We study the bistatic scattering interaction of electromagnetic pulses of short duration with targets of simple shape. The targets are spheres, either a metal sphere or a dielectric hollow sphere. The waveform incident on the target is a pulse that is designed to model the waveform emitted by a particular impulse radar system, which is then used to verify our theoretical findings. When the E-field (or H-field) of the incident wave is polarized in the plane spanned by the direction of incidence and the considered scattering direction, the corresponding bistatic form-function is referred to as the E-plane (or H-plane) bistatic form-function. These are the only cases where the bistatically scattered wave is polarized parallel to the incident wave for any bistatic angle. We also examine the nature of the features that occur in a PWD, and we demonstrate that useful signature characteristics are produced by cross-terms.

A new technique is developed to achieve focusing in range and direction (angle) (joint range- angle beamforming) on the received signals simultaneously, transmitting one wave-beamform. It is shown how this can be utilized to estimate target density functions in the range-direction coordinates in three dimensional space. It is also shown that range-angle focusing can be achieved using only one sensor via transmitting several wave-beam forms.

A method for segmenting synthetic-aperture radar (SAR) images has been developed which operates primarily in the frequency domain. Though based on a similar technique developed by Stromberg, this method has some computational advantages and may be readily adapted to optical implementation.

Fundamental results from fractal measure theory are presented and applied to form a new technique for signature analysis. A detection algorithm is generated from this theory and successfully applied to the reflectance images of a laser radar.

This paper describes extended work on a comprehensive rotational-motion estimation and compensation technique to improve the quality of range-Doppler imagery. The approach assumes that the target signature is a collection of samples in the frequency space aperture plane. Motion compensation is performed via a polar reformatting scheme that requires knowledge of target kinematics, while motion estimation is accomplished by monitoring an entropy-like function in terms of kinematic parameter estimates. When properly selected, the entropy-like function becomes an accurate measure of image sharpness, and can thus be used to control a down-hill simplex search which yields an optimum set of rotational-motion parameter estimates. Ultimately, this set is needed to produce an optimally focused target image.

Motion compensation of range-Doppler target signatures results in focused target imagery. Recently, an iterative approach based on a logarithmic entropy measure has been proposed for the motion compensation of signatures collected in the frequency domain. The effectiveness of this approach can be significantly improved by using an entropy-like function which is maximally resistant to noise and consistent with statistical boundaries. For purposes of analysis, the entropy-like function is written in terms of an information gain function (Delta) I. Several expressions for (Delta) I are tested to verify the accuracy of radial-motion parameter estimation. The effectiveness of these expressions is determined by the number of iterations required to find the minimum entropy measure, within an acceptable tolerance level for a given signal-to-noise ratio. Results show that the exponential information gain (Delta) I equals exp(1-I) yields an optimally convex entropy measure surface over a prescribed motion- parameter solution space. The surface minimum in this solution space has coordinates which are interpreted as the optimum motion-parameter estimates that can be obtained for the purpose of image focusing.

A compact library of discrete phase-only spatial filters for use in a digital or optical correlator has been constructed which discriminates between TABILS24 MV0045 and MVO294 by a factor of 10 using signal-to-clutter ratio as a measure of performance. This is a combined discrimination ratio of more than 20. This library of filters can be stored in 320 K or less of memory. These filters have been designed to give a stable recognition signal to their respective true targets over a wide range of rotation angles. More importantly, they also give strong recognition signals to target aspects which straddle images used in the filter training sets. This stability arises from a simple processing of the radar imagery before it is inserted into the correlator.

The work described in this paper focuses on recent progress in radar signal processing and target recognition techniques developed in support of WL/AARA target recognition programs. The goal of the program is to develop evaluation methodologies of hypotheses in a model- based framework. In this paper, we describe an hypothesis evaluation strategy that is predicated on a generalized likelihood function framework, and allows for incomplete or inaccurate descriptions of the observed unknown target. The target hypothesis evaluation procedure we have developed begins with a structural analysis by means of parametric modeling of the several radar scattering centers. The energy, location, dispersion, and shape of all measured target scattering centers are parametrized. The resulting structural description is used to represent each target and, subsequently, to evaluate the hypotheses of each of the targets in the candidate set.

