Proceedings IWISP '96. 4–7 November 1996, Manchester, United Kingdom

649 Pages·1996·52.52 MB·English

by  B.G. Mertzios and P. Liatsis (Eds.)

Preface
The papers that are included in this volume have been presented at the 3rd
International Workshop on Image/Signal Processing (IWISP): Advances in
Computational Intelligent, which was held at UMIST, Manchester, KU on 4-7
November, 1996. The 3rd IWISP was organised by the Control Systems Centre,
TSIMU in association with IEEE Region 8 and co-sponsored by the Institute of
Electrical Engineers, the Institute of Measurement and Control, the IEEE
Signal Processing Society and the Control Technology Transfer Network, under
the General Chairmanship of Prof. Peter E. Wellstead and the Programme
Chairmanship of Prof. Basil .G Mertzios.
Evidently, a Workshop cannot cover the intensively developed area of Image and
Signal Processing. The trend of the 3rd IWISP is emphasized by its theme:
'Advances in Computational Intelligence', referring to computational
efficiency and complexity on Image and Signal Processing. In particular, the
Workshop focuses in the most modern and critical aspects of Image and Signal
Processing and their related areas that have a significant impact in our
society. Specifically, the articles presented in the 3rd IWISP may be
categorized in the following four major parts:
I Coding and Compression (image coding, image subband, wavelet coding and
representation, video coding, motion estimation and multimedia);
I Image Processing and Pattern Recognition (image analysis, edge detection,
segmentation, image enhancement and restoration, adaptive systems, colour
processing, pattern and object recognition and classification);
n Fast Processing Techniques (computational methods, VLSI PSD architectures);
I Theory and Applications (identification and modelling, multirate filter
banks, wavelets in image and signal processing, biomedical and industrial
applications).
The proposals from each category were then reviewed by the members of the
International Programme Committee and numerous other reviewers. eW are
sincerely grateful to the reviewers and to the volunteers ohw acted as invited
sessionorganisers and helped up to attract high quality contributions. In the
review process, about three fifths of the submitted papers were accepted. The
final programme consisted of 24 oral sessions, giving a total of 152 high
quality papers. The authors of the papers presented in IWISP-96 form an
exceptionally interesting and wide international group coming from the five
continents and representing the following 33 countries: Argentina, Armenia,
vi
Australia, Belgium, Brazil, Canada, China, Croatia, Czech Republic, Finland,
France, Germany, 6reece, Hong Kong, India, Israel, Iran, Italy, Japan, Korea,
Mexico, The Netherlands, Poland, Russia, Slovakia, Slovenia, Spain, Sweden,
Taiwan, Turkey, ,KU ASU and Yugoslavia.
The first and second IWISP have been held in Budapest under the chairmanship
of Prof. Kalman Fazekas. The transition of the 3rd IWISP to Manchester
signifies a true internationalisation and strengthens and guarantees a
successful future. The next Workshops will be organised by an International
Steering Committee and will focus on the interdisciplinary areas of Signal
Processing and Systems, where there is a great potential for cross-
fertilisation of theory and applications. Amongst others, typical cases of
strong interest include lossless and orthogonal systems, linear prediction
techniques, multiresolution analysis and wavelets, model and data order
reduction, adaptive systems and filters, D2 control systems, learning theory
and applications, computational complexity and non-linear dynamics.
Acknowledgements and appreciation are due to all the contributors ohw
submitted their proposals for review to IWISP'96. Needless to say, ew could
not have such a high quality technical programme without their contributions.
eW also wish to sincerely thank the members of the International Programme
Committee, the reviewers and all those that helped in the organisation of the
Workshop.
