You are on page 1of 8

Noise Robust Coding and Iterative Decoding of

Multiple Description Image over Wireless Networks


Abstract: Motivated by recent results in multiple description image coding over wireless networks, we propose a scheme for
transmission of multiple descriptions through hybrid packet loss and additive white Gaussian noise channel. Balanced
multiple descriptions are generated by applying multiple description scalar quantizer on wavelet coefficients. Each
description is then coded into multiple bitstreams by applying !"#$ coding on wavelet trees along spatial orientations and
each stream is further compressed using arithmetic code. %se of error resilient entropy coding &E'E() is proposed in
literature for synchronization requirement of variable length codes, but E'E( is not compatible with iterative soft*in soft out
decoding of arithmetic code at the receiver. +e propose the application of E'E( assisted by state and tail bits &E'E($) in
con,unction with iterative decoding of arithmetic code at receiver for reconstructing the multiple description coded image
over packet loss and Gaussian noise channel. E-perimental results demonstrate that an additional gain of . dB in !/' is
obtained over e-isting scheme.
Kewords: E'E(, (hase*like decoder, Multiple description image coding, +ireless networks.
!" Introduction
Current advances in microelectronics and wireless
communication has enabled the development of
miniaturized, low-cost and low power computing
nodes in multihop wireless networks. Such networks
may be heterogeneous in nature and have application
in target tracking, habitat monitoring, camera sensor
network. Similarly, modern multimedia delivery
networks are hybrid in nature, i.e. consisting of in-part
packet switched network and one or more wireless
access points that serve the end user. One of the main
problems encountered in transmission of images over
such heterogeneous network is degradation of
reconstructed image quality due to packet loss. he
ob!ective of multiple description coding is to encode a
source into two or more descriptions in such a way that
the reception of an arbitrary subset of descriptions may
produce useful reconstruction, with quality of
reconstruction increasing with number of descriptions.
his has proved to be an effective way to combat the
degrading effect of packet loss.
One of the first multiple description image coder
"#$, consisting of an e%tension of &'(), was designed
based on multiple description scalar quantizer
*+,S-. "/$. 0nother group of methods for multiple
description coding use correlating transform are
proposed in "1-2$ and references therein. 3ecently Sun
et al. "4$ proposed a technique for multiple description
coding based on combining lapped transform, block
level source splitting and linear prediction.
'resently state-of-the-art image coding techniques
use wavelet transform based coding for image
compression and many multiple description image
coder has been proposed based on it. Servetto at al. "5$
proposed one of the first multiple description image
coder based on wavelet transform and demonstrated
improved performance over single description coding.
0 given image is hierarchically decomposed into
wavelet subbands and two balanced descriptions are
then obtained by applying multiple description scalar
quantizer *+,S-. on the subband coefficients. he
quantization indices of both descriptions are entropy
coded using arithmetic code and transmitted over
different channel. 6(C-based multiple description
coding was applied in "7$-"##$ for transmission of
progressively coded image over wireless channel.
6(C-based multiple descriptions have the advantage of
being able to generate arbitrary number of descriptions
from a progressive bitstream compared to source coder
based approaches. Song et al "#/$ proposed
improvement to wavelet transform based scheme of "5$
for packet loss and random error channel. he scheme
consists of S'89 coding of wavelet trees along spatial
orientation for each description which generates
multiple variable length bitstreams in contrast to
conventional S'89 coding which generates a single
progressively coded bitstream. 8n order to maintain
high compression efficiency and robust
synchronization (3(C "#1$ is applied to reorganize
these variable length bitstreams into fi%ed length slots
before transmission.
8n this research, our contribution lies in improving
the scheme of Song et al. "#/$ by adapting Chase-type
soft-input soft-output *S8SO. arithmetic decoder "#:$
for decoding of multiple descriptions transmitted over
additive white )aussian noise *0;)<. and packet
loss channels. (%isting schemes in literature suppose
the transmission media to be random bit error and
packet loss channel. his assumption is not necessarily
suitable for many binary-input continuous output
channels *e.g. 0;)< channel., where soft decision
output of channel decoder can be utilized by the source
decoder for further improving the 'S<3 performance.