Applications of wavelet decomposition techniques in target signature analysis are reviewed. A radar target classification system based on features derived from the wavelet transform of the target impulse response is considered. The performance of the proposed target recognition technique is evaluated assuming different measurement scenarios including azimuth ambiguity. The merits and limitations of the proposed scheme are presented.

Radar target classification performance is greatly dependent on how the classifier represents the strongly angle dependent radar target signatures. This paper compares the performance of classifiers that represent radar target signatures using vector quantization (VQ) and learning vector quantization (LVQ). The classifier performance is evaluated with a set of high resolution millimeter-wave radar data from four ground vehicles (Camaro, van, pickup, and bulldozer). LVQ explicitly includes classification performance in its data representation criterion, whereas VQ only makes use of a distortion measure such as mean square distance. The classifier that uses LVQ to represent the radar data has a much higher probability of correct classification than VQ.

A new two-dimensional (2-D) technique is developed to estimate the polarimetric characteristics of scattering centers that exist on radar targets. This technique uses a 2-D damped exponential model to approximate the scattering from radar targets. The validity of this model is investigated relative to the scattering characteristics that exist on the targets of interest. Simulations are shown which validate the technique.

Scattering centers extracted from radar returns are features that can be used for automatic target recognition (ATR). Most scattering center extraction algorithms fall into one of two large classes: FFT-based and model-based techniques. In this paper, we consider one FFT- based and two model-based techniques: (1) FFT with peak extraction, (2) the Burg autoregressive method, and (3) the TLS-Prony method. We present experiments that compare the algorithms for three test scenarios. We compare the algorithms in terms of scattering center accuracy and computational cost.

A model for the generation of reflectivity profiles is presented for use in a radar target recognition system. The data are assumed to come from two sensors: a high range resolution radar and a tracking radar. The object is simultaneously tracked and identified using estimation theoretic methods by comparing a sequence of received range profiles to range profiles generated from surface templates. The tracking data are used to form priors on the position and orientation of the object. The templates consist of surface descriptions comprised of electromagnetically large patches tiling the entire object. The predicted return is computed from several quantities. First, the reflectivity range profile is computed from the patches incorporating a shading function. The physical optics approximation is that patches not directly illuminated by the transmitted signal to not contribute to the return signal. Second, the reflected signal is approximated by the convolution of the transmitted signal with the range profile. Third, the receiver design yields the actual I-Q data available for processing.

Automatic target recognition usually requires a model or information from which a scene interpretation can be made. The approach of this paper is to obtain this recognition on a photograph basis. The apparent simplicity of the method only hides difficulties which are discussed in the paper. Solutions are brought in order to solve most of these and a general method is proposed. It is illustrated by some examples. Applications in multiple domains can be foreseen.

To promote user community acceptance and incorporation, an intuitive operational model and GUI-based tool were developed to aid knowledge acquisition and automatically generate the required approximate grammars. Temporal step sequences are developed by operational analysts and assigned the attributes of necessity and confirming strength. Observable detectability and confusability attributes aggregate sensor suite capabilities and are specified by the sensor system engineers. The tool guides the user in combination and correspondence of the two knowledge sources, producing either a set of usable approximate grammars together with a list of potential conflicts.

The detection of objects in high-resolution aerial imagery has proven to be a difficult task. In our application, the amount of image clutter is extremely high. Under these conditions, detection based on low-level image cues tends to perform poorly. Neural network techniques have been proposed in object detection applications due to proven robust performance characteristics. A neural network filter was designed and trained to detect targets in thermal infrared images. The feature extraction stage was eliminated and raw gray-levels were utilized as input to the network. Two fundamentally different approaches were used to design the training sets. In the first approach, actual image data was utilized for training. In the second case, a model-based approach was adopted to design the training set vectors. The training set consisted of object and background data. The backpropagation training algorithm was modified to improve network convergence and speed and used to train the network. The neural network filter was tested extensively on real image data. Detection rates were determined for varying false alarm rates in each case. The detection and false alarm rates were excellent for the neural network filters. Their overall performance was much superior to that of the size- matched contrast-box filter, especially in the images with higher amounts of visual clutter.

