Acoustic signals scattered from submerged elastic targets have produced interesting back- scattered signals in the appropriate frequency region, particularly because of the presence of resonances. Considerable work has been done for spherically and cylindrically shaped objects out to moderate frequencies for both solids and shells. Only recently, however, have researchers been reporting on results for elongated objects such as spheroidal solids and shells. Further, the higher frequency region for spheres has not been investigated extensively. This paper reports on a series of calculations that the authors have been involved with that are characteristic of elongated targets or of spherical shells at the higher frequency region. Some of the results that are discussed involve bending resonances due to obliquely incident plane waves on elongated objects and their relation to flexural resonances of beams (classes of resonances that are related to waves creeping along the longest and shortest meridians of a spheroid) flexural resonances (A₀ resonances) for shells in the time domain at coincidence frequency, and high-frequency thickness effects when scattering from shells and their prediction based on flat plate theory. Several numerical examples are shown.

This study focuses on the multidimensional inverse scattering of objects buried in an inhomogeneous elastic background structure. The medium is probed by an ultrasonic force, and the scattered field is observed along a receiver array. The goal is to retrieve both the geometry (imaging problem) and the constitutive parameters (inverse problem) of the object through an appropriate multiparameter direct linear inversion. The multidimensional inverse scattering problem being nonlinear and ill-posed, it is linearized within the Born approximation for inhomogeneous background, and a minimum-norm least-square solution to the discretized version of the vector integral formulation is sought. The solution is based on a singular value decomposition of the forward operator matrix. A priori information can be incorporated into the algorithm to enhance the accuracy and improve the resolution of the recovered object characteristics. The method is illustrated on a 2-D problem where constrained least-square inversion of the object is performed from synthetic data. A Tikhonov regularization scheme is also examined and compared to the minimum-norm least-square estimate.

The solution of an inverse scattering problem is defined in this paper as the extraction of physical information (viz., composition and shape) from the (active) echoes returned by an underwater elastic target. This pertains to the general area of target identification. In this sense, to solve an inverse scattering problem may be relatively simple in the low- or the high- frequency regions of the target's spectra, because of the number of simplifying approximations that can usually be made in these two spectral regions. However, for the intermediate or resonance region of all scatterers, the above mentioned approximations usually do not hold, and either exact solutions or novel types of approximations are required. This makes the inversion of the scattered waveforms extremely hard. A certain level of success has been achieved by exploiting the presence of certain resonance features that manifest themselves in the resonance region of the scattering cross-section (SCS) of submerged elastic targets. Although the main interest over the years has been the isolation or extraction of the resonance features contained within the SCS of a penetrable target, little has been said about what to do with these resonances once they are isolated. This resonance extraction process can be carried out theoretically, experimentally, and computationally, by means of either CW or pulsed interrogating waveforms. The paper shows some (idealized) examples to illustrate how the resonances, once they are isolated by either means, can be associated to the physical characteristics of the target that returned the echo, thus allowing for its unambiguous characterization. Means are discussed to extend the procedure, here illustrated with acoustics and sonar, to any active return from any type of sensor.

A neural network (NN) approach for the continuous recognition of undersea targets from side- scan sonar returns is presented. This approach takes more than one return at a time, thereby enabling the network to learn the spectral temporal dependencies of consecutive returns. The horizontal distance scanned at a time could be expanded by increasing the number of groups in the input layer of the network. Also, a simple and powerful stochastic algorithm called Alopex is described for the training of the network. The studies and results with approach are described.

Back-scattered echoes yield interesting information when scattering from elastic targets, particularly in the resonance region. Generally source and receiver are at the same location for a fixed orientation. If adjustments of the orientation are made over a suitable angular region then it is possible to ascertain symmetries of the target if they exist, particularly if the frequencies are varied over a suitable range of resonances. If observations of axial symmetry are made then it is possible to obtain both the dimensions of the object and the aspect ratio of the object (ratio of length to width). Once this information is obtained it is also possible to obtain information on whether the target is a solid or a shell, and in some cases what the shell thickness is as well as the material properties of the target. This is assuming that the target is in a 'free' environment; that is, the boundaries of the target are not a factor in calculation. Time domain responses for specific pulse types also yield information, and it is easy to see how a series of questions can form the basis of a scenario that can rule out certain targets or lead to a probability (confidence level) that specific targets are present.

