already available in the decoders codebook) or the semantic

parameters to animate the mouth at the decoder. The semantic

mouth animation is chosen as long as the tracked mouth shape in
the original scene has some resemblance to the textured mouth
shape. Whenever the tracker encounters a dramatic change in the
mouth shape (typically after 10-15 frames, when the NCC coeficient falls below a given threshold), the new mouth shape or the
index of an old shape is transmitted, and then the algorithm
switches back to the semantic animation until another shape
change is found. So basically small amounts of motion (as will be
the case with silent-like periods in videoconferencing sessions) are
represented with the semantic approach only, where any inbetween large motions or shape changes (for example, due to talking or smiling) are compensated for by updating the texture of the
mouth. It is worth mentioning that the semantic tracking of
mouth vertices is continued by the encoder even when the system
switches to the codebook method because positions of these vertices are needed to maintain accurate animation of the mouth in the
frames following the one with the updated mouth texture.

curves in both Figures show that the PSNR values drop gradually
in frames of un-updated mouth texture (for example, between
frames 0 and 25 of Talking), and rises back again when a shape
update takes place. The graphs also show that applying the hybrid
scheme improves the PSNR by -1 dB in regions of small motion of
which the texture is not updated. This allows similar levels of
image quality to be maintained over larger periods of the entire
sequence, and as a result reduces the annoying effects associated
with fluctuations in PSNR levels encountere,d when using the
codebook only. Of course, the quality of the modelled mouth
images using the hybrid scheme is highly dependent on the accuracy of the tracking of the mouth in the first place. Further
improvement is expected to be achieved with better future tracking
algorithms. The increase in the data rate needcd for transmitting
the semantic animation parameters is affordable because, fxst,
although 20 vertices are used to describe the mouth, only the
motion vectors of seven vertices are transmitted while the remaining vectors are linearly interpolated at the decoder. Secondly,
extra reductions in the bit rate can also be achieved through intraand inter-encoding of these parameters.
Conclusions: A switched mouth animation scheme has been implemented whereby the quality of the modelled images is improved
while still enabling the low-bit rate property of model-based coding to be exploited. Observations have shown that the hybrid
scheme offers generally better subjective quality and PSNR performance compared to a stand alone codebook- or semantic-based
animation algorithm.

35:

l b 15 20 25 3b

frame

& 4b

4k

55

24 March 1999
0 IEE 1999
Electronics Letters Online No: 19990565
DOI: lO.I049/19990565
A. AI-Qayedi and A.F. Clark (Department of Electronic Systems
Engineering, University of Essex, Colchester (704 3SQ, United
Kingdom)

rn

References

Fig. 2 PSNR comparison of codebook-only and hybrid schemes on

Talking sequence

PEARSON, D.E.:

-0-codebook only

IEEE, 1995, 83, (6), pp. 892-906

-+-

and FRIESEN, w.v.: Facial action coding system

(Consulting Psychologist Press, Inc., Palo Alto, CA, 1977)
K ~ K U E RM.,
,
and CLARK, A.F.: A model-based codec suitable for
deaf communication. Proc. Int. Symp. Very-L,ow Bit-rate Video,
1994, Paper 4.2

hybrid scheme

37

IT

Developments in model-based video coding, Proc.

EKMAN, P.,

http://peipa.essex.ac.uk/ipa/pidfaces/essex-02/

Convergence of iterative decoding

S. ten Brink

frame

E
p
J
Fig. 3 PSNR comparison of codebook-only and hybrid schemes on Miss
America sequence

-e
codebook only
-+- hybrid scheme
Experimental results: The system has been implemented and tested

on a set of standard sequences and others recorded in our own

laboratory. Figs. 2 and 3 show the relative PSNRs for the
sequences Miss America and Talking [4], respectively. The PSNRs
have been calculated for the facial region after compensating for
the global motion of the head. First the mouth is animated using
purely the codebook and secondly using our hybrid method. The
results show that both schemes give similar results when there is
high movement of the lips (for example, as in the first 15 frames
of Miss America), because only the codebook mode is active and
therefore the semantic approach has no effect. The codebook

806

A novel method for visualising the convergence behaviour of

iterative decoding schemes is proposed. Each constituent decoder
is represented by a mutual information tcmsfer characteristic
which describes the flow of extrinsic information through the soft
idsoft out decoder. The exchange of extrinsic information
between constituentdecoders is plotted in an extrinsic information
transfer chart. The concepts are illustrated for an iterative
demapping and decoding scheme.
Introduction: While many studies have concentrated on providing
asymptotic bit error bounds for parallel and serially concatenated
convolutional codes (e.g. [l]) little has yet been revealed as to the
convergence behaviour of the corresponding sub-optimal iterative
turbo decoders. In this Letter a visualisation method based on
bitwise mutual information is introduced which provides a tool for
studying the convergence behaviour of iterative decoding schemes,
yielding new design rules for improving the choice of constituent
codes. For simplicity an iterative demapping and decoding scheme
[2, 31 is considered in this Letter. The demapper takes soft values
from the channel and outputs extrinsic information E, which is
passed through a bit interleaver to become the a priori input A,
for the soft in/soft out channel decoder [4]. From the decoder
extrinsic information on the coded bits E, is fed back to the

ELECTRONICS LETTERS

13th May 1999

Vol. 35

No. 10

demapper as a priori knowledge A , to reduce the bit error rate

(BER) in further iterative decoding steps. The variables A,, E,, A,
and 4 denote log-likelihood ratios (L-values).

