Documentos de Académico
Documentos de Profesional
Documentos de Cultura
by
Saikat Majumder
Department of
Electronics and Telecommunication Engineering
Submitted in fulfillment
of the requirements of the Degree of
Doctor of Philosophy
to the
I hereby declare that the work presented in the thesis titled On Iterative De-
coding Schemes for Communication over Wireless Channels is a bonafide
record of the systematic research work done by me under the supervision of Dr.
Shrish Verma, Professor, Department of Electronics and Telecommunication Engi-
neering, National Institute of Technology Raipur, India. No part of this thesis has
been submitted elsewhere for any other degree or qualification and it is all my own
work unless referenced to the contrary in the text.
ii
Department of Electronics and Telecommunication
NATIONAL INSTITUTE OF TECHNOLOGY RAIPUR
G. E. Road, Raipur-492010, Chhattisgarh.
This is to certify that the work entitled On Iterative Decoding Schemes for
Communication over Wireless Channels is a piece of research work done by
Mr. Saikat Majumder under my guidance and supervision for the award of degree
of Doctor of Philosophy in Electronics and Telecommunication Engineering
of National Institute of Technology Raipur, Chhattisgarh, India.
To the best of my knowledge and belief, this thesis
3. Fulfills the requirement of the ordinance related to the Ph.D. degree of the
institute.
4. Is up to the standard both in respect of content and language for being referred
to the examiner.
Forwarded
Chairman, DRC
Dedicated to
The Almighty.
Acknowledgements
Throughout my Ph.D. work I came across many people whose support helped me to
complete this research work and at this moment I would like to take the opportunity
to acknowledge them. First and foremost I would like to express my deep and sincere
gratitude towards my respectable supervisor, Prof. Shrish Verma for his invaluable
guidance, constant inspiration and motivation along with enormous moral support
during my difficult phase. I am indebted to him for the valuable time he has spared
for me during this work. I express my sincere gratitude to Prof. Sudarshan Tiwari,
Director, NIT Raipur and Chairman, Doctoral Research Committee (DRC) of Elec-
tronics and Telecommunication department for encouragement and kind support.
I am grateful to my Doctoral Guidance Committee (DGC) members, Prof. A. S.
Thoke and Dr. A. Khare for their guidance and help.
I am very much thankful to the my colleges in the department, Dr. B. Acharya,
Dr. A. S. Raghuvanshi, Dr. T. Meenpal, Sri. R. Chaurasiya, Sri. A. Gupta, Dr. A.
Naugrahiya and Sri. S. Chakraborty for their continuous help and encouragement.
I am specially thankful to Dr. A. S. Raghuvanshi and Sri. Subhnakar Majumdar
for their suggestions regarding writing of the thesis. I acknowledge the help of
my colleagues in the Department of Information Technology and Department of
Computer Science & engineering in this regard. I specially acknowledge the help of
Sri. Rakesh Tripathi for numerous help during the course of Ph.D.
I acknowledge all technical and other support staffs for their cooperation. Finally,
I thank my family and parents for their support.
Saikat Majumder
Abstract
based MDC. Our innovation lie in incorporating iterative decoding between FEC
based multiple description decoder and MIMO demodulator. Next, iterative source-
channel decoding is applied in the context of two-channel quantizer based MDC.
In the proposed MDC schemes, it is demonstrated that performance of two-channel
MDC scheme can be improved by utilizing soft information in the form of received
signal log likelihood ratio (LLR) and noise resilient decoding of arithmetic codes
without the need of extra redundancy. Finally, iterative SISO decoding technique
is applied to network coding in multiple access relay channel (MARC). Proposed
algorithm enables network code and channel decoders to exchange soft information
iteratively and to achieve an improved performance compared to a hard decision
iterative decoder. This scheme is also extended to the scenario where in a MARC
setup, two user nodes cooperate with each other using distributed space-time code,
in addition to network coding.
Overall, in this work it is shown that as many conventional coding schemes can be
considered instances of product code and hence iterative decoding can be applied. In
the schemes proposed in the present work, it is shown that, in general, application of
soft and hard decision iterative decoding results in improved performance compared
to other existing schemes in literature.
******
Department of Electronics and Telecommunication
NATIONAL INSTITUTE OF TECHNOLOGY RAIPUR
G. E. Road, Raipur-492010, Chhattisgarh.
This is to certify that the work entitled On Iterative Decoding Schemes for
Communication over Wireless Channels submitted by Saikat Majumder is
hereby recommended for award of degree of Doctor of Philosophy in Electron-
ics and Telecommunication Engineering of National Institute of Technology
Raipur, Chhattisgarh, India.
Declaration ii
Acknowledgements v
Abstract vi
List of Tables xx
1 Introduction 1
1.1 The Turbo Revolution . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Genesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Turbo Code and Iterative Decoding . . . . . . . . . . . . . . . 4
1.2 Turbo-like Coding Schemes . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Concatenated Coding and Decoding Schemes . . . . . . . . . . 7
1.2.2 Joint Source-Channel Coding . . . . . . . . . . . . . . . . . . 8
1.2.3 Joint Network-Channel Coding . . . . . . . . . . . . . . . . . 9
1.3 Coding for Wireless Networks . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Channel Models . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Cooperative Diversity in Wireless Networks . . . . . . . . . . 12
ix
x
References R-1
2.1 Priority encoding for FEC based multiple description coding. P and
I indicates parity and information symbols of FEC code, respectively. 22
2.2 Block diagram of a generic baseband transceiver system with ISCD. . 26
2.3 Relay communication with three terminals . . . . . . . . . . . . . . . 30
2.4 The butterfly network: Example of network coding in wireline net-
work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 The uplink for two mobile stations (S1 and S2 ) with cooperative di-
versity. (a) Diversity obtained by simple relaying, (b) Network coded
diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
xiv
xv
4.1 n-channel FEC based multiple description coding technique for a pro-
gressive bitstream. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Encoding and iterative decoding of LDPC-RS coded FEC based mul-
tiple description scheme. . . . . . . . . . . . . . . . . . . . . . . . . 71
xvi
6.1 The uplink for two mobile stations (S1 and S2 ) with cooperative di-
versity. (a) Diversity obtained by simple relaying, (b) Network coded
diversity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.2 Multiple access relay channel with two users transmitting channel
coded information to a base station. Network coding operation is
performed at the relay node. . . . . . . . . . . . . . . . . . . . . . . 131
6.3 (a) Block diagram of encoder at mobile user nodes, (b) Network cod-
ing operation at relay node. . . . . . . . . . . . . . . . . . . . . . . . 132
6.4 Data packets from two sources are ordered in alternate rows. Total
number of rows from two sources is L1 +L2 . Remaining rows of parity
check bits are calculated and transmitted from the relay node. . . . . 133
6.5 Block diagram of the proposed decoder. . . . . . . . . . . . . . . . . . 135
6.6 EXIT chart of the proposed network-channel decoder with (15,11) RS
code as channel code (outer code) and punctured (5, 7)8 RSCC acting
as network code with code rate of 2/3. . . . . . . . . . . . . . . . . . 140
6.7 EXIT chart with (31,27) RS code and punctured (15, 17)8 RSCC act-
ing as network code with code rate of 2/3. . . . . . . . . . . . . . . . 140
6.8 EXIT chart with (15, 11) RS outer code and different RSCC codes
as network code. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.9 BER performance of the proposed scheme using (15, 11) RS code
and punctured (5, 7)8 RSCC with number of decoding iterations over
Rayleigh block fading channel. . . . . . . . . . . . . . . . . . . . . . . 143
6.10 BER performance for (31,27) RS code (15, 17)8 RSCC combination
with number of decoding iterations over Rayleigh block fading channel.143
6.11 Performance comparison of the proposed iterative decoder with dif-
ferent reference schemes. Different network codes at relay node are
evaluated for a fixed channel code at the user nodes. . . . . . . . . . 145
xix
5.1 Description size after addition of CRC bytes and reduction due to AC.108
5.2 State and tail bits after different stages in the Figure 5.12. . . . . . . 115
xx
List of Abbreviations and Symbols
List of Abbreviations
AC Arithmetic Code
ACC Accumulator
BC Broadcast Channel
BM Berlekamp-Massey
BS Base Station
CC Convolutional Code
xxi
xxii
dB Decibel
GF Galois Field
ML Maximum Likelihood
NC Network Coding
RN Relay Node
RS Reed Solomon
List of Symbols
Fading coefficient
Scalar multiplication
Belongs to
Interleaver function
Permutation function
1 De-interleaver function
Proportional to
2 Noise variance
Subset of
D Distortion function
E Expectation function
Introduction
1.1.1 Genesis
Ever since Shannons revolutionary finding regarding source and channel coding in
1948 [1], the fundamental limits of reliable communication established by him have
been the model and the driving force for the researchers in the field of coding and
information theory. The challenge in communication is that of reproducing at one
point either exactly or approximately a message selected at another point. Shannon
showed that it is possible to transmit digital data with arbitrary high reliability,
over noise-corrupted channels, by encoding the message with sufficient redundancy
and subsequently decoding at the receiver. This redundancy is typically provided
by a suitable channel code. A practical communication system designer faces two
critical questions - one is quantitative: how much redundancy is required?, and the
other one is qualitative: what is the kind of redundancy and how to achieve it? [2].
1
2
Shannon answered the first question and proved that there is a limit to the
information rate for reliable communication over noisy channel. Quantifying the
amount of redundancy required for reliable transmission allows the optimum use of
the communication resources, e.g., transmission power, channel bandwidth. Shan-
nons channel capacity theorem states that for a communication channel H, there
exists a reliable coding scheme if and only if information rate R < C(H). The
number C(H) [0, 1], called the channel capacity. However, Shannons results are
non-constructive and does not answer the second question.
The second question posed tries to investigate which coding scheme would bring
us close to the channel capacity R < C(H) in a practical and efficient way [3].
This proved to be challenging, if not impossible. There has been extraordinary
developments in design of practical codes over the decades by borrowing insights
mostly from the field of mathematics. The algebraic coding paradigm dominated
the pre 90s decades of channel coding. Algebraic coding theory is mainly concerned
with linear (n, k, d) block code over a q-ary field Fq . For the special case of q = 2 and
F2 , we have binary linear block code. A (n, k, d) block code consists of 2k n-tuple
binary codewords. The parameter d denotes the minimum distance between any
two distinct codewords, i.e. number of bits any two codewords differ. The principal
objective of algebraic coding theory is to maximize the minimum distance d for a
given code. Over a binary symmetric channel (BSC), the optimum decoding rule
chooses the valid codeword closest in Hamming distance to the received codeword.
Thus, a code with minimum distance of d can correct all patterns of (d1)/2 or fewer
channel errors. Even though binary algebraic block code can be applied on additive
white Gaussian noise (AWGN) channel, they have not proved to approach Shannon
capacity of the channel or to the channel capacity of BSC. Thus, an efficient and
practical coding scheme which approaches the capacity limit still eluded the coding
theory community even after five decades of Shannons paper.
It was not until 1993, when at the IEEE International conference on Commu-
nication (ICC) in Geneva, Switzerland, Berrou, Glavieux, and Thitimajshima [4]
stunned the research community by introducing a new class of codes, called Turbo
codes, that could achieve near-Shannon-limit performance with reasonable decod-
3
ing complexity. The novelty of Turbo code lies in using pseudorandom interleavers
in the encoding algorithm and applying carefully designed iterative decoding al-
gorithm. The interleavers allow randomness envisaged in the proof of Shannons
channel coding theorem, yet preserving enough structure to allow efficient encoding
and decoding. In the words of Daniel J. Costello, Jr. and G. David Forney, Jr. [3],
Fifty years of effort and invention have finally produced coding schemes
that closely approach Shannons channel capacity limit on memoryless
communication channels.
Invention of Turbo codes initiated the revolution in the field of error correction
coding. Today Turbo code has become part of everyday technology - these are
employed in 3G and 4G mobile telephony, several satelleite communication standards
and in IEEE 802.16 metropolitan wireless network standards.
Shortly after Turbo codes provided the initial momentum to the paradigm shift
to the coding theory research, several researchers [5, 6], with background in physics
and computer science, rather than in coding, independently discovered codes similar
to low density parity check (LDPC) codes. It was Robert Gallager who originally
invented LDPC codes and its iterative a posteriori probability (APP) decoding al-
gorithm way back in 1963 [7]. His work was largely forgotten for more than 30
years. There was little interest in 1960s and 1970s because these codes were much
too complex for the technology of that time. Following words of Forney [8] rightly
describes the limitations of those times.
As the name suggests, LDPC codes are block codes with parity-check matrices
that are sparse, or containing only a very small number of non-zero entries. Other
4
than the requirement that parity-check matrix be sparse, an LDPC code itself is no
different form any other block code. The biggest difference between LDPC codes
and classical block codes is the way it is decoded. Classical block codes are decoded
with maximum liklihood (ML) like decoding algorithm. They are usually short and
design is based on algebraic principles. On the other hand LDPC codes are decoded
iteratively using graphical representation of their parity-check matrix and are much
longer, less structured, and designed with properties of parity-check matrix as focus.
Both turbo code and LDPC code utilize the class of algorithm known as message-
passing decoding algorithm and is the main reason behind their capacity achieving
performance [9]. Wiberg [10] showed that both these classes of codes could be
understood as instances of codes on sparse graphs, and their decoding algorithm
could be understood as instances of a general iterative APP decoding algorithm
called the sum-product algorithm. MacKay [5] showed that in practice moderate
length LDPC codes (103 104 bits) could attain near-Shannon limit performance.
These results triggered a similar explosion of research on LDPC codes, which are
currently seen as competitors to turbo codes.
u v
(0)
(1)
Encoder 1 v
g2(D)/g1(D)
(1) (2)
Puncturing v ,v
P Mechanism
u' Encoder 2 v
(2)
g2(D)/g1(D)
The iterative turbo decoding system is shown in the Figure 1.2. Turbo decoder
derives its name from the principle of preheating air in an engine with its own exhaust
gases, similar to feedback provided in this decoder. Berrou et al. [4] discovered
an universal concept in soft-iterative decoding that is the key to achieving good
performance. It is the removal of the intrinsic information from the output APPs
(i)
L(i) (ul ), resulting in extrinsic APPs Le (ul ), which are then applied as a priori
(i)
input La (ul ) to the other decoder. Intrinsic information represents the soft channel
(i)
outputs Lc rl and the a priori inputs are already known prior to decoding. Extrinsic
information is the additional knowledge learned about an information bit during an
iteration. The removal of intrinsic information has the effect of reducing correlations
from one iteration to the next, thus allowing improved decoding performance with
increasing number of iterations.
In the figure, decoders 1 and 2 are maximum a posteriori (MAP) (BCJR [12])
decoders for the two constituent codes, is the same interleaver as in the encoder,
while 1 is the inverse operation of . The MAP decoder decides ul = +1 if
P (ul = +1|r) > P (ul = 1|r), and it decides ul = 1 otherwise. If the received
signal vector is represented as r, then L(ul ) is the log a posteriori probability (LAPP)
ratio defined as
P (ul = +1|r)
L(ul ) = log (1.1.2)
P (ul = 1|r)
6
P (2)
L (ul)
(1)
Le (ul) (2)
La (ul) P
-1
Decoder 1 Decoder 2
(1)
+ +- P Decision
(1) L (ul) - Lcrl(0)
La (ul)
-1
-
P (2)
++ (2)
L (ul) - Lcrl(0)
Le (ul)
The signal is transmitted over a channel with noise power spectral density N0 ,
where Lc = 4Ec /N0 is defined as the channel reliability and Ec = rEb is the energy
per channel bit. The received signal vector r is composed of three components,
(0) (1) (2)
systematic bits rl , output from encoder 1 rl , and output from encoder 2 rl .
Based on these inputs, each MAP decoder computes L(ul ), where, for example
output of MAP decoder 1 can be written as
(0)
L(ul ) = Lc rl + L(1) (1)
a (ul ) + Le (ul ) (1.1.3)
The first term in (1.1.3), called the channel value, is the log-liklihood ratio (LLR) [13]
value of systematic part of the received signal. The second term represents any a
priori information about ul provided by a previous decoder, and the third term
represents extrinsic information that can be passed on to a subsequent decoder.
Thus, the two MAP decoders exchange extrinsic information for certain number of
iterations and hard decision is performed on the LLR output of one of the decoders
to obtain the decoded codeword.
Encoding and decoding scheme of parallel concatenated codes (PCC) were discussed
in the earlier section with turbo code as an example. The component codes are often
convolutional encoders, but binary Bose-Chaudhuri-Hocquenghem (BCH) codes are
also used. The turbo encoder in the Figure 1.1 consists of two parallel concatenated
convolutional codes, but this scheme can be extended to parallel concatenation of
more than two component codes, leading to multistage turbo codes [18].
The basic structure of a serially concatenated scheme (SCC) with decoder is shown
in the Figure 1.3. The SCC encoder consists of an outer encoder (Encoder I) and an
inner encoder (Encoder II), interconnected by an interleaver. The interleaver scram-
bles the bits before they are passed on to next constituent encoder, which ensures
that even if a specific bit has been gravely contaminated by the channel, there is
chance that other constituent decoder is capable of providing more reliable informa-
tion regarding this bit. If iterative soft-decision decoding is used in the SCC decoder,
a performance similar to that of PCC may be achieved [15]. The SCC scheme is
quite general structure and many coding and detection schemes can be described as
serially concatenated scheme, such as turbo equalization [19], joint source-channel
decoding [2022], turbo multiuser detection [23], coded modulation [24]. A list of
8
u1 c1 u2 c2
Encoder I P Encoder II
e a
a L (u2) L (c1) e
L (c2) P-1 L (u1)
Decoder II Decoder I
a P e
L (u2) L (c1)
Figure 1.3: Two stage serially concatenated code: the encoder and iterative decoder
[14].
such SCC schemes with their inner and outer codes are shown in Table. 1.1. Similar
to turbo codes and PCC, a serially concatenated scheme may contain more than two
components. Figure 1.4 shows the schematic of a three-stage serially concatenated
code and its decoder. The bit input to the above mentioned encoders is ui and en-
coded output is shown as ci , where i is encoding stage. At the decoder, L indicates
the log-liklihood ratios (LLR), with superscript e and a indicating extrinsic and a
priori LLRs, respectively. Decoding is performed iteratively for certain number of
iterations, or when a correct codeword is detected.
Table 1.1: Different serial concatenated systems with iterative decoding [14].
Configuration Outer En-/Decoder Inner En-/Decoder
Serial Concatenated Code FEC En-/Decoder FEC En-/Decoder
LDPC Codes Check Nodes Variable Nodes
Turbo Equalization FEC En-/Decoder Multipath Channel/Equalizer
Turbo BICM FEC En-/Decoder Mapper/Demapper
Turbo Source-Channel Decoding Source En-/Decoder FEC En-/Decoder
Turbo CDMA FEC En-/Decoder Spreading Code/MUD
Turbo MIMO FEC En-/Decoder Mapper/MIMO Detector
Another instance of concatenated coding scheme, where turbo principle has been
successfully applied is joint source-channel decoder. Consider a classic communica-
tion chain, where source encoder (performs data compression) removes redundancy
9
u1 c1 u2 c2 u3 c3
Encoder I P1 Encoder II P2 Encoder III
e a e a
a L (u3) L (c2) L (u2) L (c1) e
L (c3) P-1 P-1 L (u1)
Decoder III Decoder II Decoder I
a P e a P e
L (u3) L (c2) L (u2) L (c1)
Figure 1.4: Three stage serially concatenated code: encoders and iterative decoders
[14].
from the source signal, followed by a channel encoder that aims to reintroduce re-
dundancy in the compressed data in a efficiently controlled manner to cope with
transmission errors. Conventional communication systems have source and channel
decoders implemented separately and are based on Shannons source-channel sep-
aration theorem [1], which holds only for idealized conditions (with infinitely long
block length codes). Therefore, joint source-channel coding has gained considerable
attention from the research community as alternative to separate source-channel
coding. The key principle of joint source-channel decoder is to exploit jointly the
residual redundancy of the source-coded stream (i.e. exploiting the inefficiency of
the compression algorithm) and the redundancy introduced by the channel code to
correct the bit errors and find the most likely source symbol.
Network coding [25, 26] is a technique to code the data packets at the nodes in a
packet network, in contrast to conventional networks where encoding is done only at
the source and intermediate nodes just forward the packets. A simple network coding
operation at a relay consists of just XORing the bits arriving from two sources and
forwarding it. Network coding can improve throughput, robustness and security of
data networks. Authors in [27,28] proposed joint network-channel coding schemes to
gain cooperative diversity in wireless networks to combat channel fading. A simple
model where diversity can be gained through network coding is the multiple-access
relay channel (MARC). Figure 1.5 depicts an example of MARC for cooperative
uplink of two mobile users (MU) with the help of an intermediate network coding
10
MU1
S1 S1
S3
S2 RN S2
MU2 BS
Figure 1.5: Uplink for two mobile users (MU1 and MU2) with the help of relay on
the MARC.
relay RN. One way to gain diversity through network coding for MARC with noisy
channels is to treat network and channel coding separately. Then, error control
coding is used for each transmission to transform the noisy channel to erasure-based
link. At the network layer, one performs network coding for the erasure-based
networks which is provided by the lower layers.
A relay can not only be used to gain diversity, but can also provide additional
redundancy which improves the performance compared to point-to-point communi-
cation if the relay has a better connection to the base station than mobile station. In
such cases also relay provide improved performance, even in the absence of fading.
The principle of joint network-channel coding is that redundancy in the network
code should be used to support the channel code for better error protection. It is
similar to source-channel coding discussed earlier where redundancy remaining after
source coding is utilized by channel code to combat noise.
transmission through wireless channel can be decomposed into one of the following
fundamental forms of channel or links.
In wireless environment, the magnitude and phase of the signal received, changes
with time and frequency. This is called fading and is caused by many factors, which
can be classified into two classes, large-scale propagation effects and small-scale
propagation effects [30]. Large scale propagation effects occur over long distances,
occur due to path loss and shadowing. On the other hand, small scale fading are more
frequent and occur due to constructive and destructive interference from different
multipath signals at the receiver. The properties of the variation of this combined
signal are normally described on a statistical basis. The received signal is said to be
in deep fade if at a given moment, all the phase shifts of these different versions of
a transmitted signal cause a destructive combination and the combined signal has
significantly reduced strength at the receiver. During fade event, communication
link suffers significantly more errors than the average error rate.
In this work, Rayleigh fading channel is considered which models many wireless
communication environments with no line of sight path between the transmitter and
receiver. The channel gain h is modeled as a complex Gaussian random process and
its two quadrature components are uncorrelated Gaussian random processes with
12
zero mean and variance 2 . The envelope of the received signal has a Rayleigh
probability distribution and its phase has a uniform distribution between /2 to
/2. The probability density function (PDF) of the envelope is given by
2
e
2 if 0
2
p() = (1.3.4)
0 if < 0
At the i-th signalling interval, the baseband equivalent of the received signal at the
front end of the receiver is then given as
yi = hi xi + ni (1.3.5)
Diversity methods are very effective way of combating multi-path fading in a commu-
nication system. By diversity we mean that multiple independently faded versions
for each transmitted signal arrive at the receiver, so that the probability of all copies
of the transmitted signal experiencing a deep fade at the same time is significantly
reduced. Diversity is achieved in different ways, some of which includes
Relay
Source Destination
MD Decoder
Channel 1 Side
D1
Decoder 1
MD Central
Source D0
Encoder Decoder 0
Side
D2
Channel 2 Decoder 2
(a)
Receiver 1
Receiver 2
(b)
Figure 1.7: Two common types of multiple-description coding, (a) Quantizer based
MDC, and (b) FEC based MDC.
produces two stream of samples which are source/channel coded and transmitted.
At the receiver when any one of the description is received and decoded, the signal
has distortion D1 or D2 . If both the descriptions are received and decoded error-
free, the resultant distortion will be D0 , where, D0 < D1 and D0 < D2 . FEC (or
channel code) based MDC creates n number of descriptions from a progressively
coded source and is obtained using erasure correction property of channel codes, as
shown in the Figure 1.7(b)).
FEC based
MDC
Multiple Soft RS decoder Hard RS decoder
Description based (Ch. 3) based (Ch. 4)
Application Coding
of Iterative MDSQ based MDC
Decoding (Ch. 5)
Network
Network coding for
Coding
MARC (Ch. 6)
tions with soft RS decoder is presented in Chapter 3. Two different algorithms for
iterative decoding of FEC based multiple descriptions with hard RS decoder are
described in Chapter 4. In Chapter 5, schemes for iterative soft decision source-
channel decoding of multiple description image over wireless channel are presented.
