This paper presents an approach for registering images obtained from sensors tuned to significantly different regions of the electromagnetic spectrum. Standard correlation techniques for registration are not appropriate because of the inherent differences between the images. We present an approach which uses surface boundaries (obtained at multiple scales via an edge extraction process) to determine the affine transformation required to bring the images into registration. The affine parameters for registration are found by a hierarchical estimation process. A least squares gradient based technique is used to directly estimate the affine parameters from the blurred edges. Results of applying the algorithm to real image data are included.

A major consideration in designing imaging systems that will be placed on moving platforms is the stability of the image. For example, the tracking or cuing of small objects requires a very steady image sequence. To provide the necessary stabilized platform typically requires the use of an inertial stabilized gimbaling system. Such gimbaling systems are both expensive and bulky. These factors are becoming increasingly important to the military community. The cost of a gimbal system can preclude its use on disposable ordinance systems and can be a significant portion of the cost of such systems. A low cost and low bulk alternative is needed. This paper describes a method for performing electronic stabilization using optical flow techniques. Electronic stabilization allows either the elimination of the inertial stabilized platform or the use of a low cost gimbal. The operating scenario available to electronic stabilization is examined and the resulting basic system requirements are derived. A model- based approach for deciding sensor motion is described. The utilization of the motion information for performing motion compensation is presented. Finally, preliminary results are given.

Passive navigation and automatic guidance using JR image irradiance tracking present a new challenge to computer vision due to a wide variety of 3D structural discontinuities and occlusions occuring in real-world scenario [1], [2], [3]. Time-varying images of real-world scenes can provide kinematical, dynamical and structural information of the world. To estimate from the image sequences, the 3D motion and structure of objects it is often necessary to establish correspondences between images. Establishing correspondences between different perspective images of the same scene is one of the most challenging and critical step in motion and scene analysis. Several approaches to the passive navigation problem have been suggested in recent literature [4]. Approaches for passive navigation using stereo image Pairs have also been reported. rllhis paper describe an image matching approach based on multiple attributes extracted from each image provide by an airborne FUR.

Under the Army Night Vision & Electronic Sensors Directorate and the Army Research Laboratory (S³I) sponsored ATR Relational Template Matching (ARTM) program, we have developed a novel model-based forward-looking infrared (FLIR) automatic target recognition (ATR) algorithm with superior performance over other model-based and template matching algorithms. The ARTM algorithm represents a significant advance over other approaches because its performance is relatively insensitive to the extreme variations of target signatures in FLIR imagery. The algorithm has been extensively evaluated in Army-sponsored competitive tests on low-depression ground target scenarios. The ARTM algorithm correctly recognizes target signatures at less than one-half the signal-to-noise level required by other approaches. Also the algorithm performs consistently over signature variations with up to 30 percent shape error.

A relational template matching ATR algorithm, a specialization of a more general algorithm design approach under the name of information-based complexity, has recently demonstrated promising results for simulated terrain board imagery. The merit of this approach is a design principle exemplified as follows. The estimation of auxiliary unknowns, e.g., boundary estimation, are not made prematurely on partial information such as low level image structure. Instead, the algorithm process considers an appropriate set of provisional boundaries which are economically represented by taking advantage of any structure present in this set. At each stage, the relative uncertainty of each boundary in the set is expressed by a probability. This uncertainty is refined by updating these probabilities by an intelligent choice of additional relevant information consisting of data and target-and-clutter model structure. This process is such that the optimal information about the auxiliary unknowns such as boundary is obtained simultaneously with the classification decision of the image region (e.g., ID). Thus, the design principle governs the intelligent choice of new information based on previous processing results, and economically represents any auxiliary information or unknowns by learning their hidden structure in the preprocessing or solution development stage. The objective of this paper is to describe the natural extension of this paradigm to multisensor imagery.

