J. A. O'Sullivan Research Projects
Joseph A. O'Sullivan Research Projects
Data storage systems have relied primarily on designs based on storing data on
tracks. On magnetic media, input data are encoded and stored as flux reversals
on tracks, with decoding being based on standard algorithms such as the Viterbi
algorithm. As data densities increase, fundamental limits for recording on tracks
are approached, and alternative data storage technologies must be considered. We
describe several approaches for the design of encoding and decoding for
two-dimensional data storage systems that have two-dimensional (2D) intersymbol
interference. This interference invalidates assumptions in the Viterbi and
related algorithms necessitating novel decoding strategies. Our approach to the
problem is two-fold: (1) to use existing equalization methods and combine them
with error-correction coding to enhance system performance, (2) perform joint
equalization and decoding using message-passing algorithms. Simulations
demonstrate the potential for such systems.
Standard image reconstruction techniques for X-ray computed tomography (CT)
often do not yield clinically useful images in the presence of highly
attenuating materials. The observed image degradation is caused by phenomena
such as aliasing, noise, scatter, and beam hardening, among others. In our
work, we adopt a stochastic model for the measured data that accounts for the
above phenomena. Images are reconstructed by maximizing the data loglikelihood
using an alternating minimization algorithm we developed. We reported that our
method successfully reduces image artifacts in several experiments on simulated
data when compared to the standard technique of filtered backprojection (FBP).
Our present research focus is on efficient and fast implementation, processing
of real data collected by clinical scanners, as well as extending our method to
volume (3-D) reconstruction.
Professor O'Sullivan's research interests include
information theory, estimation theory, and imaging science, with
applications in object recognition, tomographic imaging, magnetic
recording, radar, and formal languages. Current research projects
include: modeling and performance analysis of target recognition,
orientation estimation, and tracking using high resolution radar data;
spiral CT imaging in the presence of known high density attenuators;
physics-based capacity bounds for magnetic media; derivation
and analysis of alternating minimization algorithms;
information-theoretic analysis of steganography; and systems
integration issues in magnetic information systems.
Information Theory: Alternating Minimization Algorithms.
Much research over the past several years has been devoted to a
characterization of several types of alternating minimization
algorithms. Several alternating minimization algorithms have been
presented using a unified treatment, including
expectation-maximization algorithms, Blahut-Arimoto algorithms,
Blahut's rate-distrotion algorithm, and generalized iterative
scaling. The unified view of these algorithms has led to new
algorithms that are useful in spiral CT imaging in the presence of
high density attenuators, hyperspectral imaging, and in the general
derivation and computation of the information value decomposition. In
addition, these ideas are being applied to iterative decoding
algorithms for maximum likelihood sequence estimation in
Information Theory: Information Value Decomposition.
A new paradigm for the analysis of positive matrices and
multidimensional positive-valued functions is being explored. An
algorithm has been derived and called the information value
decomposition. One conference submission has resulted, with papers
expected shortly. The algorithm approximates an arbitrary
positive-valued matrix by a lower rank-positive valued matrix. The
lower rank approximation is closest to the original matrix in the
sense that it minimizes the I-divergence between the original matrix
and the approximation. It has been applied to hyperspectral imaging with some
success, with plans for continued research.
Information Theory: Information Hiding.
There have been several methods proposed in the literature to hide
information within host data sets. Some of the work goes under the
names steganography, (digital) watermarking, fingerprinting, and traitor
tracing. Prof. O'Sullivan and Pierre Moulin have presented
a new way to analyze these systems in a common framework and has
derived fundamental bounds on the performance of such systems. Each
system is characterized in communications terms as being designed to
send information to an ultimate arbiter, with side information in the
form of a key. The rates achievable in this communication have been
characterized in several important cases.
There is a growing recognition in the research community of the role
of information theory in imaging problems. Several activities within
the information theory community and within the imaging and automatic
target recognition communities exemplify this. These include the
establishment of the Center for Imaging Science, a paper within the
special issue of the IEEE Transactions on Information Theory
commemorating fifty years of information theory, an Information Theory
Workshop on Detection, Estimation, Classification, and Imaging, and
a special issue of the IEEE Transactions on Information Theory
on Information-Theoretic Imaging. Prof. O'Sullivan has played a role
in each of these activities and plans to continue to work on
developing and applying new ideas in this area.
Imaging Science: Automatic Target Recognition for Synthetic
Aperture Radar Data.
Comprehensive computational and analytical studies of ATR algorithms
for SAR data are being pursued. A new class of ATR algorithms based
on conditionally Gaussian models has been derived. This new class of
algorithms is being compared to exisiting algorithms over a wide range
of parameters to yeild performance-complexity tradeoffs. For
essentially all operating conditions studied to date, the new
algorithms perform achieve better performance for each complexity.
Imaging Science: Spiral CT Imaging.
Prof. O'Sullivan is a member of CT Visualization and Quantification
Team, working with B. Whiting and J. Blaine from the Electronic
Radiology Laboratory at the Washington University School of Medicine,
with J. Williamson from Radiation Oncology, and D. L. Snyder from
ESSRL. The team is developing a set of code to simulate CT data and to
implement new and known image reconstruction algorithms, in support of
multiple clinical efforts. In radiation oncology, the goal is to
estimate the position and orientation of a known high density
attenuator and to estimate the attenuation function in the vicinity of
the known attenuator. In this case, the attenuator is a radiation
brachytherapy applicator. The theory is applicable to arbitrary known
Imaging Science: Performance Bounds.
The performance os ATR systems is being analyzed and bounded using
several techniques. Working with graduate student N. A. Schmid, the
performance degradation from using estimated parameters in ATR systems
is by quantified in both the limit as the number of training samples
is large and in the limit as the number is small. Improved ATR
algorithms are being derived based on the peaking phenomenon in
Imaging Science: Hyperspectral Imaging.
Prof. O'Sullivan has been working with Profs. D. R. Fuhrmann,
D. L. Snyder, and W. H. Smith to put together a hyperspectral imaging
team. Papers have appeared over the past year that represent the
initial results from the collaboration. The goals of the research
efforts include the characterization of hyperspectral imaging sensor
data, the derivation of algorithms and performance bounds for
estimating components in the models (including endmembers), and
applications to automatic target recognition, topographic engineering,
and earth science.
Prof. O'Sullivan has a long-standing collaboration with R. S. Indeck
and M. W. Muller in magnetic recording. The work has progressed from
the micromagnetic modeling of magnetic media through the derivation of
capacity bounds for magnetic media to the design of systems based on
parameters that capture medium noise characteristics. Systems
integration issues are beign pursued that included further studies
related to capacity bounds for magnetic recording systems.
Edited June 11, 2004
Washington University in St. Louis
School of Engineering
Deptartment of Electrical and Systems Engineering