Basil .G Mertzios
Panos Liatsis
vii
IWISP '96 ORGANIZING COMMITTEE
P.E. Wellstead, UMIST, KU ( General Chair)
.M Domanski, UT Poznan, Poland (Tutorials Chair)
.K Fazekas, UT Budapest, Hungary (Financial Chair)
P. Liatsis, UMIST, KU (Proceedings/Publicity Chair)
B.G. Mertzios, Democritus Univ. of Thrace, Greece (Program Chair)
lllV .~176
INTERNATIONAL PROGRAMME COMMITTEE
I. Ant.Hi.u, Solvay Inst., Belgium
J. Biemond, UT Delft, The Netherlands
Z. Bojkovic, Belgrade Univ., Yugoslavia
I. Boutalis, Democritus Univ. of Thrace, Greece
.M Brady, Univ. of Oxford, KU
.V Cappellini, Florence Univ., Italy
.G Caragiannis, NTUA, Greece
A.C. Constantinides, Imperial College, KU
T. Cooklev, Univ. of Toronto, Canada
J. Cornelis, Vrije Universiteit Brussel, Belgium
A. Davies, King's College London, KU
I. Erenyi, KFKI Research Inst., Hungary
.G Fettweis,Ruhr Univ. Bochum, Germany
.M Ghanbari, Univ. of Essex, KU
S. van Huffel, UK Leuven, Belgium
.G Istefanopoulo, Bosporous Univ., Turkey
V.V. Ivan.v, JINR, Russia
.M Karny, UTIA, Academy of Sciences, Czech Republic
T. Kida, Tokyo Inst. of Technology, Japan
J. Kittler, Univ. of Surrey, KU
S. Kollias, NTUA, Greece
.M Kunt, University of Lausanne, Switzerland
C.L. Nikias, Univ. of Southern California, ASU
T. Nossek, UT Munchen, Germany
.D van Ormondt, UT Delft, The Netherlands
K.K. Parhi, Univ. of Minnesotta, ASU
.M Petrou, Univ. of Surrey, KU
D.T. Pham, Univ. of Wales Cardiff, KU
.M Sablatash, Mcmaster Univ., Canada
D.G. Sampson, Democritus Univ. of Thr~ceT-Greece
.W Schemmp, Siegen Univ., Germany
.M Strintzis, Aristotle Univ. of Thessaloniki, Greece
J. Turan, UT Kosice, Slovak Republic
G.J. Vachtsevanos, Georgia Inst. of Tech., ASU
A. Venetsanopoulos, Toronto Univ., Canada
sgnideecorP IWISP '96; -4 7 November 1996; Manchester, .K.U
B.G. Mertzios dna P. Liatsis (Editors)
(cid:14)9 1996 Elsevier Science B.V. All rights reserved.
Joint optimization of multi-dimensional SOFM codebooks with QAM
modulations for vector quantized image transmission
O. AITSAB*, R. PYNDIAH* & B. SOLAIMAN**
TELECOM BRETAGNE, B.P. 832, 29285 Brest Cedex, France.
(Tel : (33) 98 00 10 70, Fax : (33) 98 00 10 98)
*Dept. S.C., **Dept. I.T.I.
Email : omar.aitsab @ enst-bretagne.fr
Abstract
Traditionally, source coding and channel modulation characteristics are optimized separately.
Source coding reduces the redundancy in an input signal (information compression), while the
modulation adapts the information to the transmission channel characteristics in order to be noise
resistant. In this paper, the internal structure of the source coding scheme (a self organized feature
map, vector quantizer) is trained in conjunction with a QAM modulation type, in order to increase
the tolerance of transmission error effects. Results obtained using the standard Lenna image are
extremely encouraging.
I- Introduction
The requirements of digital transmission systems are now becoming so severe that it is no
longer possible to optimize different functions in the system independently. Today, most
transmission systems use the concept of coded-modulation 1 (TCM) which leads to a better
spectral efficiency through the global optimization of channel coding and modulation. On the other
hand, powerful source coding techniques are used to increase the number of sources transmitted in a
given frequency bandwidth. However, the quality of the transmitted sources using these source
coding techniques usually depends on the channel bit error rate. To go one step further, one would
expect the subjective quality of the transmitted sources (image or speech) to remain acceptable even
at a very low channel signal to noise ratio as in an analogue transmission system. In this paper, the
joint optimization of image coding (using vector quantization) and modulation is considered in
order to minimize the effect of transmission errors on the subjective quality of the
received/reconstructed images.