0t the encoder S'89 coded bitstreams are arithmetic
coded to further increase compression efficiency
compared to the scheme proposed in "#/$ and use of
soft channel information at the receiver improves the
noise robustness of the arithmetic decoder. Combining
(3(C with iterative source-channel decoding of
arithmetic code is not quite straightforward as (3(C
requires instantaneous hard decision decoding of the
source symbols. Our innovation lies in applying
(3(C assisted by state and tail bits *(3(CS. "#5$
with arithmetic coded S'89 trees of variable lengths
for S8SO decoding of arithmetic codes. his enables
iterative source-channel decoding of (3(C coded
fi%ed length slots at the receiver and improve 'S<3
gain significantly.
he rest of the paper is arranged as follows. 8n
section /, we describe the proposed scheme and the
constituent blocks in the transmitting system in detail.
Section 1 details the proposed soft iterative source-
channel decoding method of arithmetic coded
bitstreams reordered by (3(CS and subsequent
reconstruction of received image. 8n section :, we
present several simulation results to compare the
performance of the proposed system. 6inally, we
conclude with section =.
#" $stem Model
;e present the proposed system as shown in 6igure #.
wo balanced are descriptions are obtained by
performing +,S- after decomposing the given image
hierarchically using discrete wavelet transform
*,;.. Coefficients of wavelet trees along spatial
orientation in each descriptions are partitioned "#=$
and each partition is S'89 coded independently. C3C
parity bits are generated for predefined fi%ed length of
each S'89 coded bitstream, and then arithmetic
coding is applied to each independent bitstream for
further compression. he variable length arithmetic
codes are reorganized into fi%ed length slots using
(3(CS for synchronization requirements.
Conversion to fi%ed length slots is required to form
uniform sized packets before transmitting over packet
switching network. 0t the output of the packet loss
channel, remaining available packets are coded using
recursive systematic convolutional code *3SCC. at
wireless access points *;0'. and transmitted.
he receiver tries to decode the packets received
through the wireless channel and reconstruct the image
by combining both the descriptions. >og-liklihood
ratio *>>3. value at the input is computed from the
received signal as
/
? /
i i
y = , where
i
y
is the
output of 0;)< channel with noise variance
/
.
his >>3 value is utilized by source-channel decoder
for iterative soft decoding of arithmetic code and
3SCC. Successfully recovered S'89 coded bitstream
are decoded and wavelet trees are combined to obtain
the image in wavelet domain. 6inally reconstructed
image is obtained by inverse discrete wavelet
transform *8,;..
6igure #. @lock diagram of the proposed multiple description scheme.
#"!" DW% and MD$&
8mage is first hierarchically decomposed into < level
,;. he number of levels < is chosen depending on
the required number of wavelet trees with the
coefficients in the lowest frequency subband being
root of those trees. o create two balanced
descriptions Aaishampayan "/$ proposed +,S-,
which consists of two stages. 8n the first stage,
wavelet coefficients in all subbands are uniformly
quantized with stepsize B to obtain integer valued
quantization indices ". he second stage, inde%
assignment maps the quantization inde% field to two
complementary and possibly redundant inde% pair
fields "
#
and "
/
.
;e consider two different cases of inde%
assignment, high redundancy assignment with /
diagonals *6igure /*a.. and lower redundancy
assignment with 1 diagonals *6igure /*b... he matri%
entries in the figure represents the quantization inde% ",
which are mapped to row and column indices "
#
and "
/
.
#"#" $'I(% Coding of Wavelet %rees
8t has been demonstrated in "#=$ that ,; coefficients
of an image can be divided into 0 groups and then
each of these groups can be independently quantized
and coded so that 0 separate embedded bitstreams are
created. 0s the number of groups 0 increases, the
resilience of the coded image to transmission error also
increases, but with some decrease in coding efficiency.
