The discrimination between metallic and nonmetallic targets submerged in water or even embedded in the sea bottom is of some concern for present sonar technology. It can be achieved by processing the backscattered sonar echo from monostatic measurements. Information on the target material and inner structure (rather than merely the outer shape) is obtained only if resonances of the target are excited by the sound pulse. The 'Resonance Scattering Theory' (RST) proves an 'acoustical spectrogram' to characterize the target just as optical spectra do with atoms, molecules, etc. This analogy applies, however, to underwater objects of very small dimensions only; with realistic, full size targets typically very few resonances can be identified due to the strong radiation damping which grows with increasing frequency. It causes wide overlap of the individual resonances so that spectral analysis does not yield a reliable decomposition in the presence of an even small amount of random noise. This has been verified by measurements in a fresh water tank and in a lake. Successful signal processing turns out to be possible, however, by combining resonance excitation of the targets with signal processing in a neural network classifier. Results obtained with power spectral or time series input data to the neural net will be presented. Cylindrical and (hemi) spherical steel shells and several stones have been used as targets. The results obtained so far are quite promising so that a reliable classification is likely to be possible even under more severe operating conditions.

In this paper, we combine the matched-field method with the boundary integral equation method from inverse scattering theory to study a sound source localization problem in a shallow ocean with an unknown large inclusion. We assume that there is an unknown inclusion embedded in a shallow water waveguide. To localize a continuous wave (CW) source, we send in a number of 'mode waves', which scatter off the unidentified inclusion and are received by a hydrophone array. Combining the information of these scattered waves and the signal from the point source, we present an algorithm to estimate the location of the CW source. A numerical simulation using this method is presented.

Active sonar classification of submerged elastic structures becomes increasingly difficult when the structure is close to the bottom or surface of the sea. The backscattering cross-section (BSCS) of any target, which is relatively simpler to determine in deep waters, away from boundaries, becomes substantially distorted as the structure approaches either one of these environmental boundaries. Near these interfaces the classification methodology based on echo resonances that we have used in the past (viz., Appl. Mechanics Reviews 43, 171-208, (1990)) can no longer be used. By means of the examples of a spherical shell and an elastic solid sphere insonified by plane waves, we study the above mentioned degradation in BSCS in order to assess how distant the structure should be from these boundaries before the resonance features become discernible again in the echoes, and object recognition is again possible. Our approach is based on the method of images for the construction of the appropriate Green's functions, combined with a very involved two-body scattering formulation that determines the combined T-Matrix of two insonified objects, when the T-Matrix of each individual object is known. The method is extended to the time domain. We present form-functions in the frequency domain, as well as late-time responses in the time domain for both sphere and shell as they approach the mentioned boundaries. Boundary effects seem to be confined to a 'skin layer' bounded by R <EQ 4. Within this layer the resonance features fade and are washed out in both the frequency and time domains. The formulation uses translation operators borrowed from atomic physics.

A two point boundary value method is developed for precisely computing eigenrays and travel times between specified points in a 3-D inhomogeneous transmission channel which has irregular bottom and surface bathymetry and inclusions. The method has application to replica signal computation for use in beamformers operating in complex ocean environments.

We present in this paper two methods of resolution of the Inverse Problem for axisymmetrical targets in Underwater Acoustics. What we mean by recognition is the assessment of the radius of the target and its mechanical properties (sound velocities c_L and c_T and density (Rho) ). The first method is derived from a parametrical study of resonances in cylinders at low frequencies. This study leads to approximate expressions of resonance frequencies x^*₁. These expressions exhibit for most resonances either a dependency of x^*₁ versus the longitudinal sound velocity or versus the shear sound velocity leading to a new classification of resonances taking into account the polarization of the waves involved. The behavior of the widths of the resonances leads also to simple expressions. The results presented will be generalized from the bulk cylinder case to more complicated targets. In order to solve the inverse problem and carry out object recognition, we invert such approximate equations. The second one uses the A* algorithm of Artificial Intelligence and has been successfully applied to the recognition of elastic cylinders at high frequencies. Results will be presented.