Research has been directed toward the goal of identifying objects in a stratified ocean-elastic sediment environment. An elastic object, (called inclusion) is located in the ocean. The mathematical procedure is based on the physical idea of scattering one or more 'plane waves' off the unidentified inclusion and then trying to identify the shape of the inclusion from its far- field pattern. Based on a reciprocal identity for the acoustic-elastic waves in the stratified medium, a relation between the acoustic waves in the far-field and the shapes of the inclusion has been developed. This relation is used to obtain an approximate algorithm to reconstruct the shape of the inclusion.

This paper reports a study of three-dimensional angular distributions of scattering strength associated with the flexural and Rayleigh resonances of a solid elastic spheroid, using the extended boundary condition method. The phenomenon of focusing of scattering strength into a disclike pattern caused by an elongated rigid spheroid is also investigated. Flexural resonances are found to produce strong scattering into four almost symmetrical lobes, as much as 20 dB above the nonresonance background. Rayleigh resonances are weaker, but significant portions of the scattered energy are directed into steep angles. The significance of this, and of the disclike scattering pattern due to the elongated rigid spheroid in waveguide situations, is discussed.

Present and emerging technologies for communication, remote sensing, object tracking and identification, nonintrusive inspection of materials, etc., place increasingly greater demands on wave data gathering, data interpretation, signal processing, and predictive modeling. While large-scale computing is inextricably coupled into this process, computing alone cannot address problems of classification and identification (data inversion) without a parametrized model. With the present toward wideband signals and higher frequencies, numerically large problems in various applications and disciplines (on the scale of relevant wavelength) are encountered where previously they were not when the bandwidth was limited and the frequency moderate. In this paper, it is argued that an effective parametrization of a complex propagation, interrogation, or identification channel between a source and a receiver should be structures most effectively around 'observables' in data. Such observables are prominent features like spikes, dips, periodicities, relatively smooth oscillations, etc. In an observable- based parametrization (OBP), one seeks to relate these features to various fundamental wave processes that are operative in the signal-environmental (target) interaction. Observables in data are generated by such wave processes and by interference between them as problem parameters (frequency, aspect, time) vary. At high frequencies, a propagation and/or scattering environment may have relatively extended regular portions (or relevant wavelength scales) interrupted by more localized (in terms of wavelengths) strong perturbations that connect with other relatively regular portions. In this way, one may decompose the environment into a regime of wavelength-dependent propagating regions and scattering or coupling regions, which are assembled into a self-consistent system. By OBP, one attempts to construct wave objects which are 'good propagators' in regular regions, i.e., they describe the physics of the relevant propagation mechanism as directly as possible. If the system is such as to cause multiple scattering of propagating (traveling) waves, then these multiple effects can often be expressed 'more physically' in terms of oscillatory (standing wave) events, either in a cross section of a layered configuration, in which they synthesize traveling guided modes in the layer or throughout a propagation volume, in which case they synthesize resonant modes. Because of the fundamentally distinct and complementary roles played by traveling and oscillatory wave events, these events, and their self-consistent hybrid combinations, are cornerstones for OBP.

This paper considers the problem of clutter segmentation in fully polarimetric, high-resolution, synthetic aperture radar (SAR) imagery. The goal of segmentation is to partition an image into regions of homogeneous terrain types (grass regions, tree regions, roads, etc.). Three approaches to segmentation are examined: (1) the optimal polarimetric classifier, (2) the optimal normalized polarimetric classifier, and (3) the polarimetric whitening filter (PWF) classifier. Segmentation performance results are presented for typical high-resolution, polarimetric SAR data gathered by the Lincoln Laboratory 35-GHz airborne sensor.

An algorithm is formulated for automatic real-aperture radar target recognition that is self- adaptive with respect to both the types and the number of backscatter measurements required for an optimum classification decision. The performance of the algorithm is evaluated by means of Monte-Carlo simulation studies using cross-section data for sets of five commercial aircraft obtained from a compact radar range. Analysis is made of the resulting performance of the adaptive algorithm and compared to other techniques. Results show significant performance advantages through the use of adaptive, sequential techniques.