8n this paper, we consider each group consisting of
four spatial orientation trees with each wavelet tree
rooted at the lowest frequency subband. <odes of the
tree have either no descendents or four offspring
grouped in / / ad!acent coefficients. ;e utilize the
scheme presented in "#/$ where S'89 algorithm "#2$
is employed to encode independently i-th group of
coefficients in description , and generate variable
length bitstreams C / , # D E ,..., # , = , 0 i
,
i
. Since
each description is consisted of matri% of integer
valued quantization indices, S'89 coding is done on
each group of trees until bit-layer F to obtain integer
level accuracy.
#")" CRC and Arit*metic Coding
0fter S'89 coding of wavelet trees along spatial
orientation for each description, we append :-bit cyclic
redundancy check *C3C. to each of the 0 S'89
coded bitstreams. C3C of : bit length was used so as
to avoid unnecessary coding overhead. Since, all the
bitstreams are of different length and their embedded
nature *bits in the beginning of the bitstream are more
important than at the end., C3C is generated and
appended for the first l bits *l G #/5 bits in this paper.
in each bitstream. C3C check eliminates erroneous
wavelet trees from consideration which may have
significant impact on the 'S<3 of the reconstructed
image.
0rithmetic encoder "#4$ than maps each of the 0
bitstreams independently into binary strings ,
,
i
1
, ,..., # 0 i =
for description
C / , # D ,
. 0rithmetic
encoder has input symbol set consisting of three
symbols, namely, F, #, and (nd of @lock *(o@.. (o@
symbol is put at the end of
,
i
to mark the end of each
variable length bitstream. Since, S'89 is an efficient
compression algorithm, further compression of S'89
coded bitstreams by arithmetic code gives little
compression gain. he main ob!ective of arithmetic
coding is to provide robustness against noise, as will
be discussed in section 1.
#"+" ,R,C$% and 'acket -ormation
8n the earlier section, 0 arithmetic coded bitstreams of
variable length are generated for each description. 8f
these variable length blocks *A>@. are transmitted
consecutively, the resulting system is highly sensitive
to bit errors. One approach has been to append
synchronization code words to provide
resynchronization at the cost of e%tra added bits.
Otherwise, presence of a single error in the arithmetic
code may cause catastrophic decoding error. 0nother
approach for error robust transmission of variable
length codes is the method of (3(C "#1$, which
converts 0 A>@s into 0 fi%ed length slots *6>S..
(3(C has been applied in "#/$ for reorganizing
variable length blocks of S'89 coded bitstream
before transmission over bit flipping channel.
,ecoding (3(C in iterative source-channel receiver
is challenging because to detect the end of each A>@
in the 6>S, the A>@ subdecoder *i.e. S8SO arithmetic
decoder in our case. must be embedded into the (3(C
decoder and decision of end of block be made based
on some hard decision. Since, at the receiver (3(C
decoder lies in the iterative loop between source and
channel decoder, finding end of block based on soft
decision is difficult. his can be solved by applying
(3(CS proposed in "#5$ to the problem. (3(CS
isolates pure (3(C subdecoder from the effects of
A>@ subdecoder by coding A>@ lengths as side
*a. *b.
6igure /. *a. wo diagonal *, G /. and *b. three diagonal *, G
1. inde% assignment. 0n integer along the diagonal is mapped to
row and column indices "# and "/.
information *S8.. >et 2 be the total number of bits in 0
A>@s in a description and s
i
is the length of i th 6>S,
where,


0 i 0 2
0 2 i
0 2
0 2
i
s
<
<

=
. , mod*
. , mod* F
, ?
, ?
*#.
he encoding process consists of 0 stages. (ach
stage consists of filling up of 6>S and coding of A>@
lengths by emission of state and tail bits. 6igure 1
shows the sequence of steps to convert four A>@s to
four 6>Ss. 0t stage F *6igure 1*a.., as many bits as
possible are placed into corresponding 6>S. 0t stage n,
i-th A>@, *A>@*i. has b
i
remaining bits. searches m-th
6>S *s
m
slots left in 6>S*m.., where m G
. ., * mod* 0 n i +
and
. *n
is a predefined integer
sequence of length 0. 8f both A>@*i. has bits
remaining and 6>S*m. has blank slots, as many
remaining bits of A>@*i. as possible are placed in
6>S*m.. 0t each stage, one state bit is emitted. G #
if
m i
s b > , else S G F and
m
s
/
log tail bits are
output to code . *
i m
b s in binary. >et
. , min*
i m
b s t = . 3emaining bits and slots are
recalculated as
t b b
i i
=
and
t s s
m m
=
. 8n 6igure
1*b., after stage F, only 6>S*#. has bits left to be
placed and searches 6>S*/.. wo bits are placed in
6>S*/.. Since, no other A>@ contains unplaced bits,
the process moves to ne%t stage. 8n stage /, A>@*#.
