jxj  x=22 : If the constraint is veried
then xX0 and Px 0: Otherwise, xo0 and
Px x2 : Considering the side distortions D1 and
D2 dened by (9), the constraint is Dj  Dl p0:
The penalty is then written as


jDj  Dl j Dj  Dl 2
P1
2
for all jAf1; 2g:

10

This penalty
allows us to nd a solution
P#SB function
2
with
i1 Di si;j Di;j qi;j =si;j pDl : In this case we
say that the penalty is veried.
3.5. Solution of the problem
Considering the central distortion given by (3)
the constraints (8) and (10), the Lagrangian
functional (1) can be rewritten as
Jfqi;1 ; qi;2 g


qi;1 qi;2
;
si;1 si;2
i1
!
 
2
#SB
X
X
qi;j

lj
ai Ri;j
 Rl
si;j
j1
i1


2
X
jDj  Dl j Dj  Dl 2

mj
:
2
2
j1
#SB
X

Di s2i;0 Di;0

931

detail the derivative with respect to qi;1 :

@Jfqi;1 ; qi;2 g
@qi;1


@
qi;1 qi;2
2
Di si;0
Di;0
;
@qi;1
si;1 si;2
 
@
qi;1
Ri;1
l1 ai
@qi;1
si;1


@ jD1  Dl j D1  Dl 2

0:
m1
@qi;1
2
2

12a
12b

12c

From Eq. (4), (12a) can be derived as

8
 
>
s2i;1 1
@
qi;1
>
>
D
>
i;1
2
>
>
r
@q

1
s
s
i;1
i;1
>
< i;0 N

if mins2i;1 Di;1 ; s2i;2 Di;2

>
>
 
>
>
s2i;1 rN
@
qi;1
>
>
D
>
i;1
: s2 r 1 @q
s
N
i;1
i;1
i;0

otherwise:

s2i;1 Di;1 ;

13
In this way, the derivation of the Lagrangian
functional (12) becomes
@Jfqi;1 ; qi;2 g
@qi;1
 
Ci;1
@
qi;1
2

Di si;1
Di;1
@qi;1
1 rN
si;1
 
@
qi;1
l1 ai
Ri;1
@qi;1
si;1


@ jD1  Dl j D1  Dl 2

0; 14
m1
@qi;1
2
2
where, from (13), Ci;1 is dened by
(
1
if mins2i;1 Di;1 ; s2i;2 Di;2 s2i;1 Di;1 ;
Ci;1
rN otherwise:
15

11

Solution
of
(11)
is
obtained
when
@Jfqi;1 ; qi;2 g=@qi;1 0;
@Jfqi;1 ; qi;2 g=@qi;2 0
and @Jfqi;1 ; qi;2 g=l 0: In the following, we

Furthermore, using (9), @=@qi;1 P1 can be written as

 
8
qi;1
>
>
0
if
D
pDl
1
>
>
s
>
i;1
>
>
<
@
2
D1  Dl Di s2i;1
>
@qi;1
>
>
 
>
>
q
>
>
:
Di;1 i;1
otherwise:
si;1
16

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

932

Then, Eq. (14) becomes

@Jfqi;1 ; qi;2 g
@qi;1
 
Ci;1
@
qi;1
2

Di si;1
Di;1
@qi;1
1 rN
si;1
 
@
qi;1
l1 ai
Ri;1
@qi;1
si;1
 
@
qi;1
2
m1 E1 Di si;1
Di;1
0;
@qi;1
si;1

we have
pa;s x aejbxj
with
1
b
s

17

where, from (16), E1 2
D1  Dl if D1 > Dl or

0 otherwise.
Then, Eq. (17) yields


 
Ci;1
@
qi;1
m1 E1 Di s2i;1
Di;1
@qi;1
1 rN
si;1
 
@
qi;1
l1 ai
Ri;1
0:
18
@qi;1
si;1
Simplifying (18) and performing similar calculus
for @qi;2 permit to obtain the system (19) (for
jAf1; 2g for a two-channels scheme.
 