This concerns the estimation of physical parameters characterizing the ocean bottom from parallel time series representing (A) the instantaneous height of the water column above a section of bottom and (B) the vertical displacement of the bottom section from its long term average. Time series A, representing wind-generated waves, is modeled by a sinusoid with phase jitter. In the absence of both seismic background noise and any nonlinear behavior in the ocean bottom, time series B could be modeled by coupling the bottom through a spring and dashpot to a mass proportional to A. We created two- and three-layer adaptive networks in which series A and B (with lag) were inputs. Training consisted in subtracting the network output from the current value of series B and feeding back these errors in accordance with the appropriate formulas for gradient descent in squared error. The trained nets act as models of the way in which the bottom responds to changes at the surface.

Recent results from the Artificial Neural VIsion Learning (ANVIL) program are presented. The focus of the ANVIL program is to apply neural network technologies to the air-to-surface 3D automatic target recognition (ATR) problem. The 3D Multiple Object Detection and Location System (MODALS) neural network was developed under the ANVIL program to simultaneously detect, locate, segment, and identify multiple targets. The performance results show a very high identification accuracy, a high detection rate, and low false alarm rate, even for areas with high clutter and shadowing. The results are shown as detection/false alarm curves and identification/false alarm curves. In addition, positional detection accuracy is shown for various scale sizes. To provide data for the program, visible terrain board imagery was collected under a variety of background and lighting conditions. Tests were made on over 500 targets of five types and two classes. These targets varied in scale by up to -25%, varied in azimuth by up to 120 degrees, and varied in elevation by up to 10 degrees. The performance results are shown for targets with resolution ranging from 9 to 700 pixels on target. This work is being performed under contract to Wright Laboratory AAAT-1.

This paper uses two stages of multi-layer neural networks for character recognition. In the first stage, each neural network is trained to recognize a segment of the character image. The responses are then presented to another network where the final decision is made. The proposed method is computationally efficient, fault tolerant, has an associative memory capability, and has some of the merits of multi-decision pattern recognition techniques. The features used are gray-level representations of both typed and hand-written upper case characters. The proposed recognition scheme is tested extensively and its performance is compared with that of other non-parametric recognition methods.

A three-layer, feed-forward network was trained to classify the echoes of active sonar pulses impinging on a target in a test tank. The data set consisted of echoes recorded as the target was rotated 1.8 degrees per step. After training on 33 examples, at 5.4 degree increments, the network was typically able to decide with better than 90% accuracy whether an echo in the larger set was produced with an angle of incidence closer to end-fire or to broadside. Training time and the fractions of upstream and downstream weights that changed in the training process were observed as the number of hidden units was varied. The hidden layer consisted of binary (0 - 1) units and the learning algorithm was Rosenblatt's 'back-propagating error correction procedure'.

We use a modular object recognition system as a platform to evaluate the performance of various image processing techniques. The recognition system consists of modules for image restoration, detection, segmentation, feature extraction, invariant mapping, and classification. We are developing the system to classify objects in laser radar range imagery. The stages of the system preceding the classification stage are collectively referred to as preprocessing or early-visual processing because of the analogy with biological vision. In previous work, we presented results on invariant mapping techniques and concluded that the bi-directional log- polar mapping (BLP) method gave the best performance when evaluated within the context of an object recognition system. In the present study, we employ the BLP invariance module and use similar criteria for evaluation of several candidate image restoration and feature extraction modules. We use synthetic laser radar images of four vehicles rotated to various orientations in the field of view, scaled to various ranges, and corrupted by increasing levels of sensor noise for this evaluation. This study indicates that Markov-Random-Field image restoration and features extraction based on graded edges are a combination that provides the best recognition performance, as well as robustness to noise and discretization.

We study the scattering interaction of electromagnetic pulses of short duration with targets of either simple or complex geometrical shape using an impulse radar system. The targets of complex shape are plastic scale models of two aircraft with metallized surface. For comparison we also use a target of simple shape, in this case a metal sphere, for which we display the various signature representations, both theoretically predicted and evaluated from measured data. The form-function in the backscattering direction is determined from measured data when the targets are illuminated at a few different aspects. We extend the signature representations of targets to the combined time-frequency domain by computing and displaying pseudo-Wigner distributions of the recorded transient responses. We demonstrate that the time-frequency signature as given by the pseudo-Wigner distribution can extract and exhibit informative features in the frequency band of the incident pulse in agreement with the general time-development of resonance features. It follows that a time-frequency signature will improve the target-recognition capability ordinarily furnished by the standard form-function or radar cross-section of the considered targets.