places remaining three bits in 6>S*1., as shown in
6igure 1*d.. he 6>S are formed into packets and
transmitted over network. able 8 shows the S8 *state
and tail bits. generated for this e%ample. he S8 bits
associated with a 6>S are taken as a whole *SA>@.
and (3(C is applied to reorganize all SA>@ into fi%ed
length slots called S6>S. 8f b is the average A>@
length, it was proved in "#5$ that total fraction of S8
bits is less than

b b ? . log / *
/
+ , which in this
research is F.F/4. hus we can safely assume that S8
bits are transmitted with higher error protection with
little coding overhead and received error free.
S6>S is utilized at the receiver side to recover
variable length bitstreams back from 6>S. his is done
in two stages, consisting of subroutines 32B
reconstructor and 32B reconstructor. 3efer to "#5$ for
detailed algorithms for these two functions. 32B
reconstructor outputs data structure sseg m G Ds, ne-tC,
where s is the number of blank bits after the current
segment of 6>S*m., and ne-t points to the ne%t
segment of 6>S*m.. 8f no more sseg is associated with
current 6>S, a null is put in ne-t field. 6igure :*a.
shows sseg for the four 6>S formed in 6igure 1. Hero
in s field of sseg F indicates that no slots are left blank
in 6>S*F. after the current segment. 6or sseg #, s G -#
indicates that current A>@ fills up 6>S*#. completely
with some bits are remaining to be placed in other
6>S. 6>S*/. contains data segment from two different
A>@s as shown by two sseg. 0fter first segment, two
bit positions are left vacant and ne-t in sseg / points to
ne%t sseg. 8n the similar way sseg information is
generated for 6>S*1..
able #. State and tail bits after different stages in 6igure 1.
;ith the aid of sseg information, 32B reconstructor
algorithm rearranges bits in 6>S to original variable
length bitstreams. 32B reconstructor creates a data
structure called bseg i G
C , I , , D ne-t s s m
associated
with each A>@*i., where s and I s is the number of
blank bits in 6>S*m. before and after current segment
of A>@*i.. 6igure :*b. shows the content of bseg
structure for the given e%ample. 0t the beginning it is
assumed that all the A>@s at the receiver are empty
and are waiting to get bits from the 6>S. A>@*F. is
reconstructed with the aid of information in bseg F,
where bits are obtained from m G F-th 6>S. <umber of
bits left in 6>S*F. is four *s G :. before recovering the
A>@ segment. Since, after current segment number of
bits left is zero * F I = s ., all the bits of the 6>S*F. is
taken up to form A>@*F.. A>@*F. gets not more bits
from any 6>S because ne-t field in bseg F is null.
A>@*#. obtains four bits from 6>S*#., two bits from
6>S*/. and three bits from 6>S*1. as indicated by
bseg # in 6igure :*b.. Similarly, A>@*/. and A>@*1.
is reconstructed. One distinctive feature of 32B
reconstructor algorithm "#5$ is that can be observed
from this e%ample is that underlying variable length
code *i.e. A>@. need not be decoded to recover A>@s
back from 6>S as was required with (3(C in "#1$.
#"." R$CC ,ncoding and %ransmission
;ireless access points *;0'. constitute the last-hop
delivery link to the end user. hose packets which are
not dropped in the network reach the ;0'. hese
packets are encoded with eight state rate J 3SCC and
6igure 1. 6igure shows four variable length blocks A>@*F., A>@*#., A>@*/. and A>@*1. being rearranged into fi%ed length slots 6>S*F-
1.with (3(C?(3(CS algorithm.
n / 0 n / ! n / # $123 $-2$
-2$405 F,FF FFF FFFF
-2$4!5 # # #FF#
-2$4#5 F,#F # F#F# F#F-
-2$4)5 F,## F,FF F##FFF F##-
transmitted over wireless channel. he end user is able
to able to receive packets from one or more ;0's. ;e
assume that reception is done through two different
chain of antenna and receivers *as the number of
descriptions. and combined at last stage.