@Di;j qi;j
l a
 j i
;
19a

Ci;j
2
@Ri;j si;j
Di s
m Ej
j

i;j 1rN

X#SB

aR
i1 i i;j

qi;j
si;j


 Rl 0:

19b

Resolution of system (19) which has #SB 1

equations and #SB 1 unknowns gives us the
optimal sets of quantization steps fqi;1 g; fqi;2 g:

@Di;j
q
*
@Ri;j

s
G3=a
G1=a

and

ba
:
2G1=a

For a given subband, the coder output bit rate R

produced by the quantization step q; can be
approximated by the entropy such that
N
X
R
Prmlog2 Prm;
mN

where
Prm

mqq=2

pa;s x dx

mqq=2

is the probability of the quantization level m:

According to [37], the best decoding value, when
using the mean squared error (MSE) as distortion
measure, for the quantization level m; is
R mqq=2
mqq=2 xpa;s x dx
x#
:
Prm
Then, the MSE can be expressed as
Z mqq=2
N
X
# 2 pa;s x dx:
x  x
D
mN

mqq=2

Setting q* qi;j =si;j ; it has been shown in [24,22]

that @Di;j =@Ri;j q
* can then be calculated as

@f1;m
@f ;m
*
*
*
*
*2 0
@q* a;qf1;m a;qf0;m a;qf1;m a;q @q* a;q
m1
*2
f0;m a;q
P @f0;m
*
pa;1 q=2
ln f0;0 a; q
* 1 N
*
f0;m a; q
*
m1 @q* a; qln
2

PN 2

The proposed algorithm is based on modeling of

R and D functions as we will show in next section.

4. Model-based R and D
In each subband the probability density function of the wavelet coefcients can be approximated with generalized Gaussian. Therefore,

ln 2

20

with
@fn;m
a; q
*
@q*

"



q*
pa;1 mq*
2

n1 

1
q*
 m
pa;1 mq* 
q* n :
2
2
m

1
2

n1

21

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

933

5. Proposed algorithm

5.2. @D=@R; R and D functions

5.1. Ci;j parameter

The functions in (19) are not easily invertible.

However, for generalized Gaussian model, it has
been shown in [22] that @D=@R; R and D functions
can be tabulated in order to simplify the inversion.
The tables for these functions can be found in
[22,24] for different generalized Gaussian shape
parameters a sampled in ]0, 2].

The Ci;j parameter dened in Eq. (19) depends

on s2i;1 Di;1 and s2i;2 Di;2 : One problem is that the
distortions involved are unknown before system
(19) is solved due to their dependence on the
quantization steps fqi;1 g; fqi;2 g: We propose an
algorithm that initializes the Ci;j parameters to 1
and iteratively modies their values according to
the current s2i;1 Di;1 and s2i;2 Di;2 : More precisely, if
we dene S as the set of all possible subbands of
descriptor 1 and 2, we search the subband in S
with the highest distortion. Then we set the
correspondent Ci;j value to rN : Note that in each
iteration the subbands of both descriptors such
that Ci;j was already set are excluded from the
search set S: It is easily veried that #SB
iterations are always performed. This algorithm
is detailed in Fig. 2.

Fig. 2. Computation of Ci;j parameter.

5.3. Global bit allocation procedure

As can be seen in Fig. 3, we compute Ri;j using
the given parameters Ci;j ; lj ; mj ; rN ; Eq. (19a) and
tabulated @D=@R functions. If the debit constraint
(19b) is not veried, we recompute the Ri;j using

Fig. 3. Global bit allocation procedure.