In previous studies, simple polarimetric target and clutter models were used to derive algorithms for optimal processing of fully polarimetric measurement data. The optimal polarimetric matched filter (PMF) was derived and its performance evaluated. It was suggested that the target model used in these studies was too simple and that a more realistic target model should be developed--one which characterizes a target's polarimetric returns as a function aspect angle around the target. This is a reasonable research topic to investigate since the polarimetric properties of a target may vary significantly with viewing angle (e.g., a target imaged at broadside looks much different than a target imaged 45 degree(s) off broadside). In our previous work, targets were characterized by their polarization covariance matrices, which were calculated by averaging fully polarimetric turntable measurements of targets over 360 degree(s) of aspect. This paper investigates the variation in the target polarization covariance matrix versus aspect angle, and quantified its effect on the target-to-clutter ratio.

The design of laser radars and the determination of laser radar signatures of various vehicles require accurate data on the monostatic reflectivity of many target materials to far-field laser illumination at many different wavelengths. Modelling software is required that will take the monostatic reflectance data as an input, and combine this data with geometrical models to develop a laser radar signature. Recently a program has been started at Surface Optics Corporation to design and develop a laboratory system for the monostatic measurement of the bidirectional reflectance properties of target samples and generation of laser radar signatures from this data. This system, the Monostatic Bidirectional Reflectometer (MBR), will be flexible enough to accommodate various laser sources and detectors. The MBR will be capable of accurate measurement of the BRDF of target samples and combining this data with geometrical target models for signature predictions which will be further used for development of hardware in the loop simulations at a reduction in the cost of live-fire evaluation of various weapon systems.

The detection and recognition of a target using a Millimetric Wave (MMW) radar is limited by its low power, inherent chirp demanding large bandwidth, system losses etc. Using coherent transmitter the detection range of a target can be improved. But it calls for one order of higher technology to develop a coherent transmitter. This paper presents a MMW active radar system which uses conventional IMPATT based transmitter with limited chirp but uses coherent techniques in the receiver. Coherency between the transmitter and the receiver is maintained by memorizing the transmitted waveform and using the same as reference for the receiver. System configuration along with the hardware feasibility is discussed in this paper.

A radar scattering model was used to provide targets for a classification study of high range resolution radar signatures. Intrinsic dimensionality of these signatures was calculated using kth nearest neighbor information. Two classifier paradigms were implemented, a Gaussian classifier and a synthetic discriminant function classifier. The Gaussian correlation classifier was more robust in the presence of white Gaussian noise while the SDF approach was more robust for large angle bin size.

An adaptive neural network learning algorithm is used to estimate the location of radar targets scattering centers. The performance of the implemented spectral estimation scheme depends on learning rate, number of training vectors, and the length of each training vector. The neural network learning spectral estimation algorithms used in this study is an attractive alternative to traditional high resolution spectral estimation schemes because the number of spectral peaks depends not only on a model order assigned a priori, but also on the level of training, learning rate, and convergence. Also, neural networks are adaptive to changes in data, fault tolerant, and may be implemented using analog circuitry. Scattering features extracted using neural networks are used for target classification. The performance of the proposed target recognition scheme is compared with that of nearest neighbor based classifier.

This paper introduces a modified TLS-Prony method which uses data decimation. The use of data decimation results in the reduction in the computational complexity in a variety of applications by allowing several low order estimations to be performed rather than one high order estimation. We also present an analysis of pole variance statistics for the modified TLS- Prony method which is used to explain and quantify the characteristics of decimation. We show that using decimation we can obtain comparable performance results at a fraction of the computational cost.

The complex relationships between the vast number of system parameters make the evaluation of automatic target recognition algorithms a most complicated task. The sensor and system parameters, and environment, all play a key role in the performance of ATR systems. Consequently, a more complete performance model will require some or all of these parameters to be incorporated in the model design. Unfortunately, parameters like environment and opponent's strategy are virtually impossible to model at all but the most abstract levels. When the actual approach angle differs from the expected approach angle, several errors occur in template matching ATR algorithms and degrades the performance of ATR systems. This paper describes an effort which concentrated on the modeling of geometrically induced errors in ATR performance. Our effort was focused on the performance degradation caused by geometrically induced errors such as an aspect error. There is much merit to matched filter techniques and therefore the matched filter based analytical model is enhanced for the detection performance measure in the problem depicted above.