)" Decoding of Received Descriptions
8t is assumed that the two descriptions are received
through two orthogonal wireless channels. he
channels are orthogonal in the sense that they are
received by a single antenna by time division
multiple%ing or through two different antenna pairs by
frequency multiple%ing. Subsequent stages of
decoding are e%plained ne%t.
)"!" Iterative $ource6C*annel Decoding
8n this research, channel code *3SCC. and source code
*arithmetic code. are decoded iteratively. 8terative
decoding allows source codec to be more robust to
residual channel errors. 0lso, channel decoder can
utilize any residual redundancy of arithmetic code in
the process of decodng. ;e assume that input signal at
the @C&3 decoder K G
,...$ ,..., , "
/ # i
y y y
is
obtained as
i i i
w - y + = */.
where,
i
w
is )aussian distributed random noise with
variance
/
and
i
- is the corresponding transmitted
symbol. 8n contrast to Aiterbi decoder which outputs
hard decision, @C&3 algorithm "#7$ accepts channel
>>3 values
/
? /
i i
y = for decoding 3SCC to
obtain soft >>3 outputs as shown in 6igure =. >et L
e
be the e%trinsic >>3 calculated by subtracting a priori
input from the output of @C&3 decoder. he e%trinsic
information L
e
is in the form of packets and if the
packet has been lost *did not arrive at the receiver., its
>>3 value is taken to be zero. ;hen the last of the
packets has received, they are formed into 6>S of
predefined length. he main challenge, as discussed in
earlier section, is to rearrange *(3(C decode. soft
information 6>Ss to array of *>>3 values of
arithmetic coded. A>@ blocks M
a
G "
a
0
a
k
a
,..., , ,
#
$, where k-th A>@ block
a
k
is an
array of >>3 values of arithmetic code. his was
achieved with (3(CS algorithm described in section
/.:, aided by side information sseg.
(ach A>@
a
k
is now decoded by Chase-like S8SO
arithmetic decoder proposed by Haibi et al. "#:$. 8n a
Chase-like decoder, hard decision on each A>@ is
made to obtain a binary array . sgn*
a
k
4 = . >et
. *k 2
be the length of the array
a
k
. 4 is perturbed by
a set of test patterns
,
$
, which is a binary sequence
that contains #s in the location with minimum
magnitude in = N N
a
k

N$ N N,..., N N, "N
. * / #
a
k k2
a
k
a
k

. @y
adding this test pattern modulo-/, a new binary
sequence is obtained.
, ,
$ 4 ! = *1.
where
C ,..., / , # D q ,
and vector
$ , , , "
/ #
,
0
, , ,
p p p ! = . @y using q number of test
patterns *q G #2 in this article., the perturbed binary
sequence
,
!
may fall within the decoding sphere of
valid arithmetic code. 0rithmetic code is considered
valid if, after decoding, a (o@ symbol is detected at
the end and number of decoded symbols is correct. 6or
a valid sequence
,
!
, following metric is calculatedO
.. * log*
/
. , *
. *
#
/
/
,
k 2
i
,
i i E
,
1 !
h r d
M

=
+

*:.
where,
,
i
,
i
p h / # = is the bipolar form of
,
i
p ,
. ? / ?*
/

a
ki i
r = and *..
E
d calculates (uclidean
distance. . *
,
1 ! is the probability of source sequence
,
1
obtained after decoding
,
!
. 6inally, the decoded
6igure :. *a. 6our sseg linked list generated by 32B
'econstructor subroutine in the process of (3(CS decoding of
6>S created in 6igure 1. *b. 6our bseg linked list generated by
function 32B 'econstructor, where bseg i corresponds to
A>@*i..