The demapper transfer characteristics almost approach straight

lies. Thus it is sufficient to describe a mapping using two values:
the bitwise mutual information Zo = ZEI(ZAI= 0) given that no
other bit of the mapping is known ('no a priori knowledge at
demapper'), and I,, = ZEI(ZA, = 1) given that all other bits of the
mapped codeword are known ('perfect a priori knowledge') [2].
Keeping the mapping futed, different EdN, values just shift the
curve up and down. Keeping the EJN0 value fvced, different m a p
pings result in lines of different slope. For Gray mapping the bitwise mutual information remains almost constant with increasing
a priori knowledge la,,whereas the transfer characteristic for antiGray mapping has a steep slope, revealing the strong potential
performance improvements in an iterative demapping and decoding scheme.

q
c

0.2

01

"

'

"

"

"

0.2
0.4
0.6
0.8
1.o
output IE2 of decoder becomes input IA1to demapper

@@I

Fig. 1 Extrinsic information transfer characteristics of 16-QAM demapper for different EJNo (at code rate 1:2) and different mappings
-A- Gray mapping (EdN, = 3dB)
some 16-QAM mapping (EdNo = 3dB)
anti-Gray mapping (EdN, = 3dB)
--C some 16-QAM mapping (EdN, = 3dB)
-0-anti-Gray mapping, different EdN, values

-**

Demapper transfer characteristics: The demapper transfer characteristic is a function of the a priori bitwise mutual information ZA1
and the EdN, value of the AWGN-channel

With equiprobable binary input symbols X to the mapper (at

transmitter) the bitwise mutual information [5] is calculated as

0.2
0.4
0.6
0.8
1.o
mutual information lE2 at output of decoder, b%per binary symbol

m
Fig. 2 Extrinsic information transfer characteristics of soft idsoft out
decoder for rate 1:2 convolutional codes with different memory
--C rate 1:2, memory 2
--C rate 1:2, memory 4

-+-A-

rate 1:2, memory 6

rate 1:2, memory 8

-0
n

$1.01

'

'

'

'

'

'

'

'

'

For M bits per mapped codeword it is

and

respectively. The conditional probability distributionspa , and pEl,k

for A , and E l k are obtained by simulations. Different values of ZA,
and EdNo are considered in ZBI through changes in the distribu~.
to the strong correlation between I*, and ZE, the
tions P ~ , ,Owing
demapper characteristics can be plotted in an ZE,,ZAl diagram
(Fig. l), with the EdN, value as a parameter yielding a set of
curves. Different distributionsp a l were used to calculate the demapper characteristics without observing a notable change in the
shape of the curves. Moreover, the demapper output distributions
P ~ , can
, ~ be quite asymmetric and noncontinuous with sharp
edges, depending on the complex signal constellation and the
applied mapping. Hence, transfer characteristics based on the
mean values and variance of the input/output distributions, such
as SNR measures, would fail, whereas mutual information transfer
characteristics prove to be very robust, owing to the robustness of
the entropy measure IS].

ELECTRONICS LE7TERS

13th May 1999

Vol. 35

2 1
.-E

0.2
0.4
0.6
0.8
1.o
mutual information IA, at intput of demapper, bit per binary symbol

Fig. 3 Example of trajectory for iterative decoding at EJNo = 3.4dB,

16-QAM, anti-Gray mapping, rate 1:2 memory 4 code
trajectory of iterative decoding
-A- anti-Gray mapping (EdN, = 3.4dB)
--C conv. code, rate 1:2, memory 4

Decoder transfer characteristics: Corresponding to eqns. 2 - 4 the

decoder transfer characteristic on the coded bits is defined as

IE2

No. IO

= g(IA2)

(5)

807

It is only dependent on the bitwise mutual information IA2of the a

priori input to the decoder. Again, the probability distributions
p E 2 pa2
, for calculating ZE2,ZA2 are obtained by simulation.

