Image and Video
Compression Techniques: research on image,
video and signal compression techniques
Image and vídeo
coding using generalized bit-planes and
matching pursuits
Eduardo A. B. da Silva
(in collaboration with Rogério Caetano,
Fundação Paulo Feitoza, Brazil).
Among the state-of-the-art in wavelet
image coding one can point out the ones
that use successive approximations, due to
both compression efficiency and ease of
implementation. Initially the most
successful successive approximation
quantization methods were based on scalar
quantization. The attempts to employ
successive approximation vector
quantization methods resulted in codebooks
that were very complex to generate.
However, it has been developed a
successive approximation vector
quantization method that uses regular
lattices as codebooks, that are very easy
to generate. This has opened new avenues
of research in the area. It has been shown
that this method can be extended to
Hilbert spaces, and can be seen as a
generalization of sinal decompositions
followed by quantization and bit-plane
encoding.
Conditions for the convergence of
successive approximation vector
quantization have been determined.
However, that results have been obtained
only experimentally. Its mathematical
analysis has led to analytical results and
theorems that have set the basis for
deeper studies and the development of
image and video encoders.
Following this line, we have been
investigating the relations of successive
approximation vector quantization with
iterated function systems, since this has
been shown to be a good tool for the
analytical development of the successive
approximation theory. With the help of
this theory, along with ergodic theory and
the expansive transformations theory, more
efficient methods of vector successive
approximation have been investigated. As a
natural development of this research,
extensions of this theory to Hilbert
spaces are being investigated, leading to
a theory of signal decompositions in
generalized bit-planes. Its applications
to video compression are being
investigated, using the framework of the
matching pursuits algorithm. It has also
been investigated the dictionary design
for such decompositions, based on recent
development in the area of two-dimensional
function decomposition using geometrically
significant components, known as
ridgelets, curvelets and coutourlets.
Still based on the generalization enabled
by the decompositions using generalized
bitplanes, static image coding methods
based on the matching pursuits algorithm
are being investigated; DCT and wavelet
image coders are particular cases of
these.
Signal encoding using
multiscale recurrent patterns
Eduardo A. B. da Silva
(in collaboration with Murilo B. de
Carvalho, Universidade Federal Fluminense,
Brazil, Sergio M. M. de Faria,
Universidade de Leiria. Portugal, Nuno
Rodrigues, Universidade de Coimbra,
Portugal and Vitor Silva, Universidade de
Coimbra, Portugal).
Lossy string matching methods have
recently attracted the interest of the
data compression community, since they
provide forms of efficiently integrating
the quantization and coding operations. In
this research the use of multiresolution
techniques in recurrent pattern matching
has been investigated. With this method,
one obtains an integration of the
operations of transformation, quantization
and coding. Results show that these
methods are promising . We research
altenatives for the design of efficient
signal compressors base don this paradigm.
We also perform the theoretical
development of its relation with
orthogonal transforms. Its applications to
video coding will also under
investigation.
Stereo image
compression
Eduardo A. B. da Silva
(in collaboration with Murilo B. de
Carvalho, Universidade Federal Fluminense,
Brazil, and Carla L. Pagliari, Instituto
Militar de Engenharia, Brazil).
Stereo images provide the observer with a
sense of reality far greater than the one
provided by monocular images. In addition,
stereo information is very useful in image
analyis, including computer vision, image
segmentation, etc. The analysis of such
images needs the knowledge of the
disparity map, that contains the
information that defines pairs of
corresponding points on the two views.
From it we also obtain the
tree-dimensional image depth information.
However, there are not yet effective
methods for its computation and
representation. A lot of research has been
carried out in this area.
We investigate stereo image compression
methods using recurrent patterns. Using
this method, we expect to take advantage
of the similarities among the two views
without having to transmit the disparity
map. Encouraging results have been
obtained.
Image and video
coding using triangular mesh
approximations
Eduardo A. B. da Silva
and Gelson V. Mendonça.
In this project we develop image and video
encoders using approximations with
triangular meshes and interpolation.
Efficient forms of organizing the meshes'
parameters are being investigated,
including hierarchical representations
using rate-distortion criteria. Effective
ways of performing mesh interpolation are
also being investigated. Results show good
compression potential for the method.
Surface compression
with applications in computer graphics
Eduardo A. B. da Silva
(in collaboration with Helio Lopes,
Department of Mathematics, PUC-RIO,
Brazil).