6igure =. ,etailed e%position of iterative source-channel
decoder block in 6igure #. ,otted section outputs 6>S when last
of the packets for a description is received.
bitstream 0 corresponds to the sequence
,
!
having
the lowest metric M
,
among the valid sequences. >et
C ,..., / , # D 2 v
be the set of positions of bits which
remain unchanged among the valid sequences. hese
bits positions are most reliable bit positions in terms of
>>3 magnitude and are assigned a constant e%trinsic
information, v i p
,
i
e
ki
= ., / # * and F =
e
ki
,
otherwise. he value of 5 was set to : in this research
following "#:$. hus, after a iteration Chase-like
arithmetic decoder outputs A>@s of e%trinsic >>3 M
e
G "
e
0
e
k
e
,..., , ,
#
$, where
$ ,..., , "
/ #
e
k2
e
k
e
k
e
k
= . he M
e
is again reorganized
into fi%ed length blocks using (3(C and applied as a
priori information L
a
to the @C&3 decoder. 0fter
certain fi%ed number of iterations, let 0 be the output
of the arithmetic decoder.
)"#" Decoding of Multiple Descriptions
he 0 variable length bitstreams in 0 are checked
for C3C, and in case of check failure, corrosponding
wavelet tree is marked lost and is not utilized for
further reconstruction. 0fter S'89 decoding of the
same wavelet tree order bitstreams of two descriptions,
wavelet coefficients obtained are
#
8 and
/
8 . 8f the
same wavelet tree order bitstreams of descriptions
*corresponding to row and column inde%. arrive at the
receiver without being lost, decoding is performed by
simple matri% lookup to get the central quantization
inde%. 9owever, if some of the wavelet trees of either
description are corrupted or lost, central decoding can
not be performed correctly by matri% lookup. 8n such
cases, only available bitstream from either description
is used to perform inverse inde% assignment using
single channel inverse quantizer. hus, at the receiver,
a wavelet tree can be declared lost in three scenarios.
0 packet is lost in packet loss channel and all
the A>@ associated with it are declared lost.
<o valid arithmetic code is found in the
source-channel decoder stage.
@itstream after arithmetic decoding, which are
in fact S'89 coded A>@, fails in C3C.
;hen both descriptions of same wavelet tree are
lost, Song et al. "#/$ proposed error concealment in
wavelet domain. his error concealment is based on
the fact that wavelet coefficients in the lowest
frequency subband have similar spatial correlation
with the original image. Similar to the scheme
proposed in "#/$, one wavelet tree in S'89 has four
coefficients in the lowest frequency subband. 8f a tree
is lost, these four coefficients are also lost and error
concealment process replaces these four coefficients
with average of its neighbour coefficients whose block
have been recovered by central or side inverse
quantizer. 0ll the high frequency coefficients of the
lost tree are set to zero.
+" ,7perimental Results
;e use =#/ =#/ grey-scale >ena test image for
evaluating the performance of proposed algorithm
under following channel conditionsO *i. 0;)<
channel with no packet loss, and *ii. combined packet
loss and 0;)< channel. he given image is
hierarchically decomposed by :-scale discrete wavelet
transform, and then each of the subband coefficients is
quantized with uniform scalar quantizer of stepsize B.
6or +,S-, we consider two types of inde%
assignments, with number of diagonals , G / and , G
1, to form two descriptions. 6or each description, /=2
wavelet trees are constructed and S'89 coded to
generate /=2 bitstreams. he stepsize B was ad!usted to
obtain coding rate of F.= bpp?description after S'89
coding. :-bit C3C was generated using generator
polynomial
1 :
- - +
# /
- - + + for the first #/5 bits of
a bitstream and appended to it. his results in mere
F.FF: bps overhead in each description and will be
compensated by arithmetic coding in ne%t stage. 0ll
the bitstreams are independently arithmetic coded for
further compression, formed into 6>S using (3(CS.
Our scheme is more compression efficient as there is
no arithmetic coding stage in the scheme proposed in
"#/$. ,ata packets were formed by packing eight slots
and transmitted over packet network. 'ackets arriving
at the ;0' are encoded with 5-state 3SCC with
generator polynomial
5
. #= , #4 * . , * = G G
r
.
