The following is a representative description of AIMS' government and commercial contracts and the areas of work involved. National Institutes of Health/National Cancer Institute Electrical Impedance Tomography Electrical Impedance Tomography, or EIT, is a method of imaging human tissue based on tissue conductivity, as opposed to density which is the basis for ultrasound and X-ray methods. In EIT, a belt of simple electrodes is placed around the body or a limb. This ring of sensors provides input signal patterns that can be processed by a desktop computer to yield a grey-scale image of an interior cross-section in which ``light" and ``dark" correspond to ``more conductive" and ``less conductive." There are numerous instances within the human body where adjacent soft tissues of similar density possess markedly different conductivities. EIT is not necessarily confined to use involving humans, but is also suitable for ``imaging" the interior of any object, a suspicious package for example. For 3-dimensional imaging, the single ring of sensors can be replaced by a band of vertically stacked sensor rings.

AIMS scientists have discovered a new technique to process the electrode data which represents a quantum leap in the theoretical understanding and practical implementation of EIT. This technique leads to EIT imagery of heretofore unachievable resolution. The National Institutes of Health have recently recognized this by awarding AIMS with a Small Business Innovative Research grant to develop the technology. The most attractive features of EIT technology are the following:

• Low-cost: Comparable in cost to ultrasound imaging.
• Non-invasive: No dangerous X-rays or radioactive tracers.
• Easy operation: Equipment is safe, portable and simple to use.
• Real-time imaging: Add-on boards permit real-time diagnostic ``movies," e.g., cardio-pulmonary function.
• High-resolution: First generation EIT has resolution comparable to current ultrasound technology.
• Tomographic: Choice of cross-sectional ``slices" or full 3D images inside the body or limbs.
• Tissue mapping: Unique dielectric operation images internal organs by their electro-chemical composition, not simply by tissue density, as with X-rays and ultrasound. This has important applications in medicine as well as criminal investigation and non-destructive testing of materials and structures for faults.

   

Figure 1: A schematic of AIMS' EIT imaging system showing how a 16-electrode configuration is used to reconstruct a ``slice" of an internal conductivity distribution.

  
 
   

Figure 2: Simulated image reconstructions comparing the X-ray transform and the new EIT algorithm for a 16-electrode imaging system.

   Advanced Research Projects Agency (ARPA) Wavelet and Learning Clustering Algorithms for Automatic Target Recognition In this effort, AIMS continues to be a leading innovator in the area of automated target recognition. A fundamental difficulty of target recognition is to develop a unified methodology for target feature representation, extraction, and classification. This project achieves such a unification in a hierarchical way that takes advantage of the localization properties of the wavelet transform to generate multi-scale representations of target features. The proposed method may be called progressive recognition, in which the output of wavelet preprocessing is used as input to learning clustering algorithms which in turn provide feedback to the wavelet preprocessor. Anticipated applications of this method are:

• Sensor fusion.
• Compression, storage and progressive retrieval of FLIR, LADAR and visual target signatures, as well as high resolution radar returns for use in smart, autonomous reconnaissance, surveillance and target acquisition systems.
• Ground vehicle acoustic signature representation and compression.
• Early fault and damage detection in gear-trains, machinery and other equipment.
• Recognition of industrial objects in manufacturing.
• Underwater passive sonar target recognition systems.
• Large image database reduction and fast matching, browsing and retrieval.

   

Figure 3: Graphical interface to AIMS' wavelet transform software, being used here to develop image compression algorithms for fingerprint storage and transmission.

   Center for Night Vision/ US Army CECOM Automatic Target Recognition This project was a major one-year effort by AIMS that resulted in several advances in the state of the art of automated target recognition (ATR). The primary thrust was to perform basic research and development for next-generation, model-based target recognition systems. AIMS' methodology and algorithms have demonstrated superior performance in several controlled, blind tests. Specific innovations include:

Modeling:

• Model Construction: Model building techniques that take in CAD models, field imagery or a combination of the two with equal ease.
• Aspect VQ: Procedures, based on vector quantization, to systematically identify the minimal aspect representatives of a given target class reqeuired in a model database to achieve a desired level of performance.
• Multiresolution Polygonalization: Algorithms that allow one to compute, store and index polygonal models of target silhouettes or signatures as a function of range (or other) resolution. This provides a foundation to develop algorithms that employ progressive recognition.

Probing Methods:

• Probeset Likelihood Ratios: We have extended the probing concepts of computational geometry with a general definition of likelihood for probeset hypotheses.
• On-line Optimal Probing: We have implemented a number of new, grey-scale probing methods for viewpoint estimation, target class decisions and clutter rejection. We have also introduced new techniques that permit the on-line, real-time computation of the optimal probe sets for between-class discrimination.

   

Figure 4: FLIR image of 4 M-60 tanks at 1.5 km correctly recognized by AIMS Probing Algorithm.