In this paper, we provide insight into the development of a new set of distortion-invariant texture operators. These `circular-Mellin' operators are invariant to both scale and orientation of the target and represent the spectral decomposition of the image scene in the log-polar coordinate system. Coupled with the shift invariance property of the correlator architecture, we show that these circular-Mellin operators can be used for rotation-, scale-, and shift- invariant feature extraction. We also note that these feature extractors have a functional form that is similar to the Gabor transforms. Detailed analytical description of these operators and preliminary results are presented in this paper.

The results of a study in the uses of fractal for the automatic detection of man made objects in infrared (IR) and millimeter wave (MMW) radar imagery are discussed in this paper. The fractal technique that is used is based on the estimation of the fractal dimensions of sequential blocks of an image of a scene and then by slicing the histogram of the computed fractal dimensions. The fractal dimension is computed by a Fourier regression approach. The technique is shown to be effective for the detection of tactical military vehicles in IR, and for the detection of airport attributes in MMW radar imagery.

Based upon the requirements of Bayesian object recognition theory, this paper provides the fundamental framework to determine the joint probability density function of object regions in an IR image. This probability function contains all information about the region that is required to achieve minimum probability of error recognition. The techniques advanced here are expected to be of significant use in certain rather hostile and difficult situations such as testing piping for fault conditions within operational nuclear power plants.

The effect of varying the illumination of the input scene on the performance of classical, binary, and fringe-adjusted joint transform correlators (JTCs) has been investigated in this paper. Computer simulation results show that the fringe-adjusted JTC yields superior correlation outputs when compared to the classical and binary JTCs.

A technique is developed for classifying different objects in natural imagery by employing a wavelet transform and training a neural network on certain of the wavelet transform coefficients. The effectiveness of different choices of coefficients and neural network architectures is analyzed.

Statistical target recognition techniques perform well when the `true' target-signature data are deterministic. If the deterministic nature of the data exists and the probabilistic values of a given problem are known, then identifiers, based on Bayesian estimation theory, have great potential for solving the target identification problem. How well an identifier will perform is usually answered by Monte Carlo simulations or implementation experiments. An alternative to these performance analysis techniques is the use of information theory. Information theory has long been applied to the investigation of data compression, which deals with average distortional measures. The association of Bayes risk with distortion allows for information- theoretic tools to be applied to the statistical target identification problem. The rate-distortion function of data compression can be extended to the Bayes rate-distortion function. Derived from designer-specified risk and identifier structure, the theoretical Bayes rate-distortion function relates mutual information to the identifier performance. Mutual information is determined from the target-signature data used by the identifier and by the nature of the noise imposed upon the data. Computational efforts in determining mutual information values allow for identifier performance bounds to be extracted from the Bayes rate-distortion function.

In this paper, the method of lines (MoL) is successfully extended to solve the EM wave scattering problems of two dimensional metallic periodic surfaces. Compared with available data, the method is proven to be accurate and feasible. For different structures only a slight change of matrix elements is needed to be handled in numerical calculation. Meanwhile, the method presented here possesses the advantage of less computation. The results of groove, comb, and sinusoidal surfaces are presented.

We present connectionist classifiers that form distributed low-level knowledge representations for pattern recognition, given random feature vectors generated from dual statistically distinct sources. The classifier is a functional representation of an oscillatory network consisting of two-member clusters of model neurons whose efferent synapses may be either inhibitory or excitatory. We demonstrate the oscillatory network's performance in the context of source- dependent single speaker identification. In these tests, the backpropagation network representation learning curve began to flatten around an unacceptable error response. The oscillatory model, however, was able to discriminate accurately.