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

934

new lj : If it is veried we compute: qi;j using Ri;j

and the tabulated R functions; Di;j using tabulated
D functions; Dj using Di;j and (9); and nally Ci;j
using Di;j :
The above steps are performed #SB times. After
#SB iterations all Ci;j are computed and the
algorithm can proceed.
The last step of bit allocation is the verication
of the penalty (9). If this is not veried, we reiterate
from the beginning with a new m parameter. If the
penalty is veried, the algorithm stops and the
output of bit allocation gives the optimal quantization steps qi;j :

5.4. Bit allocation complexity

The complexity of our algorithm is the complexity of the EBWIC bit allocation presented in
[25,24]. The complexity of this coder is presented
in [24] and detailed in [21]. In this work, the
authors conclude that the highest cost of modelbased allocation method corresponds to the
computation of generalized Gaussian distribution
parameters for each subband. They need four
operations (two additions and two multiplications)
for wavelet coefcient to compute s and a
parameters. Assuming that, complexity of the
remaining part of the algorithm is lower than 1
operation for each image pixel. The authors
conclude that the complexity of such allocation
method is less than ve simple arithmetic operations for each image pixel.
We can then conclude that our bit allocation
complexity is less than ve operations for each
image pixel.

Wavelet

6. Channel model and redundancy

The two main elements which describe a channel
are the transmission rate and the channel capacity.
The transmission rate was dened by Shannon as
R Hx  Hy x for discrete or continuous channels. Where Hx is the entropy of the input and
Hy x the conditional entropy or the equivocation.
The channel capacity C is dened as the maximum of
R when the input varies over all possible collection
px: This channel capacity was dened by Shannon
for some special cases. The BSC and the Gaussian
cases. However, in most mobile radio systems, the
channel exhibits Rayleigh fading, aggravated by
typically log-normally distributed shadowing or slow
fading, resulting in a time variant channel capacity.
Lee [16] derived an estimate of the channel capacity in
Rayleigh fading environments.
In this paper we focus on binary symmetric channel
(BSC), Gaussian and Rayleigh channel models.
6.1. Binary symmetric channel
For the BSC case we have two possible symbols
each with a probability p of coming through
undisturbed, and 1  p of being changed into the
other of the pair. The capacity can be written as
C 1 p log2 p 1  plog2 1  p bits/symbol.
maxHx 1 for this channel model. Thus the
redundancy parameter presented in Section 3.3
can be estimated by
rN

maxpx HX  C
maxpx Hx

p log2 p 1  plog2 1  p:
Scalar
Quantization

Entropy
Coder

( R 1 , D1)

Scalar
Quantization

Entropy
Coder

( R 2 , D2)

( R l , Dl )
Bit
Allocation

coefficient
subband
BER

Fig. 4. General MDC scheme.

22

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

935

6.2. Additive white Gaussian noise channel

For a band-limited Additive White Gaussian
Noise (AWGN), the channel capacity can be
expressed as C=B log2 1 g bits/symbol, where
B is the channel bandwidth in symbol/s and g is the
signal to noise ratio (SNR). The SNR g is dened
as g S=N; where S is the received signal power
and N is the AWGN power within the channel
bandwidth. For this channel model, maxHx
depends on the modulation. For instance, if we
consider a QPSK modulation, maxHx 2 and
in this case the redundancy parameter in Section
3.3 is estimated by
maxpx HX  C 2  log2 1 S=N
rN

2
maxpx Hx
log2 1 S=N
:
23
1 
2

Fig. 5. Side PSNR vs. central PSNR

6.3. Rayleigh channel

In the case of Rayleigh models, an upper bound
approximation for the normalized channel capacity was introduced by Lees [16] as
C=BElog2 e e1=g e ln g 1=g
bits/symbol.
For this channel model, maxHx depends also
on modulation. For instance, if we consider a
QPSK, maxHx 2 and in this case the
redundancy parameter in Section 3.3 is written as
maxpx HX  C
rN
maxpx Hx
1 

log2 e e

N=S

e ln S=N N=S
: 24
2

7. Results
7.1. Scan-based DWT video coder
For spatial decomposition our coder uses 9-7
biorthogonal lter [1] and three level of decomposition. For temporal decomposition it uses the
(2,2) lter and performs two level of decomposition. The frames of the video sequence are
acquired and processed on the y to generate the
3D wavelet coefcients and the data are stored in

Fig. 6. BSC channel at BER 0:001: Image coded at 0:5 bpp:

PSNR 29:41 dB:

memory only until these coefcients have been

encoded [26]. No motion compensation is performed.
Fig. 1 shows the complete scan-based MDC
scheme. The bit allocation procedure, presented in
Section 3, is followed by a simple scalar quantization and the encoding of each subband uses a
context based arithmetic coder [22,39] as presented
in Fig. 4.
7.1.1. Smart arithmetic coding
In order to provide synchronization and minimize the error propagation in the case of channel

ARTICLE IN PRESS
936

M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

Fig. 7. Lena image coded at 1 bpp: BSC channel at BER 0:001 for left image and 0.01 for right image. Left image PSNR is 33:89 dB;
right image PSNR is 24:99 dB:

errors, each spatio-temporal subband is divided

into blocks. Then, arithmetic coding is performed
on each block independently. Since the block
division requires side information, the block-size
must be related to the channel BER such that high
BER implies small block-sizes. For error detection, when the number of coded coefcients is
known, it is possible to verify if the arithmetic
coder stops correctly. In case an error occurs and
the arithmetic coder is misplaced in the bitstream,
we synchronize the decoder to start at the
beginning of the next block.
It is well known that this usual method of error
detection is largely insufcient. It is not rare to
have the correct number of coefcients decoded
and the arithmetic coder well placed, but wrong
coefcients decoded. This problem is more difcult
to solve. Some methods try to use the information
of the neighbor blocks, but their effectiveness is
not optimal. Here, we propose to use some
internal information included in the arithmetic
code. The arithmetic coder works with two main
registers: the code string that represents the base of
the interval (lower bound) and the interval. With
each binary decision the current probability
interval is subdivided into two subintervals, and
the registers are modied in accordance. The states
of registers depend only on the probability of the
individual events. In our scheme, we include the n
last bits of the interval register inside the nal

bitstream (in our experiments n 8 was sufcient

for BERX0:01: The decoder simply veries if its
nal interval register correspond to the one
included by the coder inside the bitstream. If it is
not the case, it means an error occurs in the block
and then, it is discarded.
7.2. Simulations
For simulations, we use the 512
512 pixels
Lena image, 3 s of QCIF silent video and 3 s of
QCIF akiyo video. For Lena image, the total bit
rate was set to 1 bpp (i.e. Rl 0:5 bpp) or 0:5 bpp
(i.e. Rl 0:25 bpp). The silent video was compressed to 200 kbits=s (30 frames/s) and the akiyo
video was compressed to 300 kbits=s (30 frames/s).
All channel simulations were performed 10 times.
Note that mean PSNR values are computed by
averaging decoded MSE values and then converting the mean MSE to the corresponding PSNR
values. Visual results of video presents always the
frames 1, 11, 21, 31, 41, 51, 61, 71, 81 and 91 of the
video.
7.2.1. Central PSNR vs. side PSNR for still image
For 1 bpp central bit rate and Lena image,
central PSNR vs. side PSNR is plotted in Fig. 5 for
values of rN between 0 and 1. Our results are
compared to the referenced methods: [19]Ref.
[13]Ref. [31]Ref. 3. We present two different

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

937

results. In the rst one (New 1) bit allocation

procedure was performed with Cj f1; rN ; 1;
rN ; yg; resulting in a very simple MDC scheme.
In the second one (New 2), we found the best set of
Cj using the algorithm of Section 5. The New 2
method presented here provides the best results.