In Chapter 6, our contribution in applying iterative decoding to network-channel
coding in the context of cooperative space-time transmission and multiple access re-
lay channel is explained. Finally, our contributions and present potential for future
work is given in Chapter 7. A logical progression of chapters in this dissertation is
shown in the Figure 1.8.
******
Chapter 2
16
17
the same type have the same degree. On the other hand, in an irregular LDPC code,
the degrees of each set of nodes are chosen according to some distribution. Luby
et al. [42, 43] found that by using irregular graphs and optimizing degree sequences,
they could approach the capacity of erasure channel, i.e., achieve small error proba-
bilities at code rates of nearly 1 p, where p is the erasure probability. Richardson,
Urbanke et al. [44] used a technique called density evolution to design long irregular
LDPC codes that for all practical purposes achieve the Shannon limit on binary
AWGN channels.
Recently, a new class of codes, called fountain or rateless codes, were proposed
for channels whose statistics are not known a priori. Fountain codes encode a finite
length information sequence into a potentially infinite stream of encoded symbols.
The receiver then accumulates the noisy symbols until it finds that it has enough
for successful decoding [3]. Luby Transform (LT) codes of Luby [45] were one of the
first codes of this type. These were extended to create a new code called Raptor
code [46], in which an inner LT code is concatenated with an outer fixed-length,
high-rate LDPC code. Raptor codes permit linear time encoding and remove the
problem of error floors. Both these codes work well in erasure channels. Raptor code
also appear to work well in general noisy channels, like AWGN channels [47]. Since,
Raptor code is a serial concatenation of LT and LDPC code, it can be iteratively
decoded using message passing algorithm. In [48], performance of Raptor codes of
low-to-moderate rates with iterative decoding is evaluated over memoryless, as well
as correlated, fading channels. Raptor code is decoded over wireless fading channel
by performing joint channel estimation and iterative decoding at the receiver [49].
Image transmission through noisy channel has its own characteristics. Because of
limited wireless channel bandwidth, image need to be compressed, with lossless or
lossy schemes, depending on the application. But compressed image contains less
redundancy and it makes the image more vulnerable to channel errors. The conven-
19
It assumes that there are several parallel channels between source and destination,
and that some of the channels may be temporarily down or are suffering through
deep fade. Furthermore, the each path is independent and the probability that all
paths are simultaneously fading is small. These channels could be physically dis-
tinct paths formed by several nodes in a multihop wireless network, or can be virtual
channels obtained by frequency or space division multiplexing. In a simple case of
MDC consisting of two channels, a source is decomposed into two descriptions with
side rates R1 and R2 . The reception and decoding of these descriptions individually
leads to side distortions D1 and D2 , respectively. But if both the descriptions are
correctly decoded, it results in central distortion of D0 D1 (and D2 ). In general
for MDC, the region of achievable rate-distortion is bounded by [29]
R1 + R2 R(D0 )
R1 R(D1 ) (2.2.1)
R2 R(D2 )
The success and advantage of trellis coded quantization was extended to multiple
description coding in [62]. In [63], authors present a scheme for construction of
multiple description trellis coded quantizers, which use tensor product of trellises
which is applicable to multiple description coding. An integrated robust multiple
description coding scheme which uses wavelet tree coding and error resilient entropy
coding was proposed by Song et al. in [64].
Recently, problem of increasing the robustness to bit errors for two description
scalar quantizers was addressed in [65]. In [66], MDC is applied for protecting color
images with compressed sensing. In [67], dynamic multiple description wavelet based
image coding using enhanced particle swarm optimization was proposed for effective
multimedia transmission in the dynamically changing network topologies such as
mobile ad-hoc networks.
Recent advances in iterative decoding has also been applied to multiple description
coding. In order to resist packet losses and achieve good compression performance,
an entropy-constrained multiple description trellis coded quantizer was combined
with a variable length code in [68]. The proposed iterative decoding algorithm ex-
ploits the redundancy between the generated descriptions. Barros et al. proposed a
turbo cross decoding scheme for multiple descriptions in [69]. By determining the
correlation properties of multiple description block codes, a soft-in/soft-out algo-
rithm was derived which provided significant gains in terms of end to end distor-
tion. Authors were able to demonstrate additional gain by concatenating multiple
description codes and convolutional codes and performing iterative decoding be-
tween them, similar to turbo decoding principle. Wu and Chang [70] presented a
new multiple description iterative source-channel decoding scheme which allows to
exploit the source residual redundancy as well as inter-description correlation.
FEC based MDC has received considerable attention for transmission of progres-
sively coded multimedia bitstreams over fading channel. Progressive bitstream is
22
FEC codewords
Description 2 P 3 5 7 I CRC RSCC Parity
Description 3 P P P 8 I CRC RSCC Parity
Figure 2.1: Priority encoding for FEC based multiple description coding. P and I
indicates parity and information symbols of FEC code, respectively.
one in which symbols of compressed source are ordered in terms of their importance.
Wavelet based SPHIT coder [71] is example of such an algorithm generating pro-
gressive bitstream. In FEC based MDC, symbols from a progressively coded source
are encoded in such a way that contiguous bits are spread across multiple packets
instead of being in same packet [50, 51]. The design of FEC is constrained in such a
way that symbols in the begining of the stream are provided more protection than
subsequent symbols and that too must be done within the given bit budget. In FEC
based MDC, information symbols are encoded into n packets using (n, ki ), i = 1, ..., p
RS (RS) codes. RS codes are effective when errors occur in bursts and their location
in the transmitted symbol sequence is known. Each source symbols is formed by
combining several bits and typically takes the form of a byte. RS codewords are
arranged vertically, whereas, packets are obtained along horizontal rows. Stronger
error protection (more RS code parity symbols) is provided to the symbols at the
beginning of the source data stream. In the Figure 2.1 first column of FEC code-
word consists of one information symbol and n 1 parity symbols and is encoded
by (n, 1) RS code. Similarly, second and third columns are encoded with (n, 2) RS
code. Thus, source symbols are provided with varying degree of protection. This
scheme, known as priority encoding transmission (PET) was first proposed in [72].
The amount of protection provided is based on the source rate-distortion curve [29]
and channel state. Each packet is them encoded using convolutional code (after
being appended with CRC error detection) to protect it from bit errors introduced
by the wireless channel. At the receiver side, erroneous descriptions are detected
with the aid of CRC and are declared as erasured. RS code is then used to recover
23
MDC using the transform coding was first introduced by Wang et al. [82] and later
extended in [56,83,84]. The basic principle of multiple description transform coding
24
a e ch
+ n
^
s Source LSD(e) -1
LCD(u) Channel LCD (c) Symbol y
Decoder P Decoder Demapper
e a
LSD(e) LCD(u)
P
Figure 2.2: Block diagram of a generic baseband transceiver system with ISCD.
these two stages. Interleaver is also added in case of fading channel or if higher
order modulation is used. At the receiver, the demapper generates LLR-values
Lch
CD (c) from the noisy channel observations y. Subsequently, channel SISO decoder
and source SISO decoder exchange extrinsic information, LeCD (u) and LeSD (e). Note
that after deinterleaving/interleaving, extrinsic information of current stage becomes
a priori information (LaSD (e) and LaCD (u)) for the next stage. After a fixed number
of iterations, the source decoder generates estimates s of the transmitted source bits.
Some of the earlier works in this area are [100104]. In the literature ISCD has
been applied to both variable length and fixed length source codes. In [21, 105, 106]
authors exploit the efficienty of turbo decoding by integrating the variable length
codes (VLC) in a iterative decoding process. The decoding scheme considered a
VLC concatenated with a RSCC, and substantial gains were obtained with iterative
decoding. Iterative decoding by serial concatenation of VLC with a turbo code and a
LDPC code were proposed in [107109]. Bursalioglu et al. [92] proposed a lossy joint
source-channel coding scheme using Raptor code. Raptor codes [110] are a class of
rateless codes designed for transmission over erasure channel with unknown capacity.
Use of Raptor codes allows continuum of coding rates with a single basic decoding
algorithm. In recent times, Chase-like decoding [111] was applied for SISO decoding
of VLC, which in turn was applied to ISCD [112115]. Chase-like decoding have the
advantage of lower computational complexity compared to trellis based decoding
of VLC. A new scheme was proposed in [112] for joint source-channel decoding of
arithmetic codes based on maximum a posteriori sequence estimation. This scheme
performs Chase-like decoding using a priori knowledge of the source symbol sequence
and the compressed bit-stream length. In [113], Zribi et al. extend the Chase-like
SISO decoding scheme to Huffman code and consider a serial concatenation of this
decoder with channel code. Significant performance improvement was obtained with
iterative decoding between the constituent decoders.
Another area of ISCD which has received considerable attention from research
community is joint source channel coding of correlated sources. Instead of utilizing
the residual redundancy after source coding, there are several research which focus
on exploiting the memory structure of the source directly at the decoder [116]. This
28
type of joint source-channel coding has two major advantages. At the transmitter,
the information symbols are directly encoded by the channel code and no separate
source encoder is required. This simplifies the encoder structure and reduces the
power consumption. Redundancy provided by the memory structure of source im-
proves the error correction capability of the system. Designs with Hidden Markov
model source and turbo code was presented in [117119]. Design combining LDPC
codes and hidden Markov model was presented in [120, 121]. Similar improvements
were obtained with Irregular Repeat Accumulate (IRA) codes in [22] and with linear
systematic block codes [122]. Fresia et al. [123] proposed a scheme based on rap-
tor codes, which are a new class of rateless codes, for decoding Markov correlated
sources. Kim et al. [124] study joint source-channel decoding for non-binary source
samples, where source samples are modelled as the output of a multi-state Markov
chain (MC). A simplified Markov chain model was proposed where the multi-state
Markov chain is replaced with multiple number of two-state Markov chains, each of
which exploits bit-level correlation of samples.
Recently, motivated by the joint source-channel decoding (JSCD) principle of
exploiting the source redundancy, authors study the application of ISCD conceived
for distributed video coding, where the video signal is modeled using iterative
horizontal-vertical scanline model relying on a first-order Markov process [125]. Re-
search in [126] exploit the redundancy of language-based source for iterative source
channel decoding of polar code. To improve the performance of JSCD, a prac-
tical and efficient algorithm was proposed in [127], in which the performance of
the JSCD is enhanced by the additional iteration to exchange additive information
between multiple sources. An improved source APP decoding approach was pro-
posed for packetized variable-length encoded Markov sources [128]. The authors
also proposed a recursive implementation based on a three-dimensional joint trellis
for symbol decoding of binary RSCC.
29
Basics of Relaying
Relay network is a three terminal communication system, with terminals labeled the
source (S), the relay (R), and the destination (D). The objective of such a system is
to communicate information generated at S to D. The relay R aids in transmitting
information without actually being a source or sink. The signal transmitted from
the source is X and is received by relay as V . The signal transmitted from relay
is W , and the signal received at the destination is Y . Conceptually, relaying is
done in two phases. In the first phase P1, called broadcast (BC) phase, S transmits
and (R, D) receive. The second phase is multiple access (MAC) phase in which
(S, R) transmit and D receives. This distinction between phases is conceptual only
since for communication both phases may take place simultaneously. According
to [129], there are four different models of relaying based on the above two phases
of transmission.
1. P1: S transmits and (R, D) receives; P2: (S, R) transmit and D receives.
R
V
W
X Y D
S
P1 P2
First model is the most general and earlier information theoretic results on relay
channel were based on this model. In the second model D ignores the signal from S in
the BC phase, whereas, in the third model, S does not transmit in the second phase.
The fourth model is simply a multi-hop communication where relay only forwards
the received message to D and D can not directly receive from S. Unlike the first
three models, multi-hop communication does not have the benefit of diversity gain.
Next, we touch upon the many approaches to relay based cooperative diversity
proposed over the years. In case of direct or point-to-point link a single user com-
municates with a single destination. Three terminal relay channel was introduced
by van der Meulen [130], where he gave the upper and lower bound on the capacity
of relay channel. In [131], relay system was studied in the context of Aloha system.
Most prominent work in this field has been the paper of Cover and El Gamal [132].
In this work, authors evaluated the capacities of the Gaussian relay channel and an
achievable lower bound to the capacity of the general relay channel is established.
Due to the difficulty of finding new and better information theoretic results, and
practical challenges of implementing user cooperation, research interest in relaying
and user cooperation waned in early 80s [129].
One of the prominent works in recent times that received considerable attention
is [133, 134]. In this work, Sendonaris et al. proposed user cooperation as form of
diversity and show its benefit. Other important research which studied the perfor-
mance of different relaying protocols in fading environment are [135, 136]. In [136]
space-time codes were applied for obtaining cooperative diversity in relay networks.
31
Kramer et al. [137] gave a new set of insights into the information theoretic as-
pects of coding for relay channel, where authors study two basic relaying schemes -
decode-and-forward and compress-and-forward. Recent spur of activities in the field
of wireless sensor networks have also resulted in some fundamental research in relay
networks. In one of the most discussed work, Gupta and Kumar [138] studied the
transport capacity of wireless networks. The work measured the end-to-end sum
throughput of the network multiplied by the end-to-end distance. Related publica-
tions also includes [139141].
To harvest the gains predicted by information theoretic analysis, suitable coding
schemes are needed that can approach the fundamental limits of the relay channel.
Coded cooperation [142, 143] was one of the earliest techniques to obtain coopera-
tion between relay nodes using RSCC and turbo codes. In [144], Valenti and Zhao
used distributed turbo codes to approach the information-theoretic bound on outage
event probability of decode-and-forward relaying. Zhang and Duman [145] devel-
oped capacity approaching turbo coding schemes for half-duplex relay systems and
in [146] a full duplex scheme with iterative decoding was proposed. Recent ad-
vancements in LDPC codes have also been applied to relay channel. Chakrabartiet
al. [147] proposed LDPC code designs for half duplex relay channel. Their designs are
based on the information theoretic random coding scheme for decode-and-forward
relaying. This and other similar schemes [148150] utilized density evolution tech-
niques to search for optimized irregular LDPC ensembles operating at two different
rates. Nguyen et al. [151] proposed high-performance protograph-based LDPC cod-
ing scheme for half duplex relay channels.
Network coding was first proposed by Ahlswede et al. [152] to enhance the capacity
of noiseless wired network. The main idea of network coding is to allow coding at
the relay nodes, in exchange for network capacity gain [153]. This is in contrast to
conventional networks where relay or intermediate nodes just forward the informa-
tion. Increase in throughput by the use of network coding was illustrated by the
famous example of Ahlswede, as shown in the Figure 2.4. This network is referred
32
y
1 t1
y y
x y
s 3 4
x y x y
x x
2 x t2
Figure 2.4: The butterfly network: Example of network coding in wireline network
[152].
to as butterfly network, which is a multicast network with one source node and two
sink nodes. In this network, every edge represents a directed link that is capable
of carrying a single packet reliably. There are two packets, x and y, present at the
source node s, and we wish to send these two packets to both the sinks, t1 and t2 . In
this capacitated network, the desired multicast connection can be established only
if one of the intermediate nodes breaks from the traditional routing paradigm and
performs a coding operation. As seen in the figure, network coding is performed
at the node 3, which is simply binary sum or XOR operation between the bits of
two incoming packets. Thus if the input packets of node 3 are x and y, the net-
work coded output is x y. The destination can decode the received packets by
performing the same operation. Sink t1 receives packet y and recovers the packet x
by performing further XOR of y with x y. Similarly, sink t2 recovers its packets.
Thus, the butterfly network illustrates that network coding can increase throughput
for multicast in a wireline network. The whole communication of two packets is
completed in nine transmissions. Without network coding, it would require more
number of transmissions to communicate the two packets.
The pioneering work of [152] led to explosion of research on network coding.
Though most of these studies have focused on wired and error free networks, there
is some initial work [154] investigating network coding in wireless scenarios. Since
noiseless assumptions are no longer valid for wireless channel, network coding for
wireless channels may incorporate some form of error correction. Network coding
for wireless channels can utilize broadcast nature of wireless channel and should
be designed taking into consideration challenges like noise, fading and inter-symbol
interference. Schemes to obtain cooperative diversity through network coding to
33
S1 S1
x x
RN 1 BS BS
x y
RN 2 +RN
y y
S2 S2
(a) (b)
Figure 2.5: The uplink for two mobile stations (S1 and S2 ) with cooperative diversity.
(a) Diversity obtained by simple relaying, (b) Network coded diversity [28].
combat fading in wireless channel were proposed in [153, 155, 156]. Network coding
for wireless channel not only allows diversity gain, but also allows trade-off between
diversity gain and throughput. A simple model where diversity can be obtained with
network coding is shown in the Figure 2.5, where two mobile users (S1 and S2 ) are
transmitting packets to the base station (BS). With conventional relaying (Figure
2.5(a)), two mobile users will require two separate relays (or a single relay used at
two different time slots) for transmitting to the BS. If the mobile users transmit
cooperatively with the help of network coding at the relay node (RN), only one
transmission from relay is sufficient to obtain same diversity gain for both mobile
stations.
One approach to gain diversity through network coding for multiple access relay
channel (MARC) with noisy channels is to decode network and channel code sepa-
rately. The channel coding used in the physical layer transform the noisy channels to
erasure based links. At the network layer, if information packet from one of the user
is lost, it can be recovered easily by simple XOR operation between the packet re-
ceived from other mobile user and RN. However, a relay not only provides diversity,
but can also provide additional capacity in exchange for diversity. This additional
redundancy improves performance compared to point-to-point communication if the
relay has better connection to the base station than mobile station.
In the earlier discussion, application of network coding in network layer were only
considered and it was assumed that the lower layers deliver error free packets with the
34
help of channel coding. Similar to the joint source-channel coding (Sec. 2.3), Hausl
and Dupraz [28] proposed that in joint network-channel coding, where redundancy
in the network code is used to support the channel code for better error correction.
Again similar to source-channel coding, Effros et al. [157] stated that capacity can
only be achieved by treating network and channel coding jointly in a wireless relay
network. Del Ser et al. [158] proposed a joint source-channel-network decoding and
turbo equalization scheme for MARC. Focusing on densely deployed wireless sensor
networks, Del Ser et al. [159], proposed a joint source-channel-network coding scheme
for MARC. They assumed the correlateion between the information from two mobile
used to be hidden Markov model (HMM) [160] correlated and apply blind estimation
of HMM parameters at the receiver. Menghwar et al. [161,162] investigated a scheme
in which space-time coded cooperative nodes are opportunistically aided by a relay
node which performs networks coding. In [163], authors combine network coding
with full diversity channel coding to give rise to a new family of semi-random full-
diversity joint network-channel LDPC codes.
Some recent research related to network coded cooperation are as follows. The
diversity order attainable by network coded cooperation with channel coding in
a network topology with multiple sources, one relay and in the presence of fast
Rayleigh fading was analyzed in [164]. In [165], authors studied analog network
coding in the MARC with multiple relays and provided closed-form expressions of
the optimal power to be allocated to the sources and the relays. A generalized analog
network coding scheme for a non-orthogonal MARC was proposed in [166], where two
sources transmit their information simultaneously to the destination with the help
of a relay. An estimate-and-forward protocol in MARC was designed in [167] where
two correlated sources transmit their information to the common destination with
the help of a relay. The network-coded soft symbols contain redundant information
due to the correlated sources, which were compressed before being forwarded.
35
codes (RS and RSCC) exchange information between the constituent decoders and
improved decoding performance in terms of PSNR values was demonstrated [172].
Application of iterative decoding in the context of network coding is proposed
for transmission of data. A network coded relay can be considered to be a cascade
of channel code at the transmitting nodes and network code at the relay [173]. This
enables implementation of turbo-like decoding scheme and can provide significant
improvement by performing iterative decoding compared to the existing schemes
[174, 175].
This work aims to develop iterative decoding schemes for transmission of image and
data. Application of iterative decoding techniques to two different types of multiple
description coding (MDC) techniques for images have been looked into; namely, FEC
based n-channel MDC and MDSQ based two-channel MDC. Iterative decoding is
also applied for improvement in bit error rate (BER) performance in case of network
coded relay systems.
2.5.2 Objectives
From literature survey we find that in FEC based MDC techniques, hard decision
output of inner channel code is passed on to outer RS code and RS decoder does
not pass any information back to inner code. Recently, Chang et al. [80] proposed
an iterative decoding technique for FEC based MDC, in which hard decision is
exchanged between the constituent decoders. Since, hard decisions typically result
in loss of information, we propose to utilize recent development in SISO decoding of
RS codes [176] for developing soft iterative decoding scheme for FEC based MDC.
Recently Liu et al. [177] proposed LDPC-RS product codes and its decoding
algorithm in the context of cloud transmission. Cloud transmission [178] has been
proposed for achieving high degree of robustness against co-channel interference and
improving spectral efficiency. In the line of FEC based multiple description coding
literature, which uses RSCC with RS code to form a 2D product code structure, a
37
iterative decoding scheme for LDPC-RS code based MDC was developed.
Because of the enormous capacity increase multiple-input multiple-output (MIMO)
antenna systems offer, such systems gained a lot of interest in mobile communication
research [179]. MIMO systems help mitigate fading in wireless channel by provid-
ing space diversity. Space-time coded transmission over MIMO, introduced first by
Tarokh at el. [180], provide both space-time diversity and spectral efficiency. Dif-
ferential space-time block codes (DSTBC) has been shown to be a recursive code
and can be decoded using iterative techniques [181, 182]. Using iterative decoder
of [80] as outer decoder, a three stage iterative decoding technique is proposed for
differential space time coded FEC based multiple descriptions, in which DSTBC act
as innermost code.
A two-channel wavelet based MDC scheme was proposed by Song et al. in [64].
This scheme provides protection against packet loss and random bit errors, and has
potential for improvement in AWGN channel. Since, this scheme does not utilize soft
decision available at the output of AWGN channel, SISO arithmetic decoder [115]
is applied for decoding of multiple descriptions transmitted over AWGN and packet
loss channels.
In [173], Ben Slimane et al. proposed an hard decision iterative decoding tech-
nique for information transmitted by relays using network code in a MARC setup.
Based on ideas developed for RSCC-RS soft iterative decoder, we proposed a soft
decision decoder in the context of network coded relays in MARC setup. Later we
extended these ideas to the research in [162]. Menghwar et al. [161, 162] proposed
schemes which combines the benefits of space-time codes and network coding for co-
operative communication in MARC setup. We improve upon the research in [162] by
application of the concept of product code and iterative network-channel decoding.
In light of the literature survey carried out and the foregoing discussion, following
objectives are identified and addressed in this work.
In [170, 183, 184], an iterative decoding algorithm utilizing RS and RSCC was pro-
posed for FEC based MDC transmitted over wireless channels. Based on [185], RS
39
code and RSCC utilized in FEC based MDC is treated as compound code and de-
coded using iterative soft decision decoding. To obtain soft decision output, BCJR
algorithm [12] is applied for decoding RSCC (instead of Viterbi decoder as in [73,78])
and RS code is decoded by utilizing adaptive belief propagation algorithm [176] (in-
stead of hard decision Berlekamp-Massey algorithm). This allows soft information
available at the output of BCJR decoder for decoding RS coded multiple descrip-
tions and allows decoding capabilities beyond maximum separable distance (MDS)
code. Decoding is done iteratively by passing extrinsic information between both
the decoders till CRC are satisfied or maximum number of iterations are completed.
Simulation results show significant improvements in reconstructed image PSNR in
both AWGN and Rayleigh fading channels.
In [183], proposed soft iterative multiple description decoder is applied for improv-
ing the quality of image received through coded oppertunistic relay network. Multi-
terminal wireless networks like ad-hoc and sensor networks have emerged as a poten-
tial area for dissemination and collection of multimedia information. In such multi-
terminal wireless networks, cooperative communication can deliver spatial diversity
through the collaboration between users [186]. Specifically, we consider transmis-
sion over coded opportunistic amplify-and-forward (COAF) and coded opportunistic
decode-and-forward (CODF) network proposed in [187], which use RS-RSCC at the
transmitting node. Simulation result show improvement in received PSNR compared
to system which did not employ iterative decoding.