The scattering interaction of short electromagnetic pulses with a spherical target is studied. The target is assumed penetrable and is modeled as an air-filled dielectric shell. The radar cross-section (RCS) of such a target is obtained and its resonance features are analyzed. A dielectric composition makes the resonance features become very prominent compared with the case of an ideally conducting sphere. When the interrogating waveform is a pulse of short duration, the resonance features of the backscattering cross-section can be extracted within the frequency band of the incident pulse. To verify theoretical predictions, spherical targets were illuminated with short broad-band pulses using an impulse radar system. The actual shape of the pulse that is incident on the target is theoretically modeled using a digital filter design techniques, and the predicted backscattered returns of spherical targets are compared with selected echoes of the pulses transmitted by the impulse radar. The authors verify that the shape of the predicted backscattered pulse that results from the design method agrees well with the experimental findings using metal spheres of three different sizes. By means of an incident pulse of designed shape, the form-function in the backscattering radar cross-section of a dielectric target is predicted using a discrete Fourier transform (DFT) technique. It is shown that many of the resonance features of a dielectric spherical shell can be extracted from the frequency band of the incident pulse employing this method. The methodology that is developed can handle broadband pulses of any sufficiently smooth spectrum, interacting with (lossy or lossless) dielectric scatterers, and can extract the resonance features within the frequency band of the transmitted pulse. Accordingly, this methodology could also be used for assessing the performance of high-power impulse radar systems.

This paper considers the classification of radar targets using features extracted from the bispectrum of the backscattered signals. The classification performance of the bispectrum- derived features is compared with that of features extracted from the impulse response and the frequency response of the unknown target. In each case, the classification performance is evaluated using compact-range radar signal measurements of a set of five commercial aircraft. A number of scenarios of noise environment and azimuth ambiguity are considered. The case where extraneous scatterers are present in the vicinity of the measured target is also examined, and the effects of the extraneous scatterers are discussed.

Recent developments make it possible to radiate and coherently detect electromagnetic pulses consisting of a few half-cycles of a sine wave having a period on the order of 10 ps. The antennas involved are compact, typically consisting of conducting films on thin dielectric substrates having lateral dimensions of about 2 cm. Short, broadband pulse technology can be used for position sensing, object identification, and for monitoring the state of an object. Such target interrogation may have application to robotics. This possibility is explored here, and an example of the dependence of the scattered pulse on the internal state of an object is presented. The short electromagnetic pulses (SEP) permit position sensing of objects with spatial resolution on the order of 1 mm. They also permit identification of objects and the monitoring of their state. This follows because, depending on its size, shape, and the materials from which it is constructed, an object will exhibit a series of resonant frequencies at which it strongly reflects electromagnetic waves of harmonic time dependence. When illuminated by SEP, having frequency content spread over a wide bandwidth, these resonances will be excited. The resonances appear in the late-time tail of the scattered pulse. This late-time response is almost independent of the orientation of the object, and can therefore be used for identification. For targets that involve conducting bodies with aperture coupling to internal cavities the externally observed resonances can be especially useful in monitoring changes in the internal structure. The illustrative example demonstrating this aspect involves a slit- coupled closed shell with an internal load. If the shell and load are good conductors it is shown that the resulting high-Q resonances are highly sensitive to changes in the locatio of the load.

This paper concerns the design, training, test, and evaluation of a feed-forward neural network for classifying acoustic signals emitted by ships in transit. The signals were collected by an omnidirectional hydrophone. Relatively noisy surface ships, moving rapidly at medium to long range, emit signals which superficially resemble those of quieter submarines, moving more slowly and closer to the listening device. The neural network approach is motivated by an obvious analogy to the sonar classifier of Gorman and Sejnowski, who trained a neural network to classify active sonar returns from two undersea objects. The present problem can be solved by a similar network architecture, the outputs indicating which target type (if any) is present. The inputs represent the evolution of spectral densities for each of a number of time lags. Yet the number of target types and encounter geometries is far greater than could possibly be covered in any representative way by a training set comprised of real-world data. Thus, the task is to connect the network to a high fidelity, model-based digital simulator and to show that, by training on the output of the simulator, the neural network can learn to pass realistic tests. This paper describes a neural network design-and-testing exercise based on a simplistic model that captures a few of the salient features of the problem.