Boundary shape is a significant and well-modeled property of forward-looking infrared (FLIR) sensor target signatures. State-of-the-art FLIR automatic target recognition (ATR) algorithms that rely on signature shape for target detection perform well but still fall short of human performance and many DoD requirements. We find that internal signature information significantly improves detection performance. The problem is that this information is not easily modeled, especially in FLIR signatures, because the signatures exhibit significant variations dependent on a large number of unknowns. We have developed a model-driven neural network technique, called Programmed Constructive Neural Networks (PCNN), that demonstrates superior performance and generalization compared to traditional back- propagation techniques in high noise applications. We have used the PCNN technique to model internal FLIR signature information for clutter rejection. Our PCNN FLIR clutter rejection model eliminates 75% of the false alarms in a state-of-the-art shape-based algorithm with minimal detection loss. This result was achieved even on scenarios not represented in the training set.

This paper describes an approach to areas of FLIR target recognition: (1) target isolation, and (2) target classification. The method utilized for the isolation of potential target regions is based on localized texture information. The modality of the local gradient histogram is used to define both target regions and to segment these regions into subcomponents corresponding to the vehicle morphology (wheels, engine, armor, etc.). After the target regions are isolated, each region is fit with a metric (parallelogram). Each subcomponent in this region is then classified based on its shape and location within this metric. The classification is made using several neural networks with each corresponding to a specific vehicular subcomponent. The classifications of these neural networks are then used as input to another network responsible for vehicle type classification. This construct allows for azimuth and depression angle robustness of the target region, the limitations of which are discussed.

In this paper, we present a new ATR system for detecting and recognizing targets from a single IR image frame based on neural networks and Gabor functions. It uses Gabor functions to locate potential targets without prior knowledge about their type, size, and orientation. Neural networks are then used to remove false alarms and generate target identification based on information provided by Gabor functions. The new system combines Gabor functions and neural networks in a highly efficient way such that high recognition accuracy rates can be achieved under battlefield conditions. The new system has been successfully tested on hundreds of single frame IR images that contain multiple examples of military vehicles with different size and brightness in various background scenes and orientations, and very high recognition accuracy rates have been achieved.

This paper presents an expert system for target recognition from short distance infrared images. This expert system first identifies the position and orientation of the target. The number of possible positions and orientations is extremely large, therefore, the solution space cannot be fully explored. A heuristic search algorithm is used instead to guide the expert system toward the solution in an efficient manner. Bayesian estimation theory and fuzzy logics are used to derive the knowledge combination rules used by the heuristic search algorithm. These rules are generic enough to be applied to other types of expert systems and to data fusion problems. Once the position and the orientation of the target have been found, the expert system simply tries to match parts of the image of the target with templates under the same position and orientation in order to identify the target.

This paper describes a morphology-based hierarchical process for the detection and segmentation of low and high contrast targets in second generation FLIR imagery. The computational framework is based on the application of simple non-linear binary and grayscale operations that lead to real-time implementations. The process consists of two major processing steps: target-cueing/coarse-segmentation and contour refinement. Our multi-stage detection/segmentation process was applied to both real and simulated FLIR imagery. Preliminary results indicate that the developed morphology-based detector exhibits excellent detection performance for both low and high contrast targets in complex backgrounds while maintaining a low false alarm rate. Contour refinement is based on the watershed transform that is applied in a hierarchical fashion. In addition, our segmenter extracted accurate target outlines under poor conditions in which edge-based techniques or traditional watershed algorithms would have failed.

A fuzzy hierarchical FLIR ATR is proposed which more closely models the fuzziness in the FLIR data and the human decision process than the traditional ATR methods. The target and its internal hot spots are segmented out from the background by use of an iterative volume based morphological contrast peak extraction routine. The segmented regions are then represented by a set of silhouettes for each segmented blob rather than just the one `best' silhouette. For the target or foundation segment, the primary recognition feature, silhouette shape, is captured by the low frequencies of the 2-D DFT of each member of the set. The hot spots are represented both by the shape features (DFT) and by positional features. The first level of this hierarchical classification system uses an Euclidean distance figure of merit for the foundation's silhouette to assign a fuzzy classification to the target. This initial guess is then adjusted based on the internal features.