A three stage iterative decoding scheme was proposed for FEC based multiple de-
scriptions received through MIMO channel in [189]. The three decoding stages con-
sist of hard decision RS decoder, SISO MAP (BCJR) decoder for RSCC and SISO
decoder for differential space-time block code (DSTBC). In the proposed scheme,
SISO DSTBC decoder and MAP decoder exchange soft information in the form
of extrinsic information, whereas hard decision on the output of MAP decoder for
RSCC is passed on to hard decision RS decoder. The feedback provided by the
RS decoder acts as a priori information for the MAP decoder. With simulation
improvement in PSNR performance for the proposed scheme is demonstrated.
In [190], a low complexity, joint source-channel decoding scheme for arithmetic codes
is studied. We proposed a unity-rate accumulator based design for low complexity
Chase-like decoding of arithmetic codes. Chase-like decoding is a low complexity
algorithm, where a maximum a posteriori sequence estimation criterion is employed
41
for maximum likelihood decoding of variable length codes like arithmetic codes.
Earlier research proposed iterative decoding of SISO arithmetic codes with RSCC
and LDPC codes. We proposed application of unity rate accumulator as inner
encoder and decoder in the system, which improved the bit error performance of the
system by 1.25 dB, with same number of decoding iterations. Performance of the
proposed scheme was evaluated for image transmission application.
In [168], joint source-channel decoding of arithmetic codes was applied to the case
of two-channel scalar quantizer based MDC. In this scheme, multiple descriptions
of wavelet transformed image are generated and coded using SPIHT compression
algorithm for achieving robust transmission over packet-loss channel. Two differ-
ent schemes are considered in which we apply SISO Chase-type arithmetic decoder
for noise robust reception of the multiple descriptions over additive white Gaussian
channel (AWGN) in contrast to earlier works which consider reception over binary
symmetric channels. We utilize built-in arithmetic coder of SPIHT algorithm for
achieving noise robust transmission of multiple descriptions without increasing en-
coder complexity and utilize low complexity SISO decoding of arithmetic codes.
Finally, performance of these schemes were evaluated and compared to existing re-
search literature.
In this research [169], our contribution lies in improving the scheme of Song et
al. [64] by adapting Chase-type SISO arithmetic decoder for decoding of multiple
descriptions transmitted over additive white Gaussian noise (AWGN) and packet
loss channels. Scheme proposed in [64] consists of SPIHT coding of wavelet trees
along spatial orientation for both descriptions. In order to maintain high compres-
sion efficiency and robust synchronization, EREC [95] is applied to reorganize these
variable length bitstreams into fixed length slots before transmission.
In the proposed scheme, SPIHT coded bitstreams are arithmetic coded to further
increase compression efficiency and use of soft channel information at the receiver
42
improves the noise robustness of the arithmetic decoder. Combining EREC with
iterative source-channel decoding of arithmetic code is not quite straightforward as
EREC requires instantaneous hard decision decoding of the source symbols. Our
innovation lies in applying EREC assisted by state and tail bits (ERECST) [191] with
arithmetic coded SPIHT trees of variable lengths for SISO decoding of arithmetic
codes. This enables iterative source-channel decoding of EREC coded fixed length
slots at the receiver and improve PSNR gain significantly.
In the research [174], joint network-channel coding for MARC is considered when
the source node employs RS code as error correcting code and punctured RSCC
is used as network code. Encoding user information with RS code, in contrast to
simply encoding with linear block code [173], is more relevant because of widespread
prevalence of RS code in many existing standards. Our main contribution is applying
the concept of iterative soft decoder for concatenated codes [185] in the context of
network-channel decoding in MARC setup. The proposed algorithm enables network
code and channel decoders to exchange soft information iteratively and to achieve
an improved performance compared to a hard decision iterative decoder.
In [175], bit error rate performance is enhanced for a distributed space-time coded
system aided by a relay which performs network coding. Similar to earlier scheme,
the proposed scheme uses punctured RSCC as network code at the relay and RS
code as channel code. In addition both the user nodes cooperate using distributed
space-time code. Our main innovation is iterative network-channel decoding at the
receiver using the principles derived from soft decoding of concatenated RS convolu-
tional codes. The proposed algorithm enables network code and channel decoders to
exchange soft information iteratively and may yield a capacity approaching perfor-
43
2.7 Conclusion
In this chapter, the background information related to iterative decoding for codes
received over wireless channel has been presented. An exhaustive literature survey
of the state of art in modern channel codes employing iterative decoding, FEC based
MDC, quantizer based MDC, cooperative diversity in wireless networks and network-
channel coding has been given. Problem statement, objective and motivation behind
the problem statement was also stated in this chapter. In the final section, our
contribution to the research literature has been presented.
******
Chapter 3
44
45
3.1 Introduction
In the recent years, mobile wireless devices have become a dominant means of com-
munication for multimedia. Real time multimedia streaming applications require
timely delivery of information and are transmitted on the packet network on a
best effort basis using protocols like UDP or RTP, rather than a connection ori-
ented protocol like TCP/IP. To maintain necessary quality of service the receiver
should be able to tolerate packet loss using methods like multiple description cod-
ing. Each packet should also be protected using channel codes for protection against
noise. Many multimedia coders employ a progressive mode of transmission, where,
as more and more bits are received, the source can be reconstructed with better qual-
ity. Earlier works considered transmission of progressively coded image stream over
wireless channel using rate-compatible punctured convolutional codes, where more
important bits are transmitted with lower rate (stronger error correction) codes.
Multiple description coding has come up as an attractive framework for robust
transmission of image and video over packet erasure channels [56]. Recently FEC
based multiple description coding has received attention in literature because of its
ability to produce arbitrary number of balanced or equally important descriptions
from a single progressive bitstream. Initially, FEC based multiple description coding
was envisioned for packet loss channels and consisted of RS codes (but providing
unequal error protection) arranged in 2-D matrix form. Loss of any packet was
recoverable because of erasure correction property of RS codes. This scheme also
provided graceful degradation of received image quality in case of packet losses.
Though this arrangement provided protection against packet losses, it was further
adapted by Sachs et al. [73] for wireless channels by encoding each description with
convolutional code. Thus, in such an encoding scheme, RS code is used to mitigate
packet erasures, whereas convolutional code provides protection against bit errors
(caused by noise and fading in wireless channel). In the scheme of [73] and sub-
sequent research, only hard decision of channel decoder (Viterbi decoder in case of
convolutional code) is applied to FEC decoder (RS code). Due to hard decision
performed at this stage, some information is lost which could have helped in better
decoding performance. Improvement in performance can be obtained if both the
46
codes are considered a single system and could exchange soft a priori information.
In this chapter, an iterative soft decision decoding algorithm for FEC based
multiple description code is proposed for transmission of progressively coded image
over wireless channels. In the proposed scheme, inner channel decoder accepts soft
channel reliability information and provides soft extrinsic information to the outer
decoder (RS decoder). This allows iterative soft decoding of RS codes, which in
turn allows decoding capabilities beyond maximum distance separable (MDS) code.
Iterative soft decoding of RS code has been historically challenging and was recently
made possible by the works of Jiang and Narayanan [176] and others [192, 193]. In
our actual algorithm, to obtain soft decision output, BCJR algorithm (for decoding
convolutional code) is applied on the channel reliability values, instead of Viterbi
decoder as done in most of the research in literature. This soft decision is applied
to SISO RS decoder to output soft extrinsic information. Thus decoding is done
iteratively by passing extrinsic information between both the decoders till CRC
are satisfied or maximum number of iterations are completed. The algorithm is
first evaluated over AWGN and Rayleigh flat fading point-to-point link channel.
Simulation is also performed for evaluating the performance over opportunistic relay
network. For relay network, two coded opportunistic cooperative schemes proposed
in [187] are considered, the coded opportunistic amplify-and-forward (COAF) and
coded opportunistic decode-and-forward (CODF).
Rest of the chapter is organized as follows. In Section 3.2, basic principles of FEC
based MDC is introduced. In Section 3.3, the proposed soft iterative FEC based
MDC decoder is explained. Iterative decoding of FEC based MDC transmitted over
oppertunistic relay network is described in Section 3.4. Simulation results are given
in Section 3.5 and finally, Section 3.6 is conclusion.
In the earlier section, a summary of FEC based MDC was given. The system for
transmission of progressively coded image using FEC based MDC and the proposed
iterative decoding algorithm is described next. Output of source encoder, con-
sisting of progressive bitstream, are grouped into symbols belonging to GF(q) and
encoded with (n, ki ) RS code, where, i = 1, ..., p. Multiple descriptions are formed
as described in earlier section. As shown in the Figure 3.2, FEC based multiple
description packets are communicated over packet switched network (e.g. internet)
to a wireless access point (WAP) for last hop delivery to an end user. In the packet
network some of the description may be dropped (e.g. due to congestion). Packets
received at the WAP are broadcast after being appended with CRC to detect errors
and encoded with channel code like RSCC or turbo code.
At the receiver log-domain BCJR algorithm is applied to the received noisy de-
scriptions (row by row, as in the Figure 3.1). BCJR decoder [12] is a soft-input
soft-output algorithm for decoding RSCC, in contrast to Viterbi decoder which is
typically outputs hard decision. The extrinsic information out of BCJR decoder is
extracted column by column and applied to SISO RS decoder. The SISO RS de-
coder applied in this work is adaptive belief propagation (ABP) algorithm proposed
49
BCJR Decoder
Berlekamp- (row by row)
Source Massey Check CRC Wireless
Decoder RS decoder (row by row) Channel
SISO RS decoder
(column by column)
ABP Algorithm
Figure 3.2: System model showing encoding, transmission and decoding of FEC
based MDC.
in [176]. Extrinsic soft decisions of RS decoder is again applied back to BCJR de-
coder. This iterative decoding algorithm is described in detail next section. After
finite number of iterations, CRC of each description is used to identify the correct
descriptions. Incorrect or missing descriptions are declared as erased. Attempt is
made to recover the erased symbols by using conventional RS hard decision decoders,
like Berlekamo Massey (BM) algorithm [194]. Image is then reconstructed from the
correctly decoded descriptions as in a conventional FEC based MDC decoder.
[c1,j,0 , c1,j,1 , ..., c1,j,m1 , c2,j,0 , ..., ci,j,l , ..., cn,j,m1 ]T (3.3.3)
On the other hand, a description is an array of symbols along the row. Binary
representation of ith description (or row) [ci,1 , ci,2 , ..., ci,p ] is
[ci,1,0 , ci,1,1 , ..., ci,1,m1 , ci,2,0 , ..., ci,j,l , ..., ci,p,m1 ]T (3.3.5)
51
ABP Algorithm
p e
LLR ordered Lch BCJR L1 L1 Le2 Berlekamp-
RS Parity-check SPA (one
into n rows
yi(t) (descriptions)
decoder
(row by row) + -
+a matrix adaptation iteration)
Massey
RS decoder
L2
a Column by column
L1
Figure 3.3: Block diagram of iterative soft decoder of FEC based multiple description
code.
CRC will be used for detecting errors in description and is not included above to
avoid cluttering the calculations. If we assume binary phase shift keying (BPSK)
modulation, the baseband equivalent of transmitted signal with unit symbol energy
is x0i (t) = (1 2xi (t)). Signal received at the destination is
a priori LLR La1 (ri (t)), t = 1, ..., pm. BCJR algorithm calculates the a posterior
LLR values Lp1 (ri (t)). Extrinsic information to be passed on to a subsequent stage
is calculated as Le1 = Lp1 La1 for all descriptions i = 1, ..., n.
The next stage consists of an iteration of ABP algorithm for calculation of ex-
trinsic information Le2 . This stage proceeds in two parts, updating the parity check
matrix and calculation of extrinsic information. The extrinsic information of pre-
vious stage Le1 after interleaving becomes a priori information La2 and is used for
initializing each iteration in this stage. The interleaving operation reorders LLR of
bits of descriptions (rows) into arrangement given by (3.3.3). We have a vector of
LLRs of jth RS code at lth iteration as
La2 (Cjcol ) = [La2 (cj (1)), La2 (cj (2)), ..., La2 (cj (nm))]T . (3.3.10)
To update the parity check matrix H, the magnitude of a priori LLR |La2 | is sorted
in ascending order of magnitude and sorting index are stored as = [t1 , t2 , ..., tnm ].
Columns of the matrix H are then reordered according to the permutation to ob-
tain (H). Gaussian elimination then reduces the first independent n0 k 0 columns
to identity submatrix. Let this matrix be denoted as H 0 . Finally, inverse permuta-
tion is performed on the columns of H 0 as follows
H 00 = 1 (H 0 ) (3.3.11)
This extrinsic information is reordered into the form as in (3.3.5) to get a priori LLR
La1 for the MAP decoder. After Niter number of SISO iterations between RSCC and
RS code, the a posteriori LLR of each RS bit is computed as
Lp2 (cj (t)) = La2 (cj (t)) + Le2 (cj (t)) (3.3.13)
until parity checks are satisfied. It may happen that some of the RS codewords
will satisfy parity check conditions earlier than others; the extrinsic LLR of those
codewords are not calculated anymore and corresponding a posteriori LLR are saved.
To summarize the proposed algorithm for iterative decoding of FEC based multiple
descriptions can be stated as follows.
Algorithm 3.1
Definition:
yi (t): Received signal value at time slot t of i-th description.
t : Channel fading coefficient at time instant t.
2 : Noise variance.
Niter : Maximum number of decoding iterations.
: Damping coefficient.
1. Initilization
(a) Set , Niter , and LLR of the coded bits from channel observations as
Lch (xi (t)) = 2yi (t)t / 2 .
(b) Create a binary array of bits f lag of length p and initialize it with zeros.
(c) Initialize the array of binary parity check matrices, Hkn , k = 1, ..., n 1,
for (n, k) RS code.
4. Reorder the matrix containing extrinsic LLR out of step 3, where rows of the
form in (3.3.5) are transformed to a matrix containing columns of form (3.3.3).
(Each column now consists all the of bits of an RS codeword.)
(a) Sort the magnitude of LLR |La2 (Cjcol )| in ascending order and save the
ordering indices as .
(b) Select the appropriate binary parity check matrix for (n, ki ) RS code as
H = Hkn .
(c) Arrange the columns of H according to the order and systematize the
first n0 k 0 columns using Gaussian elimination. Rearrange the columns
back to their original positions to get the modified parity check matrix
H 00 according to (3.3.11).
(e) Calculate bit level reliabilities Lp2 (cj (t)) according to (3.3.13).
6. Reorder extrinsic LLR Le2 into matrix La1 consisting of rows of the form (3.3.5).
Each row La1 (ri (t)) contains a priori LLR for description i.
8. Decode the RS codes from hard decisions using BM algorithm and extract the
information symbols.
r=1 r=1
r=2 r=2
S D S D
r=b
r=R r=R
(a) Broadcasting phase (b) Relaying phase
which uses product code consisting of RS and RSCC. Channel model for both the
relay aided transmission schemes are described next. The relay network consists of
source (S), which transmits the FEC coded multiple descriptions, and destination
(D), along with a set of R relay candidates Sr , r = 1, , R. In opportunistic
cooperation network, one of the relay nodes will be selected for retransmission of
message transmitted by S. The channel model focuses on the case of slow fading
to isolate the benefit of spatial diversity obtained by relaying. All the transmissions
are assumed to be in orthogonal channels, i.e. different time slots or frequencies.
Both the coded opportunistic transmission schemes can be described by a two phase
transmission as indicated in the Figure 3.4. In the broadcast phase, source S will
broadcast message to destination D and all the relays as:
yD (t) = 1 SD xS (t) + nD (t), t = 1, 2, ..., l (3.4.16)
yr (t) = 1 Sr xS (t) + nr (t), r Sr , t = 1, 2, ..., l (3.4.17)
where, yD (t) and yr (t) are signal received at D and a relay r Sr , respectively, in the
first phase of transmission. The coded and modulated symbols are indicated as xS ,
and SD , Sr are fading coefficient of the channel from S to D and r, respectively.
In the relaying phase, one of the relay b Sr is selected as the best relay, the
retransmitted signal received at D is
yD (t) = 2 bD xb (t) + nD (t), t = l + 1, l + 2, ..., 2l (3.4.18)
where is the amplification factor. The selected relay simply amplifies the received
p
signal by a factor = 2 /(|Sr |2 1 + N0 ) and retransmits towards the destination.
In CODF, all the relays will try decoding the message received from S. Only
those relays that can decode the received message correctly will be selected for
re-transmission. If the transmission rate of the system is R bits/s/Hz, a correct
decoding at the relay r requires:
Selected relay in CODF scheme decodes the RSCC of the packets and if there is not
any error, encodes it again with RSCC. Re-encoded packets are then relayed to the
destination. At the destination iterative decoding is performed in similar manner
described in Section 3.3.2. In [187], it was shown that max criteria can enable the
CODF scheme to fully exploit the diversity gain. Sr can be an empty set when
none of the relays can decode the received message correctly. In such a scenario, S
will re-transmit its signal again in the relaying phase. Consequently, in the relaying
phase, D receives
of 11 dB is obtained.
In the next set of simulations, performance is compared to a baseline scheme
consisting of similar encoder and decoder, but with RSCC and RS code functioning
in tandem (indicated as tandem scheme), i.e. without any SISO iterative decoding.
In the tandem scheme, the RSCC is decoded using soft input Viterbi algorithm;
whereas, RS code is decoded using conventional hard-decision BM algorithm. The
proposed algorithm (indiacted as SISO CC-RS)is also compared to a FEC based
multiple description system proposed by Chang et al. [80] (indicated as reference
scheme), which utilizes iterative RS-RSCC decoder. The iterative decoder in ref-
erence scheme exchanges hard decisions between component decoders for obtaining
improved PSNR performance. For the purpose of fair comparison, the iterative
decoder of [80] is implemented under similar encoders and channel to that of the
proposed scheme. For all the subsequent simulations, to obtain improved results and
to maintain reasonable decoding complexity, we shall perform 5 decoding iterations
of the proposed algorithm.
In the Figure 3.7, performance of the proposed scheme is compared to tandem
scheme and reference schemes when descriptions are passed through AWGN channel.
As seen in the figure, the proposed SISO scheme achieves maximum PSNR value
at channel SNR of 2 dB, whereas, the tandem scheme reaches maximum PSNR at
channel SNR of 3 dB. Thus, proposed SISO CC-RS scheme provides a gain of 1 dB
compared to tandem scheme. SISO CC-RS scheme also provides significant gain in
decoded image PSNR compared to reference scheme, with maximum iteration gain
of 12 dB at SNR of 0.5 dB. Although the PSNR performance is comparable at SNR
greater than 1.5 dB, the proposed scheme greatly outperforms the reference scheme
for channel SNR values less than 1.5 dB. In Table 3.1, performance comparison of
the proposed scheme with tandem scheme is shown for other test images.
Next the comparison is performed for fading channels. It is assumed that channel
state information is available at the receiver. Figure 3.8 shows the performance over
Rayleigh flat fading channel. The channel is fast fading, i.e. fading coefficients are
varying with every symbol period. The results show that the proposed algorithm
provides significantly improved performance compared to the tandem scheme for
60
32
1 iter
30 2 iter
3 iter
4 iter
28
5 iter
10 iter
26
Average PSNR
24
22
20
18
16
14
-1.5 -1 -0.5 0 0.5 1 1.5 2
SNR (dB)
Figure 3.5: Average PSNR versus channel SNR for the case of transmission of Lena
image coded at rate 0.2 over AWGN channel. Improvement of PSNR at different
iterations of the proposed scheme is shown.
32
1 iter
30 2 iter
3 iter
4 iter
28
5 iter
Average PSNR (dB)
10 iter
26
24
22
20
18
16
14
0 0.5 1 1.5 2 2.5 3 3.5 4
SNR (dB)
Figure 3.6: Average PSNR versus channel SNR at different iterations for the case
of transmission of Lena image coded at rate 0.2 over Rayleigh flat fading channel.
61
32
30
28
26
22
20
18
16
SISO CC-RS
14 Tandem CC-RS
Ref. scheme
12
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3
SNR (dB)
Figure 3.7: Average PSNR versus channel SNR for the case of transmission of Lena
image coded at rate 0.2 over AWGN channel. Performance of the proposed scheme
is compared to two other schemes.
channel SNR < 5 dB. The improvement is more evident for lower values of channel
SNR. The results are mixed when proposed scheme is compared to reference scheme.
For the case of fast fading channel, PSNR performance of the proposed algorithm is
better for lower range of SNR values (< 2.5 dB). It can also be observed that PSNR
degradation in the proposed scheme is more graceful compared to the reference
scheme where there is an abrupt drop in PSNR at channel SNR < 3 dB. In Table
3.2, PSNR performance is compared with tandem scheme for some other test images.
Figure 3.9 depicts the PSNR performance comparison in quasi-static Rayleigh
fading channel. By quasi-static we mean the channel fading coefficient is constant for
at least the duration of transmission of a packet or RSCC. In this channel, the pro-
posed scheme outperforms both the reference scheme and baseline tandem decoding
scheme, for all values of channel SNR. The proposed soft-iterative decoding scheme
achieves maximum value of PSNR at channel SNR of 6 dB. The SNR requirement
to achieve maximum PSNR is 4 dB less compared to both the other schemes. On
the other hand, the maximum gain in PSNR compared to reference scheme is about
62
32
30
28
26
Average PSNR (dB)
24
22
20
18
16
SISO CC-RS
14 Tandem CC-RS
Ref. scheme
12
0 1 2 3 4 5 6 7
SNR (dB)
Figure 3.8: Comparison of PSNR performance of the proposed decoder with refer-
ence decoders for the case of transmission of FEC based MDC over Rayleigh fast
fading channel.
30
28
Average PSNR (dB)
26
24
22
20
SISO CC-RS
18 Tandem CC-RS
Ref. scheme
0 1 2 3 4 5 6 7 8 9 10
SNR (dB)
Figure 3.9: Comparison of PSNR performance of the proposed decoder with refer-
ence decoders over quasi-static fading channel.
63
Table 3.2: Performance advantage of proposed scheme in Rayleigh flat fading channel
Reconstructed image PSNR (dB)
Image SNR 3.0 dB 3.5 dB 4.0 dB 4.5 dB 5.0 dB
Next, performance of the proposed decoding scheme for FEC based MDC trans-
mitted over COAF and CODF relay channels is evaluated. In the Figure 3.10 and the
Figure 3.11, proposed iterative scheme is compared to baseline tandem RS-RSCC
system proposed in [187]. The baseline scheme is adapted for the case of transmis-
sion of FEC based multiple descriptions image information. The simulations are
performed for the scenario where transmission is aided by 1 to 4 relays. It is also
assumed that total energy of transmission per symbol is = 1 and 1 = 2 = 1/2.
It is also assumed that all the source-relay and relay-destination links exhibit sta-
tistically similar qualities. Simulation results show that by increasing the number
of relay candidates, significant performance gain is achieved. Figure 3.10 shows the
64
performance of the proposed and baseline scheme for COAF relay channel. It can
be seen that with the application of iterative decoding, there is marked improve-
ment in PSNR compared to baseline COAF scheme. When aided by a single relay
(R = 1), the proposed scheme provides a maximum PSNR gain of 9 dB at channel
SNR of 3 dB. Whereas, for the case of four relays, the maximum PSNR gain is 7 dB
at SNR of 2 dB, indicating reduction in gain obtained by iterative decoding. It is
because diversity gain provided by increased number of relays reduce the possibility
for further iterative decoding gain.
Figure 3.11 shows the performance results for CODF relay channel. The results
in this case also indicate the superiority of the proposed scheme over CODF relay
channel. For the case of single relay R = 1, iterative decoding results in maxi-
mum PSNR gain of 4.5 dB compared to baseline scheme at channel SNR of 3 dB.
Maximum PSNR gain for R = 4 is of 5 dB at channel SNR of 2 dB.
3.6 Conclusion
In this chapter, an iterative soft decoding scheme for FEC based multiple descrip-
tions transmitted over wireless channel has been presented. The proposed decoding
algorithm is based on exchanging extrinsic information between FEC coded descrip-
tions and channel decoder. The proposed scheme allows soft information available
at the output of BCJR decoder (for decoding RSCC) to be applied for decoding
Reed-Solomon coded multiple descriptions and allows decoding beyond MDS capa-
bility. This scheme has been tested for transmission of SPIHT coded image over
point-to-point wireless channel and wireless channels aided by relays. Specifically,
for the case of relay channel, simulations have been performed for COAF and CODF
schemes. It has been shown that the proposed method can give considerable gains
in terms of reconstructed image PSNR, with respect to iterative schemes proposed
in literature.