Object recognition in laser radar sensor imagery is a challenging application of neural networks. The task involves recognition of objects at a variety of distances and aspects with significant levels of sensor noise. These variables are related to sensor parameters such as sensor signal strength and angular resolution, as well as object range and viewing aspect. The effect of these parameters on a fixed recognition system based on log-polar mapped features and an unsupervised neural network classifier are investigated. This work is an attempt to quantify the design parameters of a laser radar measurement system with respect to classifying and/or identifying objects by the shape of their silhouettes. Experiments with vehicle silhouettes rotated through 90 deg-of-view angle from broadside to head-on ('out-of-plane' rotation) have been used to quantify the performance of a log-polar map/neural-network based 3-D object recognition system. These experiments investigated several key issues such as category stability, category memory compression, image fidelity, and viewing aspect. Initial results indicate a compression from 720 possible categories (8 vehicles X 90 out-of-plane rotations) to a classifier memory with approximately 30 stable recognition categories. These results parallel the human experience of studying an object from several viewing angles yet recognizing it through a wide range of viewing angles. Results are presented illustrating category formation for an eight vehicle dataset as a function of several sensor parameters. These include: (1) sensor noise, as a function of carrier-to-noise ratio; (2) pixels on the vehicle, related to angular resolution and target range; and (3) viewing aspect, as related to sensor-to-platform depression angle. This work contributes to the formation of a three- dimensional object recognition system.

An image restoration method that performs piecewise smooth restorations on images corrupted with very high noise levels is presented. The restoration is based on a Markov random field (MRF) model using a neural network sigmoid nonlinearity between pixels to produce a restoration with sharp boundaries while providing noise reduction. The model is currently implemented using an efficient deterministic search algorithm that typically requires less than 200 iterations in a digital computer simulation. The algorithm is able to restore images with up to 71% of the pixels corrupted with sensor noise (non-gaussian). Results from simulation indicate that the MRF restoration is capable of operating at signal-to-noise ratios 5 to 6 dB lower than the median filtering. The same model can be applied to a large number of sensor measurements (Doppler, intensity, passive infrared, range, and video sensors) by simply adjusting a single parameter. This is especially relevant for hardware implementation, since a single chip can be used for processing a wide variety of imagery. The model has a massively parallel architecture with local neighbor interactions (four nearest neighbors), and can be implemented on a large parallel computer or a custom analog VLSI chip. Implementation of the model in analog VLSI would allow video rate restoration of 512 X 512 pixel images.

Prior methods for tactical target reacquisition after a loss of track have used tree classifiers and template matchers. Examples of prior techniques include classifiers that are trained with a priori data which makes them somewhat intolerant to temporal and dynamic changes in the target pattern. Prior methods for reacquisition generally rely on proximity and area-based schemes. The disadvantage of these methods include their dependence on accurate and consistent segmentation in cluttered scenarios and their need for precise target position prediction. The random mapping network algorithm (RMNRA) offers a solution using a pattern memory and sparse feature matching technique. RMNRA assists the imaging tracker and improves tracking tenacity by reacquiring a tracked target after a loss of track has occurred. The reacquisition algorithm uses an associative memory to perform target pattern matching. The pattern matching technique is unique in that it is tolerant to some of the ambiguities that occur with classical template pattern matchers. Weighted pattern feature vectors are stored in a memory matrix to facilitate the matching of sensed and reference patterns dynamically over time. In addition, a sophisticated algorithm was designed to update the memory matrix over time to forget prior patterns as the target signature becomes stale over time and space. The algorithm has been implemented in real-time hardware and flight tested with an infrared sensor. The algorithm is discussed and results using real IR imagery are shown.

