Human-Computer Interaction (HCI)
T. S. Huang,* L. Chen, V. Pavlovic, N. Jojic, S. Chu, G. Berry, Y. Zhang, Y. Wu
U.S. Army Research Office, DAAL01-96-Z-0003; National Science Foundation, JRI-9634618; Yamaha Motor Corp. (Conducted in the Coordinated Science Laboratory)

We use the term HCI in a very broad sense to include communication between person and computer as well as communication between persons via computer. An example of the former is a person using a workstation, an example of the latter is tele-collaboration. We are investigating a variety of issues related to the use of computer vision in HCI. These include: facial feature extraction and tracking, determining 3-D head pose, facial movement modeling, analysis, and synthesis, hand gesture recognition, human body motion analysis, and person identification.

Multimedia Databases
T. S. Huang,* S. Mehrotra,* K. Ramchandran,* Y. Rui, A. She, Y. Kang, M. Naphade, T. Kristjansson, R. Wang
NSF/DARPA/NASA Digital Library Initiative Program under Cooperative Agreement 94-11318; Defense Advanced Research Projects Agency, N6601-95-C8511 (Conducted in the Coordinated Science Laboratory)

We are studying a number of challenging issues in image/video data indexing and retrieval. Of particular interest are similarity-based retrieval where similarity measures are based on image content such as color, texture, shape, and layout; mapping of high-level concepts to low-level image features; and how to deal with data and query uncertainties.

Image/Video Compression and Representation
T. S. Huang,* K. Ramchandran,* M. Gharavi-Alkhansari, H. Tao, A. Colmenarez, R. Lopez, S. Servetto
Joint Services Electronics Program, N00014-96-1-0129; Army Research Laboratory Coop. Agreement DAAL01-96-2-0003 (Conducted in the Coordinated Science Laboratory)

Our goal is to investigate image/video representation and compression schemes that are suitable for data storage, retrieval, and display. Performance criteria will be based not only on compression factors, but also on scalability, interoperability, and ease of manipulation with compressed data. Under study are fractal coding, wavelet/morphological coding, and 3-D model-based methods.

Digital Filters with Adaptive Fault Tolerance
W. K. Jenkins,* J. Jiang, C. Schmitz
Joint Services Electronics Program, N00014-96-1-0129 (Conducted in the Coordinated Science Laboratory)

This project investigates how the learning process in adaptive digital filters is disturbed by hardware failures and how to design filters and adaptive algorithms that can continue operating in the presence of such failures. Adaptive systems are capable of adjusting parameters to reduce a specified error criterion. It has been shown that whenever a hardware failure occurs that increases the error, the system will attempt to compensate for this failure by further self-adjustment. Recently this research has concentrated on compensating broader classes of hardware errors and on applying adaptive fault tolerance to adaptive filters of the infinite impulse response class.

Computationally Efficient Algorithms for Adaptive Quadratic Volterra Filters
W. K. Jenkins,* C. W. Therrien, X. Li
Joint Services Electronics Program, N00014-96-1-0129 (Conducted in the Coordinated Science Laboratory)

The structure of the input autocorrelation matrix in Volterra second-order adaptive filters for general colored Gaussian input processes has been analyzed to determine how to best formulate a computationally efficient, fast adaptive algorithm. It was shown that when the input signal samples are ordered properly within the input data vector, the autocorrelation matrix of quadratic filter inherits a block diagonal structure, with some of the subblocks also having diagonal structure. Some new results in developing and evaluating computationally efficient quasi-Newton adaptive algorithms have been obtained that take advantage of the sparsity and unique structure of the correlation matrix that results from this formulation.

Performance Bounds on Image and Video Compression
Y. Bresler,* P. Moulin, K. Ramchandran, M. Gastpar
National Science Foundation, MIP-97 07633 (Conducted in the Coordinated Science Laboratory)

Has image/video compression reached a state of saturation where more research on compression is unlikely to yield significant improvements? This research tries to shed light on this by investigating the fundamental rate-distortion performance bounds for realistic classes of image and video models. A primary goal is to uncover the performance gaps between existing commercial systems and the theoretical performance optimally attainable and to guide the generation of improved compression algorithms in the future.