******
65
32
30
26
24
22
Iterative, R=4
20 Iterative, R=3
Iterative, R=2
Iterative, R=1
18
Tandem, R=4
Tandem, R=3
16 Tandem, R=2
Tandem, R=1
14
0 1 2 3 4 5 6
SNR (dB)
32
30
28
Average PSNR (dB)
26
24
22
Iterative, R=4
20 Iterative, R=3
Iterative, R=2
18 Iterative, R=1
Tandem, R=4
Tandem, R=3
16 Tandem, R=2
Tandem, R=1
14
0 1 2 3 4 5 6
SNR (dB)
In previous chapter, an iterative soft decision decoding of FEC based multiple de-
scriptions was proposed. Improvement in PSNR performance was achieved at the
cost of higher computational complexity compared to traditional FEC based MDC
decoding techniques. The main contribution towards the complexity was from SISO
Reed Solomon decoder. In this chapter, two FEC based MDC techniques are pre-
sented which use hard decision decoder for Reed Solomon code, resulting in lesser
computational complexity. Superiority of the proposed schemes in terms of PSNR
is also compared to the existing schemes in literature.
4.1 Introduction
Recently, the cloud transmission system for next generation terrestrial broadcasting
has been envisioned [70]. This study is intended to use the radio frequency (RF)
spectrum more efficiently. Such a system is robust to co-channel interference and
immune to multipath distortion. Hence, it can reduce or eliminate TV white-space
in order to be more robust and co-exist with unlicensed devices [197]. One of the
important feature of cloud transmission system is high degree of robustness against
66
67
RS product code proposed in cloud transmission literature [70, 177, 197] is applied.
OFDM is considered for transmission over AWGN and fading wireless channels.
The proposed algorithm takes advantage of correctly decoded codeword output of
RS decoder to enhance the error correction capability of the LDPC decoder at the
next iteration of sum-product algorithm (SPA). The algorithm works by selectively
modifying the log-likelihood ratios of the identified reliable bits. It is shown that
the proposed iterative LDPC-RS product code based progressive image transmission
scheme performs significantly better than baseline LDPC-RS coding scheme without
iterative decoding. Proposed system is also compared to a similar iterative design
for transmission of progressive image using RSCC-RS product code in [80] and
demonstrate improvement in performance.
In later part of this chapter, a scheme is proposed which extends the research
in [80] to include MIMO antenna system to achieve diversity gain in fading channel.
Our innovation lies in incorporating iterative decoding between FEC based mul-
tiple description decoder and MIMO demodulator. This results in two significant
outcomes. First outcome is improved decoding performance compared to similar
existing schemes in literature for coherent scenario. Secondly, for case without CSI,
the 3 dB performance gap is significantly narrowed down compared to a coherent
receivers. Simulations are performed for cases with and without channel state in-
formation (CSI) at the receiver.
Rest of the chapter is organized as follows. In Section 4.2, iterative decoding
scheme for LDPC-RS coded FEC based MDC is presented. In this section itself,
proposed encoding and decoding scheme is given and in Section 4.2.3, performance
of iterative decoder for LDPC-RS coded MDC is given. Iterative decoding of differ-
ential space-time coded FEC based MDC is presented in Section 4.3, with simulation
results given in Section 4.3.4. Finally, Section 4.4 is the conclusion.
69
RS codewords
5 7 I LDPC Parity
Frequency
Description 3 P P P 8 I LDPC Parity
Time
Figure 4.1: n-channel FEC based multiple description coding technique for a pro-
gressive bitstream.
In this section, proposed encoding scheme for FEC-based MDC is described which
use RS code as FEC code and LDPC code as channel code. An (n, k) RS code is a
maximum distance separable (MDS) code if dmin = n k + 1, where dmin is referred
to as minimum distance of the code. An (n, k) MDS code is a construction where
k information symbols are encoded into n channel symbols and reception of any
(ndmin +1) of the n channel symbols enables k information symbols to be recovered.
RS codes are typically decoded by Berlekamp-Massey (BM) algorithm [194]. In
RS code, each symbol is of m bits and source symbols for RS code are formed
by grouping m bits into a symbol. Let S = [S1 , S2 , ...], where Si GF (2m ), be
such a stream of symbols from a progressively coded source. The source can be
reconstructed progressively from the prefixes of the symbol stream S, while an error
renders the subsequent symbols useless. For example, if there is an error in symbol
i, source can be decoded using the preceding symbols [S1 , S2 , ..., Si1 ] and symbols
subsequent to Si cannot be utilized in the process of decoding.
A FEC-based multiple description code is obtained by applying unequal FEC
to different parts of symbol stream with the level of protection depending on the
relative importance of the information symbols and producing n equally important
70
where P r(x) the probability that x number of source bits are received without
encountering error. E(D) can also be expressed in terms of number of correctly
received description. Let P (i) denote the probability that number of correct de-
scriptions is i, then
1 1
kX
E(D) = D(0) (4.2.3)
i=0
p p kl+1 1
( ! !)
X X X
+ D mkp Pr P (i) (4.2.4)
l=1 i=1 i=kl
71
Wireless
Channel
Row by row operation
^ ~ ^
Source C RS Decoder C X Log-SPA+WBF Lch OFDM
Decoder (Col. by col.) P-11 Decoder P-12 Demodulator
^ ~I LI
C LLR
P1 modification
Figure 4.2: Encoding and iterative decoding of LDPC-RS coded FEC based multiple
description scheme.
The model of proposed system is shown in the Figure 4.2. In FEC based MDC
literature, encoding process consist of RS codes along the column and convolutional
code as rows [73, 74]. With our proposed application of LDPC code instead of
systematic convolutional code makes this coding scheme relevant to newer research
in the field of cloud transmission and terrestrial video broadcasting [177, 197]. Next
the proposed scheme is described in detail for the system model considered in this
paper.
The encoder, as depicted in top part of the Figure 4.2, consists of a progressive
source encoder (e.g. SPIHT) followed by RS encoder. The RS code based multi-
ple description encoder, described in earlier section, generates n equally important
descriptions C. The outer interleaver 1 is applied on each description so as to dis-
tribute the bits in a correctly decoded symbol throughout a packet or description.
1 is a bit-level interleaver in contrast to the symbol-level interleavers used in FEC
based multiple description coding literature. Each description is then individually
encoded with systematic LDPC code and bit level random inner interleaver 2 is
applied on the LDPC coded packet. 2 is a bit-level random channel interleaver
whose function is to distribute the effect of fading throughout the packet. The en-
72
where, t = 1, ..., mL/R and R is the rate of LDPC code. Standard log-SPA is used
in this scheme for decoding of LDPC code. In the proposed decoding scheme, SPA is
performed row by row. Given high computational complexity of SPA decoder, mixed
set of decoding technique shall be applied for LDPC code [177]. After each step of
SPA, if the results do not satisfy the parity constraints, parallel weighted bit-flipping
(PWBF) [200] is applied on the a posteriori LLR obtained from SPA. One step of
SPA includes several iterations of SPA decoder. PWBF and similar algorithms are
effective in trading off between error performance and decoding complexity. PWBF
is based on its bit-flipping function which provides each bit in the hard-decision
vector with a measure of its reliability. Unreliable bits are then identified and
flipped. If the flipped hard-decision vector satisfies the parity check matrix, we stop.
Otherwise, bit-flipping function is recalculated for each bit and bit-flipping process
is continued. Though SPA is the best performing decoder for LDPC code over
continuous output channels, PWBF has the advantage of faster convergence. It has
been demonstrated that PWBF converges in about five iterations with performance
close to that of standard belief-propagation algorithm [200]. Next computations
associated with SPA is explained.
For the sake of clarity of notations, let L(t) = Lch (xi (t)) be the initial LLR of the
t-th variable node, H is the LDPC parity check matrix. Then, for each H(s, t) = 1,
set L(t s) = L(t), where L(t s) is the message sent from variable node t to
check node s. Outgoing message from check node s transmitted towards variable
73
nodes is computed as
Y 1
L(s t) = 2 tanh1 tanh{ L(t0 s)} (4.2.6)
2
t0 N (s){t}
where, N (s) are neighbouring variable nodes of s. Each variable node computes the
outgoing message L(t s) as
X
L(t s) = L(t) + L(s0 t) (4.2.7)
s0 N (t){s}
where, N (t) are neighbouring check nodes of t. For each variable node t = 1, 2, ..., n,
a posteriori LLR Lp (t) is calculated as
X
Lp (t) = L(t) + L(s t) (4.2.8)
sN (t)
When hard decision is performed on a posteriori LLRs of the bits of i-th description,
the estimate of bits of LDPC codeword in i-th description is
1 if Lp (t) < 0
xi (t) = (4.2.9)
0 else
The decoded LDPC codeword Xi = [xi (1), xi (2), ..., xi (pmn(1) /k (1) )]. If Xi satisfies
the parity check constraints, any further decoding for description i is halted; else,
PWBF is applied on a posteriori LLR Lp (t) obtained from last SPA stage. If PWBF
succeeds in decoding, the hard decision decoded codeword is saved as Xi .
Irrespective of the correctness of the decoded description, the bits within each
description are deinterleaved by the outer deinterleaver 1
1 . As already mentioned,
correctness of mc bits in LDPC code, and can be utilized to improve the reliability
of the remaining m(p c) bits in a description. In the subsequent iterations, mc
reliable information bits of LDPC code are spread using 1 and is utilized to modify
the channel LLR input to the SPA decoder. The iteration loop is run until all the
codewords are successfully decoded or maximum number of iterations is reached.
The implementation of the proposed scheme proceeds as given in Algorithm 4.1.
In the algorithm, it is assumed that the output of the OFDM demodulator and
deinterleaver 1
2 is given as Lch , which is a matrix consisting of rows of LDPC
codeword LLR Lch (xi ), stacked one over another, and is of the form similar to the
Figure 4.1. In the algorithm, step 1 initializes the count of decoding iterations. Step
2 runs the complete iterative decoding algorithm for maximum of Niter iterations.
One decoding iteration of LDPC code is run in Step 2.1 using SPA or PWBF.
Decoding status of LDPC and RS code are stored in flag arrays FLDP C and FRS ,
respectively. The flag is changed from 0 to 1 if codeword is decoded successfully.
In step 2.1.1, SPA is applied on a row of LDPC codeword only if FLDP C [i] 6= 1;
i.e. i th row of LDPC code has not already been decoded correctly. Updated edge
LLRs L(s t) and a posteriori LLR Lp (xi (t)) are saved for subsequent decoding
iterations. If after current iteration of SPA, a valid codeword is detected, FLDP C [i] is
set to 1. Otherwise, at step 2.1.2, PWBF is applied on Lp (xi (t)). If this results is a
valid codeword, FLDP C [i] is set to 1. This process is repeated for all n descriptions.
In step 2.2, if for all i, FLDP C [i] = 1, further LDPC-RS decoding iterations are
stopped and program jumps to step 3.
From the saved hard decisions X of LDPC decoder, systematic part I is extracted
(step 2.3). The p columns of I consist of interleaved bits of p RS codewords, which
are decoded with hard decision BM algorithm in step 2.5. If decoding is successful,
columns in C = 1
1 (I) is replaced with decoded codeword, otherwise erroneous
codeword is left as it is. For the correctly decoded RS codeword j, FRS (j) is set
to 1. If all the RS codewords are decoded successfully, the algorithm comes out
of iteration loop (while loop) at step 2.6. At step 2.7, channel LLR matrix Lch is
modified according to LLR modification algorithm (Algorithm 4.2) which accelerates
decoding convergence. Iterations are run till any of the escape conditions are satisfied
75
1. Initialize: l = 1
2. While (l Niter ), do
A. Apply niter iterations of SPA to L(t) = Lch (xi ) and save the
calculated a posteriori LLR as Lp (xi (t)). Save the edge LLRs for
subsequent iterations.
2.2 If for all i [1, n], FLDP C [i] = 1, exit while loop.
I
2.3 Extract the matrix I from X = , where I is the systematic part of
P
X = [X1 , ..., Xn ]T .
2.4 Obtain C = 1
1 (I), where C = [C1 , ..., Cj , ..., Cp ] is the 2D matrix
2.6 If for all j [1, p], FRS [j] = 1, exit while loop.
2.7 Modify channel LLR Lch using soft value modification algorithm (Algo-
rithm 4.2).
2.8 l = l + 1.
i. For each bit b [(j 1)m + 1, jm] in the selected RS symbol, set
if Ii (b) = 0
max
L0I (i, b) =
max if Ii (b) = 1
0
LI
2.2 Modified channel LLR is Lch =
LP
uses 32 bit cyclic redundancy check (CRC) for detection of errors in descriptions.
Thus, in [80], number of information symbols transmitted per inner codeword (con-
volutional code) is 32 bytes, whereas, remaining four bytes are dedicated for CRC.
Because of this, maximum PSNR achievable by scheme in [80] will be less com-
pared to the proposed scheme. In this paper, simulation results were obtained for
transmission over AWGN, Rayleigh fading channels and frequency selective fading
channels.
In the Figure 4.3 to 4.6, we show results of the simulations carried out on SPIHT
coded, standard 8 bits-per-pixel (bpp) 128 128 Lena image transmitted with a
rate of 2 bpp. The algorithm is run for maximum of Niter = 5 number of iterations,
whereas each step of LDPC decoder consists of 25 SPA iterations. In the simulations,
peak signal-to-noise ratio (PSNR) is used as a measure of reconstruction quality and
is plotted against average received channel SNR. For each simulation point, 300 trials
have been performed in order to average statistical variations.
Figure 4.3 compares PSNR performance of the proposed scheme for transmission
over AWGN channel. With the proposed scheme, maximum PSNR of 34.4 dB is ob-
tained at channel SNR of -2 dB. In contrast to this, the tandem scheme achieves this
maximum at channel SNR of -1.25 dB. For AWGN channel, the reference scheme
achieves maximum PSNR of 32.84 dB at channel SNR of -0.6 dB. Thus, the pro-
posed scheme achieves maximum PSNR performance for 0.75 dB less received power
compared to tandem scheme and 1.4 dB less compared to the reference scheme.
Next, simulations performed over Rayleigh fading channels are presented. Figure
4.4 compares the PSNR performance of the proposed scheme with tandem and ref-
erence scheme with Rayleigh channel parameter fD Ts = 0.01, where fD is Doppler
frequency and Ts is sampling time. It is evident that the proposed system outper-
forms both the reference schemes significantly in terms of average PSNR perfor-
mance. LDPC-RS iterative scheme achieves maximum PSNR at channel SNR of 9
dB, which is 3 dB less compared to tandem scheme. In this case, reference scheme
reaches maximum PSNR of 32.84 dB at channel SNR of 10 dB. The gap with respect
to proposed LDPC-RS iterative scheme is now 1 dB. It can be observed that perfor-
mance of reference scheme outperforms tandem scheme for transmission over fading
79
35
30
20
15
LDPC-RS iterative
LDPC-RS tandem
Ref. scheme
10
-2.4 -2.2 -2 -1.8 -1.6 -1.4 -1.2 -1 -0.8 -0.6 -0.4
SNR (dB)
Figure 4.3: Comparison of PSNR performance as function of channel SNR for 128
128 Lena image transmitted over AWGN channel.
35
30
20
15
LDPC-RS iterative
LDPC-RS tandem
Ref. scheme
10
5 6 7 8 9 10 11 12
SNR (dB)
Figure 4.4: Performance comparison of the proposed scheme with two reference
schemes for Lena 128 128 image transmitted over Rayleigh fading channel with
fD Ts = 0.01.
35
30
Average PSNR (dB)
25
20
LDPC-RS iterative
15
LDPC-RS tandem
Ref. scheme
5 6 7 8 9 10 11 12 13
SNR (dB)
Figure 4.5: Performance comparison of the proposed scheme with two reference
schemes for Lena 128 128 image transmitted over Rayleigh fading channel with
fD Ts = 0.05.
81
35
30
20
Iterative in 'COST207RAx6'
Tandem in 'COST207RAx6'
15 Iterative in 'COST207TUx6'
Tandem in 'COST207TUx6'
3 4 5 6 7 8 9 10
SNR (dB)
Figure 4.6: Performance comparison of the proposed scheme with LDPC-RS tandem
scheme for Lena 128128 image transmitted over two COST 207 frequency selective
channels.
4.3.1 Preliminaries
Figure 4.7 illustrates the actual realization of n-channel FEC based MDC considered
in this section. Similar to earlier section, RS codes are applied along the columns.
Descriptions are formed along the rows and appended with CRC. Each description
is then channel coded with recursive systematic convolutional code (RSCC) before
transmission.
For calculating first symbol C1 (at time i = 1), the reference symbol taken
is C0 = D. In (4.3.12), G(f (di )) is one of the symbols in the group G and f (di )
is a rule or function that maps bits pairs di {00, 01, 10, 11} to one of the four
symbols in the set G. Group property ensures that Ci is a codeword if Ci1 C is
a codeword. Xi is used to indicate the transmitted DSTBC symbol. The S rows
83
The basic transmitter is shown in the Figure 4.8. Multimedia information is encoded
by a progressive source encoder (e.g. SPHIT or JPEG 2000) to produce a bitstream
u. The contiguous source bits are grouped into RS code symbols of m bits, i.e.
GF (2m ). As explained in 4.3.1, these symbols are encoded with (n, k) RS code with
level of protection depending on relative importance of information symbols. If the
RS codewords are arranged as columns in the Figure 4.7, descriptions or packets are
obtained by extracting symbols along the rows. Each descriptions is encoded, row-
by-row, using recursive systematic convolutional code (RSCC) to produce bitstream
r. Bits in a RSCC coded description are interleaved and fed into DSTBC encoder
block for transmission through wireless channel.
Transmission takes place over a channel with T transmit and U receive antennas.
At time slot i, the transmitting antennas transmits S rows of STBC symbol Xi over
the channel. The demodulated baseband signal at a receive antenna at time i is
r
yi = Xi H i + N i (4.3.14)
T
where Hi = [htu (i)] is the T U channel gain matrix, Ni is AWGN matrix with
variance of each element being 2 = N0 /2, and Xi is transmitted code from (4.3.12).
At the receiver, three stage serially concatenated iterative decoding scheme is
applied where information is exchanged to and fro between DSTBC demodulator
(inner decoder), RSCC decoder (middle decoder) and RS decoder (outer decoder).
84
Packet RSCC
Source u RS Code c r d DSTBC
Encoder (column by formation Encoder P
column) (row by row) (row by row)
Encoder
X
Figure 4.8: Encoder block diagram consisting of cascade of FEC based multiple
description coder and DSTBC.
p
Le(d) La(r) L (r) BM RS Estimated
DSTBC RSCC APP ^
r ^
-1 Hard Decoder c bits
APP P Decoder
y (row by row) Decision (Column by
Decoder column)
^ a
L (d ) Le(r) La(r)
H
Channel
estimate Soft value
P modification
Figure 4.9: Iterative decoding of FEC based multiple descriptions, consists of three
concatenated stages.
The serially concatenated decoder is shown in the Figure 4.9. Inner and middle
decoder use the a posteriori probability (APP) decoding algorithm, whereas outer
decoder uses the hard decision BM algorithm. The three stages form two iterative
decoding loops: inner decoding loop and outer decoding loop, as explained in next
section.
In the iterative decoder of the Figure 4.9, the inner decoding loop consists of DSTBC
APP decoder and RSCC APP decoder. DSTBC decoder receives signal yi at time
i and outputs extrinsic information Le (di ), which is then deinterleaved and applied
to RSCC APP decoder as a priori input La (ri ). Extrinsic output Le (ri ) is applied
back to DSTBC APP decoder, while a posteriori output Lp (ri ) acts as input for
outer loop. Both the decoders in inner loop use BCJR algorithm for SISO decoding.
Outer decoding loop consists of performing hard decision on output Lp (ri ) of RSCC
decoder and hard decision RS decoding by BM algorithm. RS code being block code,
it is not difficult to identify those RS codewords (columns in the Figure 4.7) which
85
were decoded successfully. Flags FRS (j), j = 1, ..., p indicating successfully decoded
RS codewords is passed to the soft value modification block, which modifies the a
priori input to RSCC APP decoder. Both the decoding loops are run for predefined
number of iterations. Number of inner and outer iterations are indicated by LI and
LO , respectively. Next, each of these blocks are explained in more detail.
To keep the calculations simple, a system with two transmit and one receive antenna
as described in (4.3.1) is being considered. Generalization to multiple transmit and
receive antennas is straight forward. At time i, if Xi = Ci were transmitted, (4.3.14)
can be written as r
yi = C i Hi + N i (4.3.15)
T
where yi = [y1i , y2i ]T is a vector of two discrete received signals. Received signal
probability conditioned on transmitted code and channel gain is then
1
p(yi |Ci , Hi ) exp( kyi Ci Hi k2 ) (4.3.16)
2 2
The APP decoder utilizes the conditional probability in (4.3.16) for optimal decod-
ing. The decoder calculates APP or log-likelihood ratio (LLR) of each bit
P (dji = 1|Y)
Lp (dji ) = log (4.3.17)
P (dji = 0|Y)
where dji is the j-th bit of the i-th transmitted DSTBC symbol and Y is the entire
received sequence Y = [y1 , ..., yi , ...]. In this specific example, j = 1, 2, indicating
two bits per symbol in BPSK group.
As mentioned earlier, DSTBC encoding can be represented with a trellis and
can be decoded with usual trellis based decoding methods like Viterbi and BCJR
decoder. Trellis diagram for code set defined by (4.3.10) and (4.3.12) is shown in
the Figure 4.10(a), where state of the encoder Si = Ci . There are M transitions
per state, where M = |G| is the cardinality of set G. Because of optimal and SISO
nature of BCJR algorithm, it is used here for APP decoding of DSTBC. The set of
all state transitions (m0 , m) corresponding to input bit dji = b be is defined by the
trellis (Figure 4.10(a)), where
Figure 4.10: (a) Basic trellis structure of BPSK group DSTBC, (b) Section of ex-
tended trellis of BPSK group DSTBC for P = 2 having 64 states.
These are the set of transition at time i corresponding to input Gi , where Gi = f (di )
has been obtained by simple mapping from a pair of bits. Probability of transition
from state m0 to m at time i is
trellis and i (m) = m0 i+1 (m0 )i+1 (m, m0 ) is obtained by backward recursion. Us-
P
where, La (dji ) is a priori LLR obtained from the middle decoder. APP LLR for
each bit is then calculated as
i (m0 , m)
P
p (m0 ,m):dji =1
L (dji ) = P (4.3.22)
(m0 ,m):dji =0 i (m0 , m)
87
In (4.3.21), fading channel gain coefficients or CSI are required for calculation of
transition metric between states. Similar to DPSK, channel state information is not
essential for decoding DSTBC successfully. If current channel gain Hi is not known,
they can be estimated using linear prediction [204] using previous channel estimates
Hi .
P P
X 1X
Hi = wn Hin = wn CH
in yin (4.3.24)
n=1
2 n=1
where, P is the prediction order and w = [w1 , ..., wP ]T are predictor coefficients
obtained by solving Wiener-Hopf equation Rw = r. Specifically for the case of
T = S = 2 and channel gain being constant for a minimum duration of a STBC
symbol (or two bit periods), matrix R is given as
r0 r2 . . . r2(P 1)
r2 r0 . . . r2(P 1)+2
..
.. .. .. ..
(4.3.25)
. . . . .
r2(P 1) r2(P 1)+2 . . . r0
and r = [r2 , r4 , ..., r2P ]T . The coefficients rn are autocorrelation coefficients of the
channel process. For Rayleigh flat fading channels, the autocorrelation is rn =
J0 (2fD Ts n) with r0 = 1 + 2 2 , where fD is maximum Doppler frequency and Ts is
sampling period. J0 is first order Bessel function.