Fig. 8. Gaussian channel at BER 0:001: Image coded at

0:5 bpp: PSNR 31:27 dB:

7.2.2. Results for still image

Figs. 69 show some visual results of transmission of Lena image at 1 and 0:5 bpp over BSC and
Gaussian channels with BER 0:001 and 0.01.
The mean PSNR we obtained in the case of
0:5 bpp and BER 0:001 was 28:62 dB and for a
BER 0:01; it was 24:79 dB:
As mentioned in Section 1.1, we can nd great
amount of works dedicated to MDC for onoff
channels but very reduced amount of work

Fig. 9. Lena image coded at 1 bpp: Gaussian channel at BER 0:001 for left image and 0.01 for right image. Left image PSNR is
34:90 dB; right image PSNR is 25:49 dB:

Table 1
Mean PSNR (dB) results of QCIF silent video, coded at
200 kbits=s: Channel transmission at BER 0:01; same rN for
the two cases (with and without noise)
Channel

BER

BER

BER

2

10

2

10

Table 2
Mean PSNR (dB) results of QCIF silent video, coded at
200 kbits=s: Channel transmission at BER 0:001; same rN for
the two cases (with and without noise)
Channel

BER

BER
3

BER
3

10

10

33.17
31.54
31.54
31.54

31.50
30.08
31.45
31.45

40.72
39.11
39.11
39.11

40.00
39.07
38.99
39.04

103

41.51
40.09
40.09
40.09

41.15
40.09
40.04
40.060

2

10

UMTSIndoor
31.47 25.57 39.06 38.56 40.06 39.76
UMTSPedestrian 31.47 28.02 39.06 38.67 40.06 39.89
UMTSVehicular 31.47 26.26 39.06 38.58 40.06 39.68

Gaussian
UMTSIndoor
UMTSPedestrian
UMTSVehicular

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

938

Y
35

30

PSNR

25

20

15

10

0
1

7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94
Frame

Fig. 10. Mean PSNRs of each frame for Y component for Gaussian channel at BER 0:001; pink: no noise; dark blue: proposed
MDC; yellow: SDC; light blue: SDC TC:
U

45
40
35

PSNR

30
25
20
15
10
5
0
1

7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94

Frame

Fig. 11. Mean PSNRs of each frame for U component for Gaussian channel at BER 0:001; pink: no noise; dark blue: proposed
MDC; yellow: SDC; light blue: SDC TC:

dedicated to MDC for channels as BSC or

Gaussian. In [36] the authors constructed an
MDC for networks with packet lost and/or bit
errors. They provide a mechanism for bit allocation between the redundancy in terms of FECs and
redundancy that is meant to correct for packet
loss. In this work, they present same results with
Lena image compressed at 0:5 bpp: When BSC
transmission for a BER 0:001; the PSNR is
27:5 dB and for a BER 0:01 the PSNR obtained
is 20:2 dB: Showing that our MDC outperforms
this one. Moreover, the results of [36] were

obtained with seven levels of decomposition and

assuming the LL band uncorrupted.

7.2.3. Results for video

For video, we present results for Gaussian and
UMTS channels. UMTS channel2 presents three
different models: indoor, pedestrian, and vehicular. They are dened in [7].
2
The authors wish to thank France Telecom for providing a
UMTS simulator.

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

939

V
45
40
35

PSNR

30
25
20
15
10
5
0

Frame

Fig. 12. Mean PSNRs of each frame for V component for Gaussian channel at BER 0:001; pink: no noise; dark blue: proposed
MDC; yellow: SDC; light blue: SDC TC:

35

30

25

PSNR

20

15

10

94

91

88

85

82

79

76

73

70

67

64

61

58

55

52

49

46

43

40

37

34

31

28

25

22

19

16

13

10

Frame

Fig. 13. Mean PSNRs of each frame for Y component for UMTS Indoor channel at BER 0:01; pink: no noise; blue: proposed
MDC; yellow: SDC case.

To compare with, we present results obtained

with the proposed MDC when subject to noise or
not. We present also some results obtained when
transmission of a singular description coder (SDC)
with a similar codec. In this case the video is
compressed at 200 kbits=s: We also use an SDC
with a Turbo Coder SDC TC with a bit rate of
200 kbits=s including the channel rate.
In Tables 1 and 2 are presented the mean PSNR
for BER 0:001 and 0.01 for the proposed MDC.