Self-references holographic imaging techniques with nonuniform illumination coding can be used for high resolution imaging and tracking of complex targets from either ground-based or space-based platforms. Various image reconstruction algorithms were discussed in the context of a self-reference holographic imaging technique which we had recently proposed. In this paper the image reconstructions are considered in the presence of signal generated shot noise and detector noise, taking into account the finite resolution of the detector array. Examples of performance of linear and nonlinear image reconstruction algorithms are presented. This discussion includes the consideration of the design and selection of illuminating wavefronts and provides a framework for evaluation of system requirements and its limitations.

One of the more useful techniques to emerge from AI is the provision of an explanation modality used by the researcher to understand and subsequently tune the reasoning of an expert system. Such a capability, missing in the arena of statistical object recognition, is not that difficult to provide. Long standing results show that the paradigm of Bayesian object recognition is truly optimal in a minimum probability of error sense. To a large degree, the Bayesian paradigm achieves optimality through adroit fusion of a wide range of lower informational data sources to give a higher quality decision--a very 'expert system' like capability. When various sources of incoming data are represented by C++ classes, it becomes possible to automatically backtrack the Bayesian data fusion process, assigning relative weights to the more significant datums and their combinations. A C++ object oriented engine is then able to synthesize 'English' like textural description of the Bayesian reasoning suitable for generalized presentation. Key concepts and examples are provided based on an actual object recognition problem.

Land mine detection and extraction from infra-red (IR) scenes using real-time parallel processing is of significant interest to ground based infantry. The mine detection algorithms consist of several sub-processes to progress from raw input IR imagery to feature based mine nominations. Image enhancement is first applied; this consists of noise and sensor artifact removal. Edge grouping is used to determine the boundary of the objects. The generalized Hough Transform tuned to the land mine signature acts as a model based matched nomination filter. Once the object is found, the model is used to guide the labeling of each pixel as background, object, or object boundary. Using these labels to identify object regions, feature primitives are extracted in a high speed parallel processor. A feature based screener then compares each object's feature primitives to acceptable values and rejects all objects that do not resemble mines. This operation greatly reduces the number of objects that must be passed from a real-time parallel processor to the classifier. We will discuss details of this model- based approach, including results from actual IR field test imagery.

To extract and recognize topographic symbols from topographic contour maps is a practical problem in the field of pattern recognition. In this paper, we try to put forward an effective method to solve this problem. By using this method, the symbols are first entracted from the topographic contour maps by eliminating the background of the symbols, and then they are recognized by a neural net, the Hamming net. Experimental results show the effectiveness of this method.

Machine recognition of partially occluded objects is essential for any realistic automatic object recognition (AOR) system. This important problem however, has been mostly ignored by the researchers and developers of AOR systems. Due to this lack of attention very little work is done even in the formulation of the problem and much less for its solution. In this paper I attempt to analyze the occlusion problem, define its various categories, and to present an approach for its solution in some of these categories. I also present some of the empirical results of the implementation of the approach on real imagery. These results have been very encouraging so far. However, more work is definitely needed to be done for the resolution of this problem.

Exact elastodynamic theories for fluid loaded elastic shells are only amenable to solution for simple geometries. If separation of variables methods or the extended boundary condition techniques can not be used then alternative numerical techniques are very time consuming and filled with numerical pitfalls. On the other hand the more tractable approximate shell theories (used in place of the exact theory) do not appear to work in general for fluid loaded targets. In particular, flexural resonances are predicted below coincidence frequency where they should not be observed. Further, the lowest symmetric modes are over-estimated when compared to the exact theory predictions. In addition, the presence of water borne waves is not observed from thin shell theories while they are present in exact calculations. Our interest is to develop a thin shell theory that can account for all of the above features correctly and thus can be used as a better approximate theory for future work. We thus construct shell theories which include the various shell features without undue approximation. We include the usual kinetic energy terms, contributions from a generalized Hooke's law, and rotary inertia terms. We then include a complex impedance (not the usual real component). This construction allows us to approximate the exact results in greater detail than earlier theories for three dimensional scatterers. Detailed comparative numerical examples are presented.

The T-matrix or extended boundary condition equations due to Waterman have led to a method that can predict scattering from axially symmetric targets over a broad frequency range. Several studies have been performed for spheroids with interesting results. Interest in scattering from cylinders with hemispherical end caps has prompted us to analyze recent data from Le Havre. Much of the phenomena predicted by the T-matrix calculations are confirmed by the data we analyze. In addition, bending resonances are predicted by the theoretical calculations and are compared with predictions from beam theory. The systematics contained in the calculations strongly suggest that the excited waves are related to Rayleigh-type resonances studied on shells.