Using the channel estimate in (4.3.24), Nguyen and Ingram [181] calculated the
transition metric i (m0 , m) between states as
P
X
a 1 1X
exp{ dji L (dji ) 2 kyi w n CH 2
in yin k } (4.3.26)
j
2pre 2 n=1
2
PP
The variance term in (4.3.26) is 2pre = r0 {1 n=1 wn (r2n /r0 ).The symbols
Ci1 , ..., CiP are replaced according to per-survivor principle [205]. For this pur-
pose, MAP BCJR algorithm is now carried over expanded trellis with M P states and
88
The inner decoder passes extrinsic information Le (dji ) to next stage, which on dein-
terleaving, acts as a priori information La (ri ) for the RSCC APP decoder. The
APP decoder for RSCC is conventional SISO BCJR decoder [12]. It calculates the
a posteriori LLR of output as Lp (ri ) and binary hard decision is performed on it
where sgn is the signum function. The hard decision ri passed on to outer RS
decoder, irrespective of the correctness of the descriptions. Correctness of the de-
scription is checked using CRC and if description i satisfies CRC, flag FRSCC (i) is
set to 1. CRC block is not shown in the Figure 4.8 and the Figure 4.9 to avoid clut-
tering the the figures. The extrinsic information of the coded bits, which is returned
back to inner decoder, is
Le (ri ) = Lp (ri ) La (ri ) (4.3.28)
Outer RS Decoder
As stated before, for RS codeword l having nkl parity symbols, the RS decoder (i.e.
BM algorithm) can correct up to n kl erasures. Since, k1 k2 ... kp due to
unequal error protection, it is very likely that the correctly decoded RS codewords
(columns in the Figure 4.7) would be placed contiguously at the left end of the
product code. Similar to scheme in Section 4.2, correctly decoded RS codewords are
indicated with value 1 in flag array FRS .
If c is the number of correctly decoded RS codewords, we know that there are
mc correct bits in each description, where m is the number of bits per RS code
89
i. For each bit b [(j 1)m + 1, jm] in the selected RS symbol, set
if Ii (b) = 0
a max
L (r)(i, b) =
max if Ii (b) = 1
35
30
20
(L I , LO) = 4
15 (L I , LO) = 3
(L I , LO) = 2
(L I , LO) = 1
10
0 0.5 1 1.5 2 2.5 3 3.5 4
SNR (dB)
Figure 4.11: PSNR as function of channel SNR for different number of global itera-
tions in Rayleigh fading channel.
encoder uses (n, ki ) RS code, where n = 128 and the description size is 32 RS code
symbols (i.e. p = 32). All the descriptions are CRC coded. RSCC of rate 1/2 and
generator polynomial (15, 17)8 is applied for channel coding. While decoding, it is
assumed that any error in a description is detected by CRC and the description will
be treated as erasure.
The goal of our simulation is to see how the proposed receiver with RS-RSCC-
DSTBC iterative decoding compares with reference MDC iterative decoder pro-
posed in literature. For comparison, baseline scheme proposed in [80] consisting of
RS-RSCC iterative decoder is being considered. For fair comparison, RSCC coded
descriptions in baseline scheme are encoded with Alamouti STBC [206]. Perfor-
mance is evaluated in terms of peak-signal-to-noise ratio (PSNR) as function of
average channel SNR. In all results, we shall compare the channel SNR required to
achieve maximum PSNR, unless mentioned otherwise. There are two set of computer
simulation results.
In the first set of simulations, the effect of iteration on the improvement in
PSNR curve is studied when CSI available at the receiver. For this purpose the
91
35
30
L I = 4, L O = 4
Figure 4.12: PSNR as function of channel SNR for different values of inner and
outer loop iterations in Rayleigh fading channel.
proposed scheme is tested over Rayleigh fast fading channel, i.e. fading coefficients
are assumed to be constant for at least the duration of STBC symbol. Figure 4.11
shows improvement in PSNR curve with number of iterations. The iterations are
global in the sense that one iteration of inner loop is followed by one iteration of
outer loop and so on. For clarity, one such global iteration is indicated in the figure
as (LI , LO ) = 1. With two iterations(indicated as (LI , LO ) = 2 in the figure), the
transmission power required to achieve maximum value of PSNR is about 1.5 dB
less compared to PSNR for one iteration ((LI , LO ) = 1). Maximum value of PSNR
is obtained for four iterations ((LI , LO ) = 4) and channel SNR of 1.4 dB. It can
also be observed that rate of improvement of PSNR as function of channel SNR
(as indicated by the slope of curve) improves with number of iterations. This may
be attributed to the acceleration of convergence provided by turbo-like iteration
decoding architecture.
Next set of simulations were performed for local iterations: LI number of it-
erations of inner loop is followed by LO number of iteration of outer loop. This
allows us to isolate the decoding gain provided by the individual decoding loops.
92
35
30
20
L I = 4, L O = 4, w/ CSI
15 L I = 4, L O = 1, w/ CSI
w/o CSI, P = 2
w/o CSI, P = 3
Baseline w/ CSI
10
0 0.5 1 1.5 2 2.5 3
SNR (dB)
4.4 Conclusion
In section 4.2, a novel method has been proposed for iterative decoding of LDPC-
RS based MDC transmitted over fading channel. The proposed scheme employs
RS codes to form multiple descriptions and each description is channel coded with
LDPC code. The descriptions are OFDM modulated and transmitted over wireless
channel. At the receiver, soft decision output of OFDM demodulator is applied
to LDPC decoder. An iteration of proposed decoder consists of one step of sum-
product algorithm (for decoding LDPC code) and RS decoder. Successfully decoded
bits in RS codeword are applied back to improve the decoding performance of inner
LDPC code in subsequent iterations. Decoded LDPC codewords in turn improve the
decoding of RS code. Experimental results showed the superiority of the proposed
scheme in comparison to similar decoding schemes in literature.
94
34
32
30
26
24
22
20
18 (L I ,L O) = 4, w/ CSI
w/o CSI, P = 2
16 w/o CSI, P = 3
Baseline w/ CSI
14
0.5 1 1.5 2 2.5 3
SNR (dB)
Figure 4.14: PSNR performance of proposed and baseline scheme in Jakes channel
with fD Ts = 0.005.
******
95
Table 4.1: Performance advantage of proposed scheme in Jakes fading channel with
fD Ts = 0.01 with CSI
PSNR (dB)
Image SNR 0.8 dB 1.0 dB 1.2 dB 1.4 dB
96
97
5.1 Introduction
As mentioned before, MDC generates multiple descriptions of the source in such a
way that each description independently describes the source with certain fidelity.
The goal is to be able to decode an error-free output when all the descriptions are
available and decode a slightly distorted output when only some of the descriptions
are available. FEC based MDC is one of such technique discussed in earlier chapters.
In this chapter, we explore MDSQ based two-channel MDC which generates two
equally independent descriptions. Vaishampayan [31] proposed the first constructive
approach towards a multiple description scheme by using MDSQ and considered a
special case, namely the balanced descriptions. Creating multiple equally important
descriptions may be desirable in MDC systems when there is no need to differentiate
the importance of one description from another.
The MDSQ of Vaishampayan works as follows. The signal is sampled using a
regular scalar quantizer and a sample X is assigned a quantization index I. Then,
multiple description index assignment is done, mapping each index I to a codeword
index pair (I1 , I2 ) from codebook shown in the Figure 5.1. Figure 5.1(a) shows the
staggered index assignment matrix, where cells of the quantizer correspond to the
quantization index I numbered from 0 to 8. The row and column indices (I1 and
I2 )of the matrix form the indices of the two descriptions 1 and 2, respectively. If
both the side index values (I1 , I2 ) are available at the receiver, the central decoder
can recover the original index I back. If only one index (I1 or I2 ) is available,
side decoder estimates sample X as the expected value by keeping available index
fixed. For example, in the Figure 5.1(a), if only I1 = 2 is known, I2 is obtained by
calculating expectation of I2 = 4 and I2 = 5. Also, the quality of reconstruction at
the side decoders depend on the number of diagonals in the index assignment matrix.
The number of unoccupied cells in the matrix constitute redundancy, indicating the
need of further compression by entropy coder. Figure 5.1(b) shows index assignment
matrix with three diagonals and having lower redundancy.
In this work, MDSQ is applied to wavelet coefficients. A given image is de-
composed into subbands, and then a uniform scalar quantizer is applied to each of
the subband coefficients, producing a quantized field. MDSQ index assignment on
98
I1 I1
0 1 2 3 4 0 1 2 3 4
0 0 0 0 2
1 1 2 1 1 3 4
I2 2 3 4 I2 2 5 6 8
3 5 6 3 7 9 10
4 7 8 4 11 12
Figure 5.1: Multiple description index assignment: (a) 2 diagonal staggered quanti-
zation cells; (b) 3 diagonal higher spread quantization cells.
u
(1)
Arithmetic c
(1)
u
(2)
RSCC x
encoder P encoder
e (1)
L (c )
a (1)
L (u )
e (2) + n
L (u ) SISO arithmetic P-1 RSCC MAP
decoder decoder
e (1) P a (2) y
L (c ) L (u )
A low complexity SISO arithmetic decoder based on Chase-like decoding [111] was
proposed in [115]. The system consists of a finite alphabet source, an AWGN channel
and an arithmetic decoder as shown in the Figure 5.2. The source generates packets
of L symbols u(1) = [u11 , ..., u1L ]. Each packet is then compressed using arithmetic
encoder [207] and the resulting binary stream c(1) = [c11 , ..., c1M ] is of variable length
M . c(1) is interleaved to produce u(2) and acts as input to RSCC encoder. BPSK
modulated output of RSCC encoder x is transmitted over an zero-mean AWGN
channel of noise variance 2 .
Output of the channel at the receiver can be given as y = x + n. At the de-
coder, the received signal is iteratively decoded by exchanging extrinsic information
between SISO arithmetic decoder and the MAP BCJR decoder. As shown in the
Figure 5.2, extrinsic output Le (u(2) ) of BCJR decoder is deinterleaved to produce
a priori information La (c(1) ) for the arithmetic decoder. Extrinsic output Le (c(1) )
of the arithmetic decoder is fed back as a priori information La (u(2) ) to the BCJR
decoder.
At the receiver, the arithmetic decoder is based on the MAP principle. A sum-
mary of the low complexity Chase-type SISO decoding algorithm [115] for AC is
given next.
1. Apply hard decision over a priori input La (c(1) ) = [1 , 2 , ..., M ] to get c(1) .
Bit reliabilities [|1 |, |2 |, ..., |M |] are calculated as magnitude of the LLR
values.
100
(i) (i)
3. Generate 2q test patterns t(i) = [t1 , ..., tM ], 0 < i 2q , each of length M . Each
test pattern has maximum weight q with all the possible bit combinations in
the q least reliable positions.
4. Form test vectors z(i) = c(1) t(i) , 0 < i 2q , where, is bit-wise XOR
operation. Each test vector is decoded using classical arithmetic decoding.
5. If the decoded sequence (for i-th test vector) has exactly L decoded symbols
and correct end-of-block (EoB) symbol is decoded (length-valid sequence),
calculate its distance metric as:
M 2
(i) 1 X j (i)
D = 2 hj + log P (u(i) ) (5.2.1)
2 j=1 Lc
(i) (i)
where, hj = 1 2zj , and Lc is channel reliability value. P (u(i) ) is the a
priori source probability of decoded sequence corresponding to z(i) .
6. The decoded bit stream, denoted as u(1) , is the sequence with highest metric.
7. To calculate the extrinsic output value, we find the set of bits positions J =
[j1 , j2 , ...] for which bit values are same for all length-valid candidate sequences.
It is clear that the positions j J where we have same bit for all candidate
sequences are the most reliable bits. The extrinsic value assigned to those bits
(1) (1)
are Le (cj ) = (1 2zj ) for all j J and Le (cj ) = 0 for all j
/ J. The
constant is determined empirically to obtain best error performance.
u
(1)
Arithmetic
(1)
c u
(2)
RSCC c
(2)
u
(3)
Rate 1 x
encoder P encoder
P ACC
e (1)
(1)
L (c )
a (2)
L (u )
e
L (c )
a
L (u )
(2) e (3)
+ n
L (u ) SISO arithmetic -1 P-1
P RSCC MAP ACC MAP
decoder decoder decoder
e (1) P a (2) e (2) P a (3) y
L (c ) L (u ) L (c ) L (u )
have used this property of accumulator to obtain iteration gain in cascade of stages
comprising of SISO arithmetic decoder. Proposed scheme is described next.
The block diagram of the proposed scheme using rate-1 accumulator (ACC) is
shown in the Figure 5.3. The first two stages are as in the section 5.2.1, with the
difference that binary output of RSCC is scrambled again and applied to rate-1
accumulator. The rate-1 accumulator is simply a RSCC with generator polynomial
(3, 2)8 and with no systematic bits transmitted.
At the receiver noise corrupted version of the code is decoded by soft-out BCJR
decoder (for accumulator) to produce extrinsic information Le (u(3) ). It is then de-
interleaved and applied to RSCC MAP decoder. Extrinsic output of RSCC decoder
Le (u(2) ) is unscrambled and applied as input to SISO arithmetic decoder described
earlier. Extrinsic LLR Le (c(1) ) at the output of arithmetic decoder is applied as
a priori information La (u(2) ) to RSCC MAP decoder. Extrinsic output of RSCC
MAP decoder is Le (c(2) ) and interleaved input to ACC decoder is La (u(3) ). This
process is performed iteratively until convergence is reached.
Simulations were conducted to evaluate the error rate performance of the proposed
technique. In this section, numerical results on packet error rate (PER) are discussed
and compared to the results obtained in [115]. We consider a memoryless binary
source with P (0) = 0.2 and P (1) = 0.8. Each source packet length is of 512
bits, with each packet ending with an EoB symbol. The packets are encoded with
102
100
5 iter, ACC-CC-AC
10 iter, ACC-CC-AC
20 iter, ACC-CC-AC
10 iter, CC-AC
10-1
PER
10-2
10-3
10-4
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Eb /N0 (dB)
nentially with memory of the code and hence at the decoder side our scheme is less
complex computationally because of using two memory 1 decoders rather than one
memory 3 decoder.
Next the performance of the ACC-CC-AC scheme was evaluated for an SPIHT
coded image transmission system and compared with baseline system. In the consid-
ered system, the Lena test image is compressed using the SPIHT encoder, without
AC stage, generating output of rate RS bpp. SPIHT coded stream was formed into
packets of 512 bits. The source packets are further encoded with proposed scheme
as shown in the Figure 5.3. At the receiver, iterative decoding as described earlier
was applied with Chase decoding parameter q fixed to 4 bits. Simulations were done
to obtain average PSNR over 500 image transmissions for both the cases.
Simulation results are given in the Figure 5.5 and the Figure 5.6 for source bit
rates 0.4 bpp and 1 bpp, respectively. In the figures, PSNR performance of the
ACC-CC-AC based image transmission system is compared to the results obtained
by the CC-AC based baseline system. It can be seen in the Figure 5.5 that at
Eb /N0 of 3.25 dB the proposed ACC-CC-AC system gives a PSNR gain of 10 dB
for RS = 0.4 bpp. On the other hand, for RS = 1.0 bpp, at Eb /N0 of 3.5 dB there is
a gain of 12 dB. All these results were obtained for 5 number of decoding iteration
for both the schemes.
36
34
32
Average PSNR (dB)
30
28
26
24
22 ACC-CC-AC
CC-AC
20
1.5 2 2.5 3 3.5 4 4.5 5 5.5
Eb /N0 (dB)
Figure 5.5: PSNR versus Eb /N0 plot for proposed scheme applied to SPIHT image
coding at RS = 0.4 bpp.
40
38
36
Average PSNR (dB)
34
32
30
28
26
24
22 ACC-CC-AC
CC-AC
20
1.5 2 2.5 3 3.5 4 4.5 5 5.5
Eb /N0 (dB)
Figure 5.6: PSNR versus Eb /N0 plot for proposed scheme applied to SPIHT image
coding at RS = 1 bpp.
105
Noisy Channel
S1 Iterative Source-
SPIHT
Decoder channel decoder
Error Inverse MDSQ Y1
IDWT Concealment Quantizer Inverse index
Assignment
Reconstructed SPIHT S2 Iterative Source- Y2
image Decoder channel decoder
Figure 5.7: Proposed two channel MDC system with multiple wavelet tree coding.
The proposed image transmission system is shown in the Figure 5.7. The image
is hierarchically decomposed using DWT. The wavelet coefficients are quantized
using uniform scalar quantizer and two descriptions are generated by MDSQ index
assignment. SPIHT coding of wavelet trees is applied to each of the descriptions.
Coding of wavelet trees [209] is different from usual SPIHT in the sense that each
spatial orientation tree in the wavelet domain is individually encoded using SPIHT.
For example, if there are T wavelet trees in each description, there will be T SPIHT
coded bit-streams and they can be independently decoded. All T SPIHT coded
streams are appended with 16-bit CRC to detect any error. This allows SPIHT
decoding of received bit-streams, only if, they are received error-free. Wavelet tree
coding will be explained in next section. Each SPIHT coded spatial orientation
tree is further compressed with AC. AC coded bit-streams are channel coded using
RSCC and accumulator (ACC-CC-AC scheme in Section 5.2.2). Coded bits are
transmitted over AWGN channels after BPSK modulation.
The receiver iteratively decodes T ACC-CC-AC coded bit-streams from each
description. It is assumed that there is no loss of packet in the channel and syn-
chronization is perfect at the receiver. Decoding of ACC-CC-AC coded bit-streams
was described earlier in Section 5.2.2. Next the decoded bit-streams are checked for
106
errors and only streams which satisfy CRC are allowed for SPIHT decoding stage.
Error detection is essential at this stage because of progressive nature of SPIHT
coded bit-stream. Any bit error encountered at the beginning of the bit-stream will
significantly distort the image. After SPIHT decoding of error-free bits-treams, we
may encounter three scenarios at this stage:
1. Quantization index for wavelet tree order of both the descriptions are recov-
ered.
2. Quantization index for wavelet tree order of only one description is recovered.
3. Quantization index for wavelet tree order of none of the descriptions are re-
covered.
For Case 1 and 2, inverse MDSQ is applied to available quantized index from one or
both descriptions to obtain a single field of quantization index. Estimate of actual
value of wavelet coefficients are obtained by application of inverse quantization on
recovered quantization index. For case 3, when both the descriptions are erroneous,
a complete wavelet tree is lost and error concealment [64] needs to be carried out.
It is well known that in the lowest frequency subbands, wavelet coefficients are
correlated to the neighboring coefficients. On the other hand, this is not true for
higher frequency subbands. Therefore, error concealment for wavelet coefficients
is carried out in lowest frequency subband only. A single wavelet tree in SPIHT
has four coefficients in the lowest frequency subband. For a missing block, each
of the four missing coefficients will be replaced by the average of its neighboring
coefficients. For higher frequency subbands, coefficients are simply set equal to
zero for lost data blocks. Finally, inverse discrete wavelet transform is applied to
reconstruct the received image.
After the generation of two balanced descriptions by MDSQ, multiple wavelet tree
coding is applied to produce multiple bit-streams to enhance error resilience capa-
bility of the descriptions. Figure 5.8(a) shows the spatial orientation tree structure
107
(a) (b)
Figure 5.8: Wavelet tree in a 3 level DWT and its corresponding image content.
of Said and Pearlman [71]. This structure defines the spatial relationship of wavelet
coefficients in a hierarchical pyramid. The wavelet tree is rooted at the lowest fre-
quency subband. Each node in the tree has either four offspring or no descendant
at all. The offspring are grouped in a 2 2 adjacent coefficients. Its descendants
correspond to the pixels of same spatial orientation in next finer level of pyramid. If
all the coefficients in a tree are grouped together, they form a square block of image
as shown in the Figure 5.8(b). They are frequency components for a specific image
area with the same block size at the corresponding position.
For each wavelet tree, SPIHT algorithm was applied to encode them indepen-
dently. Independent coding of wavelet trees has the advantage that any error in one
SPIHT coded stream does not impact other SPIHT coded bit-streams. Thus, error
resilient capability of a description is increased with increase in number of wavelet
trees, with little decrease in coding efficiency.
Experiments were conducted to evaluate the performance of the proposed scheme for
transmission of gray-scale image. It was demonstrated that PSNR performance of
MDC can be improved by utilizing soft output of AWGN channel and by employing
108
SISO arithmetic decoder. Simulations were performed for both high redundancy, 2
diagonal (D = 2) index assignment and lower redundancy, 3 diagonal (D = 3) index
assignments. We assume that both the descriptions are transmitted through AWGN
channel are decoded at a central decoder.
In the considered system, test image was decomposed by 9-7 wavelet filters and
number of decomposition levels was set to four. Quantization was performed with
stepsize uniformly throughout resolution levels. Two image descriptions were
obtained by applying MDSQ index assignment on the quantization indices. SPIHT
coding was applied to each of the descriptions until bit-layer 0 to get integer level
accuracy. was selected such that bit rate per description is 0.5 bpp. With four
levels of decomposition and each wavelet tree being rooted in a 2 2 pixel block
in lowest subband, number of wavelet trees in a description is 256. Each SPIHT
coded wavelet tree was appended with 16-bit CRC and an EoB symbol. The size
of a description after SPIHT coding is given in Table 5.1. Since number of wavelet
trees are fixed and 16-bit CRC is appended to each stream, CRC bytes increase the
size of a description by 3.1 % only. Each stream was compressed using AC and four
such streams are formed into a transmission frame. Size of description after AC
and percentage reduction is also shown in Table 5.1 for different test images. Thus,
in the proposed scheme, data bytes increase due to CRC had been compensated
by reduction due to AC. Finally, data frames were encoded with RSCC and rate-1
accumulator, as described in Section 5.2.2.
Table 5.1: Description size after addition of CRC bytes and reduction due to AC.
D=2 D=3
decoding between AC and RSCC (CC-AC scheme in Section 5.2.1) for both the
received descriptions. This scheme was designated as CC-AC in the figures next.
To evaluate the contribution of AC in iterative decoding, another baseline scheme
was considered (labeled as Baseline in the figures) which did not use AC and hence
no iterative decoding was performed between AC and RSCC. For fair comparison,
description size after SPIHT coding was kept same for all the schemes.
Since, compression performance of the three test images do not differ signifi-
cantly, we considered Lena image for subsequent simulations. PSNR performance of
the three schemes for 2 diagonal and 3 diagonal MDSQ were obtained and is shown in
the Figure 5.9 and the Figure 5.10, respectively. For 2 diagonal case, ACC-CC-AC,
CC-AC and baseline schemes achieve maximum PSNR of 35.6 dB at channel Eb /N0
of 2.75 dB, 3.75 dB and 5 dB, respectively. The proposed ACC-CC-AC scheme
provides a gain of 1 dB over CC-AC scheme and 2.25 dB over baseline scheme in
terms of Eb /N0 . For D = 3, the maximum average PSNR is obtained at channel
SNR of 2.75 dB, 4 dB, and 5.25 dB respectively. Here, the gain of ACC-CC-AC
scheme over CC-AC scheme is 1.25 dB and over baseline scheme is 2.5 dB. Thus,
ACC-CC-AC outperforms both the reference schemes and requires significantly less
transmit power.
The baseline scheme achieves maximum PSNR at Eb /N0 = 5 dB for D=2 and
at Eb /N0 = 5.25 dB for D=3. Three diagonal MDSQ requires 0.25 dB more power
compared to two diagonal MDSQ and can be attributed to lower redundancy of
three diagonal MDSQ. This is not the case with ACC-CC-AC scheme which achieves
maximum PSNR at Eb /N0 = 2.75 dB for both D=2 and D=3.
40
35
Average PSNR (dB)
30
25
20
ACC-CC-AC
15 CC-AC
Baseline
Figure 5.9: PSNR performance of the proposed scheme (D = 2) with SPIHT coding
of multiple wavelet trees.
40
35
Mean PSNR (dB)
30
25
20
15 ACC-CC-AC
CC-AC
Baseline
10
1.5 2 2.5 3 3.5 4 4.5 5 5.5
Eb /N0 (dB)
Figure 5.10: PSNR performance of the proposed scheme (D = 3) with SPIHT coding
of multiple wavelet trees .
111
resilience capability for each description. The spatial orientation trees in the wavelet
domain are individually encoded using SPIHT. Since, all these SPIHT coded bit-
streams are of different lengths, to maintain robust synchronization, authors in [64]
adopted error resilient entropy coding (EREC) [95] to reorganize these variable
length bit-streams into fixed length data slot before multiplexing and transmission.
Error concealment is applied at the receiver to alleviate image degradation due to
loss of both descriptions of a wavelet tree.
In this research, our contribution lies in improving the scheme of Song et al. by
adopting Chase-type SISO arithmetic decoder for decoding of multiple descriptions
transmitted over additive white Gaussian noise (AWGN) and packet loss channels.