We compare with the case without noise. We can

see that the performances of the proposed MDC in
the presence of noise are slightly the same than
without noise, especially for the U and V
components. For the Gaussian case and BER
0:001; same simulations with the SDC TC
simulator gives the following PSNR: Y PSNR
28:66 dB; UPSNR 36:50 dB and VPSNR
38:12 dB: We can see that the proposed MDC
provides a gain of 3 dB over a standard method.

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

940

Y
35

30

PSNR

25

20

15

10

94

91

88

85

82

79

76

73

70

67

64

61

58

55

52

49

46

43

40

37

34

31

28

25

22

19

16

13

10

Frame

Fig. 14. Mean PSNRs of each frame for Y component for UMTS Pedestrian channel at BER 0:01; pink: no noise; blue: proposed
MDC; yellow: SDC case.

Y
35

30

PSNR

25

20

15

10

94

91

88

85

82

79

76

73

70

67

64

61

58

55

52

49

46

43

40

37

34

31

28

25

22

19

16

13

10

Frame

Fig. 15. Mean PSNRs of each frame for Y component for UMTS Vehicular channel at BER 0:01; pink: no noise; blue: proposed
MDC; yellow: SDC case.

Figs. 10, 11 and 12 present the mean PSNR of

different frames for Y, U, and V, respectively,
when transmission over Gaussian channel with a
BER 0:001: We can conclude that the proposed
MDC and the SDC TC are able to recover from
channel losses. However, the MDC presents better
PSNRs since the SDC TC is designed for a
worst case transmission. In these gures, we can
see that the Y component is the most sensitive to

noise. In the rest of the paper, we present only

results for this component.
Figs. 1315 present the mean PSNR of different
frames for Y when transmission over UMTS
channel for the three different models, respectively. The BER in this case is 0.01.
Finally, we show some visual results in Figs. 16
and 17 for Gaussian transmission for BER 0:001:
These images allow to compare the proposed MDC

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

with the SDC TC: Fig. 18 represents silent video

for UMTS transmission for BER 0:01:

8. Conclusion
In this paper, we proposed an MDC scheme
for noisy time-varying channels. It includes

941

scan-based DWT and an efcient bit allocation

procedure that dispatches source redundancy
between the different channels. The amount of
redundancy is estimated based on the channel state
and the a priori channel model. The scan-based
DWT allows the development of a stripe-based
MDC that takes into account channel changes in
time.

Fig. 16. Gaussian channel at BER 0:001: Two left columns using the proposed MDC and two right columns with the SDC TC:

ARTICLE IN PRESS
942

M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

Fig. 17. Gaussian channel at BER 0:001: Two left columns using the proposed MDC and two right columns with the SDC TC:

We show that the proposed MDC is a

simple alternative for real-time video transmission
where methods that use error control schemes
such as FEC or ARQ are not suitable for delay
reasons.
With MDC, a long burst error or even the loss
of an entire description does not have a catastrophic effect, as long as not all bitstreams
experience failures simultaneously. Thus, one

could even use fewer error control bits for each

bitstream.

Acknowledgements
The authors want to acknowledge the anonymous reviewers for their advice which improved
the quality of the paper.

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

943

Fig. 18. UMTS Pedestrian channel at BER 0:01: Two left columns using the proposed MDC and two right columns with the SDC.

References
[1] M. Antonini, M. Barlaud, P. Mathieu, I. Daubechies,
Image coding using wavelet transform, IEEE Trans. Image
Process. 4 (8) (1992) 10531060.
[2] J. Apostolopoulos, S. Wee, Unbalanced multiple description video communication using path diversity, in: International Conference on Image Processing, IEEE, Greece,
2001.

[3] N. At, Y. Altunbasak, Multiple description coding for

wireless channels with multiple antennas, in: GLOBECOM, San Antonio, TX, 2001, pp. 20402044.
[4] V. Chande, N. Farvardin, H. Jafarkhani, Image communication over noisy channels with feedback, in: Data
Compression Conference, Snowbird, UT, 1999.
[5] Q. Chen, T. Fisher, Robust quantization for image coding
and noisy digital transmission, in: Data Compression
Conference, Snowbird, UT, 1996, pp. 312.