We present a study of the scattering of sound waves by a neutrally buoyant, submerged, spherical shell that has heavy masses internally mounted by springs to the shell. The presence of the attached oscillating masses substantially changes the scattering cross-section of the otherwise empty shell. Furthermore, if the compliance (i.e., the spring constants) of the mounts are varied -- all else being the same -- then the resulting cross section also varies substantially, often by more than an order of magnitude. It is seen that if the incident plane wave impinges upon the shell at its North pole, where the masses are mounted, then the spring-mass effect on the scattering response is strongest.

In this article we examine time evolution of the a posteriori probability distribution for the source location in an elementary passive sonar model. The approach employs Bayesian inversion to obtain the a posteriori distributions. The resulting multimodal distributions reflect the nonlinear relationship between source location and array output. This probabilistic approach allows us to include a priori information, utilize ocean acoustic modeling, and make direct use of increased data integration time.

Target discrimination and recognition are dependent on the understanding of the dynamic acoustic information processed by a given sonar array. Visualization of the true near-field acoustic energy distribution of an active sonar array would be of great value to the understanding of the effects of structural defects in the array and its interface to the water medium. Knowledge of the projected near-field acoustic energy would also enhance the analysis of returned signals for homing and target recognition. Holographic interferometry has presented itself as a viable and useful method for the realization of this type of information.

The Resonance Scattering Theory (RST) was developed for acoustic scattering in 1977 and shortly thereafter was extended to the resonance scattering of elastic waves, and later on to the scattering of electromagnetic waves (radar). Resonance amplitudes in acoustic scattering from submerged elastic objects, which the RST describes in a mathematical fashion, are a dominant feature in the acoustic echoes which can be used to effect an `acoustic resonance spectroscopy' (as pointed out by Andre Derem). Such a spectroscopy can be used for characterizing the target as to its size, shape, and composition, and can ultimately lead to a procedure or a mechanism for (automatic) target recognition.

By including the effect of fluid loading for the thin spherical shell in a proper manner, so- called shell theories can predict the water borne pseudo-Stoneley waves described extensively in the literature. Shell theories give reasonably good results for the motion of a bounded elastic shell by using the assumption that various parts of the shell move together in some reasonable manner. Without proper fluid loading, however, shell theories do not predict the pseudo-Stoneley resonances observed in nature and predicted by exact theory. With proper fluid loading, as well as rotary inertia and translational and rotary kinetic energy terms, a shell theory can exactly predict these water borne resonances. These resonances are predicted by the shell theory and compared with results from exact elastodynamical calculations.

Active acoustic classification of underwater targets is an important problem of current interest. The echo return of such objects has components related to such distinct dynamical elements as the specular (geometric) return, reradiation of elastic resonances, diffracted (or Franz) waves, etc. The arrival times and spectral content of these components will generally give information about the structure and geometry of the scatterer. This information is reflected to the time- frequency structure of the echo return. In this paper, a comparative study is presented of the time-frequency analysis capability of a number of tools in applications to the echo structure of finite elastic cylinders. The time-frequency tools considered include the Wigner-Ville distribution, the Choi-Williams distribution, the Gabor transform, and the continuous wavelet transforms. The comparison is based on echo returns that have been synthesized from numerical T-matrix solutions to the associated free-field scattering problems.

The computation of the scattered field (acoustic, electromagnetic, and elastic) from a wide class of objects can be treated in a consistent manner using the extended boundary condition (EBC) or T-matrix method (originally derived by Waterman). A basic assumption for the objects using this technique is that it is located in free space i.e., volume void of boundaries. This method can still be used when the object is placed in a waveguide, however, it requires coupling the free space solution with the waveguide solution. In this paper the basic derivation for the far-field solution of the scattered field due to the scattering of a guided wave from an object in a waveguide is outlined. This method is then applied to a problem of physical interest which pertains to acoustical scattering from three-dimensional elongated impenetrable objects in an underwater acoustic waveguide.