A new algorithm which can skeletonize both black-white and gray pictures is presented. This algorithm is based on distance transformation and can preserve the topology of the original picture. It can be extended to 3-D skeletonization and can be implemented by parallel processing.

This paper discusses an approach for applying IR target modeling to aid model-based automatic target recognition (ATR) algorithms. The paper also presents results based on experiments with real long-wavelength IR data. The algorithm uses an IR thermal prediction model to approximate the (long-wave IR) expected target signature. The algorithm then uses the predicted signature or some features based on it to locate or classify the target within the image.

The detection and identification of vehicles in forward looking infrared (FLIR) images presents a number of problems. Not only is the view of the vehicle influenced by its orientation and the position of the sensor, but it is also affected by the environment. To counter these effects, we have developed a set of algorithms which use models of both the vehicles of interest and the environment to aid detection and identification. This algorithm set consists of an area-of-interest locator, an object segmenter, and a template matcher. These algorithms are controlled by a process which uses information about the vehicles and the environment to select input parameters for the processing algorithms. This article contains a description of the detection and identification system algorithms, the performance characterization of the individual algorithms, and the temporal fusion performance prediction results.

This paper presents new techniques for the texture classification of regions based on edge co- occurrence matrices and discrete Hermite functions which are used to describe them. The paper briefly defines co-occurrence matrices and how they can be used to describe the relationship of edges around a pixel. Texture is interpreted as a measure of the edginess about a pixel and is described by edge co-occurrence matrices. The texture of the region is characterized by an orthogonal decomposition of the co-occurrence matrix using 2-dimensional discrete Hermite functions. The coefficients of this decomposition provide a low order feature vector which can be used for texture classification. The coefficients of the Hermite functions used in the decomposition of the co-occurrence matrix are analyzed by two neural network classifiers: the multilayer perceptron and the cascade correlation. Experiments have been performed for the training and validation of the networks on two types of terrain (grass and trees) taken from FLIR images during a low level approach to a bridge.

This paper describes an approach for employing an Assumption Truth Maintenance System (ATMS) to support a tri-service algorithm development environment. A truth maintenance system provides a mechanism for tracking assumptions and logical propositions, and recognizing contradictions. The higher order logic embodied by the truth maintenance system adds value and functionality to the tool set for applications programming. The first section of this paper describes an implementation of an ATMS in the Army's MAXIMIZE (matrix for algorithm exploration, optimization and evaluation) system. The latter part of the paper describes preliminary results of an integrated truth maintenance system to assist image understanding.

It is widely recognized that processors developed for many different uses in the near future must be faster, more versatile, easier to use, less costly, and -- in many cases -- smaller than those currently available. The required increase in functionality is associated with many factors, including: (1) the increasing complexity of the algorithms being developed for various applications (e.g., aided/automatic target recognition -- ATR); (2) the development of larger sensor arrays for both focal plane arrays (FPAs) and synthetic aperture radars (SARs); (3) real-time performance needs for processing-intensive systems deployed on a wide variety of platforms; (4) more-stringent requirements on sensor resolution; (5) the need for effective sensor fusion to handle data available from multiple sources; (6) increased intra- and inter- platform communications; and (7) hardware/software development and implementation costs. This paper discusses ongoing and required work related to component development and electronics packaging critical to meeting future processing requirements within reasonable size, weight, and power constraints.