ARTICLE IN PRESS
944

M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945

[6] Q. Chen, T. Fisher, Image coding using robust quantization for noisy digital transmission, IEEE Trans. Image
Process. 7 (1998) 496505.
[7] ETSI (European Telecommunications Standards Institute), Selection procedures for the choice of radio
transmission technologies of the UMTS, tr 101 112
v3.2.0 Edition, April 1998.
[8] J. Garcia-Frias, J. Villasenor, An analytical treatment of
channel-induced distortion in entropy coded image subbands, in: Data Compression Conference, Snowbird, UT,
1997.
[9] V. Goyal, J. Kovacevic, Optimal multiple description
transform coding of gaussian vectors, in: International
Conference on Image Processing, Chicago, IL, 1998.
[10] V. Goyal, J. Kovacevic, R. Arean, M. Vetterli, Multiple
description transform of coding of image, in: International
Conference on Image Processing, Chicago, USA, 1998.
[11] V. Goyal, J. Kovacevic, M. Vetterli, Multiple description
transform coding: Robustness to erasures using tight
frame expansions, in: Proceedings of the International
Symposium Information Theory, Cambridge, MA, 1998,
p. 408.
[12] C. Hsu, A. Ortega, M. Khansari, Rate control for robust
video transmission over burst-error wireless channels,
IEEE J. Select. Areas Commn. 17 (5) (1999) 117.
[13] W. Jiang, A. Ortega, Multiple description coding via
polyphase transform and selective quantization, in: Proceedings of the SPIE Visual Communications and Image
Processing Conference (VCIP), San Jose, CA, 1999.
[14] N. Kamaci, Y. Altunbasak, R. Mersereau, Multiple
description coding with multiple transmit and receive
antennas for wireless channels: the case of digital modulation, in: GLOBECOM, San Antonio, TX, 2001,
pp. 32723276.
[15] K. Kintzley, An application of multiple description scalar
quantizers to speech coding on correlated fading channels,
Masters Thesis, Electrical Engineering Department, Texas
A&M University, College Station, TX, August 1995.
[16] W. Lee, Estimate of channel capacity in Rayleigh fading
environment, IEEE Trans. Veh. Technol. 39 (1990) 187189.
[17] S. Lin, D. Costello, M. Miller, Automatic repeat request
error control schemes, IEEE Comm. Mag. 22 (12) (1984)
517.
[18] H. Liu, M. Zarki, Performance of h263 video transmission
over wireless channels using hybrid ARQ, IEEE J. Select.
Areas Comm. 15 (9) (1997) 17751785.
[19] A. Miguel, A. Mohr, E. Riskin, Spith for generalized
multiple description coding, in: International Conference
on Image Processing, Vol. 3, IEEE, New York, 1999,
pp. 842846.
[20] A. Mohr, E. Riskin, R. Ladner, Graceful degradation
over packet erasures channels through forward error
correcting, in: Proceedings of the IEEE Data Compression
Conference (DCC), Snowbird, UT, 1999.
[21] C. Parisot, Allocations bases mod!eles et transforme en
ondelettes au l de 1eau pour le codage dimages et de

[22]

[23]

[24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

[32]

[33]
[34]

[35]

[36]

[37]