Spatio-Temporal Magnetic Resonance Imaging
Y. Bresler,* Z.-P. Liang
National Institutes of Health, PHS 1 R21HL62336-01 (Conducted in the Coordinated Science Laboratory)

Dynamic magnetic resonance imaging (D-MRI) involves, in a broad sense, producing a time series of images of a time-varying or moving 3-D object with both high spatial and temporal resolutions. Such capability is key to cardiac imaging, functional brain mapping, and interventional MRI. However, to date, more than 20 years after the introduction of MRI, this objective remains elusive. Our goal is to establish a new unified theoretical framework for D-MRI that will provide optimal procedures to sample (k, t)-space and to incorporate valid spatio-temporal constraints into the dynamic imaging process. We also expect to develop practical data-acquisition pulse sequences that match both theoretical and physical MRI system constraints and associated computational algorithms for image reconstruction and to evaluate experimentally their feasibility for cardiac imaging.

VLSI Adaptive Equalizers for Equalizing Magnetic Recording Channels
W. K. Jenkins,* I. Li
Joint Services Electronics Program, N00014-96-I-0129 (Conducted in the Coordinated Science Laboratory)

This project is investigating the design of an adaptive equalizer-on-a-chip for the equalization of magnetic recording channels. A design based on combining a residue number system architecture with a block LMS adaptive algorithm is being evaluated for its potential to a design that achieves sufficiently high operating speeds for magnetic disk applications, while having simple enough circuit requirements to be fabricated as a monolithic VLSI component. Special attention is being devoted to the management of short word length finite precision arithmetic and its effect on the learning characteristic of the equalizer. This project involves both VLSI design and fabrication.

Channel Equalization with Adaptive Filtering and the Preconditioned Conjugate Gradient Algorithm
W. K. Jenkins,* R. A. Soni
Joint Services Electronics Program, N00014-96-1-0129 (Conducted in the Coordinated Science Laboratory)

Communication system performance is often degraded by imperfections of the channel. When additive noise and nonideal channel characteristics are unknown prior to transmission, adaptive equalizers are used to compensate for these imperfections and improve overall performance. For highly correlated received sequences, the convergence rate of the equalizer is a strongly limiting factor. This project aims to develop novel schemes employing preconditioned conjugate gradient (PCG) optimization for channel equalization. Results have been obtained to illustrate that, compared to an LMS equalizer, the PCG equalizer provides significantly improved performance for algorithms which minimize the mean squared error and constant modulus error criteria.

Adaptive and Optimal Time-Frequency Methods for Nonstationary Signals
D. L. Jones,* M. L. Kramer, B. Krongold, A. Rao, L. Qian
U.S. Office of Naval Research, N00014-95-1-0674 (Conducted in the Coordinated Science Laboratory)

New adaptive and statistically optimal time-frequency analysis methods are being developed for improved processing of nonstationary signals. The class of problems for which time-frequency-based detection is being characterized and optimal kernels for detection are being derived. New adaptive time-frequency representations for high-resolution visual characterization of signals are also under development. These methods are being applied to problems in condition assessment for machinery monitoring and fault detection, mine classification, and transient detection and analysis.

Energy Partitioning Using Overdetermined Basis Decompositions
D. L. Jones,* B. Krongold
U.S. Office of Naval Research, N00014-95-1-0907 (Conducted in the Coordinated Science Laboratory)

This research project is developing signal processing methods based on overdetermined basis decompositions for estimating the relative energies of individual components of complex signals and for component separation and recovery. Such an approach can decompose a signal with multiple overlapped, nonorthogonal components onto different basis elements, thereby separating them in situations in which standard filtering approaches or orthogonal basis decompositions cannot.