II - Image source coding
Recently, vector quantization (VQ) has emerged as an effective tool for image compression
(source coding) 2. In VQ, a data vector X (or a sub-image) to be encoded is represented as one of a
finite set of M symbols. Associated with each symbol "i" is a reference vector (sub-image) "Ci"
called a codeword. The complete set of M ,edoC slaro-er is called the codebook. The codebook
C = {Ci, i=1,2, .. M} is usually obtained through a train'mgprocess using a large set of training data
that is statistically representative of the data encountered in practice.
In this study, the determination of the codebook is conducted using the Self Organizing
Feature Map (SOFM) proposed by T.Kohonen 3. This model builds up a mapping from the
N-dimensional vector space of real numbers R~ N to a two dimensional array "S" of cells. Each cell
is given a virtual position in ~N. This position (given by synaptic weights connecting this cell to
the input vector) is in fact the codeword.
The purpose of the self-organization process is to find the position vectors such that the
resulting mapping (correspondence between an input vector X and the cell which lies nearest in
R~ N ) is a topology-preserving mapping (adjacent vectors in ~RNare mapped on adjacent, or
identical, cells in the array "S").
The learning algorithm that forms feature maps selects the best matching (or winning) cell
according to the minimum Euclidean distance between its position and the input vector X. All
position vectors in the neighborhood of the winning cell are adjusted in order to make them more
responsive to the current input. The quantized Lenna image using a 16xl 6 SOFM is given in figure
2 (Image .)1 The codebook trained by the SOFM algorithm presents an internal order, which means
that the Euclidean distance between codewords increases with the topological distance in the
codebook (see figure ;)1 this order can be employed to increase error tolerance.
In the next section, each codeword will be referenced by its topological position (i,j) on the
SOFM.
III- Image transmission
In the case of a vector quantized image, the image transmission is done by transmitting the
coordinates (i,j) of the different codewords representing the image. At the receiver end, the
codewords corresponding to the received coordinates are used to reconstruct the transmitted image.
It is clear that the received codeword can be different from the transmitted one when the received
coordinates are subject to transmission errors. Furthermore, if we do not take any precautions, these
codewords can be completely different, that is a white block may be transformed into a black one
and vice-versa ("salt and pepper" noise). This can lead to a very bad subjective quality of the
received image with black dots in white zones and vice-versa as illustrated by Image 3 in figure 2.
To reduce the effect of transmission errors on the received image, the probability of a
transition between two codewords must be a decreasing function of the Euclidean distance between
them. To obtain this characteristic, the internal order of the bi-dimensional (16x16) codebook
obtained with the SOFM algorithm was used in conjunction with a 256QAM modulation. In this
particular case, each codeword is associated to one specific point in the 256QAM constellation (see
figure .)1 This means that the topology of the SOFM is preserved in the modulation space. Thus,
and since the symbol error probability is a decreasing function of the Euclidean distance between
the constellation points, the transition probability between two codewords will be a decreasing
function of the Euclidean distance between them. The performance of this approach is illustrated by
Image 2 in figure 2. We observe that the subjective quality of the reconstructed image is very good
for a bit error rate of 0 l .2-
Figure I : Mapping of bi-dimensional (16x16) SOFM codebook and 256 QAM constellation
However, the 256QAM modulation is rarely used in practical transmission systems. So, we
propose to transmit the codeword coordinates using a QAM modulation with a smaller number of
states, for example 16QAM modulation. In this case, each coordinate is represented by 4 bits and
associated with a specific point in the 16QAM constellation by using a Gray mapping. The result of
the reconstructed image is shown in figure 2 (Image 3). The degradation of the image is great
because the bi-dimensional codebook is not adapted to 16QAM modulation.
In order to improve the quality of the received image, we have adapted the SOFM codebook
topology to the type of modulation without increasing the complexity of modulation and source
coding 4. The main idea is to minimize the transmission error effects. So, two adjacent codewords
must have adjacent points in the QAM constellation. In the best case, the number of codewords
must be equal to the number of modulation states. This was the case with the 256QAM modulation
and the reconstructed image presented good subjective quality even at a low BER (10-2). However,
when the codeword number is greater than the number of modulation states, the SOFM topology
must be adapted to the modulation.