  

Multi-Sensor:

• Probe Extensions: Alternative sensors, imaging or otherwise, induce probesets that can be handled by our methods in a natural way. In particular, our definition of probeset likelihood is sensor independent.
• Model Extensions: We have shown how the polygonalization methods can be adapted to any one-dimensional signal, and for example, have pointed out the similarities that exist between polygonal signatures and mmWave radar returns.
• Learning: Since our off-line model/parameter tuning methods are based on general vector quantization theory, our methods extend naturally to LADAR, TV, mmWave, SAR, i.e., any signal space.
• Multi-look: Due to the set theoretic formulation of the optimal probeset notion, extending the algorithm to include multiple views of a single target is simple.

Motion:

• Moving Targets: Our notion of probeset extends immediately to the case where the targets are in motion since it induces a natural definition of state that allows the direct insertion of standard filtering methods such as Kalman or extended Kalman filtering, or nonlinear filtering.
• Moving Sensors: Again, the extension is clear-cut and can be handled in a natural way using a Lagrangian formulation for the dynamics.

ATR Metrology Standards:

• Algorithm Quality: We have clearly delineated the inadequacies of the metrics currently used to assess the quality of ATR algorithms and have defined the notion of relative PEC statistics (Performance-Efficiency-Complexity) as a remedy.
• Database Difficulty: We have also outlined a method that permits the direct comparison of ATR databases and to rate their relative difficulty.

   

Figure 5: Graphical interface to AIMS ATR Probing Algorithm used for algorithm visualization.

   Naval Research Laboratory Algorithms and Software Systems for Sensor Integration Because future ship sensor systems are facing increasingly complex and dynamic environments, there is a critical need to develop advanced technology for next generation shipboard systems. A cornerstone of such technology is the development of efficient multi-sensor multi-target tracking systems which are able to function on board ships and operate as part of a large network which includes other ships or platforms equipped with similar systems.

AIMS engineers are currently developing algorithms and software systems for sensor integration. Development of a sensor fusion testbed is currently underway on high-end graphics Silicon Graphics workstations. The figure below illustrates the sensor integration interface and its implementation of a ``bearings only'' tracking algorithm.
 
 

  

Figure 6: Tracking algorithm based on multi-sensor input implemented in graphical software system developed by AIMS engineers.

   Naval Research Laboratory Specific Emitter Identification (SEI) The main objective of this on-going project is to provide a passive method of identifying specific radar emitters based soley on measurement of the pulses which they emit.

Passive identification of radar emitters in emitter dense and electromagnetically cluttered environments is a technology which would find both commercial and military applicability. In particular, the ability to identify radar emitters based solely on the passive measurement of pulses which they emit provides a basis for at least three major areas of direct benefit: (i) target recognition/combat id, (ii) target tracking, (iii) global and tactical situational awareness. In a real battle scenario in which there are mobile radar emitting targets such as ships, aircraft, and radar guided missile platforms any one of these capabilities is essential. Moreover, such technology may be employed in both offensive and defensive weapon systems. In a defensive role a passive emitter identification technology could provide automatic and accurate warnings of threats and potential danger. In an offensive role, a passive emitter identification technology could be used as a key component in a missile guidance system intended to neutralize radar emitting targets.
 
 

  

Figure 7: Schematic of a pulse classification system.

  

A crucial issue which is being addressed at AIMS is the development, analysis, and comparison of representations of emitter signals. Particular emphasis is being placed on determining representations which uniquely and robustly characterize emitter pulse behavior in the face of noise, clutter, and parameter agility. To this end, AIMS engineers have been developing and testing wavelet and wavelet-like representations of radar pulses.

Another large focus of the effort concerns the development of learning algorithms. An important requirement for an overall pulse recognition system is that it be endowed with the ability to ``learn'' fast and efficiently. In terms of learning, the fundamental specification is to learn, by example, the defining characteristics of a particular radar.

AIMS' approach to SEI may be extended to problems of generic pattern recognition. Potential areas include speech recognition, speaker identification, face recognition, finger print identification, and the recognition of tumors and growths in medical image processing applications.
 
 

  

Figure 8: Wavelet processing software developed for use in SEI.

   Naval Research Laboratory / Department of the Navy Electronic Countermeasures Effectiveness Monitoring This ongoing series of projects involves another aspect of ship self-defense, namely the lethality assessment of potential threats, and the monitoring of the effectiveness of countermeasures taken against these threats. Straightforward solutions to these problems are confounded by the adverse affects of radar multipath. AIMS engineers have developed fast regularization filtering and advanced signal processing techniques that demonstrate significant promise in combatting these problems. Ongoing efforts in this area include software development for a graphical, sensor-fusion simulation testbed, as well as more advanced algorithms for multipath cancellation.
 