Model-based Tomographic Imaging Methods
Z.-P. Liang,* C. P. Hess
National Science Foundation, BES 95-02121, MIP 94-10463

The mathematical basis of tomographic imaging is conventionally rooted in the well-established Fourier or radon transform theories, so that image quality is mainly dependent on how the data space is sampled. In practice, physical and temporal constraints often prevent a sufficient coverage of the data space, resulting in various image artifacts, such as Gibbs ringing, resolution degradation, and various motion effects. This project is aimed at overcoming these problems by developing new model-based imaging techniques that can effectively incorporate a priori information into the imaging process. Application of these techniques to cardiac imaging and functional brain mapping is also addressed.

Artificial Neural Networks
Z.-P. Liang,* T. S. Huang,* Y. Zhang, H. Pan
Joint Services Electronics Program, N00014-96-1-0129

The primary goal of this project is to develop new neural network architectures and learning algorithms useful for multisensory data fusion, recognition of time-varying patterns, and automatic image segmentation. To achieve this goal, work is being carried out to develop a new neuronal model with both regular and modulatory inputs, a new wavelet-based multichannel network architecture, and a dynamical system-based learning rule. Practical issues of hybrid processing with both neural network models and statistical models such as the hidden Markov model are also being investigated in this project.

Automatic Segmentation of Brain Images
Z.-P. Liang,* J. Ji, Z. Fu
National Science Foundation, BES 94-10463

After two decades of active research, automatic image segmentation remains one of the most challenging problems in image processing and computer vision. This project is aimed at developing a prototype pattern recognition system for automatic segmentation of brain images. This system contains components for multiscale processing, pattern generation, and neural network learning. We expect that the computational principles used in building this system will be useful for solving other practical pattern recognition problems.

Radar Imaging of Runways during Aircraft Landing
D. C. Munson, Jr.,* J. A. Lee
Rockwell International (Conducted in the Coordinated Science Laboratory)

We are investigating synthetic aperture radar (SAR) as a means of imaging runways through fog and cloud cover from an approaching aircraft. Current radars with traditional signal processing are incapable of providing the resolution required at long ranges, because of the wide beam widths of the antennas employed. Our approach uses the changes in angular aspect of points in the airport scene, provided by the motion of the aircraft, to produce high-resolution imagery from return signals collected by a conventional radar.

Monitoring Lateral Stability of Rail
E. J. Barenberg* (Civil & Environ. Engr.), D. C. Munson, Jr.,* Y. Ding
Association of American Railroads (Conducted in the Coordinated Science Laboratory)

Thermal expansion and contraction of railroad track can cause severe rail stress, resulting in breakage. The objective of this project is to develop an automated rail inspection system to measure changes in position of rail, which can in turn be used to calculate stress within the rail. Our current approach uses differential GPS, with carrier-phase measurement, to achieve high accuracy. We are working with our industrial partner, Rockwell-Collins, to improve accuracy by incorporating a smoothness constraint on the trajectory along a rail. Future work will address the calculation of rail stress and identify how this information can be used in an operational setting.

Design and Optimization of Passive and Active Radar
R. E. Blahut,* Y. Bresler,* W. C. Chew,* P. Moulin,* D. C. Munson, Jr.,* M. Brandfass, A. Lanterman, R. Venkataramani, S. Voloshynovskyy, Y. Wu, S. Xiao
Defense Advanced Research Projects Agency, AFOSR F49620-98-1-0498 (Conducted in the Coordinated Science Laboratory)

The goal of this project is to develop a new class of widely applicable passive and active radar imaging algorithms and associated theory, using state-of-the-art physics-based modeling, advanced statistical inference techniques, and recent advances in computational methods and hardware. In particular, this project emphasizes passive imaging and characterization of aircraft using reflected commercial TV and radio signals. We plan to validate our new theory and algorithms on such data, in close collaboration with our industrial partner, Lockheed-Martin.