Existing schemes in literature suppose the transmission media to be random bit
error and packet loss channel. This assumption is not necessarily suitable for many
binary-input continuous output channels (e.g. AWGN channel), where soft deci-
sion output of channel decoder can be utilized by the source decoder for further
improving the PSNR performance. In the proposed scheme, SPIHT coded bit-
streams are arithmetic coded to further increase compression efficiency and use of
soft channel information at the receiver improves the noise robustness of the arith-
metic decoder. Combining EREC with iterative source-channel decoding of AC is
not quite straightforward as EREC requires instantaneous hard decision decoding
of the source symbols. Our innovation lies in applying EREC assisted by state and
tail bits (EREC-ST) [191] with arithmetic coded SPIHT trees of variable lengths
for SISO decoding of AC. This enables iterative source-channel decoding of EREC
coded fixed length slots at the receiver and improve PSNR gain significantly.
The proposed system is shown in the Figure 5.11. This scheme is in continuation
of the MDC scheme presented in Section 5.3, except EREC encoding and decoding
stage. In this design, EREC encoder and decoder blocks are replaced by EREC-ST
blocks to create equal length packet from variable length SPIHT coded bitstreams.
Two balanced descriptions are obtained by performing MDSQ after decompos-
ing the given image hierarchically using discrete wavelet transform. Coefficients of
112
S1 A1
I1 SPIHT Verify Iterative Source-
Decoder CRC channel decoder
Conceal Inverse Y1
Inverse
IDWT Errors Quantizer MDSQ S2 A2
Reconstructed SPIHT Verify Iterative Source-
image I2
Decoder CRC channel decoder
Y2
Figure 5.11: Block diagram of proposed two-channel MDC scheme utilizing iterative
source-channel decoding and EREC-ST.
wavelet trees along spatial orientation in each descriptions are partitioned and each
partition is SPIHT coded independently. CRC parity bits are generated for prede-
fined fixed length of each SPIHT coded bitstream, and then AC is applied to each
independent bitstream for further compression. The variable length AC are reor-
ganized into fixed length slots using EREC-ST for synchronization requirements.
EREC-ST is a recently developed variation of EREC in which start (S) and tail (T)
information of variable length bitstream are encoded separately. Conversion to fixed
length slots is required to form uniform sized packets before transmitting over packet
switching network. At the output of the packet loss channel, remaining available
packets are coded using RSCC at wireless access points (WAP) and transmitted.
The receiver decodes the packets received through the wireless channel and re-
constructs the image by combining both the descriptions. Log-likelihood ratio (LLR)
value at the input is computed from the received signal as i = 2yi / 2 , where yi is
the noisy symbol output of AWGN channel with noise variance 2 . This LLR value
is utilized by source-channel decoder for iterative soft decoding of AC and RSCC.
Successfully recovered SPIHT coded bitstream are decoded and wavelet trees are
combined to obtain the image in wavelet domain. Finally reconstructed image is
obtained by inverse discrete wavelet transform. Function of individual blocks are
113
described next.
Image is first hierarchically decomposed into N level DWT. The number of levels
N is chosen depending on the required number of wavelet trees with the coefficients
rooted in the lowest frequency subband. MDSQ on DWT coefficients produce two
complementary and possibly redundant index pair fields (2-D matrix for each de-
scription) I1 and I2 . Similar to earlier section, we consider two and three diagonal
index assignment method for MDSQ.
The quantized DWT coefficient matrix Ij , j 1, 2 is divided into K groups and then
each of these groups are independently SPIHT coded so that K separate embedded
bitstreams for each description are created. Each group consists of four spatial
orientation trees with each wavelet tree rooted at the lowest frequency subband.
Nodes of the tree have either no descendents or four offspring grouped in 2 2
adjacent coefficients. SPIHT algorithm is employed to encode independently i-
th group of coefficients in description j and generate variable length bitstreams
Sij , i = 1, ..., K, j 1, 2. Since each description is consists of matrix of integer
valued quantization indices, SPIHT coding is done on each group of trees until
bit-layer 0 to obtain integer level accuracy.
CRC and AC
After SPIHT coding of wavelet trees along spatial orientation for each description,
4-bit CRC is attached to each of the K SPIHT coded bitstreams. Following [64],
CRC of 4 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), CRC is generated
and appended for the first l bits (l = 128 bits) in each bitstream. CRC check
eliminates erroneous wavelet trees from consideration which may have significant
impact on the PSNR of the reconstructed image. Arithmetic encoder then maps
114
each of the K bitstreams independently into binary strings Aji , i = 1, ..., K for
description j 1, 2.
If variable length blocks (VLB) are transmitted consecutively, the resulting system
can be severely effected by bit errors because of loss of synchronization. One ap-
proach is to append synchronization code to provide resynchronization at the cost
of extra added bits. Otherwise, presence of a single error in the AC may cause
catastrophic decoding failure. Another approach for error robust transmission of
variable length codes is the method of EREC, which converts K VLBs into K fixed
length slots (FLS).
Decoding EREC in iterative source-channel receiver is challenging because to
detect the end of each VLB in the FLS, the VLB subdecoder (i.e. SISO arithmetic
decoder in our case) must be embedded into the EREC decoder and decision of end
of block be made based on some hard decision. Since, at the receiver EREC decoder
lies in the iterative loop between source and channel decoder, finding end of block
based on soft decision is difficult. This can be solved by applying EREC-ST to the
problem at hand. EREC-ST isolates pure EREC subdecoder from the effects of
VLB subdecoder by coding VLB lengths as side information (SI). Let L be the total
number of bits in K VLBs in a description and si is the length of i-th FLS, where,
dL/Ke for 0 i < mod(L, K)
si = (5.4.2)
bL/Kc for mod(L, K) i < K.
Figure 5.12: Figure shows four variable length blocks VLB(0), VLB(1), VLB(2) and
VLB(3) being rearranged into fixed length slots FLS(0-3)with EREC/EREC-ST
algorithm.
are output to code sm bi in binary. Let t = min (sm , bi ). Remaining bits and slots
are recalculated as bi = bi t and sm = sm t. In the Figure 5.12(b), after stage
0, only FLS(1) has bits left to be placed and searches FLS(2). Two bits are placed
in FLS(2). Since, no other VLB contains unplaced bits, the process moves to next
stage. In stage 2, VLB(1) places remaining three bits in FLS(3), as shown in the
Figure 5.12(d). The FLS are formed into packets and transmitted over network.
Table 5.2 shows the ST (state and tail bits) generated for this example. The
ST bits associated with a FLS are taken as a whole (SVLB) and EREC is applied
to reorganize all SVLB into fixed length slots called SFLS. If b is the average VLB
length, it was proved in [191] that fraction of ST bits per VLB is (2 + dlog2 be)/b,
which in this research is 2.7 percent. Thus, it can be safely assumed that ST bits are
transmitted with higher error protection with little coding overhead and received
error free.
Table 5.2: State and tail bits after different stages in the Figure 5.12.
n=0 n=1 n=2 SVLB SFLS
FLS(0) 0.00 000 0000
FLS(1) 1 1 1001
FLS(2) 0.10 1 0101 010-
FLS(3) 0.11 0.00 011000 011-
116
Figure 5.13: (a) Four sseg linked list generated by SVLB Reconstructor subroutine
in the process of EREC-ST decoding of FLS created in Figure 3. (b) Four bseg
linked list generated by function VLB Reconstructor, where bseg i corresponds to
VLB(i).
SFLS is utilized at the receiver side to recover variable length bitstreams back
from FLS. This is done in two stages, consisting of subroutines SVLB reconstructor
and VLB reconstructor. Refer to [191] for detailed algorithms for these two func-
tions. SVLB reconstructor outputs data structure sseg m = s, next, where s is the
number of blank bits after the current segment of FLS(m), and next points to the
next segment of FLS(m). If no more sseg is associated with current FLS, a null is
put in next field. The Figure 5.13(a) shows sseg for the four FLS formed in the
Figure 5.12. Zero in s field of sseg 0 indicates that no slots are left blank in FLS(0)
after the current segment. For sseg 1, s = 1 indicates that current VLB fills up
FLS(1) completely with some bits are remaining to be placed in other FLS. FLS(2)
contains data segment from two different VLBs as shown by two sseg. After first
segment, two bit positions are left vacant and next in sseg 2 points to next sseg.
In the similar way sseg information is generated for FLS(3).
With the aid of sseg information, VLB reconstructor algorithm rearranges bits
in FLS to original variable length bitstreams. VLB reconstructor creates a data
structure called bseg i = {m, s, s0 , next} associated with each VLB(i), where s
and s0 is the number of blank bits in FLS(m) before and after current segment
of VLB(i). The Figure 5.13(b) shows the content of bseg structure for the given
117
example. At the beginning it is assumed that all the VLBs at the receiver are empty
and are waiting to get bits from the FLS. VLB(0) is reconstructed with the aid of
information in bseg 0, where bits are obtained from 0-th FLS. Number of bits left
in FLS(0) is four (s = 4) before recovering the VLB segment. Since, after current
segment number of bits left is zero (s0 = 0 ), all the bits of the FLS(0) is taken up to
form VLB(0). VLB(0) gets not more bits from any FLS because next field in bseg 0
is null. VLB(1) obtains four bits from FLS(1), two bits from FLS(2) and three bits
from FLS(3) as indicated by bseg 1 in the Figure 5.13(b). Similarly, VLB(2) and
VLB(3) is reconstructed. One distinctive feature of VLB reconstructor algorithm is
that can be observed from this example is that underlying variable length code (i.e.
VLB) need not be decoded to recover VLBs back from FLS as was required with
EREC.
Wireless access points (WAP) constitute the last-hop delivery link to the end user.
Those packets which are not dropped in the network reach the WAP. These packets
are encoded with eight state rate 1/2 RSCC and transmitted over wireless channel.
The end user is able to able to receive packets from one or more WAPs. We assume
that reception is done through two different chain of antenna and receivers (as the
number of descriptions) and combined at last stage.
In this block, channel code (RSCC) and source code (arithmetic code) are decoded
iteratively as described in Section 5.2.1, but with the constraint of fixed length
packets. Iterative decoding allows source codec to be more robust to residual channel
errors. Also, channel decoder can utilize any residual redundancy of AC in the
process of decoding. The proposed iterative decoder is shown in the Figure 5.14.
We assume that input signal at the BCJR decoder Y = [y1 , ..., yi , ...] is obtained as
y i = xi + w i (5.4.3)
118
sseg
a
A Chase-like L Packet RSCC
FLS to +
arithmetic VLB sequence + Decoder
e Y
decoder to FLS G _ (BCJR)
a
G
VLB to FLS to
FLS packet
L
e (EREC) sequence
EREC-ST
The K variable length bitstreams in A are checked for CRC. In case of check failure,
corresponding wavelet tree is marked lost and is not utilized for further reconstruc-
tion. After SPIHT decoding of both descriptions, wavelet coefficients obtained are
obtained as I1 and I2 . If wavelet tree order of both descriptions (corresponding to
row and column index) arrive at the receiver without being lost, decoding is per-
formed by simple matrix lookup to get the central quantization index. However, if
some of the wavelet trees of either description are corrupted or lost, central decoding
can not be performed correctly by matrix lookup. In such cases, available bitstream
from either description is used to perform inverse index assignment using single
channel inverse quantizer. Thus, at the receiver, a wavelet tree can be declared lost
in three scenarios:
1. A packet (or FLS) is lost in network and all the VLB associated with it are
declared lost.
When both descriptions of same wavelet tree are lost, error concealment in wavelet
domain as described earlier is applied.
A Grey scale Lena image of size 512 512 is used for evaluating the performance
of proposed algorithm under following channel conditions: (i) AWGN channel with
no packet loss, and (ii) combined packet loss and AWGN channel. The given image
is hierarchically decomposed by 4-scale discrete wavelet transform, and then the
subband coefficients are quantized with uniform scalar quantizer of step-size . For
MDSQ, two types of index assignments are being considered, with number of diag-
onals D = 2 and D = 3, to form two descriptions. For each description, 256 wavelet
trees are constructed and SPIHT coded to generate 256 bitstreams. was adjusted
120
to obtain coding rate of 0.5 bpp/description after SPIHT coding. 4-bit CRC was
generated using generator polynomial x4 + x3 + x2 + x for the first 128 bits of a
bitstream and appended to it. This results in mere 0.78 percent overhead in each
description and is compensated by AC. All the bitstreams are independently arith-
metic coded for further compression, formed into FLS using EREC-ST. Proposed
scheme is more compression efficient compared to the MDC scheme of [64] as there
is no arithmetic coding performed in the later scheme. Data packets were formed
by packing eight slots and transmitted over packet network. Packets arriving at the
WAP are encoded with 8-state RSCC with generator polynomial (Gr , G) = (17, 15)8 .
In the simulation results discussed next, three schemes are compared. As base-
line, we consider scheme of Song et al. [64] with 256 bitstreams per description,
which is compared to the proposed scheme using EREC-ST and iterative soft de-
cision decoding of AC. To investigate the gain provided by iterative decoding, we
also compare with a similar scheme utilizing EREC-ST, but with conventional hard
decision decoding of AC. We first consider transmission over AWGN channel with-
out any packet drop. Figure 5.15 and Figure 5.16 shows the performance of the
proposed system for 2 diagonal and 3 diagonal index assignments, respectively. The
results are obtained by averaging 100 trials for each channel condition.
In the Figure 5.15, PSNR performance is compared for MDC schemes utilizing
2-diagonal index assignment. For hard decision decoding of AC aided by EREC-ST
(without soft iterative decoding), we can observe 1.5 dB improvement in average
PSNR at Eb /N0 of 3.5 dB compared to baseline scheme. With iterative decoding
remarkable improvement is observed in PSNR performance. After 1 iteration of soft
decoding of AC, there is a maximum gain in PSNR of 3.5 dB at Eb /N0 of 3.5 dB.
The maximum improvement after 4 iterations is 7 dB, which is observed at relatively
noisy SNR region of Eb /N0 = 3 dB. Proposed scheme, after four iterations, achieves
maximum PSNR of 35.6 dB at Eb /N0 of 3.5 dB. The baseline scheme achieves this
maximum value at Eb /N0 of 5 dB; thus, requiring 1.5 dB more signal power to
achieve same PSNR performance.
The Figure 5.16 compares the PSNR performance for the case of 3-diagonal in-
dex assignment. Compared to baseline scheme, proposed scheme with hard decision
121
decoder (hard decoder with EREC-ST) achieves maximum PSNR gain of 1.25 dB
at Eb /N0 of 3 dB. With four iterations of soft-decision decoding, this gain increases
by 6 dB. For four iterations of the proposed scheme, maximum PSNR of 36 dB is
obtained at channel Eb /N0 of 4 dB. The baseline scheme achieves this maximum
value at Eb /N0 of 5.5 dB; thus, requiring 1.5 dB more signal power to achieve same
PSNR performance. In the Figure 5.17, we give examples of reconstructed images in
order to illustrate the improvement in visual quality. It shows improvement in im-
age quality with the proposed decoder (Figure 5.17(a)) compared to baseline scheme
(Figure 5.17(b)) at Eb /N0 of 3 dB. These results demonstrate that the proposed im-
provement in MDC scheme makes it significantly noise robust compared to existing
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. The
objective is to evaluate the degradation of PSNR value for different number of packet
lost. We conduct experiments with the proposed system only as its superiority
compared to baseline system has been demonstrated in the earlier set of simulations.
The results for different wireless channel conditions are shown in the Figure 5.18 for
D = 2 and the Figure 5.19 for D = 3.
It 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.
The degradation is more evident in high Eb /N0 region (> 3 dB) whereas, there is
smaller effect of packet loss in low Eb /N0 region. These results demonstrate excellent
capability of the presented system to combat packet loss and noisy channel via
integrated design.
5.5 Conclusion
In Section 5.2, an accumulator based source-channel decoding scheme has been pro-
posed for transmission of arithmetic code over noisy channel. The proposed scheme
consisting of cascade of unity rate accumulator, two state RSCC and arithmetic
encoder has been compared with a similar system consisting of eight state RSCC
122
38
36
34
Average PSNR
32
30
28
38
36
34
Average PSNR
32
30
28
SISO w/ ERECST 4 iter.
SISO w/ ERECST 1 iter.
26 Hard dec. w/ ERECST
Baseline scheme
Figure 5.16: Performance of proposed scheme for transmission over AWGN channel
(with no packet loss) for D = 3.
123
Figure 5.17: Image recovered by proposed scheme (a) compared with recovery by
baseline system (b) at Eb /N0 of 3 dB and D = 3. PSNR of images (a) and (b) are
33.6 dB and 27 dB, respectively.
and arithmetic encoder. Simulation results show that the proposed scheme performs
better in terms of packet error rate for same overall code rate and decoding iterations
compared to the baseline system. Also because of reduction in number of states in
BCJR decoder, encoding and decoding complexity is reduced, even though there is
an extra accumulator stage.
In Section 5.3, iterative source-channel decoding techniques described in Section
5.2 has been applied for transmission of two-channel MDSQ based MDC over noisy
channel. Proposed scheme utilize low-complexity Chase-type SISO arithmetic de-
coder to improve the performance of MDC in the presence of noise. This results
in efficient utilization of soft information at the output of noisy channel and hence,
achieve noise-robustness. There is no significant increase in complexity at the en-
coder because arithmetic coder is a part of many bit plane coders like SPIHT and
EBCOT in JPEG 2000. At the decoder side, an iterative source-channel decoder
is employed which use SISO arithmetic decoder for decoding of descriptions. With
simulations, it has been demonstrated that proposed scheme performs better in
terms of decoded image PSNR compared to baseline scheme.
In Section 5.4, developed iterative decoding scheme has been extended to hybrid
124
36
4.0 dB
34 3.5 dB
3.0 dB
2.5 dB
32
Average PSNR (dB)
30
28
26
24
22
20
0 1 2 3 4 5 6 7 8 9 10
Total packets lost
36
4.0 dB
3.5 dB
34 3.0 dB
2.5 dB
32
Average PSNR (dB)
30
28
26
24
22
0 1 2 3 4 5 6 7 8 9 10
Total packets lost
channel consisting of combined packet loss channel and AWGN channel. Multiple
descriptions are obtained by applying two-channel MDSQ on image decomposed by
2-D discrete wavelet transform. The resulting descriptions are SPIHT and arith-
metic coded to further increase compression efficiency. Receiver for each description
consist of SISO decoder for variable length arithmetic codes, which exchange ex-
trinsic information with SISO RSCC decoder with the aid of EREC-ST algorithm.
Simulation results demonstrate considerable improvement of performance by the
proposed schemes over existing scheme in literature.
******
Chapter 6
Iterative Decoding of
Network-Channel Codes
Cooperative communication using relay nodes is one of the most effective means
of exploiting space diversity for low cost nodes in wireless network. In cooperative
communication, besides communicating their own information, users also relay the
information of other users. In this chapter we investigate two schemes where coop-
eration is achieved using a common relay node which performs network coding. Its
objective is to provide space diversity for two information nodes transmitting to a
base station.
The proposed schemes use Reed-Solomon error correcting code for encoding the
information at the user nodes and punctured convolutional code as network code,
instead of XOR based network coding. In the first scheme, network coding is ap-
plied in multiple-access relay channel (MARC) scenario and source nodes do not
communicate with each other directly. In the second scheme, both the source nodes
exchange information and together form virtual antenna array for space-time coded
cooperation, in addition to network coded cooperation. We propose iterative soft de-
coding of joint network-channel code by treating it as a concatenated Reed-Solomon
convolutional code.
126
127
6.1 Introduction
The increasing popularity of ubiquitous computing has motivated the deployment
of wireless communication devices that require ad hoc communication. One such
application of pervasive computing is in wireless sensor networks (WSN), which is
used to monitor the physical world parameters like temperature, radiation levels,
image, video etc. The WSN consists of large number of low cost wireless sensor
nodes and each node itself has very limited computational capability. The nodes
are further constrained by limited power of transmitters. The main challenge of
wireless sensor networks is to achieve proper balance between transmit/processing
power and quality of service. However, such multi-terminal systems are limited by
impairments due to wireless channels, such as fading, and interference.
Spatial diversity has been accepted as one of the most effective ways to combat
fading over wireless channels. This has inspired extensive research on multiple-
input-multiple-output (MIMO) over the past decade [153]. However, achieving full
diversity gain requires multiple antennas to be placed sufficiently apart, which may
be difficult for small mobile terminals. Because of these practical limits, alternate
approaches, such as relaying, distributed antenna system (DAS) and user coopera-
tive transmission [129] have been proposed to provide spatial diversity. There are
many cooperative strategies to achieve efficient node cooperation, such as amplify
and forward [136], decode and forward [210], and coded cooperation [116,211]. Based
on these basic schemes, various other improvements have been proposed in recent
years.
The idea of network coding was first proposed by Ahlswede et al. [152] to enhance
the capacity of the noiseless wired network. The core idea of network coding is to
allow simple coding capability at relaying nodes, in exchange for network capacity
gain. It has been shown that considerable capacity gain can be obtained as such
over wired communication links [152]. This pioneering idea has inspired considerable
research efforts in communication and computer networking communities. Although
the noiseless assumption used in [152] is no longer valid in wireless communication,
the wireless medium does provide some desirable characteristics that facilitate the
application of network coding, e.g., broadcasting without additional cost [153].
128
S1 S1
x x
RN 1 BS BS
x y
RN 2 +RN
y y
S2 S2
(a) (b)
Figure 6.1: The uplink for two mobile stations (S1 and S2 ) with cooperative diversity.
(a) Diversity obtained by simple relaying, (b) Network coded diversity.
cost.
Focus of our research is MARC, where two user nodes obtain cooperation through
a fixed relay node performing network coding. This system has advantage of diver-
sity gain, as well as increase in system throughput. The benefit of network coding in
MARC schemes has been demonstrated in literature. Authors in [161,162] proposed
schemes which combines the benefits of space-time codes and network coding for
cooperative communication. The authors use simple XOR based network coding at
the relay for obtaining diversity. Du et al. [212] investigated a cooperative strategy
based on parity check network coding. Their study revealed that a successful design
should employ the most effective extra check bits to make full use of information
in relayed bits to help decode the message from two users. Wei Fang et al. [191]
proposed a joint network-channel coding scheme based on distributed turbo code for
multiple-access relay channel, using decode-and-forward protocol. The problem of
transmitting correlated binary sources over MARC was considered in [213], where a
joint source-channel-network decoding technique was proposed to fully exploit the
correlation between sources.
In this research, we consider joint network-channel coding for MARC when the
transmitting node employs RS code error correcting code and punctured recursive
systematic convolutional code (RSCC) is used as network code. Encoding user in-
formation with RS code, in contrast to simply encoding with linear block code [173],
is more relevant because of widespread prevalence of RS code in many existing stan-
dards. Our main contribution is application of iterative soft decision decoding in
the context of network-channel decoding in MARC setup. The proposed algorithm
enables network code and channel decoders to exchange soft information iteratively
and to achieve an improved performance compared to a hard decision iterative de-
coder. The convergence behaviour of the network-channel decoder is analyzed using
EXIT (extrinsic information transfer) chart and we also compare the computational
complexity of the proposed algorithm. We demonstrate using bit error rate (BER)
simulations that the proposed scheme performs better compared to existing design
proposed in [173] and conventional XOR based network coding.
In later part of the chapter, we enhance the bit error rate performance of a dis-
130
tributed space-time coded system aided by a relay which performs network coding.
We improve upon the research in [162], which uses simple XOR based network code,
by application of the concept of product code and iterative network-channel decod-
ing. The proposed scheme uses punctured RSCC as network code at the relay and
RS code as channel code. Our main innovation is iterative network-channel decod-
ing at the receiver using the principles derived from soft decoding of concatenated
RS-convolutional codes [185].
Rest of the chapter is organized as follows. In Section 6.2, proposed iterative
network-channel decoding scheme for MARC is presented. In Section 6.3, a iterative
network-channel scheme is presented in which source nodes cooperate with each
other using Alamouti space-time block code, in addition to network coding at the
relay. Finally, Section 6.4 is conclusion.
We consider the scenario shown in the Figure 6.2, with two mobile user nodes
communicating information to a common base station in a MARC setup. Basically,
the system consists of two mobile sources MU1 and MU2 transmitting binary data
sequences S1 and S2 respectively, towards the relay node (RN) and a base station
(BS). Both user nodes have a symmetric positioning with respect to RN and BS.