The Signal Processor Packaging Design (SPPD) program was a technology development effort to demonstrate that a miniaturized, high throughput programmable processor could be fabricated to meet the stringent environment imposed by high speed kinetic energy guided interceptor and missile applications. This successful program culminated with the delivery of two very small processors, each about the size of a large pin grid array package. Rockwell International's Tactical Systems Division in Anaheim, California developed one of the processors, and the other was developed by Texas Instruments' (TI) Defense Systems and Electronics Group (DSEG) of Dallas, Texas. The SPPD program was sponsored by the Guided Interceptor Technology Branch of the Air Force Wright Laboratory's Armament Directorate (WL/MNSI) at Eglin AFB, Florida and funded by SDIO's Interceptor Technology Directorate (SDIO/TNC). These prototype processors were subjected to rigorous tests of their image processing capabilities, and both successfully demonstrated the ability to process 128 X 128 infrared images at a frame rate of over 100 Hz.

Texas Instruments (TI) is developing the Aladdin computer under contract with the U. S. Army Communications and Electronics Command (CECOM). The program is sponsored by the Defense Advanced Research Projects Agency (DARPA) and the U. S. Army Night Vision and Electro-Optics Directorate (NVEOD). Processors currently available for today's advanced weapons systems are limited in their real-time processing capabilities and are generally specific to a selected mission. The lack of availability of a high performance, general purpose, programmable processor in a small volume applicable to a variety of weapons systems applications creates a high non-recurring development cost for each new program. Automatic target acquisition systems are specifically in need of processors that meet the required real- time processing throughput of a variety of algorithms within very restrictive volume constraints. The objective of the Aladdin program is to develop a very high performance miniature processor that fits within a 75 cubic inch cylindrical volume, and is easily programmable to provide the ability to detect, recognize, identify, and locate the optimal aimpoint of a target or targets.

Alliant Techsystem's Aladdin processor is a real-time automatic target recognition processor that provides 2 GFLOPS of 32-bit floating point operations and 1 GOP of 16-bit fixed point operations simultaneously in a soup-can size package only 4.5 inches in diameter by 6.0 inches long. It is highly modular, and can be scaled up to 50 GOPS + 8 GFLOPS to satisfy target recognition and tracking needs using sensors as large as 2048 X 2048 pixels suitable for helicopters and aircraft, or down to 66 MFLOPS on a 2 square inch multichip module for a minimum hardware configuration suitable for missiles and smart munitions. A 6U VME chassis configuration is also available. All hardware configurations are Ada- programmable and share a common Ada software development environment and Ada runtime system for cost-effective life-cycle operation. This paper describes the hardware architecture of the processor along with the Ada runtime system, the software development environment, and the software programming methodology. Typical system applications are included to illustrate the versatility and cost effectiveness of the processor.

A heterogeneous parallel-processing computer architecture is being developed for embedded real-time interpretation of images and other data collected from sensors on mobile platforms. The Advanced Target Cueing and Recognition Engine (ATCURE) architecture includes specialized subsystems for input/output, image processing, numeric processing, and symbolic processing. Different specialization is provided for each subsystem to exploit distinctive demands for data storage, data representation, mixes of operations, and program control structures. The characteristics of each subsystem are described, with the Image Processing Subsystem (IPS) used to illustrate how the design is driven by careful analysis of current and projected computational requirements from many applications. These considerations led to a programming model for the Image Processing Subsystem in which images and their subsets are the fundamental unit of data. The processor implementation incorporates a scalable synchronous pipeline of processing elements that eliminates many of the bottlenecks found in MIMD and SIMD architectures.

The advent of highly advanced multi-chip module (MCM) technologies makes it feasible to consider higher performance computer architectures optimized for a particular packaging approach. In particular, the High Density Interconnect (HDI) technology and related patterned overlay technologies allow for high bandwidth topologies to be constructed which are beyond the reach of many other packaging technologies. Through the use of the recently developed three-dimensional extensions of the HDI process, even more aggressive computer designs can be undertaken. Variations of the physical topologies that can be constructed with the two- and three-dimensional HDI processes provoke new paradigms in future high-performance computer construction.