vid!eos, Ph.D. Thesis, University of Nice, Sophia Antipolis,

France, http://www.i3s.unice.fr/parisot/, January 2003.
C. Parisot, M. Antonini, M. Barlaud, Ebwic: a low
complexity and efcient rate constrained wavelet image
coder, in: International Conference on Image Processing,
Vancouver, Canada, 2000.
C. Parisot, M. Antonini, M. Barlaud, 3D scan-based
wavelet transform for video coding, in: MMSP, IEEE,
New York, 2001.
C. Parisot, M. Antonini, M. Barlaud, 3D scan-based
wavelet transform and quality control for video coding,
EURASIP Special Issue Multimedia Signal Process. 2003
(1) (2003) 5665.
M. Pereira, M. Antonini, M. Barlaud, Channel adapted
multiple description coding scheme using wavelet transform, in: International Conference on Image Processing,
IEEE, Rochester, NY, USA, 2002.
M. Pereira, M. Antonini, M. Barlaud, Multiple description
coding for noisy-varying channels, in: Proceedings of the
IEEE Data Compression Conference (DCC), Snowbird,
UT, USA, 2003.
D. Redmill, N. Kingsbury, Still image coding for noisy
channels, in: International Conference on Image Processing, Austin, TX, 94, 1994, pp. 9599.
A. Reibman, H. Jafarkhani, Y. Wang, M. Orchard, R.
Puri, Multiple description video coding using motioncompensated prediction, in: International Conference on
Image Processing, IEEE, Kobe, Japan, 1999.
J. Rogers, P. Cosman, Robust wavelet zerotree image
compression with xed length packetization, IEEE Signal
Process. Lett. 5 (5) (1998) 105107.
A. Said, W. Pearlman, A new, fast, and efcient image
coded based on set partitioning in hierarchical trees,
IEEE Trans. Circuits Systems Video Technol. 6 (3) (1996)
243250.
S. Servetto, K. Ramchandran, V. Vaishampayan, K.
Nahrstedt, Multiple description wavelet based image
coding, in: International Conference on Image Processing,
IEEE, Chicago, USA, 1998.
S. Servetto, K. Ramchandran, V. Vaishampayan, K.
Nahrstedt, Multiple description wavelet based image
coding, IEEE Trans. Image Process. 9 (5) (2000) 813826.
C. Shannon, A mathematical theory of communication,
Bell System Technical J. 27 (1948) 379423, 623656.
P. Sherwood, K. Zeger, Progressive image coding for
noisy channels, IEEE Signal Process. Lett. 4 (1997)
189191.
P. Sherwood, K. Zeger, Error protection for progressive
image transmission over memoryless and fading channels,
IEEE Trans. Comm. 46 (12) (1998) 15551559.
S. Somasundaram, K. Subbalakshmi, Exploiting path
diversity and forward error correction for robust transmission of images, in: PCS, Saint-Malo, France, 2003,
pp. 259262.
G. Sullivan, Efcient scalar quantization of exponential
and laplacian random variable, IEEE Trans. Inform.
Theory 42 (1996) 13651374.

ARTICLE IN PRESS
M. Pereira et al. / Signal Processing: Image Communication 18 (2003) 925945
[38] N. Tanabe, N. Farvardin, Subband image coding using
entropy coded quantization over noisy channels, IEEE J.
Select. Areas Comm. 10 (1992) 926943.
[39] D. Taubman, M. Marcellin, Jpeg2000: Image Compression
Fundamentals, Standards, and Practice, Kluwer International Series in Engineering and Computer Science,
Vol. 642, Kluwer Academic Publishers, Boston, MA, 2001.
[40] V. Vaishampayan, Vector quantizer design for diversity
systems, in: CISS, Baltimore, MD, 1991.
[41] V. Vaishampayan, Design of multiple description scalar
quantizers, IEEE Trans. Inform. Theory 39 (3) (1993) 821834.
[42] V. Vaishampayan, Design of entropy-constrained multiple
description scalar quantizers, IEEE Trans. Inform. Theory
40 (1) (1994) 245250.

945

[43] V. Vaishampayan, S. John, Interframe balanced-multipledescription video compression, in: Packet Video Workshop, Columbia University, NY, 1998.
[44] Y. Wang, M. Orchard, A. Reibman, Multiple description
image coding for noisy channels by paring transform
coefcients, in: IEEE 1997 First Workshop on Multimedia
Signal Processing, Princeton, NJ, 1997.
[45] Y. Wang, Q. Zhu, Error control and concealment for
video communication: a review, Proc. IEEE 86 (5) (1998)
974997.
[46] S. Yang, V. Vaishampayan, Low delay communication for
Rayleigh fading channels: an application of the multiple
description quantizer, IEEE Trans. Comm. 43 (1995)
27712783.