The representation of visual information in human striate cortex is of significance to machine vision. Invariance to certain geometrical transformations in the field-of-view may be provided by the computational anatomy of the visual cortex. For example, there is evidence that the retino-cortical mapping is closely approximated by a log-polar transform. When combined with the foveation response, log-polar mapping can provide a basis for translation, rotation, and scale-invariant perception. There is also evidence that the visual system is sensitive to the spatial frequency content of its input. Although a Fourier transform is physiologically implausible, some authors have suggested its use for invariant object recognition because the magnitude of the Fourier transform is shift invariant. The Fourier transform magnitude operation followed by log-polar mapping can also provide a basis for translation, rotation, and scale-invariant perception. Both of these image-transform (feature mapping) algorithms give mathematical invariance to translation, rotation, and dilation. For an automatic recognition system, however, the feature mapping module has to be robust to discretization error, noise, and possible obscuration. Robustness considerations led to the development of the bi- directional log-polar mapping (BPM) algorithm. The BPM algorithm overcomes the pixel- dropout problems associated with conventional approaches to log-polar mapping. The authors evaluate several feature mapping models, both biologically and mathematically inspired, for their effect on recognition performance when embedded in a neural-network-based, object- recognition system. The modular recognition system, consisting of image restoration, detection, segmentation, feature extraction, invariant mapping, and classification, is being developed to classify objects in laser radar range imagery. Synthetic laser radar range images of four vehicles rotated in the field-of-view, scaled to various ranges, and corrupted by increasing levels of sensor noise are used for this evaluation. This study indicates that feature mapping based on the bi-directional log-polar map provides translation, rotation, and scale- invariant recognition capabilities as well as robustness to noise and discretization.

In this paper a high-order bidirectional associative memory (HOBAM) based image recognition system and a dynamically reconfigurable multiprocessor system that achieves real- time response are reported. The HOBAM has been utilized to recognize corrupted images of human faces (with hats, glasses, masks, and slight translation and scaling effects). In addition, the HOBAM, incorporated with edge detection techniques, has been used to recognize isolated objects within multiple-object images. Successful recognition rates have been achieved. A dynamically reconfigurable multiprocessor system and parallel software have been developed to achieve real-time response for image recognition. The system consists of Inmos transputers and crossbar switches (IMS C004). The communication links can be dynamically connected by circuit switching. This is the first time and the transputers and crossbar switches are reported to form a low-cost multiprocessor system connected by a switching network. Moreover, the switching network simplifies the design of the communication in parallel software without handling the message routing. Although the HOBAM is a fully connected network, the algorithm minimizes the amount of information that needs to be exchanged between processors using a data compression technique. The detailed design of both hardware and software are discussed in the paper. Significant speedup through parallel processing is accomplished. The architecture of the experimental system is a cost-effective design for an embedded system for neural network applications on computer vision.

A hybrid electro-optic image processor has been developed for automatic target recognition using an acousto-optic correlator and digital electronics. The optical system performs the computationally intensive correlation operation on the large 2-D input scenes. The electronics provide the decision-making capability and also perform part of the normalization needed for increasing the peak-to-sidelobe ratio in cluttered scenes. The system is able to analyze each correlation plane and apply a real-time template selection algorithm to accommodate scale or rotation changes of the target. A demonstration of the current system capabilities is presented using a terrain board with several different types of stationary and moving model vehicles.

A new distortion-invariant optical correlation filter used to produce easily detectable correlation peaks in the presence of noise and clutter, and to provide better intraclass recognition, is presented. The new minimum noise and correlation energy (MINACE) filter produces sharp correlation peaks while maintaining an acceptable SNR in the correlation plane output in the presence of false targets (discrimination) and noise (clutter and background). The MINACE filter achieves this by use of a new tighter bound to the correlation plane energy problem. New test results are presented to show that this approach provides improved recognition (and reduced true class training set size), good discrimination, and improved performance in noise and clutter. Filter size requirements for such advanced distortion- invariant filters are also addressed.

Imagery from actual sensors was used for automatic object recognition with a binary phase- only filtering (BPOF) optical correlator. The correlator was primarily tested for applications that involve objects with a nonrepeatable signature. Digital image processing techniques were used to threshold gray-level images and approximate their boundaries with polygons to limit variations of input imagery before correlation. Correlation responses increased and worked best when relatively short line segments were used in the approximation. The best results were obtained when a convex polygon was drawn to the outward-most points of the object. This is equivalent to extracting specific shape features of the object. The method presented here reduced the sensitivity of a BPOF to changes in an object's appearance when an object varied in different or unknown ways.