We study all the resonances generated on elastic shells for a ka from 0 to 500 for steel and aluminum for a thickness of 5%. We observe the lowest order symmetric and antisymmetric model or Lamb resonances, waterborne and pseudo-Stoneley resonances and the higher order Lamb modes A_i and S_i where i equals 1, 2, 3 . . .. We plot some of the phase velocities of some of the relevant resonances out to a ka of 500 and indicate simple expressions that predict the onset of each of the resonances. We demonstrate by use of partial wave analysis that the new expressions that predict the onset (critical frequencies) of the highest order Lamb modes are reliable. Further, interesting phenomena occur at the inception of some of the resonances and we discuss some of those cases.

In the investigation of the scattering of high frequency waves from spherical shells it has been observed that it is possible to excite very large return signals at frequencies for which the integral/half integral wavelengths (due to the shear/dilatational wave of the material) are equal to the shell thickness. The large returns are shown to be attributable to resonance effects associated with the S₁ symmetric Lamb wave, the upper limit of the resonance being predicted by the flat plate limit of the shell thickness. Following a brief sketch of the pertinent types of elastic waves, the role of the S₁ symmetric Lamb wave in producing these large return signals is demonstrated using partial wave analysis.

Computational efficiencies can accrue from the processing of compressed imagery, due to an inherently reduced data burden. Since certain encryption schemes yield compressed ciphertext, computations over the range space of compressive or encryptive transformations can, in principle, exhibit computational advantages over the processing of uncompressed data or plaintext. We have recently elucidated theory fundamental to the processing of compressed and encrypted imagery, and have proposed general techniques for computation over well- known compressive formats. In this introductory paper, we analyze the efficiency of generalized operations over compressed data, with emphasis upon functions common to image and signal processing. Complexity theory derived from principles of sparse matrix processing is employed in the prediction of a critical compression ratio (CCR). Compression exceeding the CCR is required to achieve computational speedup within a given transformational regime. Additionally, given the compression ratio of a transform, as well as an image operation, we can predict the speedup of the corresponding operation over the transform's range space. Furthermore, we propose a novel computational paradigm which is based upon a network of transformations, and given optimization algorithms which determine the time-optimal computational path through such a network.

Several studies have discussed how the granulometric pattern-spectrum moments can provide good texture discrimination within images. Because textural images are modeled as random processes, the moments of an image's pattern spectrum are random variables, and knowledge of their distributions is key to the classification procedure. Both exact and asymptotic discriptions of the mean and variance distributions have previously been found under the assumption that the texture elements are nonoverlapping. The present study employs computer simulations to address the situation where the elements are not disjoint. The image is generated by Monte Carlo techniques with the predefined set of primitives, openings are calculated, and the pattern spectrum is found. It is seen that the pattern-spectrum mean remains close to its theoretical distribution.

This research presents an approach to integrate a priori information to the path metric of the LINK algorithm. The zero-crossing contours of the $DEL²G are taken as a gross estimate of the boundaries in the image. This estimate of the boundaries is used to define the swath of important information, and to provide a distance measure for edge localization. During the linking process, a priori information plays important roles in (1) dramatically reducing the search space because the actual path lies within +/- 2 (sigma) _f from the prototype contours ((sigma) _f is the standard deviation of the Gaussian kernel used in the edge enhancement step); (2) breaking the ties when the search metrics give uncertain information; and (3) selecting the set of goal nodes for the search algorithm. We show that the integration of a priori information in the LINK algorithms provides faster and more accurate edge linking.

In this paper we provide the results of an empirical investigation of iterative maximum entropy spectrum estimation. The Lim-Malik algorithm is compared to the Burg algorithm on a number of analytic as well as practical one-dimensional signals. We study the convergence of the Lim-Malik algorithm and suggest some criteria to assure convergence in the one- dimensional case.