The GE Distributed Application Environment (GEDAE) is a graphical programming tool that facilitates the development of parallel distributed data flow applications on a heterogeneous network of processors. GEDAE provides a simple Motif interface, allowing the user to create, edit, control, and monitor hierarchical data flow graphs. GEDAE's simple programer interface makes it possible to easily create primitive function boxes by encapsulating existing software in a C language shell. GEDAE maintains much of the efficiency of special purpose code while hiding the details of network interconnection and data flow. GEDAE's data flow control software automatically takes advantage of any data parallelism or pipelining inherent in the constructed data flow graph. Data flow is efficient because objects cast or transfer themselves differently, depending on the type of data transfer, to optimize communications efficiency. An autocode generation capability provides the ability to automatically generate source code for multiprocessors and to efficiently support fine-grained data flow graphs. Currently, GEDAE is being enhanced to allow applications developed in a workstation environment to be mapped to embedded processors, for efficient execution.

The development of large-scale systems with real-time characteristics repeatedly exceeds budget bounds and schedule constraints. The problem is exacerbated with the implementation of distributed computer architectures. Nonlinearities and blockings are injected into system performance that seldom had to be addressed before the advent of parallel and distributed processing. To assist in the design phase of system development, Science and Technology Associates, Inc. (STA) has devised a paradigm which consists of three primary components; these are representation, optimization, and assessment. The representation component is an object oriented scheme that encapsulates resources, behavior and the connectivity within a design. Optimization allows the designer to automatically allocate system work, such as software processing, to resources in an optional manner. Assessment provides quantitative guidance on the performance of each system element. By rapidly iterating the tool triad, a feedback control process for design occurs which verifies system logic and performance prior to prototyping or the writing of software code.

This paper discusses a practical solution for supporting the deployment of data flow graphs onto the Loral/Rolm Computer Systems, Inc. vector processing multi-processor architecture. It outlines the support software (both workstation hosted and target system hosted) that is required to design, debug, and maximize deployed data flow graph performance on the multiprocessor architecture. The deployment process guarantees real-time deadlines, minimizes run time scheduling overhead, and minimizes designer partitioning input. It is known that determining effective run time data flow graph node schedules for multi-processor architectures is an NP-complete class of problem not well suited to real-time systems. Loral/Rolm Computer Systems, Inc.'s vector processing toolset recognizes this problem and this paper discusses a prescheduling and pre-assignment approach for partitioning data flow graphs to available hardware resources. In particular the toolset components (which are based upon an enhanced data flow graph language) of workstation pre-assignment, prescheduling, run time gross allocation and local compute element dispatching are discussed in detail.

We present a novel approach for implementing and optimizing an Automatic Target Recognition (ATR) algorithm for Synthetic Aperture Radar (SAR) imagery using the Princeton Engine (PE), a general purpose massively parallel single instruction multiple data (SIMD) machine. This approach was developed in the Algorithm Understanding Laboratory (AUL), a unique facility which is chartered to assist algorithm developers through high-speed implementation and near real-time visualization, and is located within the National Information Display Laboratory (NIDL). The PE architecture automatically provides a high speed-up directly proportional to the width of the image being processed, thereby reducing the train/test cycle times of ATR algorithms from days and hours down to minutes. Given this speed-up, the user can now train the system to classify a set of objects and then test it rapidly, thus tightening the train/test loop. With our approach, one can operate on the entire image, retaining useful image information until the very last stage in the algorithm.

The high-resolution direction-of-arrival (DOA) estimation algorithms are studied to develop architecture for real time applications. Methods for DOA estimation for wideband sources proposed by Buckley and Griffiths and MUSIC algorithm for narrowband sources proposed by Schmidt have been selected for hardware implementation. These algorithms have been simplified and generalized into one common programmable algorithm. It is then parallelized and is executed in a pipelined fashion. A parallel architecture has been designed for this generalized algorithm.