Since radio terminals cannot transmit and receive simultaneously, the nodes are
assumed to operate in half-duplex mode and are using orthogonal channels. The
relay node decodes the intercepted packets, re-encodes them and performs network
coding on the information received from both the information nodes. Network coded
packets are then forwarded to the base station in the second phase of transmission.
Decoding is performed at the BS by combining the packets received from both
the user nodes and relay. The detailed protocol showing transmitting nodes, their
channel assignments, and the respective transmitted messages is shown in Table 6.1.
In the next paragraphs we describe the encoding operations at source and relay in
detail and decoder is discussed in subsequent subsections.
131
MU1
S1 S1
S3
S2 RN S2
MU2 BS
Figure 6.2: Multiple access relay channel with two users transmitting channel coded
information to a base station. Network coding operation is performed at the relay
node.
Let Fq denote the finite field of size q. It assumed that Fq is an extension field
of F2 as q = 2m . The message vector for an (n, k) RS code is given as U (,p) =
(,p) (,p) (,p)
[U1 , U2 , ..., Uk ] Fqk , where n and k are the length and dimension of the
code, respectively, and n = q 1. The superscript {1, 2} denotes the mobile
user MU and variable p denotes the pth codeword or packet. The RS codeword p of
source MU- is obtained as
(,p) (,p)
C (,p) = U (,p) G = [C1 , C2 , ..., Cn(,p) ] Fqn (6.2.1)
where G is the generator matrix of the RS code. In practice, the bits from each
(,p)
source are grouped into m bit symbols, Ui Fq , for i = 1, ..., k and coded with
(n, k) RS code. The encoding at MU nodes is shown in the Figure 6.3(a). The RS
code has dual functions; first, it is efficient against burst errors, since a sequence
of m + 1 consecutive bit errors can affect at most two code symbols. Second, RS
code aids in iterative joint network-channel decoding as discussed in next section.
132
(a)
MU1 RS Decode
packets & encode Row RSCC Parity Sr
ordering (column by column)
BPSK
MU2 RS Decode
packets & encode
(b)
Figure 6.3: (a) Block diagram of encoder at mobile user nodes, (b) Network coding
operation at relay node.
At each source node, L codewords are generated, grouped into a frame and inter-
leaved with . The stream is then formed into matrix of size L n, where each
row forms a packet. Each packet is thus composed of randomized symbols from dif-
0(,p) 0(,p) 0(,p)
ferent RS codewords and is represented as C 0(,p) = [C1 , C2 , ..., Cn ] Fqn
where p now indicates packet number of each source. The symbol array C 0(,p) are
0(,p) 0(,p) 0(,p)
converted into bit array c0(,p) = [c1 , c2 , ..., cmn ] F2mn . The translated bits
are modulated and broadcast to BS and relay.
The function of the relay node is to perform network coding on the incoming packets
from two sources. In case of simple XOR based network coding, one packet from
source MU1 is linearly combined, using bit-by-bit XOR operation, with packet from
source MU2 and the resulting bits are transmitted to the BS. Resulting network
code at relay node has rate of 2/3. The Figure 6.3(b) shows the proposed encoding
operation at the relay node. The relay node overhears transmission from both the
MU nodes during the first phase of transmission, decodes and reencodes them. The
reencoded packets are ordered into matrix S of size L mn bits, as shown in the
Figure 6.4, where L = L1 + L2 . Here, we have the first L rows belonging to the
cooperating users.
From these packets, remaining N L rows are generated using network code and
forwarded by the relay. In this research we use RSCC as network code instead of
133
m x n bits
Packet 1, MU1
}
Packet 1, MU2
Packet 2, MU1
Packet 2, MU2
L1 + L2 S
:
:
:
Packet L1, MU1
}
N
Parity node packets P
Figure 6.4: Data packets from two sources are ordered in alternate rows. Total
number of rows from two sources is L1 + L2 . Remaining rows of parity check bits
are calculated and transmitted from the relay node.
XOR based network code. RSCC of rate L/N is applied on each column of L bits,
from which N L parity check bits are obtained. The parity check bits forms matrix
P of size (N L) mn. It is to be noted that relay retransmits only the parity
part of RSCC. Puncturing may be applied on the parity bits to attain necessary
code rates. In this paper, every alternate row of the parity bits (P in the Figure
6.4) is punctured to obtain code rate of 2/3. Any different puncturing pattern may
be applied, but this may result in different performance. Network code is obtained
from these parity check bits and each row is transmitted as packet to the BS. Thus,
each row encounters a different channel and hence, bits in a row suffer from same
amount of fading.
Channel Model
The wireless channel is assumed to be flat Rayleigh fading channel and the fading
coefficient is constant for the duration of one packet transmission. Let the p-th
received packet at the BS from link l be denoted as ypl = [yp1
l l
, yp2 l
, ..., ypn ], for l
1, 2, r and p = 1, ..., Ll . Thus, the received signal from link l and packet p is
l
ypi = hlpi slpi + zpi
l
, i = 1, 2, ..., mn (6.2.2)
134
where, slpi is the transmitted symbol of unit energy. The fading coefficients hlpi
are zero mean complex valued Gaussian random variable with Rayleigh-distributed
envelope. The channel is assumed to experience slow Rayleigh fading such that
fading coefficients are nearly constant over one codeword interval. Gaussian noise
l
zpi experienced by the i th symbol of link l has double-sided power spectral density
N0 /2.
The block diagram of the proposed iterative soft network-channel decoding algorithm
is shown in the Figure 6.5. At the receiver, RSCC and RS code are considered a
concatenated code structure and is decoded iteratively [185]. The next stages con-
sists of iterative soft decoding process in which MAP decoder is applied along the
columns for SISO decoding of RSCC and adaptive belief propagation (ABP) algo-
rithm [176] for SISO decoding of RS codes along rows. Assuming BPSK modulated
transmission, we denote the LLR of i-th received bit slpi of packet p from link l as
l (p, i) and is calculated as
l
l (p, i) = (|ypi + hlpi |2 |ypi
l
hlpi |2 )/N0 . (6.2.3)
Channel state information hlpi is assumed to be available at the base station. The
LLR of received packets are stacked over one another in the form the matrix ch
as given in the Figure 6.4. ch consists of alternate rows of LLRs of packets from
MU1 and MU2, while last (N L1 L2 ) rows are LLR of packets received form RN.
al , el and pl denote the a priori LLR, extrinsic LLR and the a posteriori LLR,
respectively and are related as
el = pl al (6.2.4)
e
-1 La1 RS ABP L1
Stack e
P1 Decoder
P1
y
1
G1 rows G1
ch e S1
BCJR Gr
2
y Calculate G2 G Isolate Combine
yr LLR decoder rows rows
a e S2
Gr Gr G2 a e
L2 RS ABP L2
P2-1 Decoder P2
deinterleaved to obtain La1 and La2 . Each row in La1 and La2 constitute the a priori
LLRs of an RS codeword and can be decoded independently by applying ABP
algorithm. After an iteration of ABP on all the rows, extrinsic information Le and
a posteriori information Lp for = 1, 2 is obtained.
With Lp , BM algorithm [194] is performed to decode the RS codes. The re-
trieved codes are represented as S . Le1 and Le2 are then interleaved and rows are
combined into a priori LLR matrix ar for the next iteration of MAP decoder. The
decoding terminates once all the RS codewords satisfy parity check constraints or
the maximum number of iterations is reached.
One of most popular MAP decoder used for decoding RSCC is BCJR algorithm.
We will briefly summarize the BCJR algorithm employed as network decoder in the
proposed scheme. Let ch (ci ) = ch (i, j), i = 1, ..., N , be the j-th column of ch
corresponding to a RSCC codeword c = [c1 , ..., ci , ..., cN ].
The BCJR decoder computes the a posteriori LLR p (ci ) with ch (ci ), i = 1, ..., N
and a priori LLR a (ci ) = ar (i, j), i = 1, ..., L as input, where a (ci ) is the j-th
column of ar . The a posteriori LLR is at the output of BCJR algorithm is defined
as
p P (ci = +1|y)
(ci ) = log . (6.2.5)
P (ci = 1|y)
BCJR decoder obtains a posteriori LLR by incorporating the trellis of the code:
0
P
p S + p(si1 = s , si = s, y)/p(y)
(ci ) = log P 0
, (6.2.6)
S p(si1 = s , si = s, y)/p(y)
136
where S is the set of all the states of trellis, si S is the state of the encoder at
time i. S + is the set of ordered pairs (s0 , s) corresponding to all state transitions
(si1 = s0 ) (si = s) caused by input ui = +1, and S is similarly defined for
ui = 1. Using Bayes rule, (6.2.5) can be written as
p P (y|ci = +1) P (ci = +1)
(ci ) = log + log (6.2.7)
P (y|ci = 1) P (ci = 1)
= e (ci ) + a (ci ), (6.2.8)
where, the second term a (ci ) represents a priori LLR and was applied as input
to the decoder. The term e (ci ) = p (ci ) a (ci ) is the extrinsic information is
obtained by subtracting a priori input from the output of the decoder. Extrinsic
LLR for column j is saved as er (i, j) = e (ci ), i = 1, ..., L and er is passed on to the
next decoder stage.
The next stage consists of ABP algorithm for SISO decoding of RS codes. ABP
algorithm was explained in Section 3.3 of Chapter 3. In this section, we summarize
the important steps of ABP algorithm in the context of this research.
The extrinsic LLR matrix e of previous stage after deinterleaving becomes
matrix La for this stage. For simplification of notations, let the LLR of bits in i-th
row (or i-th RS codeword) at iteration t be denoted as
where, (t) = [(t) (c1 ), ..., (t) (cj ), ..., (t) (cmn )] acts as a priori information for an
RS codeword. To update the parity check matrix H, |(t) (cj )| is sorted in ascending
order of magnitude and sorting index are stored as = [j1 , j2 , ..., jmn ]. Columns of
the matrix H are then reordered according to the permutation to obtain (H).
Gaussian elimination then reduces the first independent m(n k) columns to iden-
tity submatrix. Let this matrix be denoted as H 0 . Finally, inverse permutation is
performed on the columns of H 0 as follows:
H 00 = 1 (H 0 ). (6.2.10)
137
Le is then interleaved using and rows of LLR are combined back into ar . The
combined LLR matrix ar is of size (L1 + L2 ) mn and is of form given in the Figure
6.4. The iterations are run a predetermined number of times or until parity checks
are satisfied.
In this section, we analyze the complexity incurred by the proposed iterative decod-
ing algorithm. Since, encoding operation at the user nodes and encoding/decoding
at relay is not different for both proposed and baseline scheme, we concentrate only
on the decoding complexity in this paper. The following calculations are based on
analysis done by Li Chen [185]. As discussed in earlier sections, the iterative de-
coder consists of two constituent decoders BCJR and ABP algorithm. An iteration
of BCJR decoding algorithm for RSCC requires O(12(L1 + L2 )) floating point
operations, where is the number of states of the RSCC coder. SISO decoding
of RS code, consists of three stages, which includes Gaussian elimination, belief
propagation or sum-product algorithm and RS hard decision decoding. Gaussian
elimination process requires O(mn(mn km)2 ) binary operations. If is the aver-
age row weight of adapted parity check matrix H 00 , each iteration of sum-product
138
floating point operations. In the above equation, (L1 + L2 ) is the total number of
RS codewords received from both the mobile users. Number of binary operations
required is
O(Niter (L1 + L2 )mn(mn km)2 ) (6.2.15)
In this section we utilize extrinsic information transfer (EXIT) chart [14, 216], to
analyze the iterative decoding behavior of the proposed network-channel decoder.
It allows us to visualize the decoding trajectory of iterative decoder components
and prediction of BER performance. This in turn facilitates selection of suitable
constituent codes for the scheme without time consuming Monte Carlo simulations
for evaluating the performance of the scheme.
A fundamental assumption of EXIT chart analysis is that extrinsic information
passed from one SISO decoder to other is a Gaussian distributed random variable.
The LLR a of a priori input for uncoded information x {+1, 1} is modeled as
a = a x + na , (6.2.17)
where, na is a Gaussian random variable with zero mean and variance a2 . The
variance must satisfy the condition a = a2 /2. The mutual information between a
and x is defined as
1 X 2pa (|x)
I(x, a ) = pa (|x) d, (6.2.18)
2 x=1 pa (|x = 1)pa (|x = +1)
0.9
0.8
(I(2)
a
)
0.7
I (1)
e
0.6
0.5 RS (15,11)
Eb/N0 = 2 dB
Eb/N0 = 3 dB
0.4
0 0.2 0.4 0.6 0.8 1
I (1)
a
(I(2)
e
)
Figure 6.6: EXIT chart of the proposed network-channel decoder with (15,11) RS
code as channel code (outer code) and punctured (5, 7)8 RSCC acting as network
code with code rate of 2/3.
0.95
0.9
(I(2)
a
)
0.85
I (1)
e
0.8
0.75 RS (31,27)
Eb/N0 = 3 dB
Eb/N0 = 4 dB
0.7
0 0.2 0.4 0.6 0.8 1
I (1)
a
(I(2)
e
)
Figure 6.7: EXIT chart with (31,27) RS code and punctured (15, 17)8 RSCC acting
as network code with code rate of 2/3.
141
0.9
0.8
I (1)
e
(I(2)
a
)
0.7
Figure 6.8: EXIT chart with (15, 11) RS outer code and different RSCC codes as
network code.
phase shift keying (BPSK) modulated signal transmitted over Rayleigh block fading
channel, i.e. the channel fading coefficient is assumed to be constant for the duration
of one packet. RN is assumed to be near the source nodes, MU1 and MU2 and MU-
RN links are assumed to be AWGN channel. In this research, number of packets
transmitted by both the source is L1 = L2 = 20.
First we investigate the iterative convergence behavior of the proposed design.
The Figure 6.9 shows BER performance of the proposed network-channel decoder
with (15, 11) RS code as channel code and punctured (5, 7)8 RSCC as network code.
It can be observed that iteration gain is obtained for Eb /N0 values higher than 2 dB,
as predicted in EXIT chart of the Figure 6.6. Error rate decreases with increase in
iterations and there is no significant improvement in BER after 10 iterations. The
Figure 6.10 shows the performance of the iterative decoder with (31, 27) RS code
and (15, 17)8 RSCC combination, in which iteration gain starts for Eb /N0 3 dB.
Best BER performance is achieved for 10 iterations or more. This corresponds to
EXIT chart in the Figure 6.7, as explained earlier.
Next we look at the interaction between channel and network code. For this pur-
143
100
10-1
10-2
BER
10-3
1 iter
10-4 2 iter
3 iter
5 iter
10 iter
10-5
0 1 2 3 4 5 6
Eb /N0 (dB)
Figure 6.9: BER performance of the proposed scheme using (15, 11) RS code and
punctured (5, 7)8 RSCC with number of decoding iterations over Rayleigh block
fading channel.
100
10-1
10-2
BER
10-3
10-4
1 iter
2 iter
10-5 3 iter
5 iter
10 iter
10-6
0 1 2 3 4 5 6 7 8 9
Eb /N0 (dB)
Figure 6.10: BER performance for (31,27) RS code (15, 17)8 RSCC combination
with number of decoding iterations over Rayleigh block fading channel.
144
pose, we fix the channel code at the user nodes to (15,11) RS codes and look at con-
volutional codes with different generator polynomials applied as network code. To
compare the performance with XOR based network code and the reference scheme,
the convolutional code at the relay is punctured to get rates close to 2/3. The pro-
posed decoding algorithm is run for 10 iterations. With all the links having same
average received SNR, the Figure 6.11 shows average bit error rate of the MARC as
function of Eb /N0 . For comparison we also show the results obtained in the reference
scheme [173] with a similar network-channel decoding scheme. It is observed that
the average bit error probability of the proposed scheme has improved considerably
compared to the reference system. In contrast to the proposed scheme, the refer-
ence scheme uses BCH code as both channel and network code. For instance, using
punctured (23, 35)8 RSCC as network code, a gain of nearly 1.5 dB can be observed
at a bit error rate of 103 , as compared to (63, 39, 9) BCH code based network
coding. Compared to (31, 21, 5) BCH code based network code, gain is even more
significant (of 2.5 dB) at bit error rate of 103 . Whereas, the gain of the proposed
scheme with (5, 7)8 RSCC network code with respect to (63, 39, 9) and (31, 21, 5)
BCH code based network code is 1 dB and 2 dB, respectively. For the purpose of
comparison, the Figure 6.11 also shows the performance of the baseline scheme with
XOR based network code.
Next we simulate another type of interaction between channel and network code
where network coding at the relay node is kept fixed and channel code is varied.
The Figure 6.12 illustrates the bit error rate of the proposed scheme with punctured
(5, 7)8 RSCC as network code. The channel codes considered are (15,7), (15,11) and
(7,4) Reed Solomon (RS) codes. It can be observed that relative gain in performance
is higher than that of baseline scheme consisting of relatively complex (31,21,5) BCH
code as network code, and channel codes being (15,7), (15,11), (7,4) BCH code. For
example, to achieve bit error rate of 104 the relative gain of the proposed coding
scheme with user channel code of (15,7) RS code is about 2 dB better compared to
corresponding reference scheme with (15,7) BCH code. On the other hand the gain
is 3 dB for (15, 11) block code used as channel code at the user nodes. But the gain
is reduced to 2 dB for (7,4) block codes used as channel code at the two user nodes.
145
100
NC w/ (5,7) RSCC
NC w/ (23,35) RSCC
NC w/ (15,17) RSCC
10-1 Ref. w/ (63,39,9)
Ref. w/ (31,21,5)
Ref. w/ XOR
Average BER
10-2
10-3
10-4
0 1 2 3 4 5 6 7 8 9 10
Eb /N0 (dB)
Figure 6.11: Performance comparison of the proposed iterative decoder with different
reference schemes. Different network codes at relay node are evaluated for a fixed
channel code at the user nodes.
100
CC w/ RS (15,7)
CC w/ RS (15,11)
CC w/ RS (7,4)
10-1
Ref. w/ (15,7)
Ref. w/ (15,11)
Ref. w/ (7,4)
10-2
Average BER
10-3
10-4
10-5
10-6
0 1 2 3 4 5 6 7 8 9
Eb /N0 (dB)
Figure 6.12: Performance comparison with different channel codes at the user nodes
and fixed network coding at relay.
146
10-2
10-3
BER
10-4
10-5
User 1
User 2
10-6
0 1 2 3 4 5 6 7 8 9 10
1 (2 = 10 - 1 )
Though, all the simulations performed in this paper use BPSK modulation, this can
be easily extended to higher order modulations.
The Figure 6.13 shows the performance of the proposed scheme in asymmetric
SNR scenario. The information at both sources are encoded with (15,7) RS code
and network coding operation is done at the relay with (15, 17)8 RSCC as mother
code. Similar to scenario in [173], we study a case where sum of SNRs for the direct
links of the users is assumed constant, that is, 1 + 2 = 10 dB. Average received
SNR of the relay-destination link is assumed constant with 3 = 8 dB. It can be
observed that because of coupling between the two users at relay node, SNR of one
link affects the BER performance of other link also. When one of users has bad link,
both users will experience bad link performance regardless of the other users link
quality. Thus, best performance is obtained with the proposed scheme when SNR
of both the links are almost equal.
147
MU1
S1
S1
RN
S2
MU2 S2 BS
(a)
MU1
S3
S5
RN
S4
MU2 BS
(b)
Figure 6.14: Space-time and network coded cooperation. (a) In first phase, user
nodes broadcast information to relay, base station and other user node. (b) Second
phase consists of retransmission by user nodes and relay node.
We consider the scenario shown in the Figure 6.14, with two mobile users (MU1
and MU2) nodes communicating information to a common BS. The two MU nodes
cooperate with each other using Alamouti space-time block code. In addition, a
dedicated relay node performs network coding on the information received from
both the information nodes. We further assume that all the nodes are using or-
thogonal channels with relay node operating in half-duplex mode. Similar scenario
is presented in [162] for evaluating the performance improvement due to additional
network coding node. But instead of simple XOR based network coding, we consider
convolutional encoder as network code and evaluate the performance of the system
in iterative decoding scenario. Next we describe the channel coding operation at the
MU nodes and relay.
148
MU1 RS Decode
packets & encode Sr
Row RSCC Parity
ordering (column by column)
BPSK
MU2 RS Decode
packets & encode
(b)
Figure 6.15: (a) Block diagram of encoder at mobile user nodes, (b) Network coding
operation at relay node.
The Figure 6.15(b) shows the encoding operation at the relay node. The relay
node overhears transmission from both the MU nodes during time slot t = 1, 2,
decodes and reencodes them. The reencoded packets are ordered into matrix of
size (L1 + L2 ) mn bits. The rows from the two cooperating sources are arranged
alternately as shown in the Figure 6.4. In this research, we use RSCC as network
149
code instead of XOR based network code. RSCC of rate (L1 + L2 )/N is applied on
each column of length (L1 + L2 ) bits, and N L1 L2 parity bits are generated.
Puncturing may be applied on the parity bits to attain necessary code rates. Network
code Sr is obtained from these parity check bits and each row is transmitted as packet
to the BS. Thus each row encounters different channel and bits in a row suffer from
same amount of fading.
The transmission of the message is accomplished in two phase or five time slots. The
first two phases or four time slots constitute transmission by distributed Alamouti
code. In the first phase (Figure 6.14(a)), the mobile users MU1 and MU2 broadcast
their messages S1 and S2 over wireless channel, respectively. This being broadcast
phase, the transmitted messages are received by the BS, relay and the other user.
Table 6.2: Transmission schedule at the mobile user and relay nodes.
Time slot (t) 1 2 3 4 5
Transmitter MU1 MU2 MU1 MU2 RN
Message S1 S2 S2 S1 Sr
In the next multiple access phase, the cooperating users MU1 transmits S2 at
time slot t = 3 and MU2 transmits S1 at time slot t = 4 to the BS. The relay node
decodes the signal received from both the users, reencodes them using RS code.
Bits from both the sources are combined, network coded and transmitted to the
destination as Sr . The scenario in second phase is illustrated in the Figure 6.14(b).
As mentioned earlier, all the nodes transmit in orthogonal channels (separate time
slots), the Table 6.2 details the channel assignment for different transmitting nodes.
The signal received at the BS after two phases of transmission and cooperative
Alamouti relaying from two MS, in matrix vector notation, is
1 1
y h h2 s n
pi = 1 pi + 1 (6.3.19)
2
ypi h2 h1 s2pi n2
150
where, slpi is the i-th bit of packet p transmitted from source MU l for l = 1, 2,
l
and ypi is the signal received corresponding to it. Also, hl , l = 1, 2 denotes complex
multiplicative fading coefficients for MU-BS channel and is assumed to be constant
for at least one codeword duration. It is worth mentioning again that each row of
bits of matrix in the Figure 6.4 undergoes different fading. Utilizing (6.3.19), log-
liklihood ratios (LLR) of the received packets from the two sources are obtained as
1 , 2 , respectively, using soft-output Alamouti decoder [201].
On the other hand, signal received at the relay node is
r
ypi = hr srpi + nr (6.3.20)
e
-1 La1 RS ABP L1
Stack e
P1 Decoder
P1
1 G1 rows G1
y S1
2 Alamouti G2 G
ch
BCJR Ge Isolate Combine
y soft output decoder rows rows
r a e S2
y Gr G G2 a e
L2 RS ABP L2
Calculate LLR
P2-1 Decoder P2
The Figure 6.17 shows the EXIT characteristics of the proposed decoder with (31,25)
RS code. The inner decoder (decoder 1) consists of the cooperative Alamouti space-
time decoder and punctured (7, 5)8 convolutional code acting as network code with
overall code rate of 2/5. The (Ia , Ie ) curves are plotted with inner decoder 1 for
average channel Eb /N0 of 6, 7 and 8 dB. Inverse EXIT characteristics (Ie , Ia ) of
decoder 2 (outer RS decoder running JN algorithm) is also shown in the figure. It
shows, at 6 dB, the tunnel starts to open between EXIT curve of decoder 1 and
decoder 2, and at 7 dB, the tunnel is completely open. Therefore, the decoder bit
error rate (BER) cliff is expected to start at 6 dB and can be verified in the Figure
6.19. Similarly, EXIT chart for the proposed system with (15,7) is given in the
Figure 6.18, where turbo-cliff starts at 5 dB as shown in the Figure 6.20.