In computer graphics, the objects
generated can be defined by the
specification of the geometry of the
surfaces, as well as their properties. The
more realistic is the computer graphics
model, the larger is the amount of data
needed to represent them. There are
several techniques for the representation
of the data correnponding tho the
resulting surfaces. Such data consists
basically of two parts: the topology of
the surface and the position of its
vertices. However, in general the data
compression methods employed for such data
are trivial from the signal and image
processing points of view. In this project
the Institute will develop techniques of
signal compression applied both to the
meshes' topology and to their vertices.
Turbo-coded
quantization
Eduardo A. B. da Silva,
Gelson V. Mendonça (in collaboration with
José F. L. de Oliveira).
Data compression systems based on the
turbo-coded quantization (TCQ) can achieve
rate-distortion performances close to the
optimum. They are based on the turbo-coded
modulation, using the well-known duality
between coded modulation and quantization.
In this work, we investigate the use of
"turbo-coded quantization", that takes
advantage of the excellent performance of
turbo codes in order to generate an
efficient data compression system.
Preliminary results using random signals
are promising. Image coders based on this
paradigm are also being investigated.
Joint source-channel
coding
Eduardo A. B. da Silva
(in collaboration with Weiler Alves
Finamore, CETUC/PUC-RIO, Brazil).
In applications where one needs to
transmit images/vídeo through channels of
low signal to noise ratios, one has to
insert redundant bits in order to the
error probability remain within acceptable
limits. On the other hand, one wants to
use a source coding technique in order to
reduce redundancy as much as possible.
There is therefore a clear compromise
among the bits spent with the increase in
redundancy (channel coding) and the bits
saved with the reduction of the source
redundancy (source coding).
Recently, there has been increased
interest in this problem, due to, among
other things, the need to transmit video
in wireless systems. A popular approach is
joint source-channel coding. We
investigate this problem, using techniques
that combine vector quantization and error
control coding, as well as wavelet and
lapped transforms. Techniques using
combinatiosn of turbo-quantization and
turbo-modulation are also continue under
investigation.
Lossless video
compression
Eduardo A. B. da Silva.
In television studios, it is sometimes
necessary to compress videos so that there
is no loss. There are several lossless
still image coders. However, little has
been done in lossless video compression.
The few encoders to date treat the video
as a sequence of still images. In this
line of research we develop lossless video
encoders that exploit the redundancy among
frames.
Wyner-Ziv theorem
applied to video coding
Eduardo A. B. da Silva.
The Wyner-Ziv theorem is related to the
problem of source coding with side
information present only at the decoder.
It says that the rate-distortion function
of a source X is the same irrespective of
the side-information being present only at
the decoder or at both the encoder and
decoder. Therefore, if, in video coding,
one considers the motion vectors as side
information, then, according to the
Wyner-Ziv theorem, it is possible, at
least in theory, to design video encoders
where the motion estimation is carried out
only at the decoder, and yet to obtain
performances as good as the conventional
ones. Recently video coding algorithms
based on the Wyner-Ziv paradigm have been
proposed. In this line of research we
investigate video encoders based on the
H.264 standard using the Wyner-Ziv
paradigm. Algorithms using turbo codes as
well as LDPC (low density parity check
codes) are being considered.
Error concealment
methods in video compression
Eduardo A. B. da Silva
(in collaboration with Ricardo L. de
Queiroz, Universidade de Brasília,
Brazil).
Error concealment methods are of paramount
importance, specially in error-prone
environments, as in video transmission in
wireless systems, as mobile phones. In
this work we investigate error concealment
techniques to be applied to
state-of-the-art video coding methods, as
the H.264 standard.
Compression of
electric power systems signals
Eduardo A. B. da Silva,
Paulo S. R. Diniz (in collaboration with
Marco A. M. Rodrigues, CEPEL, Brazil).
Power systems signal analysis constitute a
powerful tool for the determination of
failures on a power system, as well as its
diagnosis. However, in order to do so,
large amounts of data has to be stored and
transmitted from remote stations to bases
where they can be properly analyzed. This
requires efficient methods to compress
such data so that the accuracy of the
analysis made upon them is not impaired.
In such applications, the signals are
formed by a combination of specific
waveforms, what makes them suitable to
compression using the matching pursuits
algorithm. In this line or research, we
investigate the application of such
methods for compression and analysis of
power systems signals.
