IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 4, APRIL 1998 593
[15] D. Bhandari, C. A. Murthy, and S. K. Pal, “Genetic algorithm with
elitist model and its convergence,” Int. J. Pattern Recognit. Artif. Intell.,
vol. 10, pp. 731–747, 1996.
[16] S. Bandyopadhyay, C. A. Murthy, and S. K. Pal, “Pattern classification
with genetic algorithms,” Pattern Recognit. Lett., vol. 16, pp. 801–808,
1995.
[17] C. A. Murthy and N. Chowdhury, “In search of optimal clusters using
genetic algorithm,” Pattern Recognit. Lett., vol. 17, pp. 825–832, 1996.
[18] S. Daly, “The visual difference predictor: An algorithm for the assess-
ment of image fidelity,” in SPIE Conf. Human Vision, Visual Processing
and Digital Display III, San Jose, CA, 1992, pp. 2–15.
[19] C. A. Murthy and S. K. Pal, “Histogram thresholding by minimizing
gray level fuzzyness,” Inform. Sci., vol. 60, pp. 107–135, 1992.
[20] X. Ran and N. Farvardin, “A perceptually motivated three-component
image model—Part I: Description of the model,” IEEE Trans. Image
Processing, vol. 4, pp. 401–415, 1995.
[21] L. Thomas and F. Deravi, “Region-based fractal image compression
using heuristic search,” IEEE Trans. Image Processing, vol. 4, pp.
832–838, 1995.
An Error Resilient Scheme for Image
Transmission over Noisy Channels with Memory
Philippe Burlina and Fady Alajaji
Abstract— This correspondence addresses the use of a joint source-
channel coding strategy for enhancing the error resilience of images
transmitted over a binary channel with additive Markov noise. In this
scheme, inherent or residual (after source coding) image redundancy
is exploited at the receiver via a maximum a posteriori (MAP) channel
detector. This detector, which is optimal in terms of minimizing the
probability of error, also exploits the larger capacity of the channel
with memory as opposed to the interleaved (memoryless) channel. We
first consider MAP channel decoding of uncompressed two-tone and bit-
plane encoded grey-level images. Next, we propose a scheme relying on
unequal error protection and MAP detection for transmitting grey-level
images compressed using discrete cosine transform (DCT), zonal coding,
and quantization. Experimental results demonstrate that for various
overall (source and channel) operational rates, significant performance
improvements can be achieved over interleaved systems that do not
incorporate image redundancy.
Index Terms—Channels with memory, DCT coding, error resilience,
joint source/channel coding, MAP decoding, unequal error protection.
I. INTRODUCTION
We address the problem of the reliable communication of images
over bursty channels. Traditional approaches to the design of visual
communication systems over noisy channels rely on Shannon’s
Manuscript received January 24, 1996; revised May 16, 1997. This work
was supported in part by the Natural Sciences and Engineering Research
Council (NSERC) of Canada. Parts of this work were presented at the 1995
International Symposium on Information Theory and the 1996 International
Conference on Image Processing. The associate editor coordinating the
review of this manuscript and approving it for publication was Dr. Christine
Podilchuk.
P. Burlina is with the Institute for Advanced Computer Studies and the
Electrical Engineering Department, University of Maryland, College Park,
MD 20742 USA (e-mail: burlina@cfar.umd.edu).
F. Alajaji is with the Department of Mathematics and Statistics and the
Department of Electrical and Computer Engineering, Queen’s University,
Kingston, Ont. K7L 3N6, Canada.
Publisher Item Identifier S 1057-7149(98)02465-8.
source-channel coding separation principle [9], resulting in what is
known as tandem source-channel coding schemes. The optimality
of this design principle holds only asymptotically; i.e., when no
constraints exist on coding/decoding complexity and delay [9]. An
alternate approach lies in joint source-channel coding (JSSC): this
strategy includes techniques such as maximum a posteriori (MAP)
detection, channel optimized vector quantization, or adaptive source-
channel rate allocation. JSSC has recently received increased attention
(e.g., [5], [7], [11]), and has been shown to outperform tandem
schemes when delay and complexity are constrained. Most of the
work on joint source-channel coding of images [5], [7], [11] has
dealt with memoryless channels, disregarding the fact that real-world
communication channels—in particular, mobile radio or satellite
channels—often have memory.
In this work, we investigate the problem of MAP detection of
images transmitted over a binary Markov channel. The MAP detector
fully exploits the statistical image characteristics in order to efficiently
combat channel noise. It also exploits the larger capacity of the
channel with memory as opposed to the interleaved (memoryless)
channel. We first describe MAP detection schemes that directly utilize
the inherent image redundancy in uncompressed binary images and
bit-plane encoded grey-level images. The amount of needed overhead
information and the performance degradation when the decoder has
imperfect knowledge of the channel parameters are considered.