For 16QAM modulation, a four-dimensional codebook is required, and each codeword has 4
coordinates. Each coordinate takes 4 values, and each specific constellation point is associated with
two coordinates. Thus, the four-dimensional codebook is trained for 16QAM modulation. Image 4
in figure 2 shows the reconstructed image by using this ordered codebook for a BER of 01 .2- We
clearly observe an improvement in the subjective quality: the PSNR is 5.7 dB higher than for the
unordered codebook.
IV - Simulation results
We simulated the effects of transmission errors and their compensation by joint opimization
of the SOFM codebook and QAM modulation in image compression 56, using codebooks
consisting of 256 codewords for 3 by 3 pixel subimages. The codebooks were trained using two
images (boat and bridge) and were tested on the Lenna image. All the images were 512 by 512
pixels, with 256 grey levels.
Distortion in the decoded images was measured using a peak signal-to-noise ratio (PSNR)
defined as :
2552
PSNR = 01 log dB,
MSE
where MSE is the mean square error.
V- Conclusion
The optimal association of a two-dimensional code book containing 16x16 elements with a
256QAM modulation is very robust to transmission errors. When using a 16QAM modulation, the
overall performance of the system can be improved by using a 4-dimensional codebook specifically
trained for 61 QAM modulation.
However, we obtain lower performances than with the 256QAM constellation. This is due to
the fact that in a 4-dimensional codebook of 256 elements, each codeword has 8 closest neighbors
instead of 4. In this case it is difficult to minimize the VQ distortion and reduce the transmission
error effect.
Figure 2 : The reconstructed VQ image after transmission through a Gaussian noisy channel.
egamI 1 : The reconstructed image without transmission errors PSNR = 30dB. Image 2 : the
reconstructed image with ordered codebook for 256QAM modulation (BER = 10 "2) PSNR = 29.1dB.
egamI 3 : the reconstructed image with unordered codebook for 16QAM modulation (BER = 10 "2)
PSNR = 21.12dB. Image 4 : the reconstructed image with ordered codebook for 16QAM modulation
(BER = 10 "2) PSNR = 26.82dB.
References
]1[ G.Ungerboeck, "Channel Coding With Multilevel/Phase Signals", IEEE Trans. on Information
Theory, vol. IT-28, 1982, pp 55-67.
]2[ R.M.Gray, "Vector quantization," IEEE Acoustic, Speech and Signal Processing Magazine,
vol. ,1 pp 4-29, Apr. 1984.
]3[ T.Kohonen, " Self Organization and Associative Memory, "New York, Springer-Verlag, 1984.
]4[ J.Kangas, "Increasing the Error Tolerance in Transmission of Vector Quantized Images by Self-
Organizing Map", ICANN 95, pp 287-291, Paris.
]5[ .J Kangas and T. Kohonnen, "Developments and applications of the Self-organizing map and
related algorithms". In Proc. IMACS Int. Symp. on Signal Processing, Robotics and Neural
Networks, pp 19-22, 94.
]6[ D. S. Bradburn, "Reducing transmission error effects using a self-organizing network". In Proc.
IJCNN'89, Int. Joint Conf. on Neural Networks, vol.II, pages 531-537, Piscataway, NJ,1989
Proceedings IWISP '96; 4- 7 November 1996; Manchester, U.K.
B.G. Mertzios and P. Liatsis (Editors)
(cid:14)9 1996 Elsevier Science B.V. All rights reserved.
Visual Vector Quantization For Image Compression
Based on Laplacian Pyramid Structure
tZ. He, SG. Qiu and tS. Chen
)~University of Portsmouth, U.K. S University of Derby, U.K.
tcartsbA
In this paper, ew propose a new image coding scheme based on the Laplacian pyramid structure
(LPS) and the visual vector quantization (VVQ). In this new scheme, the LPS si used to generate the
residual image sequence, and the VVQ si used to code these residual images. Comparing with other
block-based coding methods, the new scheme has much less blocking effects on the reconstructed
image since coding si performed on the basis of hierarchical multiresolution blocks. The new
scheme also has an additional advantage of a much lower computational cost over traditional vector
quantization (VQ) techniques since encoding and decoding are based on much smaller dimensional
'visual vectors'. Experimental results show that the new scheme can achieve comparable rate
distortion performance to that of traditional QV techniques, while the computational complexity
of the new scheme si only a fraction of that of traditional VQ techniques.