ECG signal
compression
Eduardo A. B. da Silva
(in collaboration with Murilo B. de
Carvalho, Universidade Federal Fluminense,
Brazil, Jurandir Nadal, PEB-COPPE/UFRJ).
In this theme this Institute investigates
the application of wavelet transforms to
ECG signal compression. Techniques based
on multiscale recurrent patterns are under
investigation.
Image Processing
Digital Watermarking
Eduardo A. B. da Silva
(in collaboration with Murilo B. de
Carvalho, Universidade Federal Fluminense,
Brazil).
With the ever-growing use of multimedia
systems, copyright protection of data in
multimedia systems is becoming more and
more important. The use of digital
watermarking has come to fulfill these
needs. In the last few years there has
been a significant growth in the amount of
research in this subject. One of the main
problemas involving digital watermarks is
that they have to be robust, that is, any
image modification that can erase the mark
will render the image useless. This
Institute intends to investigate the use
of multiscale recurrent patterns in order
to insert robust watermarks.
Non-linear
decompositions using morphological
operators
Eduardo A. B. da Silva
(in collaboration with Marcos Craizer,
Department of Mathematics, PUC-RIO,
Brazil).
Non-linear decompositions using the
"lifting" technique have already been
described in the literature. In this
topic, we investigate the use of
morphological operators in such
decompositions. In thi research, we devise
efficient design methods for such
decompositions. We also investigate
morphological operators that can lead to
maximally decimated decompositions in
which the bands generated satisfy shape
criteria. In addition, a theoretical
structure is under development in order to
both formally analyze and design them.
Image analysis and
pattern recognition
Eduardo A. B. da Silva
(in collaboration with Gelson Vieira
Mendonça, COPPE/UFRJ, Alexandre Pimentel
Mendonça, Instituto Militar de Engenharia,
Brazil. and Sergio Rodrigues Neves, IPQM,
Brazil).
Image analysis has a fundamental role in
systems that rely on visual information
for some sort of decision making. One of
the first steps in image analysis is
segmentation, that consists in identifying
image areas that contain objects of
interest. Several segmentation techniques
have been investigated. One application in
which we have been working is on target
segmentation in infrared images, using
wavelets combined with mathematical
morphology.
Another important operation in image
analysis is pattern recognition. We
investigate the use of linear
discriminative filtering to identify
patterns in images. A theoretical analysis
of the problem regarding it as an image
restoration problem has been leading to
encouraging results.
Detection and
classification of tumors in mammographic
images
Eduardo A. B. da Silva
(in collaboration with Wagner C. A.
Pereira, PEB-COPPE/UFRJ, Brazil).
Mammography analysis leads to a
significant increase in success in early
detection of breast cancer. The usual
diagnostic procedure is the following:
when the mammography shows the possibility
of a malign tumor, the pacient is sent to
a biopsy. One wishes both to minimize the
number of biopsies due to false tumor
detection as well as the number of
undetected malign tumors. Normally
mammographies have high resolution,
exceeding the human visual capabilities.
Therefore, it is desirable to have
automatic mammography analysis methods
that can decrease the amount of false
positives and false negatives. We
investigate the use of mathematical
morphology in the detection and
classification of tumors in mammographies,
as well as of independent component
analysis together with neural networks.
Sinal Processing
Redundant
decompositions
Eduardo A. B. da Silva,
Paulo Sergio Ramirez Diniz (in
collaboration with Lisandro Lovisolo,
Universidade do Estado do Rio de Janeiro,
Brazil).
As described above, signal decompositions
in redundant dictionaries have found an
increasing number of applications. An
example is the matching pursuits
algorithm, that we investigate both for
compression of images and power systems
signals. The mathematical theory
underlying these decompositions is the
frame theory . In this line we make a
theoretical investigation of frames that
generate decompositions that can be
efficient for both signal compression and
analysis.
Filter design for the
computation of electric power systems
parameters
Eduardo A. B. da Silva
(in collaboration with Marco A. M.
Rodrigues, CEPEL, Brazil).
In power systems, several measurements can
be made in order perform both signal
analysis and fault detection. Oscilography
deals with both such signal's acquisitions
and analysis. In oscilography, it is
extremely important that there can be good
estimates of the magnitude and phase of
the electrical quantities. Fourier filters
are often used for that. However, their
performance tends to degrade when there is
small variations on the frequency of the
power source. This Institute will continue
to investigate the design of digital
filters that can be good substitutes for
Fourier filters, minimizing the frequency
variation problem.