The performance of the joint Fourier transform correlation (JFTC) algorithm is improved by enhancing the frequency domain horizontal edges for the case where the reference and input scene are in the lower half and upper half, respectively, of the space domain. The enhancement of frequency domain horizontal edges eliminates any vertical edges which corresponds to cross correlation between objects in the input scene. Consequently the correlation peak intensity between similar objects is increased and the cross correlation peaks are reduced.

An experimental model-based automatic target recognition (ATR) system, XTRS, for recognizing 3-D vehicles in real or synthetic, groundbased or airborne, 2-D laser-radar range and intensity images is discussed. The key to recognition is a new generic matching engine that compares image events (e.g., silhouettes) and their constituent primitives to some appearance-model (AM) hierarchy, which describes how 3-D--possibly articulated--objects appear in the imagery. The authors describe each of the system's components; i.e., the even characterization (for extracting events from images and decomposing the events into primitives), the models, the matching, and the control. XTRS may also perform low-level information fusion by constructing feature-indicating interest images and combining them into an overall interest image used for locating an event. Examples of processing and performance results are given for real CO₂ groundbased laser-radar imagery. The authors also report on a performance evaluation of XTRS based on synthetic data where importance parameters (e.g., range to target) are systematically varied. Overall, more than 1500 range and intensity image pairs have been used throughout XTRS's development.

A good deal of effort in developing ATR systems involves using component algorithms from existing systems. In many cases the developers modify existing algorithms and link them with newly developed ones in order to create a new system. A significant amount of time is also spent in performing evaluations and diagnostics on such systems. In the case of model-based ATR, the developers are always interested in enhancing the computation speed so they attempt various modifications and architectures at the system as well as the algorithm level. The proper tools, are critically needed that will allow ATR developers to perform rapidly and efficiently many of the above tasks. The model-based inconically reconfigurable object recognition system (MIRORS), which provides many of the tools for solving such problems, is described.

A method for object recognition using a multisensor model-based approach has been developed. The sensor algorithm research expert system (SARES) is a sun-based workstation for model-based object recognition algorithm development. SARES is a means to perform research into multiple levels of geometric and scattering models, image and signal feature extraction, hypothesis management, and matching strategies. SARES multisensor fusion allows for multiple geometric representations and decompositions, and sensor location transformations, as well as feature prediction, matching, and evidence accrual. It is shown that the fusion algorithm can exploit the synergistic information contained in IR and synthetic aperture radar (SAR) imagery yielding increased object recognition accuracy and confidence over single sensor exploitation alone. The fusion algorithm has the added benefit of reducing the number of computations by virtue of simplified object model combinatorics. That is, the additional sensor information eliminates a large number of the incorrect object hypotheses early in the algorithm. This provides a focus of attention to those object hypotheses which are closest to the correct hypothesis.

Automatic object recognition, with its diverse applications in numerous fields of science and technology, is permeating many aspects of military and civilian industries. This paper gives an overview of the issues confronting the automatic object recognition field and the approaches being used to address these issues.

With the projected increase in sophistication of potential threats and their associated countermeasures, the next generation of military operations cannot rely on a single sensor for threat information. The prudent choice of sensor suite in a given setting will avail a wider variety of statistically independent measurements, thus providing more resources for the tracking and classification of targets of interest. One basic problem in data fusion/target classification is to find reliable ways to correlate tracks originating from the same target using the attributes of individual sensor tracks. That practically all sensors provide some form of spatial track data accounts for the fact that spatial data fusion is the most commonly used data fusion technique. An algorithm for correlating all tracks from different sensors based on their spatial characteristics is presented in this paper. The technique used here is an extension of the multiple-hypothesis technique for tracking multiple targets in a cluttered environment. In this multiple-hypothesis correlation approach, all feasible correlation hypotheses are considered and maintained for at least a short period of time. The likelihoods for these hypotheses to be correct are evaluated and updated with the arrival of new data. The unlikely ones are periodically pruned with the most highly probable ones being retained. By using Kalman filtering techniques, the state estimates of each of the fusion hypotheses that survive have a smaller error covariance than any of the tracks from which they were derived.