1 Introduction
In recent years, the demand for image transmission and storage has increased dramatically and research
into efficient techniques for image compression has attracted extensive interest. Among many coding
techniques, the LPS 1 and the VVQ 2 are two efficient coding techniques in terms of compression
ratio, fidelity and computational expense.
In this paper, we propose a new image coding scheme by combining the LPS and the VVQ, which
inherits the advantages of the both techniques. In this new scheme, the LPS is employed to generate
the residual image sequence and the VVQ is used to code these residual images. Experimental results
show that the new scheme can achieve comparable rate distortion performance to that of traditional
VQ techniques, while the computational cost of the new scheme is much lower since the encoding
and deoding are based on much smaller dimensional 'visual vectors'. Because the coding operation is
performed on the basis of hierarchical multiresolution blocks, the new scheme has much less blocking
effects on the reconstructed image than that of traditional VQ techniques.
The remaining of the paper is organized as follows. Section 2 summarizes the LPS and the VVQ
system for coding Laplacian residual images is described in section 3. Section 4 discusses the image
reconstruction. Section 5 presents experimental results and section 6 gives some conclusion remarks.
2 The Pyramid Structure
The generation of the pyramid structure includes the generation of the Gaussian pyramid and the
generation of the Laplacian pyramid. The process is illustrated in Fig.1.
Gaussian Pyramid Generation
The original image Go of size M (cid:141) N pixels becomes the level 0 of the Gaussian Pyramid. Upper level
images are generated by using the reduction function R(.)1 defined in (1), iteratively.
Gl(i,j) - ~ y w(m,n).Gl_l(2i + m, 2j + n) O < l < i, O < i < Ml, O < j < Nl. (1)
m---2 n---2
L is the number of levels in the pyramid, Ml and Nl are the dimensions of the lth level, and w(m, n)
are weighting kernels. Fig.2 shows a 5-level Gaussian pyramid of "Lena".
Laplacian Pyramid Generation
The reverse of the reduction function R(.)is the expansion function E(.)1 defined in (2). Let the
n,ZG be the result of expanding Gl n times. Then
Figure :1 Pyramid Structure Generation Figure :2 5-Level Gaussian Pyramid of "Lena"
2 2
Gt,n(i,j)- 4 ~ ~ w(m,n).Gl,~_a( i~- ' m j- 2 n ) O<l<_L, O<_n<_l, (2)
m---2 n=-2
0_ i < Ml-n, O<_j<Nl-n
The Laplacian pyramid is a sequence of residual images Io, Ix, " ," IL, each being the difference of two
adjacent levels of the Gaussian pyramid. Thus
Ii-Gz-Gl+l,l for 0___l_L-1 (3)
IL--GL
Fig.3 shows a 5-level Laplacian pyramid of "Lena" image generated by Eqn.(3).
Figure :3 5-Level Laplacian Pyramid of "Lena" Figure :4 VVQ Image Coding System
3 Visual Vector Quantization
The design of the VVQ coding system consists of the design of the coding-book used in coding phase
and the design of the decoding-book used in decoding phase.
Design of Coding-book
The residual image Il of size Mlx zN is divided into Pl x Ql blocks of size ml x ,zn where Pl =
Ml/ml, QI = N~/n~. The horizontal and vertical derivatives ]2[ of each block are calculated as
1
~I,(4p+i,4q+j).gh(i,j) O<_p<P,, O<_q<O, (4)
Oh(p, q)- ~- o=i o=j
Dv(p,q) - ~- ~ Ii(4p + i,4q + j). gv(i,j) 0 _< p < Pl, 0 _< q < Ql (5)
i=0 j=0
where the values of the kernels hg and vg can be written collectively in matrix form as
1 1 -1 -1 1 1 1 1
Gh -- 1 1 -1 -1 Gv - 1 1 1 1
1 1 -1 -1 -1 -1 -1 -1
1 1 -1 -1 -1 -1 -1 -1
The horizontal and vertical derivatives of a block are used to form a "visual vector", ,hD( Dr), to
represent the block. The visual vectors of all blocks of residual image Il are partitioned into Nc clusters
using the competitive learning [3], and these cluster centers are used to comprise the coding-book for
residual image Il.