The MAP detection approach is then validated for systems employ-
ing image compression. The residual redundancy of quantized low-
frequency discrete cosine transform (DCT) coefficients is exploited
via unequal error protection (UEP) and MAP decoding. Experimental
results show that the proposed schemes exhibit very good perfor-
mance, in spite of their low complexity (which primarily resides
in the MAP decoder). Specifically, significant gains over systems
not exploiting image redundancy can be achieved, at relatively low
overall transmission rates.
II. C
HANNEL MODEL
Consider a binary channel with memory described by
Y
i
=
X
i
8
Z
i
,
for
i
=
1
;
2
;
111
where
X
i
;Z
i
and
Y
i
represent, respectively, the
input, noise and output of the channel. The input and noise sequences
are assumed to be independent from each other. The noise process
f
Z
i
g
is a stationary ergodic Markov process described in [2], with
channel bit error rate (BER) denoted by
, where
2
[0
;
1
=
2)
,
and correlation parameter denoted by
0
(the noise correlation
coefficient is given by
1+
). When
=0
, the channel reduces to the
memoryless binary symmetric channel (BSC). The channel transition
and marginal probabilities
Q
(
z
n
j
z
n
0
1
) Pr
f
Z
n
=
z
n
j
Z
n
0
1
=
z
n
0
1
g
and
Q
(
z
n
) Pr
f
Z
n
=
z
n
g
, are given by
Q
(0
j
0)
Q
(1
j
0)
Q
(0
j
1)
Q
(1
j
1)
=
1
1+
1
0
+
1
0
+
and
Q
(1) =
=1
0
Q
(0)
. Note that this Markov model is general;
it can represent any irreducible first-order two-state Markov chain.
The channel capacity is given [2] by
C
=1
0
H
(
Z
2
j
Z
1
)=1
0
(1
0
)
h
b
1+
0
h
b
1
0
1+
where
h
b
(
1
)
is the binary entropy function. The capacity is monoton-
ically increasing with
(for fixed
) and monotonically decreasing
with
(for fixed
). Note that for fixed
,as
!1
;C
!
1
:
1057–7149/98$10.00 1998 IEEE
594 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 4, APRIL 1998
(a) (b) (c)
(d) (e)
Fig. 1. MAP detection of two-tone Lena over the Markov channel with
=
0
:
1
. (a) Binary Lena. (b) Lena received,
=0
:
(c) MAP detected using
second-order model. (d) Received uncoded Lena,
=10
:
(e) Decoded Lena: adaptive scheme.
III. UNCOMPRESSED IMAGES
A. Image Models, MAP Detection, and Image Redundancy
Consider a two-tone image
U
=
[
U
i;j
]
of height
J
and width
K
,
where
U
i;j
=0
or
1
;i
=1
;
111
;J; j
=1
;
111
;K
. We assume that
the image satisfies a causal second-order Markov property such that
any pixel at location
(
i; j
)
depends on the pixels at locations
(
i
0
1
;j
)
and
(
i; j
0
1)
. When the image is explored lexicographically, it can
be represented as a second-order Markov process
1
f
X
n
g
where
Pr
f
X
n
=
x
n
j
X
n
0
1
=
x
n
0
1
;
111
;X
1
=
x
1
g
=Pr
f
X
n
=
x
n
j
X
n
0
1
=
x
n
0
1
;X
n
0
K
=
x
n
0
K
g
for
n>K
. Note that this model is completely specified by four
transitional distributions. We also consider the following special cases
of
f
X
n
g
: the first-order Markov chain and nonuniform independent
and identically distributed (i.i.d.) models [4].
Consider the problem of transmitting the binary second-order
Markov source
f
X
n
g
over the Markov channel. The optimal detec-
tion technique that minimizes the sequence probability of decoding
error is the sequence MAP method [3]. More specifically, if
Y
n
=
y
n
=(
y
1
;y
2
;
111
;y
n
)
denotes the received binary sequence at the
output of the channel, the MAP detector “guesses” the transmitted
1
General Markov random field (MRF) models [8] are not used here, since
MAP estimation for these models would require computationally intensive
algorithms such as simulated annealing. We therefore restrict ourselves to
causal models that are easily implemented via sequential decoding algorithms.