;e first consider transmission over 0;)< channel
without any packet drop. ;e compare the performance
of the proposed system to the baseline scheme of Song
et al. "#/$ with /=2 bitstreams per description under
similar channel conditions. 6igure 2 and 4 shows the
performance of the proposed system for 1 diagonal and
/ diagonal inde% assignments, respectively. 6or hard
decision decoding of arithmetic code aided by
(3(CS *without iterative decoding., we can observe
#.= d@ improvement in average 'S<3 at (
b
?<
F
of 1.=
d@ compared to baseline scheme, in both 1 diagonal
and / diagonal case. ;ith iterative decoding
remarkable improvement is observed in 'S<3
performance. 0fter # iteration of soft decoding of
arithmetic code, there is a ma%imum gain in 'S<3 of
1 d@ at (
b
?<
F
of 1.= d@ in 1 diagonal case. he
improvement is 1.= d@ compared to baseline scheme
for / diagonal case. he ma%imum improvement after
: iterations is observed at relatively noisy region of
(
b
?<
F
G 1 d@. he gain in 'S<3 is 2 d@ and 4 d@
respectively for 1 and / diagonal cases, compared to
baseline system under 0;)< channel. 8mprovement
in 'S<3 saturates after : iterations. he results are
obtained by averaging #FF trials for each channel
condition. 8n 6ig. 5, we give e%amples of reconstructed
images in order to illustrate the improvement in visual
quality. 6igure 5*a. shows improvement in image
quality compared to baseline scheme *6igure 5*b.. at
(
b
?<
F
of 1 d@. hese results have demonstrated that
the proposed improvement in +,C scheme makes it
significantly noise robust compared to e%isting
schemes in literature.
Second set of simulations were carried out to
demonstrate the performance of our proposed scheme
over channels with both packet loss and random noise.
he ob!ective is to evaluate the degradation of 'S<3
value for different number of packet lost. ;e conduct
e%periments with the proposed system only as its
superiority compared to baseline system has been
demonstrated in the earlier set of simulations. he
results for different wireless channel conditions are
shown in 6igure 7 for , G 1 and 6ig. #F for , G /. ;e
obtain the results after averaging /FF trials of
transmission over each channel condition. 8t is evident
from the figures that proposed scheme is able to
achieve graceful degradation of image quality as more
and more packets are lost in packet network. he
degradation is more evident in high (
b
?<
F
region *P 1
d@. whereas, there is smaller effect of packet loss in
low (
b
?<
F
region. hese results demonstrate e%cellent
capability of the presented system to combat packet
loss and noisy channel via integrated design.
." Conclusion
8n this paper, we proposed a multiple description
image transmission scheme over hybrid channel
consisting of combined packet loss channel and
0;)< channel. 8t was shown that robustness to
packet loss and channel noise can be improved by
utilizing the soft channel information at the receiver.
;e achieved this by iterative decoding between Chase-
like S8SO arithmetic decoder and @C&3 decoder, with
(3(CS stage in-between. Significant gain in 'S<3
is obtained over e%isting scheme with lesser number of
bits transmitted.
6igure 4. 'erformance of proposed scheme for transmission over
0;)< channel *with no packet loss. for , G /.
6igure 7. 'erformance of proposed scheme *, G 1. for
transmission over packet loss channel at different values of
(b?<F for wireless channel.
6igure 2. 'erformance comparison of proposed scheme for
transmission over 0;)< channel *with no packet loss. for 1
diagonal inde% assignment case *, G 1..
*a. *b.
6igure 5. 8mage recovered by proposed scheme *a. compared
with recovery by baseline system *b. at (b?<F of 1 d@ and , G 1.
'S<3 of images *a. and *b. are 11.2 d@ and /4 d@, respectively.
References
"#$ Aaishampayan A. 0., Q0pplication of multiple
description codes to image and video
transmission over lossy networks,R in !roc. .
th
"nt. +orkshop !acket 3ideo, @risbane, 0ustralia,
#772.
"/$ Aaishampayan A. 0., Q,esign of multiple
description scalar quantizers,R "EEE $rans.
"nform. $heory, vol. 17, pp. 5/#-51:, +ay #771.