Target detection theory is developed for 3-D pulsed imager operation of a coherent laser radar in a downlooking scenario. Generalized likelihood-ratio tests (GLRTs) and receiver operating characteristics (ROCs) are presented for range-only and joint-range-intensity processors. This work extends previous studies in three ways: (1) fine-range information is included; (2) maximum-likelihood estimation of an unknown range plane is performed; and (3) connections to Markov random field preprocessing are established.

The theory of track-while-image operation of a coherent laser radar was initiated by Enders, who considered the tracking of a spatially-resolved, speckle target using a sequence of intensity, range, or joint range-intensity images in an uplooking (background-free) radar scenario. The present paper extends Enders' work to a downlooking scenario, in which every pixel in the image frame contains a reflecting element--either target or background. Linear least-squares trackers using intensity, range, and joint range-intensity frame centroids are studied, with tracking performance established through Riccati-equation evaluations and computer simulations.

A low-power diode laser is used in a range sensor for spacecraft rendezvous and docking. The sensor consists of a transmitter, a retroreflector mounted on the other spacecraft, and a receiver mounted along with the transmitter. This is the first laser-based payload for Indian spacecraft.

This paper introduces a novel approach for target detection in synthetic aperture radar (SAR) imagery utilizing 3-D morphological filtering. The basic concepts in the 3-D morphology are described. The utilization of 3-D filtering to extract distinctive vehicle signatures from a SAR image is presented. A working detection system utilizing these techniques is described. A standard detection system is described and its performance compared to that of the system utilizing 3-D morphological filtering.

Low-level computer vision routines for simple-shaped (rectangular and circular) object recognition have been parallelized and implemented in order to run on a portable, standalone, multiple transputer network. The aim is to provide a flexible, though powerful, computational capability to an open architecture for a multispectral computer vision system applied to both visual and IR bands. Among the developed routines are: image registration routines, rotation, scaling and matching of the images taken from visual band and IR band cameras; filters, linear and NOS filters; transforms, Fourier and Hough. The parallelization of the routines allow them to be run on an arbitrary number of transputers, i.e., on the field with a limited number of transputers for dedicated online tests, or in laboratory for more precise results through intensive offline processing. The main application is the optical survey of masonry structures through the analysis of their surface from the moisture point of view (i.e., old historical buildings in the region of Venice). Recognition of geometrically and physically known objects, like thermal sensors and radiometric and geometric references, represents a crucial initial step toward these tasks.

Translational, rotational, and scaling invariant (TRSI) pattern recognition is a third-order problem encountered frequently in real-world applications. But neither traditional image processing/pattern recognition algorithms nor artificial neural networks have yet provided satisfactory solutions for this problem after years of study. Recent research has shown that a higher-order neural network (HONN), of order three with built-in invariances, can effectively achieve TRSI pattern recognition. For an N X N image, the memory needed to store the connections is proportional to N⁶. This huge memory requirement limits the HONNs application to large-scale images. To solve this problem the authors first adapt edge detection and log-spiral mapping algorithms to preprocess the image so that the problem is converted into a second-order one. This reduces the HONN memory requirement to O(N⁴). Second, the authors modified the second-order HONN architecture to further reduce the memory size to O(N²). Synthetic and real images with resolution 256 X 256 have been used for simulation. The training samples are noise free, and the testing samples are rotated, translated, scaled, or noise-corrupted versions of the training patterns. Simulation results show that this system can indeed achieve TRSI pattern classification. In addition, its high robustness to noise and pattern deformation makes it very useful for real-world applications.

This paper describes the results of using video feedback to support the calculation of the radon transform and angular correlation. Application of this type of optical processor to the extraction of features from synthetic aperture radar imagery of ships is described. By theoretical analysis and experimental evaluation of the optical architecture outputs (using some digital post processing) it is shown that the internal structure of objects, their primitive dimensions, and even their boundary can be obtained. In particular, the power of angular correlation to extract object length, width, area, aspect ratio, orientation and boundary from suitably thresholded images is shown. Alternative techniques to extract the object boundary based on angular correlation are discussed, including direct optical computation as well as digital deconvolution. Simple optical shapes, such as squares, rectangles, and triangles were used to initially compare the optical architecture outputs with digital baseline calculations. In addition, the test and evaluation of the optical processor using simple synthetic models of ship data is discussed. The results of a study that uses object primitives (derivable from angular correlation) in conjunction with the Radon transform (along the longitudinal axis) to classify ships using a backpropagation neural network also are described. A discussion of these results is presented pertaining to their broader application to optical parts inspection and to feature extraction from imaging infrared surveillance sensors. The practical implementation of such a processor in compact form using lenslet array optics also is discussed.