Fig. 2 shows the transfer characteristics for convolutional codes

of rate 1.2 with memory 2, 4, 6 and 8. It is remarkable that all
curves cross at a single point: (0.5, 0.5). For arbitrary code rates R
this point t u m s out to be at (0.5, R).
Extrinsic information transfer chart: Connected through interleavers, the extrinsic output of the demapper becomes the a priori

input to the decoder ZA2 = ZE,,and the extrinsic output of the

decoder becomes the a priori input to the demapper ZA = IE2.This
exchange of extrinsic information is accounted for in the extrinsic
information transfer chart (EIT chart) by plotting the demapper
and decoder characteristics into a single diagram.
Fig. 3 shows an example of an iterative decoding trajectory: the
E d N , value of the channel has raised the demapper curve just high
enough to open a tunnel for the trajectory. This matches with the
BER(EdN,) plot of [2] where the turbo cliff is at -3.3dB. The
BER floor is determined by the intersection of demapper and
decoder curves on the very right side of the EIT chart. Note that
mutual information (decoder output on the information bits, not
shown) and BER are connected through a bound given by the
converse of the coding theorem [5]. Both demapper and decoder
characteristics are obtained separately, and not in conjunction
with a particular iterative decoding scheme; for verification, the
trajectory of Fig. 3 has been evaluated by means of simulation.
Clearly, the trajectory closely matches the demapper and decoder
Characteristics. For short interleavers the trajectory would tend to
diverge from the characteristics after a few iterations. From the
EIT chart some design guidelines become apparent, for example,
the larger the code memory, or the steeper the slope of the demapper curve, the later the turbo cliff in the BER(EdN,) chart, but the
lower the BER floor.

Evolutionary pro ramming with

accumulated evo ution information

Doo-Hyun Choi
In natural evolution, the individuals of species accumulate
successive slight variations in their genes and the accumulated
evolution information is d e n t e d by thelr offspring. A new
evolutionary programming algorithm adopting this concept is
presented. Similar to natural evolution, the algorithm uses the
accumulated evolution information inherited from its parent. The
information is obtained through many generations of the
evolution and speeds up the convergence of the evolutionary
prograWnng. The eficiency and robustness of the proposed
algorithm has been verified through benchmark testing.
Introduction: Evolutionary algorithms mimic thle process of natural evolution. The aim of evolutionary programming (EP) is to
achieve intelligent behaviour through simulated evolution [11. EP
is suitable for combinatorial and real-valued function optimisation. The most signifcant advantage of using evolutionary searching lies in the increase in flexibility and adaptability to the task at
hand, in combination with robust performance and global search
characteristics [2].
In natural evolution, when one part varies through the law of
correlation and the variations are accumulated through natural
selection, other modifications of the most unexpected nature will
appear [3]. In this Letter, a new EP algorithm is presented in
which the evolution information has been accumulated into the
genes of individuals. In the algorithm, the accumulated evolution
information not only speeds up of the convergence but also finds a
better optimum solution.

Conclusions: It has been shown that the transfer characteristics

based on mutual information can facilitate the design of iterative
demapping schemes. The flow of extrinsic information was visualised in the EIT chart to provide insight into the turbo cliff
position and BER floor. Initial results for other concatenated
codes [l, 61 are encouraging.
Acknowledgment: This work was camed out in a joint project with

Bell Laboratories, Lucent Technologies, Pagoda Park, Swindon,

Wiltshire SN5 7YT, United Kingdom.

evaluations

Fig. 1 Evolution of averaged minimum cost for Sphere function

0 IEE 1999
Electronics Letters Online No: 19990555
DOI: 10.1049/el:19990555

22 February 1999

proposed EP
........... SAEP
_ - - _ AEP

S. ten Brink (University of Stuttgart, Room 2.333, Institute of

Telecommunications, Dep. 0408, Pfaffenwaldring 47, 70569 Stuttgart,
Germany)
E-mail: tenbrink@inue.uni-stuttgart.de

References
1

2
3
4

s., DIVSALAR, D., MONTORSI, G., and POLLARA, F.: Serial

concatenation of interleaved codes: performance analysis, design,
and iterative decoding, IEEE Trans., 1998, lT-44, (3), pp. 909-926
TEN BRINK, s., SPEIDEL, J., and YAN, R.-H.: Iterative demapping and
decoding for multilevel modulation. Proc. IEEE GLOBECOM98,
Sydney, Australia, 1998, pp. 579-584
TEN BRINK, s., SPEIDEL, J., and YAN, R.-H.: Iterative demapping for
QPSK modulation, Electron. Lett., 1998, 34, (15), pp. 1459-1460
ROBERTSON, P., VILLEBRUN, E., and HOEHER, P.: A comparison of
optimal and sub-optimal MAP decoding algorithms operating in
the log domain. Proc. ICC 95, June 1995, pp. 1009-1013
HAMMING, R.w.: Coding and information theory (Prentice-Hall,
New Jersey, 1986)
BERROU, c., GLAVIEUX, A., and THITIMAJSHIMA, P.: Near Shannon
limit error-correcting coding and decoding: Turbo-codes. Proc.
ICC93, May 1993, pp. 10641070
BENEDETTO,

808

-101

10

$ .
2M)

400
800
evaluations

800

10

Fig. 2 Evolution of averaged minimum cost for Rosenbrock function

proposed EP
........... SAEP
- - _ _ AEP

Proposed algorithm: The proposed algorithm introduces two

parameters, distance variables and accumulated evolution information variables, into the, genes of the individuals of the EP. The
individual is represented as follows:

ELECTRONICS LETERS

13th May 1999

Vol. 35

No. IO