"1$ ;ang K., Orchard +., 3eibman 0., Q+ultiple
description coding using correlating transforms,R
"EEE $rans. "mage !rocessing, vol. #F, pp. 1=#-
122, /FF#.
":$ )oyal A. S., Sovacevic &., Q)eneralized multiple
description coding with correlating transforms,R
"EEE $rans. "nf. $heory, vol. :4, pp. /#77-///:,
/FF#.
"=$ Shelil S., @ekka 3. (. 9., 3ouvaen &. +.,
Q+ultiple description wavelet-based image
coding using correlating transforms,R 1E% *
"nternational 6ournal of Electronics and
(ommunications, vol. 2#, pp. :##-:#4, /FF4.
"2$ Shelil S. et al., Q8mproved multiple description
wavelet based image coding using subband
uniform quantization,R 1E% * "nternational
6ournal of Electronics and (ommunications, vol.
2=, pp. 724-74:, /F##.
"4$ Sun ). at al., Q+ultiple description coding with
prediction compensation,R "EEE $rans. "mage
!rocessing, vol. #5, no. =, pp. #F14-#F:4, /FF7.
"5$ Servetto S. ,. et al., Q+ultiple description
wavelet based image coding,R "EEE $rans.
"mage !rocessing, vol. 7, no. =, pp. 5#1-5/2,
/FFF.
"7$ Sachs ,. )., 0nand 3., 3amchandran S.,
Q;ireless image transmission using multiple
description based concatenated codes,R in !roc.
!"E, San &ose, C0, &an. /FFF, vol. 174:.
"#F$ Chang S., Cosman '. C., +ilstein >. @.,
Q'erformance analysis of n-channel symmetric
6(C-based multiple description coding for
O6,+ networks,R "EEE $rans. "mage
!rocessing, vol. /F, no. :, pp. #F2#-#F42, /F##.
"##$ Chang S., Cosman '. C., +ilstein >. @.,
Q8terative channel decoding of 6(C-based
multiple-description codes,R "EEE $rans. "mage
!rocessing, vol. /#, no. 1, pp. ##15-##=/, /F#/.
"#/$ Song ,., Cao >., Chen C. ;., Q3obust multiple
description image coding over wireless networks
based on wavelet tree coding, error resilient
entropy coding, and error concealment,R 6ournal
of 3isual (ommunication and "mage
'epresentation, vol. #7, pp. 1##-1#7, /FF5.
"#1$ 3edmill ,. ;., Singsbury <. )., Qhe (3(CO
0n error-resilient technique for coding variable-
length blocks of data,R "EEE $rans. "mage
!rocessing, vol. =, no. =, pp. =2=-=4:, #772.
"#:$ Haibi S. et al., Q&oint source?channel iterative
arithmetic decoding with &'() /FFF image
transmission application,R E%'1"! 6ournal in
ignal !rocessing, +ay, /F#/.
"#=$ Creusere C. ,., Q0 new method of robust image
compression based on the embedded zerotree
wavelet algorithm,R "EEE $rans. "mage
!rocessing, vol. 2, no. #F, pp. #:12-#::/, #774.
"#2$ Said 0., 'earlman ;. 0., Q0 new, fast, and
efficient image codec based on set partitioning in
hierarchical trees,R "EEE $rans. on (ircuits and
ystems for 3ideo $echnology, vol. 2, no. 1, pp.
/:1-/=F, #772.
"#4$ aubman ,. S., +arcellin +. ;., &'() /FFF T
8mage Compression 6undamentals, Standards
and 'ractice, Sluwer 0cademic 'ublishers,
@oston, /FF/.
"#5$ 6ang K., Q(3(C-based length coding of
variable-length data blocks,R "EEE $rans. on
(ircuits and ystems for 3ideo $echnology, vol.
/F, no. #F, pp. #1=5-#122, /F#F.
"#7$ @ahl 3. et al., QOptimal decoding of linear codes
for minimizing symbol error rate,R "EEE $rans.
"nformation $heory, vol. /F, no. /, pp. /5:-/54,
#74:.
6igure #F. 'erformance of proposed scheme *, G /. for
transmission over packet loss channel at different values of
(b?<F.