sequence
^
x
n
according to
^
x
n
= arg max
x
2f
0
;
1
g
Pr
f
X
n
=
x
n
j
Y
n
=
y
n
g
:
(1)
It can be shown [4] that (1) is equivalent to
^
x
n
= arg max
x
2f
0
;
1
g
log(
Q
(
x
1
8
y
1
)
P
(
x
1
))
+
K
k
=2
log(
Q
(
y
k
8
x
k
j
y
k
0
1
8
x
k
0
1
)
P
(
x
k
j
x
k
0
1
))
+
n
k
=
K
+1
log(
Q
(
y
k
8
x
k
j
y
k
0
1
8
x
k
0
1
)
2
P
(
x
k
j
x
x
0
1
;x
k
0
K
))
:
(2)
The sequence MAP detector described in (2) can be implemented
using the Viterbi algorithm. Here,
x
k
denotes the state at time
k
;
the trellis will hence have two states, with two branches leaving and
entering each state. For a branch leaving state
x
k
0
1
at time
k
0
1
and entering state
x
k
at time
k
, the path metric is
0
log(
Q
(
y
k
8
x
k
j
y
k
0
1
8
x
k
0
1
)
P
(
x
k
j
x
x
0
1
))
;
for
k
K;
and
0
log(
Q
(
y
k
8
x
k
j
y
k
0
1
8
x
k
0
1
)
P
(
x
k
j
x
x
0
1
;x
k
0
K
))
;
for
k>K:
The surviving path for each state is the path with the smallest cumu-
lative metric up to that state. The sequence MAP decoder observes
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 4, APRIL 1998 595
the entire received sequence
y
n
in order to estimate
x
1
;x
2
;
111
;x
n
.
In this scheme, channel protection is achieved by utilizing the
natural source redundancy as well as the channel noise correlation.
The total redundancy contained in the source
f
X
n
g
is
T
=
1
0
H
1
(
X
)
, where
H
1
(
X
)
is the source entropy rate. This total
redundancy can be written as
T
=
D
+
M
[3], where
D
1
0
H
(
X
1
)
denotes the redundancy due to the nonuniformity of the
marginal distribution (
Pr
f
X
1
=
x
g
), and
M
H
(
X
1
)
0
H
1
(
X
)
,
denotes the redundancy due to the memory of the process. The type
and amount of redundancy exhibited by an image is important since it
dictates the behavior of the MAP detector. If
M
D
, the process
tends to behave like a symmetric Markov source. This results in
a mismatch situation (cf. [3, Section V]) that prevents the decoder
from fully exploiting the channel noise correlation (when the channel
capacity increases, the performance of the MAP detector deteriorates.)
If the redundancy due to the nonuniformity of a process is high
relative to its redundancy in the form of memory
(
D
M
)
,
then the process behaves like a nonuniform i.i.d. source and no
such mismatch occurs [3, Sec. IV]. Images and facsimile documents
exhibit very different types and degrees of redundancy.
2
Furthermore,
redundancy varies within images themselves since images are hardly
stationary sources. This observation suggests the use of an adaptive
scheme, as will be proposed next.
B. Two-Tone Image Detection
We start by modeling the two-tone images according to the second-
order causal Markov chain. Image lines are each represented as a
Markov chain with transitional probabilities computed empirically,
and transmitted uncompressed in a lexicographic fashion over the
Markov channel. At the receiver, the sequence MAP decoder is
implemented according to (2). While the 2-D Markov model is
appealing, since it closely captures the spatial dependency specific to
image sources, simulation results suggest that the use of this model
often results in a mismatch between the source and the channel
[4]. This leads us to conclude that when images are modeled by
a second-order Markov chain and sent over the binary Markov
channel, the best performance is obtained when
=0
; i.e., when
the channel is fully interleaved and transformed into a memoryless
channel (BSC). Fig. 1(a)–(c) show the binary Lena image transmitted
over the interleaved channel with BER
=0
:
1
. The resulting average
decoding bit error probability is 0.039.
We next consider MAP decoding when the image is modeled as
a first-order Markov chain. Since images are nonstationary, image
lines can be classified in two ways: (a) Lines for which neither
M
nor
D
are dominant, in which case no mismatch occurs. (b) Lines
having
M
D
, which are likely to result in mismatch. We hence
employ an adaptive encoding system on the image lines that takes
into consideration the line redundancy. Each image line, modeled as
a first-order Markov chain, is processed as follows: The empirical
distributions for the line are computed. If
M
<T
3
D
, for some
threshold
T
, we transmit the image line over the channel and MAP
decode it using the line statistics and first-order Markov assumptions.