The effective flux incident upon the detectors of the sensor, after is has been corrected for atmospheric effects, is a nonlinear function of the emissivity of the target for that channel and the temperature of the target. The sensor system cannot separate the contribution from the emissivity and the temperature that constitute the flux value. In this paper, we describe a method that estimates the water temperature from thermal data. This method is then tested with remotely sensed data obtained from NASA's Thermal Infrared Multispectral Scanner (TIMS)--a 6 channel thermal sensor. Since this is an under-determined set of equations i.e. there are 7 unknowns (6 emissivities and 1 temperature) and 6 equations (corresponding to the 6 channel fluxes), there exist theoretically an infinite combination of values of emissivities and temperature that can satisfy these equations. However using some realistic bounds on the emissivities, bounds on the temperature are calculated. These bounds on the temperature are refined to estimate a tighter bound on the emissivity of the source. An error analysis is also carried out to quantitatively determine the extent of uncertainty introduced in the estimate of these parameters. This method is useful only when a realistic set of bounds can be obtained for the emissivities of the data. In the case of water the lower and upper bounds were set at 0.97 and 1.00 respectively. A set of images obtained with the TIMS are then used as real imagery data. The data was acquired over Utah Lake, Utah, a large freshwater lake near Salt Lake City, in early April 1991. It will be used to identify water temperatures for detection of underwater thermal, saline, and fresh water springs. An image entirely consisting of water is analyzed. The temperatures of the pixels are calculated to an accuracy of less than 1 deg. K. The error histograms of the temperature estimates are also calculated.

An accuracy requirement of +/- 0.011 degrees in the declination measurement of a remotely imaged munition cannot easily be satisfied using conventional imaging instrumentation. A dedicated, digital, data acquisition system capable of extensive self diagnostics is developed. Based on internal testing, the developed system is expected to meet design goals during a formal certification process.

This paper describes the development of an information-theoretic image measure for sensor evaluation under the contract to the United States Air Force. While current approaches are based on human perception models, a need exists for evaluation of sensors for ATC/ATR systems. Such an evaluation should be performed in terms of the probabilities of detection/identification and false alarms independent of the idiosyncrasies of the specific ATC/ATR algorithms. Such an approach based on the information-theoretic content of images for the target vs. background separability is being developed and applied to evaluating sensors using the Tower Test data collected at the Wright Laboratories.

A previous paper showed that the spatial photocount average performed over a photon-limited image by a photon counting mask, and the subsequent histogram manipulation produce images which can be easily recognized by an observer. The results obtained were demonstrated as fairly good for images with a number of illuminated pixels greater than 0.8%. The aim of this paper is to determine the minimum number of illuminated pixels required for a reasonably good scene reconstruction.

Autosophy is an emerging science dealing with self-assembling structures, such as crystals or living trees. It provides a new mathematical theory of 'learning', a new post Shannon 'Information Theory' and algorithms for growing self-assembling 'Information Networks' in a memory. Autosophy methods can extract the 'True Information' from images, resulting in orders of magnitude 'lossless' images compression for transmitting aerospace sensor data and for storing these images in data bases. Robot vision is simplified by the peculiar Omni Dimensional storage method in which each image pattern or fragment is stored only once. Autosophy methods may succeed, where conventional image processing has failed, in providing electronic vision comparable to our own.

The aim of this presentation is to address a new theoretic approach to the problem of radar/radiometric imaging (PRI) stated and solved as an ii-posed inverse problem on i mite—dimensional signal subspaces. The statistical experiment design theory (EDT) —based technique for RPI inverse problem solution is developed and discussed. The technique exploits the EDT methodology to Ci) design appropriate finite-dimensional model of RRI experiment in the terms of projection schemes for inverse problem solution, (ii) provide proper matching of object image and observat ion data subspaces ; and C I i i ) determine good estimates of an object from the limited number of wavefield measurements using statistical restoration technique. We also discuss issues concernini the available control of some additional "decrees of freedom" while such RRI experiment is conducting.

An accuracy requirement of +/- 0.011 degrees in the declination measurement of a remotely imaged munition cannot be satisfied using classic image processing techniques. A dedicated, extensible image analysis system is developed using conventional high level programming languages and targeted for general purpose computational platforms. Based on internal testing, the developed system is expected to meet design goals during a formal certification process.

An abstraction of the structural pattern recognition problem is presented in which matches between observed and known segmented patterns are formed by minimizing a match distance function. Match distance is defined in terms of a fidelity measure of the reconstructions of pattern vectors from their segmented representations. Using this formulation it is shown that when reconstruction error is used to formulate match distance and the reconstruction distance measure is metric, then the structural classifier will be forced to form a trivial match between elements of the segmented patterns. Convergence properties of matching algorithms based on modified distance measures are described.