   

Figure 9: AIMS developed new processing techniques that significantly enhance signal subspace methods applied to direction of arrival estimation in the presence of multipath

  
 
   

Figure 10: This sequence of figures shows the steps involved in estimating the look-angle of an anti-ship cruise missile. The accuracy of the estimated (dotted) compared to the true waveforms was considered unobtainable prior to AIMS' ongoing effort.

   Naval Research Laboratory / Department of the Navy Active Radar Decoy Performance Evaluation This project is one of several AIMS projects related to the important problem of ship self-defense.

Navy's electronic warfare community has demonstrated a need for low speed, low altitude, small unmanned aerial vehicles (UAVs) for use as ship decoys. UAVs that are launched from a container on a ship's deck and that carry electronic payloads away from the ship's location can provide highly effective offboard countermeasures against hostile radars and radar-guided weapons. By creating realistic appearing radar images of the ship, these UAVs can decoy the incoming missiles.
 
 

  

Figure 11: The FLYing Radar Target (FLYRT) design configuration (Figure reproduced from NRL Publication 217-4810.)

  

AIMS has been tasked with a multiyear effort to evaluate the fidelity with which these decoys mimic the ship and hence confuse the threat. AIMS' proprietary tools, e.g., the Learning Vector Quantization (LVQ) classifier, have been used as the benchmark discriminator. Considering ever increasingly complicated threat scenarios in which missiles may employ pulse to pulse frequency agility or integration schemes, AIMS engineers are devising experiments to assess decoy effectiveness as well as to determine additional degrees of freedom which may need to be built into decoy designs.

Naval Research Laboratory / Department of the Navy Modeling Support for Unconventional Techniques for Anti-Ship Missile Defense A difficult threat for ship EW systems is the antiship cruise missile (ASCM). Typically, such threats fly at a low altitude and with low probability-of-intercept (LPI) emission signatures. This often results in late detection and little time to respond. This effort investigates a new defense technique which attacks a key ASCM sensor.
 
 
   

Figure 12: AIMS has introduced the powerful methodology of control system synthesis and analysis under energy and observability constraints to this important ECM problem.

  

As part of this effort, AIMS has been tasked by the Navy to develop distributed sea clutter and multipath models for radar simulation. As opposed to the point scattering models these distributed scatterer models have been demonstrated to be more realistic. Based on sound principles in electromagnetics, these models also employ stochastic features to emulate the temporal characteristics of monostatic as well as bistatic backscatter from ocean surfaces. Working code has been delivered to the Navy, and preparation of a software package is in progress which shall operate with the accepted nomenclature of environmental parameters such as wind speed and sea states.

National Security Agency CELP Speech Compression With the advent of multimedia communications and the continuing integration of communications networks with computers, software, and databases, the compression of voice data is crucial for these rapidly advancing technologies not to be overwhelmed by the volume of the data involved. Dr. J.S. Baras, Chief Scientist of AIMS, has co-invented a record-breaking, real-time speech-coding algorithm (patent pending) that achieves toll-quality speech reproduction at a rate of 4.8 kilobits/second with below 5 MIPS complexity, which is 25% of the complexity of current methods having comparable fidelity. AIMS engineers are at work with the National Security Agency on improvements to the algorithm and a prototype hardware implementation using low-cost digital signal processing chips. The potential applications for this product are far-reaching and include the following:

• Digital answering machines.
• Compressed storage of audio ``help files" for software annotation.
• Automated speaker-recognition.

Raster-to-Vector Document Conversion The vast majority of the world's existing collection of engineering schematics, architectural blueprints, scientific diagrams and maps are only available on paper. At the same time, there is an exponentially increasing trend away from hardcopy formats that is to a large extent due to the widespread use of powerful workstations and software for Computer-Aided Design (CAD). As a consequence, there is a pressing requirement to transfer hardcopy graphics data to magnetic media and to do so using data formats which are readable by CAD software programs.

The purpose of raster-to-vector conversion is to make the transfer of drawings from paper to CAD-suitable formats. This means that the data must be transferred as ``vectorized" collections of geometric objects (lines, circles, etc.) and character data rather than simply binarized or grey-scale raster images. This type of conversion will make the converted drawings usable by CAD, database, and data transmission software, and will thus make the conversion more than an archiving process, but a true transition of the drawings to the CAD environment.

AIMS has identified a number of technical areas in which we can make significant contributions to the raster-to-vector conversion problem. We offer superior filtering algorithms which can greatly enhance the conversion process. These include both weak continuity and morphological multiresolution filtering. AIMS engineers have also developed fast algorithms for the creation of general masks, using parameterized curves, B-splines and combinatorial composition techniques, as well as new, fast algorithms for zooming and resizing images. We also have the latest algorithms for character and geometrical object recognition in highly noisy environments which substantially augment or supersede the Hough Transform.