Otherwise, if
M
T
3
D
, we first convert the redundancy in
the symmetric Markov source
f
X
n
g
from the form of memory
into redundancy in the form of nonuniform distribution via the
transformation, as follows [4]:
3
V
n
=
X
n
8
X
n
0
1
;n
=1
;
2
;
111
;K
.
We then transmit
V
K
directly over the Markov channel, and MAP
decode it as
^
V
K
using i.i.d. source assumptions. The decoded
binary image stream is reconstructed using
^
X
k
=
^
V
k
8
^
X
k
0
1
;
2
Computational studies that quantify natural redundancy inherent in two-
tone images are reported in [4]
3
This is essentially equivalent to differential encoding for binary sources.
TABLE I
P
ERCENTAGE OF OVERHEAD FOR BINARY LENA.
K
=
512;
R
=2
.
TABLE II
B
INARY LENA:ROBUSTNESS RESULTS FOR ADAPTIVE MAP DECODING SCHEME
IN
PSNR (dB);
T
=10
;
d
=
DESIGN BER;
a
=
ACTUAL BER;
d
=
DESIGN CORRELATION PARAMETER;
a
=
ACTUAL CORRELATION
PARAMETER. (a) ROBUSTNESS WITH BER
(
d
=
a
=10)
.
(b) R
OBUSTNESS WITH CORRELATION PARAMETER
(
d
=
a
=0
:
1)
(a)
(b)
k
=1
;
2
;
111
;K
with
^
X
1
=
^
V
1
. To prevent error propagation,
packetization is used by grouping source samples into blocks. An
example of adaptive MAP decoding
(
T
= 10)
of Lena over a very
noisy channel with high noise correlation
(
=0
:
1
and
= 10)
is shown in Fig. 1(d) (received as if it were not protected) and (e)
(MAP decoded). A 4.68 dB peak signal-to-noise ratio (PSNR) gain is
achieved by the adaptive MAP decoder over the case when no MAP
decoding is done. Detailed performance evaluation of this scheme for
various images is reported in [4].
C. Overhead Information
As in all joint source-channel coding schemes, it is assumed that
the image statistics are available at the decoder. This can be achieved
by transmitting them along with the image using a forward error-
correcting code.
4
We assume that a rate
1
=R
convolutional encoder
is used to protect the source statistics. If the channel is very noisy,
we might need to use a more powerful convolutional code. This
can be achieved by increasing the number of states of the code or
increasing
R
.If
l
denotes the number of accuracy digits for each
source parameter, then the percentage of overhead information is
equal to
% Overhead
=
mR
d
log
2
(10
l
0
1)
e
K
where
K
is the image width and
m
is the number of source statistics
per line (
m
=4
for the second-order Markov model,
m
=2
for
the first-order model, and
m
=1
for iid model). The amount of
overhead needed for the Lena two-tone image is presented in Table I
for
R
=2
and
l
=1
;
2
.
4
Note that we can avoid transmitting overhead information about the source
statistics by using training images to estimate the statistics of the source.
This approach is justifiable in applications where the images belong to a
particular class—e.g., in the transmission of medical magnetic resonance
images (MRI’s).
596 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 4, APRIL 1998
(a) (b)
(c)
Fig. 2. Transmission of grey Lena using MAP detection of bit-plane encoded images;
=
0
:
1
;
=10
. (a) Original Lena. (b) Received uncoded;
PSNR
=
14.45 dB. (c) Decoded Lena; PSNR
=
19.53 dB.
D. Robustness Under Imperfectly Known Channel Statistics
Until now we have assumed that the channel statistics
(
and
)
were known a priori at the receiver. We investigate here the
robustness of the MAP decoding system when these parameters are
not known perfectly. This may occur due to inadequate estimation of
the channel parameters, particularly when the channel is time-varying
(e.g., mobile radio channels). Simulation results using the adaptive
MAP decoding scheme for the transmission of Lena are displayed in
Table II. In Table II(a), we present PSNR results when the receiver
misestimates the BER
with the correlation parameter
=10
.In
Table II(b), we provide PSNR results when the receiver misestimates
the correlation parameter
with the channel BER
=0
:
1
. We can
conclude that the MAP scheme is not very sensitive to errors in
estimating
or
, provided that we do not design
or
to be zero
when the actual parameter is nonzero.