Object recognition is by nature a combinatorially explosive problem, and the combinatorics is especially worse when recognizing a 3-D object from a single 2-D image. This is because of the large search space consisting of all the possible viewing angles, translation parameters and object models. Thus there is a need to develop methods to reduce the search combinatorics. Towards this end we have developed methods to reduce search combinatorics in target recognition in ATR applications. Some of these methods make use of domain specific information leading to what are called strong search methods in artificial intelligence. Other methods use more general methods leading to what are called weak methods. In this paper we discuss one of the strong methods we have developed.

Signature modeling for multispectral applications requires careful consideration of several factors to insure a self-consistent description of the scene as viewed by the various sensors. These factors include geometric descriptions, scenario dynamics, and environmental variables. Recent work at the Georgia Tech Research Institute (GTRI) addressed several of these issues as part of a research program into object recognition for Air Force applications. One phase of this research produced a flexible and computationally efficient software environment for the generation of multispectral signatures of objects. A related effort on multispectral scene simulation addressed the complementary issue of creating a common background scene for use in dual mode (radar and infrared wavebands) applications. The paper discusses both programs and presents results from the signature and simulation models.

This paper presents a neural network model to emulate the ability of the human visual system to detect changes in heading direction, i.e. curvilinear motion. The network consists of three layers. The input to the network is a two-dimensional velocity field, and the output is a signal representing the magnitude and the direction of the rotational component in the flow. The first layer of the network computes local differences vectors of the velocity field to define the orientation of the translational field lines. The second layer of the network extracts the instantaneous heading direction from the translational component of the velocity field. And the third layer determines the rotational component of the velocity field. The magnitude of perceived curvilinear motion is directly proportional to the magnitude of the rotational component. The simulation results match psychophysical data of four human subjects at both slow (2.0 m/s) and fast (26.4 m/s) locomotion speeds. The biological feasibility of this neural network is supported by finding in biological vision systems.

Inherent limitations of the biological mechanism of vision have resulted in elegant algorithms to accomplish the required information processing. Biological and psychophysical evidence have led to hypothesizing just such an algorithm for the processing of motion information. The Temporal Frequency Spectrum (TFS) model for motion perception uses a transformation process analogous with those Fourier processes sometimes used in the study of spatial vision. The TFS can serve as a metric for describing complex moving scenes and as a model for human motion perception. This model can accomplish motion perception even when the viewer is in motion. Human psychophysical evidence is provided consistent with a TFS mechanism and the use of a TFS mechanism in machine vision is considered.

This paper outlines the human factors issues inherent in the transmission of video data over a low bandwidth secure radio channel, necessary for battlefield teleoperation of remote platforms. Bandwidth allocation for image transmission for low data rate operations is 16 kilobits per second (Kb/s), as opposed to the 63 megabits per second (Mb/s) required for standard video transmission. Radically degraded imagery affects remote driver ability. Observations of the remote driving task have pointed out areas in which further research is indicated. Human factors issues in image-compression will be investigated to favorably impact the design of enhanced image-compression techniques and future human factors research.

A computational system was developed to integrate both Fourier image processing techniques and biologically based image processing techniques. The Fourier techniques allow the spatially global manipulation of phase and amplitude spectra. The biologically based techniques allow for spatially localized manipulation of phase, amplitude and orientation independently on multiple spatial frequency scales. These techniques combined with a large variety of basic image processing functions allow for a versatile and systematic approach to be taken toward the development of specialized patterning and visual textures. Current applications involve research for the development of 2-dimensional spatial patterning that can function as effective camouflage patterns and masking patterns for the human visual system.

In this paper, the lnr-(theta) coordinate transform method is used in the joint transform correlator (JTC) so that the rotation and scale variant pattern can be recognized. The distortion invariant JTC is implemented in real time by using the LCLVs. Initial experimental results are given in this paper.

The disturbance of the uneven lighting is one of the difficult problems in image binarization. Usually, it is solved by the method of binarization in partitioned windows. But the size of the window may often bring about the problem of homogenous windows, i.e., windows contain only background or object pixels, and this can seriously affect the results of the binarization. In this paper, while using Otsu's automatic threshold selection method during the binarization in partitioned windows, we put forward a discriminating method which is based on pyramid data structure of Lorentz information measure. By using the method, the problem is solved perfectly.