This paper outlines the use of space-filling curves in transform image compression. Specifically, a space-filling Hilbert curve is used for mapping the two-dimensional image into a suitable one-dimensional representation. Compared to simple raster-scans, this topological mapping is spatially non-disruptive and tends to preserve local pixel correlations in the original two-dimensional image. Standard transform coefficient reduction and coding techniques can then be applied to the one-dimensional representation for the purposes of data compression. The advantages of the one-dimensional coding, in terms of computational cost and subjective image quality, are also discussed.

Photointerpretation is a computationally expensive process. An opportunistic knowledge-based approach to the problem has been developed and proven effective for target recognition on the ground and in the air. The system hypothesizes about the identity of regions in an image based on their geometric and intensity properties. Each region looks for support for its interpretation based on its spatial relationships with other hypotheses. In this paper parallel implementation using a supporting spatial representation on an eight node Hypercube (a coarse-grained, local memory, message passing machine) is described. The algorithm would be equally suitable for a local area network of workstations. A quadtree data structure representing each image region can be compactly and efficiently coded into a few integers; hence, passing regions between computer nodes becomes a reasonable algorithmic device.

A new technique for recognizing two-dimensional objects independent of scale and orientation is presented. This technique's performance on real imagery of tactical military targets, both occluded and nonoccluded, is evaluated. This study shows that this method is robust with respect to partial occlusion. The relatively small storage requirements and fast search time make it an attractive candidate for real-time applications.

Boundary segmentation has long been a problem for automatic target recognizers. Its performance is crucial because it serves as the front end to the entire system. The authors examine and compare the characteristics and capabilities of four segmentors: the Boundary Contour System, the Meyer Line Finder, the Canny, and the Sobel Edge Detector. The first three models are 'smart' systems, that is, they have some 'higher level' processing capability, while the Sobel is a simple operator. In addition, the Boundary Contour System is neural based while the remaining three are conventional. The performance of each segmentor is evaluated with respect to the following image metrics: signal-to-noise, contrast, resolution, and the following boundary characteristics: spatial frequency, edge orientation. Both computer and terrain board modeled infrared imagery is used. Performance is quantified through both segmentation accuracy measures and visual fidelity.

The problem of edge detection in images has been to locate those pixels which contain an edge, i.e., a jump in the intensity level. Recently, a number of researchers have turned their attention to obtaining subpixel accuracy in estimating edge location. This work has generally focused on matching a parametric surface to the region in the image where an edge occurs. In this paper a two-dimensional maximum likelihood approach is used to obtain the most likely position and orientation of the edge. Comparisons show that the maximum likelihood approach gives considerably less error in locating the position of the edge over a wide range of noise conditions and blurring which may be unknown to the estimator.

The paper presents a new approath for edge nhanement. It onstruts an edge nhanernent's function which r1ies on grey information attached to th local area in the 2—dimensional image. This function can enhance dgs in the image and determine the direction of the edges .Mathematical analysis show that the commonly used operator of Sobel is a special case of this method. It's proved by the microcomputer image processing system that the method work well for a variety of images. In addition. it can be implemented by parallel processing and be easy implemented by hardware. The speed of implementation also demonstrates that it is practical for many applica tions.

Automatic target detection in IR imagery is a very difficult task due to variations in target brightness, shape, size, and orientation. In this paper, the authors present a contrast, size, and orientation invariant algorithm based on Gabor functions for detecting targets from a single IR image frame. The algorithms consists of three steps. First, it locates potential targets by using low-resolution Gabor functions which resist noise and background clutter effects, then, it removes false targets and eliminates redundant target points based on a similarity measure. These two steps mimic human vision processing but are different from Zeevi's Foveating Vision System. Finally, it uses both low- and high-resolution Gabor functions to verify target existence. This algorithm has been successfully tested on several IR images that contain multiple examples of military vehicles with different size and brightness in various background scenes and orientations.