152
0.95
0.9
0.85
I (1)
e
(I(2)
a
)
0.8
0.75
RS (31,25)
0.7 Eb/No = 6 db
Eb/No = 7 db
Eb/No = 8 db
0.65
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(1) (2)
I a (Ie )
Figure 6.17: EXIT chart of the proposed iterative soft decoder utilizing (31,25) RS
code (outer decoder) as channel code.
0.9
0.8
0.7
(I(2)
a
)
0.6
I (1)
e
0.5
0.4
RS (15,7)
0.3
Eb/No = 5 db
Eb/No = 6 db
0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(1) (2)
I a (Ie )
Figure 6.18: EXIT chart of the proposed iterative soft decoder utilizing (15,7) RS
code as channel code.
153
100
1 iter
5 iter
10 iter
10-1 20 iter
10-2
BER
10-3
10-4
10-5
1 2 3 4 5 6 7 8 9
Eb /N0 (db)
Figure 6.19: BER performance of the proposed scheme using (31,25) RS code with
number of decoding iterations over Rayleigh block fading channel.
100
1 iter
2 iter
3 iter
5 iter
10-1
10-2
BER
10-3
10-4
0 1 2 3 4 5 6 7 8 9 10
Eb /N0 (db)
Figure 6.20: BER performance of the proposed scheme using (15,7) RS code with
number of decoding iterations over Rayleigh block fading channel.
154
XOR based network coding at RN. Overall code rate in both the cases is 2/5.
It can be observed that proposed scheme benefits from iterative decoding of
network and channel codes, where there is an improvement of 5.25 dB in Eb /N0
for achieving BER of 104 . Performance of both the schemes can be improved if
the link between relay node and the base station is better. The Figure 6.21 also
illustrates BER for both the schemes with SNR of RN-BS link 10 dB higher than
MU-BS link. For the proposed scheme, there is an improvement of 0.5 dB, compared
to equal SNR case, to achieve BER of 104 . For the reference scheme, improvement
in this case is 0.65 dB. Performance of the reference scheme without RS coding at
the MU, XOR network coding at RN and cooperative space-time transmission is
also shown for reference. Stronger RSCC used as network code can improve the
error performance of the scheme as illustrated in the Figure 6.22.
6.4 Conclusion
In section 6.2, an iterative network-channel decoder has been proposed for MARC
setup. The received signal at BS is considered output of a product code, though
encoding is done by different codes at geographically isolated locations. In this 2-D
product code structure, rows are formed by the channel coded packets from two
users, and columns formed by network code at the relay. Proposed scheme con-
sists of channel coding performed by RS codes at the user nodes and punctured
convolutional code acting as network code at the common relay. Iterative soft de-
cision decoding is performed at the receiver by exchanging extrinsic information
between RS and convolutional codes. Considerable improvement in bit error rate
performance is obtained with the proposed scheme when compared to the base-
line network-channel scheme. The EXIT analysis of the proposed decoder has been
presented for validating the decoding convergence.
In section 6.3, the work on network-channel decoder has been extended to space-
time-network coded cooperation over wireless network. The proposed scheme use
RS code as channel code and punctured RSCC based network code, with space-
time cooperation between source nodes. First stage of the receiver generates soft
156
100
10-1
10-2
BER
10-3
10-4
Iter RS-CC
Ref RS-XOR
Iter RS-CC, +10db
10-5 Ref RS-XOR,+10db
Ref. XOR
Coop Alamouti
10-6
0 2 4 6 8 10 12 14 16 18
Eb /N0 (dB)
Figure 6.21: Performance comparison of the proposed iterative decoder with a ref-
erence scheme employing XOR based network code. BER performance are also
compared for the case when SNR at relay is 10 dB higher than direct links.
100
(2,3)8
(5,7)8
10-1 (15,17) 8
10-2
BER
10-3
10-4
10-5
10-6
1 2 3 4 5 6 7 8 9
Eb /N0 (dB)
Figure 6.22: Performance of the proposed iterative scheme using (31,25) RS code
for different RSCC used as network code.
157
decisions from space-time decoder and is applied to a cascade of BCJR decoder and
SISO RS decoding stage. Improvement in error rate performance has been obtained
with proposed scheme and compared to scheme using XOR based network code. The
EXIT analysis of the proposed decoder has been presented to validate the decoding
convergence results.
******
Chapter 7
7.1 Conclusion
A FEC based MDC decoding scheme using SISO RS decoder has been proposed.
The scheme utilizes soft information available at the output of inner RSCC decoder
for SISO decoding of outer RS code. After an iteration, RS decoder in turn generates
extrinsic information, which is applied back to RSCC (BCJR) decoder. Performance
of the proposed scheme [170, 183] has been compared to similar reference scheme
[80] in literature for transmission of SPIHT coded image. In case of COAF and
CODF relay channels, performance is compared to baseline system based on scheme
proposed in [187].
In the simulations performed with setup given in Section 3.5, PSNR gain
158
159
For the simulation setup given in Section 4.2.3, proposed iterative LDPC-
RS scheme in AWGN channel achieves maximum PSNR performance for 0.75
dB less SNR compared to tandem scheme (similar scheme without iterative
decoding) and 1.4 dB less compared to reference scheme in [80].
improvement has also been shown in Rayleigh channel with fading parameter
fD Ts = 0.05.
In case of COST 207 Rural Area and Typical Urban, 6-tap frequency selective
channel, proposed LDPC-RS scheme has advantage in channel SNR of 4.5 dB
in achieving maximum PSNR compared to tandem scheme.
With the simulation setup given in Section 4.3.4, proposed three stage DSTBC-
RSCC-RS iterative decoding scheme, after four global iterations, shows an
improvement of 2.6 dB of channel SNR in achieving maximum value of PSNR.
After four inner and outer decoding loops (local iterations), proposed DSTBC-
RSCC-RS scheme shows an improvement of 2.4 dB in terms of channel SNR. It
is observed that if the code redundancy is not fully exploited by inner decoding
loop, it can be utilized to a certain extent in the outer decoding loop. Since the
outer decoding loop utilize hard decisions, it does not provide gain comparable
to a inner soft decision loop.
In the first part of Chapter 5, source-channel decoding scheme has been proposed
[190] based on concatenated accumulator-convolutional code-arithmetic code (ACC-
CC-AC) and compared to reference scheme [115]. Using the knowledge derived from
this scheme, source-channel decoding techniques are applied to two-channel MDSQ
based multiple description coding of image [168, 169] and compared to a reference
scheme by Song et al. [64]. In the proposed receiver, soft channel information is
used to improve the noise robustness of the arithmetic decoder and aid in iterative
soft decision decoding. Receiver for each description consist of SISO decoder for
variable length arithmetic codes, which exchange extrinsic information with SISO
RSCC decoder with the mediation of EREC-ST algorithm. Finally, available wavelet
coefficients are combined to obtain the reconstructed image.
AWGN channel. For 3-diagonal case, this gain is 5.85 dB at Eb /N0 of 3 dB.
Over AWGN channel, proposed iterative MDC decoder (2-diagonal case) achieves
maximum PSNR at Eb /N0 of 3.5 dB, which is 1.5 dB less compared to the ref-
erence scheme. In 3-diagonal index assignment case, proposed iterative MDC
decoder achieves maximum PSNR at Eb /N0 of 4 dB, which is again 1.5 dB
less compared to the reference scheme.
In Chapter 6, soft iterative network-channel decoder has been proposed for two dif-
ferent scenarios. First scheme [174] consists of channel coding performed by RS codes
at the source nodes and punctured RSCC acting as network code at the common
relay. At the decoder, received signal is considered output of a 2-D product code,
though encoding is done by component codes in geographically isolated locations.
Iterative soft decision decoding is performed at the receiver by exchanging extrin-
sic information between SISO decoders for RS codes. When compared to scheme
proposed by Ben Slimane et al. [173], following conclusions are drawn.
For the simulation setup given in Section 6.2.5, proposed scheme with (15,11)
RS code as channel code and (23, 35)8 RSCC as network code, a gain of 1.5
dB in Eb /N0 is obtained at bit error rate of 103 , as compared to (63, 39, 9)
BCH code based network coding in [173].
It is also observed that because of coupling between the two users at relay
node, SNR of one link affects the BER performance of other link also. When
one of users has bad link, both users will experience bad link performance
regardless of the other users link quality. Thus, best performance is obtained
with the proposed scheme when SNR of both the links are almost equal.
In the second part of this research, ideas developed in Section 6.2 has been
extended to scheme which combines the benefits of space-time codes and network
coding in MARC setup [175]. Similar to earlier scheme, the proposed scheme uses
punctured RSCC as network code at the relay and RS code as channel code. In addi-
tion, both the user nodes cooperate using distributed space-time code. Performance
is compared to baseline scheme developed out of work in [162].
In the simulation setup with (31, 25) RS code and (7, 5)8 RSCC, iterative
decoding gain is obtained for Eb /N0 6 dB. With (15, 7) RS code as channel
code, iteration gain starts for Eb /N0 5 dB.
With both proposed and baseline scheme are composed of (31, 25) RS code
and (7, 5)8 RSCC, proposed scheme benefits from iterative decoding of network
and channel codes, where there is an improvement of 5.25 dB in Eb /N0 for
achieving BER of 104 .
Thus, in this dissertation, it has been shown that iterative decoding principles ap-
plied to multiple description coding schemes result in improvement of reconstructed
image PSNR. It is also demonstrated through simulation that iterative soft decision
decoding techniques applied to network-channel coding schemes result in improve-
ment of BER performance. The proposed schemes has been compared to existing
165
schemes in literature and results obtained validate the superior performance of the
proposed schemes.
******
References
[3] D. J. Costello and G. D. Forney, Channel coding: The road to channel ca-
pacity, Proceedings of the IEEE, vol. 95, no. 6, pp. 11501177, June 2007.
[4] C. Berrou and A. Glavieux, Near optimum error correcting coding and de-
coding: turbo-codes, IEEE Transactions on Communications, vol. 44, no. 10,
pp. 12611271, Oct 1996.
R-1
R-2
[9] W. Ryan and S. Lin, Channel codes: classical and modern. Cambridge Uni-
versity Press, 2009.
[10] N. Wiberg, Codes and Decoding on General Graphs, ser. Linkoping studies in
science and technology: Dissertations. Linkoping University, 1996.
[22] S. Majumder and S. Verma, Joint source-channel decoding of ira code for
hidden markov source, in 1st International Conference on Recent Advances
in Information Technology (RAIT). IEEE, 2012, pp. 220223.
[27] Y. Chen, S. Kishore, and J. Li, Wireless diversity through network coding,
in IEEE Wireless Communications and Networking Conference, 2006. WCNC
2006., vol. 3, April 2006, pp. 16811686.
[28] C. Hausl and P. Dupraz, Joint network-channel coding for the multiple-access
relay channel, in 2006 3rd Annual IEEE Communications Society on Sensor
and Ad Hoc Communications and Networks, vol. 3, Sept 2006, pp. 817822.
[32] H. Holma and A. Toskala, HSDPA/HSUPA for UMTS: high speed radio access
for mobile communications. John Wiley & Sons, 2007.
[33] E. D. et al., 3G evolution: HSPA and LTE for mobile broadband. Academic
press, 2010.
[36] D. Divsalar and F. Pollara, Hybrid concatenated codes and iterative decod-
ing, Telecommunications and Data Acquisition Progress Report, vol. 130, p. 1,
1997.
[48] B. Sivasubramanian and H. Leib, Fixed-rate raptor codes over rician fading
channels, IEEE Transactions on Vehicular Technology, vol. 57, no. 6, pp.
39053911, Nov 2008.
[49] S. Majumder and S. Verma, Joint channel estimation and decoding of raptor
code on fading channel, Indian Journal of Computer Science and Engineer-
ing, vol. 4, no. 2, pp. 168173, 2013.
[50] P. Sherwood and K. Zeger, Progressive image coding for noisy channels,
IEEE Signal Processing Letters, vol. 4, no. 7, pp. 189191, July 1997.
[52] C. Lan, T. Chu, K. R. Narayanan, and Z. Xiong, Scalable image and video
transmission using irregular repeat-accumulate codes with fast algorithm for
optimal unequal error protection, IEEE Transactions on Communications,
vol. 52, no. 7, pp. 10921101, July 2004.
[55] C. Zhong and J. P. Havlicek, LDPC codes for robust transmission of im-
ages over wireless channels, in Thirty-Fifth Asilomar Conference on Signals,
Systems and Computers, vol. 1, Nov 2001, pp. 797800.
[64] D. Song, L. Cao, and C. W. Chen, Robust multiple description image coding
over wireless networks based on wavelet tree coding, error resilient entropy
coding, and error concealment, Journal of Visual Communication and Image
Representation, vol. 19, no. 5, pp. 311319, 2008.
[65] S. Dumitrescu and Y. Wan, Bit-error resilient index assignment for multi-
ple description scalar quantizers, IEEE Transactions on Information Theory,
vol. 61, no. 5, pp. 27482763, May 2015.
[71] A. Said and W. A. Pearlman, A new, fast, and efficient image codec based
on set partitioning in hierarchical trees, IEEE Transactions on Circuits and
Systems for Video Technology, vol. 6, no. 3, pp. 243250, 1996.
[74] R. Puri and K. Ramchandran, Multiple description source coding using for-
ward error correction codes, in Conference Record of the Thirty-Third Asilo-
mar Conference on Signals, Systems, and Computers, vol. 1, Oct 1999, pp.
342346.
[81] A. Bais, T. Dey, and N. Sarshar, Unequal channel protection of multiple de-
scription codes for wireless broadcast applications, Signal Processing: Image
Communication, vol. 27, no. 6, pp. 625636, 2012.
[97] B. Hochwald and K. Zeger, Tradeoff between source and channel coding,
IEEE Transactions on Information Theory, vol. 43, no. 5, pp. 14121424,
1997.
[99] M. Fresia and G. Caire, Combined error protection and compression with
turbo codes for image transmission using a JPEG 2000-like architecture, in
IEEE International Conference on Image Processing, Oct 2006.
[102] J. Hagenauer and N. Gortz, The turbo principle in joint source-channel cod-
ing, in Proceedings of IEEE Information Theory Workshop. IEEE, 2003,
pp. 275278.
[103] J. Hagenauer and R. Bauer, The turbo principle in joint source channel de-
coding of variable length codes, in Proceedings of IEEE Information Theory
Workshop. IEEE, 2001, pp. 3335.
R-12
[107] K. Lakovic and J. Villasenor, On reversible variable length codes with turbo
codes, and iterative source-channel decoding, in Proceedings of IEEE Inter-
national Symposium on Information Theory, 2002, pp. 170.
[109] K. Lakovic, T. Tian, and J. Villasenor, Iterative decoder design for joint
source-channel LDPC coding, in International Conference on Computer as a
Tool, vol. 1, Nov 2005, pp. 486489.
[111] D. Chase, Class of algorithms for decoding block codes with channel mea-
surement information, IEEE Transactions on Information Theory, vol. 18,
no. 1, pp. 170182, 1972.
[118] , Joint turbo decoding and estimation of hidden markov sources, IEEE
Journal on Selected Areas in Communications, vol. 19, no. 9, pp. 16711679,
Sep 2001.
[119] G.-C. Zhu and F. Alajaji, Joint source-channel turbo coding for binary
markov sources, IEEE Transactions on Wireless Communications, vol. 5,
no. 5, pp. 10651075, May 2006.
[120] L. Yin, J. Lu, and Y. Wu, LDPC-based joint source-channel coding scheme
for multimedia communications, in 8th International Conference on Commu-
nication Systems, vol. 1, Nov 2002, pp. 337341.
R-14
[122] N. Sun, J. Wu, and G. Zhou, Distributed joint source-channel code for spatial-
temporally correlated markov sources, in IEEE International Conference on
Communications, June 2013, pp. 49054910.
[124] H. Kim, D. Har, Z. H. Mao, M. Sun, and H. N. Lee, Efficient joint source-
channel decoding of multi-state markov sequences, IET Communications,
vol. 6, no. 9, pp. 10381044, June 2012.
[132] T. M. Cover and A. E. Gamal, Capacity theorems for the relay channel,
IEEE Transactions on Information Theory, vol. 25, no. 5, pp. 572584, 1979.
[134] , User cooperation diversity. part II. implementation aspects and perfor-
mance analysis, IEEE Transactions on Communications, vol. 51, no. 11, pp.
19391948, 2003.
[138] P. Gupta and P. R. Kumar, The capacity of wireless networks, IEEE Trans-
actions on Information Theory, vol. 46, no. 2, pp. 388404, Mar 2000.
[139] F. Xue and P. R. Kumar, Scaling laws for ad hoc wireless networks: an infor-
mation theoretic approach. Now Publishers Inc, 2006.
[140] L.-L. Xie and P. R. Kumar, A network information theory for wireless com-
munication: scaling laws and optimal operation, IEEE Transactions on In-
formation Theory, vol. 50, no. 5, pp. 748767, May 2004.
R-16
[142] T. E. Hunter and A. Nosratinia, Coded cooperation under slow fading, fast
fading, and power control, in Conference Record of the Thirty-Sixth Asilomar
Conference on Signals, Systems and Computers, vol. 1, Nov 2002, pp. 118122.
[144] M. C. Valenti and B. Zhao, Distributed turbo codes: towards the capacity
of the relay channel, in IEEE 58th Vehicular Technology Conference, vol. 1,
Oct 2003, pp. 322326.
[145] Z. Zhang and T. M. Duman, Capacity approaching turbo coding for half
duplex relaying, in Proceedings of International Symposium on Information
Theory, Sept 2005, pp. 18881892.
[148] P. Razaghi and W. Yu, Bilayer low-density parity-check codes for decode-
and-forward in relay channels, IEEE Transactions on Information Theory,
vol. 53, no. 10, pp. 37233739, Oct 2007.
[150] J. P. Cances and V. Meghdadi, Optimized low density parity check codes
designs for half duplex relay channels, IEEE Transactions on Wireless Com-
munications, vol. 8, no. 7, pp. 33903395, July 2009.
[153] Y. Chen, S. Kishore, and J. Li, Wireless diversity through network coding,
in IEEE Wireless Communications and Networking Conference, vol. 3. IEEE,
2006, pp. 16811686.
[166] S. Wei, J. Li, W. Chen, L. Zheng, and H. Su, Design of generalized analog
network coding for a multiple-access relay channel, IEEE Transactions on
Communications, vol. 63, no. 1, pp. 170185, Jan 2015.
[167] Y. Li, J. Li, Z. Lin, Y. Li, and B. Vucetic, Compressive soft forwarding in
network-coded multiple-access relay channels, IEEE Transactions on Vehic-
ular Technology, vol. 64, no. 5, pp. 21382144, May 2015.
[169] , Iterative soft decoding of multiple description image over wireless chan-
nel, Journal of Computer and Communications, vol. 2014, 2014.
[170] , Iterative decoding of FEC based multiple description codes for image
transmission over wireless channel, in Proceedings of IEEE 21st National
Conference on Communications (NCC), Feb 2015.
[171] R. Perkert, M. Kaindl, and T. Hindelang, Iterative source and channel decod-
ing for GSM, in Proceedings of IEEE International Conference on Acoustics,
Speech, and Signal Processing, vol. 4, 2001, pp. 26492652.
[177] B. Liu, Y. Li, B. Rong, L. Gui, and Y. Wu, LDPC-RS product codes for
digital terrestrial broadcasting transmission system, IEEE Transactions on
Broadcasting, vol. 60, no. 1, pp. 3849, March 2014.
[180] V. Tarokh, N. Seshadri, and A. R. Calderbank, Space-time codes for high data
rate wireless communication: performance criterion and code construction,
IEEE Transactions on Information Theory, vol. 44, no. 2, pp. 744765, Mar
1998.
[183] S. Majumder and S. Verma, Iterative decoding of FEC based multiple de-
scription codes over fading channel and opportunistic relay networks, Inter-
national Journal of Applied Engineering Research, vol. 10, no. 19, pp. 39 917
39 924, 2015.
[194] S. Lin and D. J. Costello, Error control coding. Pearson Education India,
2004.
[195] S. Lipschutz and M. Lipson, Schaums outline of theory and problems of linear
algebra. Erlangga, 2001.
[199] J. Song and K. J. R. Liu, Robust progressive image transmission over ofdm
systems using space-time block code, IEEE Transactions on Multimedia,
vol. 4, no. 3, pp. 394406, Sep 2002.
[200] X. Wu, C. Zhao, and X. You, Parallel weighted bit-flipping decoding, IEEE
Communications Letters, vol. 11, no. 8, pp. 671673, August 2007.
[203] C. Schlegel and A. Grant, Differential space-time turbo codes, IEEE Trans-
actions on Information Theory, vol. 49, no. 9, pp. 22982306, Sept 2003.
[208] K. Anwar and T. Matsumoto, Very simple bicm-id using repetition code and
extended mapping with doped accumulator, Wireless Personal Communica-
tion, no. 67, pp. 573584, 2012.
[209] C. D. Creusere, A new method of robust image compression based on the em-
bedded zerotree wavelet algorithm, IEEE Transactions on Image Processing,
vol. 6, no. 10, pp. 14361442, Oct 1997.
[210] B. Zhao and M. C. Valenti, Distributed turbo coded diversity for relay chan-
nel, Electronics Letters, vol. 39, no. 10, pp. 786787, May 2003.
[212] B. Du and J. Zhang, Parity-check network coding for multiple access relay
channel in wireless sensor cooperative communications, EURASIP Journal
on Wireless Communications and Networking, vol. 2010, p. 6, 2010.
R-24
Experience
Educational Qualifications
S. No Institute Degree Year
1. North Eastern Regional Institute B. Tech 2004
of Science & Technology (Electronics and Communication)
2. National Institute M.Tech (Electronics 2006
of Technology Calicut Design & Technology)
A-1
Publications and Outcomes
J1. [188] Saikat Majumder, Shrish Verma, Iterative Decoding of LDPC-RS Coded
FEC Based Multiple Description Coded Images . Communicated
J2. [189] Saikat Majumder, Shrish Verma, Three Stage Iterative Decoding of Dif-
ferential Space-time Coded FEC-based Multiple Descriptions . Communicated
J3. [172] Saikat Majumder, Shrish Verma, Iterative Channel Decoding of FEC-
Based Multiple Descriptions using LDPC-RS Product Codes . International
Journal of Applied Engineering Research, vol. 11, no. 9, 2016.
J4. [183] Saikat Majumder, Shrish Verma, Iterative Decoding of FEC-based Mul-
tiple Description Codes over Fading Channel and Opportunistic Relay Net-
works . International Journal of Applied Engineering Research, vol. 10, no.
19, 2015.
J5. [174] Saikat Majumder, Shrish Verma, Improved Iterative Decoding of Network-
Channel Codes for Multiple-Access Relay Channel . International Scholarly
Research Notices, vol. 2015, 2015.
J7. [169] Saikat Majumder, Shrish Verma, Iterative Soft Decoding of Multiple
Description Image over Iterative Soft Decoding of Multiple Description Image
A-2
A-3
J8. [168] Saikat Majumder, Shrish Verma, Noise Robust Transmission of Mul-
tiple Description Image over Wireless Networks . International Journal of
Engineering Science and Technology, vol. 5, no. 11, 2013.
J9. [190] Saikat Majumder, Shrish Verma, Accumulator Aided Decoding of Low
Complexity SISO Arithmetic Codes with Image Transmission Application .
International Journal of Computer Applications, vol. 79, no. 7, 2013.
J10. [49] Saikat Majumder, Shrish Verma, Joint Channel Estimation and Decoding
of Raptor Code on Fading Channel . Indian Journal of Computer Science and
Engineering, vol. 4, no. 2, 2013.
Conference proceedings
C1. [170] Saikat Majumder, Shrish Verma, Iterative Decoding of FEC Based Mul-
tiple Description Codes for Image Transmission Over Wireless Channel . Pro-
ceedings of IEEE 21st National Conference on Communications (NCC 2015),
Feb. 27-March 1, Mumbai, 2015.
C2. [184] Saikat Majumder, Shrish Verma, Improved Decoding of Multiple De-
scription Codes over Wireless Broadcast Channels . Proceedings of Emerging
Research in Computing, Information, Communication and Applications (ER-
CICA 2014), 1-2 August, Bangalore, 2014.
C3. [22] Saikat Majumder, Shrish Verma, Joint Source-Channel Decoding of IRA
code for Hidden Markov Source . Proceedings of IEEE 1st Conference on
Recent Advances in Information Technology (RAIT 2012), Dhanbad, 2012.