E. Bit-Plane Encoded Grey-Level Images
For illustrative purposes, we herein consider the application of
the MAP decoding method to bit-plane encoded images. In bit-plane
coding, each plane is traditionally compressed using binary image
coding techniques [6]. This method is very sensitive to channel errors
and typically yields low compression ratios leaving little room for
TABLE III
MAP-UNC V
ERSUS UNC: AVERAGE PSNR (IN dB) OF DECODED LENA OVER
MARKOV CHANNEL WITH BER
AND
CORRELATION PARAMETER
.RESULTS
AVERAGED OVER 30 EXPERIMENTS.
R
IS THE OVERALL RAT E I N B/PIXEL
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 4, APRIL 1998 597
TABLE IV
MAP-UEP I V
ERSUS ML-IL-UEP I: AVERAGE PSNR (IN dB) OF DECODED
LENA OVER MARKOV CHANNEL WITH BER
AND CORRELATION PARAMETER
.
R
ESULTS AVERAGED OVER 30 EXPERIMENTS.
R
IS THE OVERALL RAT E I N B/PIXEL
protection against channel noise. Consider instead the problem of
directly sending the uncompressed bit-planes modeled as Markov
sources over the Markov channel. As in the case of two-tone images,
we use an adaptive MAP detection scheme taking into account the
source and the channel statistics applied on each bit-plane image
explored in a lexicographic fashion. Experimental results are shown
in Fig. 2 for the Lena grey-level image. Significant improvements
over the received images are achieved. For
=
10
and
=0
:
1
,
gains in excess of 5 dB are achieved.
IV. C
OMPRESSED IMAGES
MAP decoding of uncompressed images relies on the significant
intrinsic source redundancy to help combat channel noise. Since
source coding schemes are not ideal, they always leave some residual
redundancy in their output bitstream that can similarly be exploited
at the receiver. A challenging issue lies in the use of the limited
redundancy residing in compressed images for channel protection.
A. Image Compression Scheme
Standard visual compression methods such as Joint Photographers
Expert Group (JPEG) and Motion Pictures Expert Group (MPEG)
are fragile to channel errors. Errors corrupting the compressed data
contribute unequally to the final distortion of the reconstructed image
or video stream. This observation justifies the use of unequal error
protection. We propose to improve the error resilience of compressed
images by designing several schemes that combine UEP and MAP
detection. Our objective is to characterize the effectiveness of these
methods for various levels of image compression.
Consider the case of JPEG encoded images, or that of MPEG1/2
or H.261/3 encoding of intraframes. These schemes incorporate
DCT coding, quantization and entropy coding. Clearly, the most
fragile module lies in the variable-length coding (either Huffman
or arithmetic), for which the occurrence of an error produces cat-
astrophic error propagation and total loss of the packet until the next
TABLE V
MAP-UEP II
VERSUS ML-IL-UEP II: AVERAGE PSNR (IN dB) OF DECODED
LENA OVER MARKOV CHANNEL WITH BER
AND CORRELATION PARAMETER
.
R
ESULTS AVERAGED OVER 30 EXPERIMENTS.
R
IS THE OVERALL RAT E IN B/PIXEL
synchronization occurs. Error resilience in this case consist in the
reliable reception of synchronization messages or the packetization
of VL codes.
5
Since the synchronization issue is outside the scope of this work,
we consider instead a compression scheme similar in spirit to the
above cited standards with the exclusion of entropy coding. More
specifically, our image compression scheme is as follows: The image
is first subdivided into 8
2
8 blocks, and for each of these blocks the
DCT is computed. The resulting 64 DCT coefficients are uniformly
quantized using one of the quantization matrices proposed in [10]
derived from psychovisual thresholds. The coefficients are then
ordered in a zig-zag fashion. While the basic JPEG scheme would
Huffman encode the resulting stream on the basis of the coefficients’
amplitude and leading run-lengths of zeros, we proceed with zonal
coding and conversion to a binary bitstream. For zonal coding, we
use the first 15 zig-zag scanned coefficients. The retained quantized
coefficients are then converted to binary using a folded binary code
(FBC) representation. The bit rates used for converting each quantized
coefficient are those proposed for zonal coding in [10].
B. Channel Coding Schemes
Error resilience is provided by combining UEP and MAP
detection. Because of the high-energy compaction property of
the DCT for highly correlated sources [6], most of the signal
information is concentrated in the lower spatial frequencies. The DC
coefficient is the most important DCT coefficient since it measures
the average value of each block. An error in the DC coefficient
typically results in blocking artifacts. These artifacts are often
resolved through additional channel protection or postprocessing
techniques that employ edge-preserving smoothing operators on
the decoded image. However, traditional channel protection or
error-concealment operations disregard the source characteristics.
We propose instead to use MAP detection of channel encoded DC
5
This issue is given much attention in current standardization efforts of
MPEG4.