The experiments on the perceptual contrast of two gratings with the same structure and different qualifications, such that the light and dark slits of one grating flicker simultaneously and the light and dark slits of the other grating flicker alternately, show that within a certain flickering frequency range, the perceptual contrast of the alternately flickering grating is greater than that of the simultaneously flickering one. The fact leads to the conclusion that the lateral inhibition which derives the enhancement of perceptual contrast is out of phase with the input stimulus. And the high-pass filtering properties in both spatial domain and temporal domain for the human vision originate from the same mechanism--the postponed lateral inhibition.

Adjusting components of the display system to reproduce colors observed in the original scene is difficult and expensive. We know that a color can be matched in hue (H), luminance (L), and saturation (S) by a mixture of three primary colors, red (R), green (G), and blue (B). The color senses to the mixture are various for different observations as the senses are psychophysical characters. In this paper we describe a psychophysical method of transforming the ideal primary colors X, Y, and Z making up the colors in the original scene into the colors R, G and B of the display system. The R, G, and B value corresponding to the X, Y, and Z values come from statistical data of standard observations. By means of a polynomial regression analysis, the transform relating the X, Y, and Z to the R, G, and B can be reached on those standard observations. Sufficient observation data and effective color samples are essential to find the psychophysical relation between the two sets of colors. Adaptation of a color lookup table to enhance the fidelity of color reproduction with the method is easier and more straightforward. This method can avoid the obstacles which occur in hardware methods of color reproduction.

The auxiliary antenna approach to sidelobe/mainbeam discrimination is a classic technique based on comparing the signal strength from a small co-boresighted auxiliary antenna with the signal from the main antenna. A piecewise linear boundary is usually drawn in the main channel/auxiliary channel signal space to perform the discrimination. The Automated Remote Tracking Station (ARTS) program has a requirement to automatically acquire and track a satellite when the initial ephemeris information is only approximately known. This requirement is satisfied by using an auxiliary antenna approach to detect the mainbeam intercept, while the antenna is scanning. Once the mainbeam is detected, standard monopulse tracking is engaged to complete the acquisition and lock onto the satellite. The ARTS system requirements for mainbeam/sidelobe lock can be treated as probabilities of detection and false alarm. This system requirement can be visualized as constraint surfaces in the main channel, auxiliary channel, and aperture power signal space. The classic approach to designing a decision boundary in the presence of an unwanted or unknown parameter (aperture power) is to project the constraint space so that it is independent of that parameter. The decision boundary can then be drawn in the projected space. When this approach was applied to the ARTS requirement, it yielded several decibel (dB) of increased dynamic range, at low signal strengths, over conventional sidelobe discrimination boundary designs.

This paper presents techniques in image simulation and computer graphics intended to generate models to be used for a better understanding of SAR images; only geometric aspects are considered. A fractal relief model is used as a basis for creating simulated SAR pictures. The technique employed is bidimensional Fourier filtering. The relief aspect is controlled by the filter shape. A shading model and algorithm is implemented for the simulation of the coherent lighting in SAR. The model is based on the physics of coherent radiation scattering and propagation. The parameters of the surface are controlled to modify the speckle statistics. The 3D model is presented in a computer graphics station. Due to the high dimension of the scene, heuristic methods are developed for fast ray tracing. Different incident radiation positions and scene projections are available allowing interactive visualization of the synthetic image.

In order image objects to be recognized efficiently, it is possible to use statistical encoding methods and information theory results and then to analyze only typical elements groups on the picture. As the typical groups revealing by means of complete sorting is impossible, then one can suppose that the visual analyzer selected such group during the evolution and fixed them in its own mechanism. Hubel's fields situated in visual cortex and responding to linear differently oriented segments are such determined mechanisms. In this paper the original methods of line segments distinguishing as the typical groups for subsequent recognition are investigated. The further development consists of low (1 %) contrast lines detection. The special logical multidimensional filters were created for this purpose. They counted up the line peculiarities on the discrete raster. When this, the every sloping line (within the certain angle sector) is given as a set of some vertical (horizontal) concrete length segments. For example, the slope relatively vertical from 0° to 18° is composed with the vertical segments. Their length cannot be less than 3. The line bent within the range 18° - 26° is built by segments of 3 and 2 element length. In order to enhance the determination of the line head and tail coordinates the cross scanning is applied while filtering. The obtained results were used for space photos analysis with the aim of roads, rivers and other prolongated object detection.