The use of model-based approaches in the detection and identification of man-made objects is encountered frequently in the literature today. These approaches generally depend on very high resolution imagery, and must first be cued to an approximate location of these objects. The approaches are also highly computationally intensive and, therefore, cannot be relied upon to perform broad sweeps in either a real time scenario or during training stages. It is therefore necessary to preprocess imagery with techniques that are less expensive in computer processing and more general in their approach to detection. With these requirements in mind we examine a modified form of the chord transform in the detection of man-made objects in an image. In addition, since the chord transform is an O(N⁴) algorithm, we explore the possibility of a parallel implementation of this approach in a SIMD architecture. The resulting mechanism is capable of quickly identifying straight lines, right angles, parallel lines, and arcs in an image. These primitives are indicative of man-made objects in an image.

Focal plane array (FPA) families for a new generation of Forward Looking Infrared (FLIR) systems are currently being developed under the direction of the U.S. Army. The new FPA will have an impact on the performance of Automatic Target Recognizers (ATRs) which perform target detection and identification. Described in this paper are the results of an experiment designed to study the effects of various infrared FPA parameters on the image quality of a FLIR. The research was conducted using a computer simulation of a FLIR imaging sensor without the use of specific ATR algorithms. The study determined the amount of image degradation and information loss caused by specific FPA parameters. Among the parameters studied were detector size, detector geometry, noise, detector response nonuniformity, and array sampling effects, most of which were investigated at three different ranges.

In this paper, different methods are presented for the detection of important directions on thermal infrared images. These methods are compared to the well-known Hough method which performs poorly for these types of images. These algorithms are especially designed for an automatic target recognition system, but could as well be used for other types of application.

In pattern recognition systems the basic task is to match scenes obtained by a sensor at different times. Over the years several algorithms were developed for automatic pattern recognition systems for target acquisition. How good are they? Why do they work? Why do they not work under a different environmental condition? These are questions to which we seek answers. In this paper, a methodology is developed to characterize algorithm performance when subjected to scene geometrical distortions to give insight into the questions posed above. The methodology used is a parameter variation techniques where by perturbations in sensor range and roll angle are made and target probability of detection and location determined. In this investigation actual infrared images are used. The sensitivity analysis methodology developed is illustrated with an example. Application of the generic methodology developed can serve as a very useful tool in testing and evaluating automated pattern recognition systems used in industrial, scientific, commercial, medical, civil defense, and national defense applications.

The paper presents a simple but effective feature assessment scheme that can be employed for a quick optimal evaluation of the individual discrimination potentials of a large number of features. The approach, Class Overlap Region Partitioning Scheme (CORPS), can be used either as a stand alone tool or as a front end to more complex combinatorial feature selection procedures such as branch and bound and genetic algorithms. The approach has the flexibility to permit imposition of a bias on the evaluation in favor of reducing either of the two possible types of errors in a binary decision process, for example false alarm or leakage in a target detection problem. Details of the associated algorithmic and operational procedures are furnished to facilitate wide usage of this new tool.

We present a new method for automatic thresholding of thermal images when the target is a landing aircraft. The raw thermal image is subjected to clip-low operations based on a sequence of threshold pairs. Each threshold pair is derived from the median and upper hinge (in Tukey's sense) of the sub-sampled raw image. The inter-centroid distance for each resulting image pair in the sequence along with the corresponding sample population are used to select the desired threshold. Experimental results on real FLIR imagery are presented.

Matrix specialized computer is described in briefand the approach to its design is analyzed as well implementation with application of HIC and the ques-tions of designing of highly efficient software. Extension of a number of problems to be solved at present by the systems, operating in real time mode, makes it necessary to form qualitatively new computational structures, providing capacity at 1-2 orders higher than capacity of all-purpose computers. Among such tasks are algorithms, producing single-type processing over larger size of information in real time mode . Formation of computer aids for solution of this task is rather complicated . Hereinafter one of the solutions of this problem is proposed.