We wish to develop a scheme that will enable us to predict the near field that results from the interaction of a submerged object with a guided wave for a general object in a general waveguide. The method will be based on the extended boundary condition (EBC) method (T- matrix) of Waterman and a new normal mode method that allows one to decompose the normal-mode solutions into a spherical representation suitable for operating on the spherical tensors that arise from the T-matrix method. Of particular importance is the fact that the resulting formulation allows one to couple the resulting near-field solution to an outgoing normal-mode series that leads to the general waveguide solution.

In the process of scattering from submerged elastic shells, it is possible to excite many types of resonances. Among these are the lowest order symmetric and antisymmetric Lamb modes, and the waterborne waves, such as the pseudo-Stoneley resonances and the higher order symmetric and antisymmetric Lamb modes Si and Ai (i equals 1, 2, 3, ...). The frequencies at which these originate are referred to as critical frequencies. We establish simple rules to determine the frequencies at which the resonances originate as a function of shell thickness and material properties.

This article uses the formalism of Hackman and Sammelmann [J. Acous. Soc. Am., 80(5), pp. 1447-1458] to calculate the scattering from an elastic shell in a shallow water waveguide. This formalism is an exact solution of the Helmholtz Integral Equation. Some of the phenomena associated with placing an elastic spherical shell in a shallow water waveguide [J. Acous. Soc. Am., 82(1), pp. 324-336] are splitting of the free field resonance spectrum, generation of depth dependent frequency shift of the resonance spectrum, and the generation of super resonances.

The Helmholtz-Poincare Wave Equation (H-PWE) arises in many areas of classical wave scattering theory. In particular it can be found for the cases of acoustical scattering from submerged bounded objects and electromagnetic scattering from objects. The extended boundary integral equations (EBIE) method is derived from considering both the exterior and interior solutions of the H-PWEs. This coupled set of expressions has the advantage of not only offering a prescription for obtaining a solution for the exterior scattering problem, but it also obviates the problem of irregular values corresponding to fictitious interior eigenvalues. Once the coupled equations are derived, they can be obtained in matrix form by expanding all relevant terms in partial wave expansions, including a biorthogonal expansion of the Green function. However some freedom of choice in the choice of the surface expansion is available since the unknown surface quantities may be expanded in a variety of ways so long as closure is obtained. Out of many possible choices, we develop an optimal method to obtain such expansions which is based on the optimum eigenfunctions related to the surface of the object. In effect, we convert part of the problem (that associated with the Fredholms integral equation of the first kind) an eigenvalue problem of a related Hermition operator. The methodology is explained in detail and examples are presented.

We describe an end-to-end Automatic Target Recognition (ATR) system for recognizing targets in Synthetic Aperture Radar (SAR) imagery. The system heavily relies on the use of functional template correlation, a technique recently developed by the authors for applying machine intelligence directly at the pixel level through the use of functional templates (FTs). Targets are detected using a CFAR-like, circularly-symmetric FT. They are recognized with azimuth-dependent FTs that deal with the fact that the appearance of an object in SAR imagery changes significantly with the direction of radar illumination relative to the object. FTs were specifically designed for ISAR data at 19 deg depression angle. Excellent recognition results were obtained when these FTs were blindly applied to over 20,000 ISAR images covering depression angles from 18 to 32 deg. When the same FTs were applied to 255 airborne stripmap SAR images at 22 deg, good recognition results were obtained with no false alarms. Although the paper deals primarily with fully polarimetric data, the ideas presented readily apply to single- or dual-polarization SARs.

The calculation of moment invariant is a simple and repetitive algebraic operation which is invariant to translation, rotation, and scale change of an object. But because of the dash between the vast computation for the moment itself and the limitation of the memory space and the speed of the microprocessor it is very difficult to achieve a real-time calculation of the moment invariant. In this paper, we advanced a new fast algorithm of 2-D moment invariant based on image projection, by means of projection transformation we can compress the information of a 2-D image into 1-D information. Thus, on one hand, the amount of computation and data size are decreased greatly, on the other hand, projection transformation as merely an operation of additions is easier to achieve on hardwares now.

Based on the acoustic inhomogeneous wave equation, the forward and inverse scattering problem in ultrasound tomography is characterized in terms of the Lippman-Schwinger equation. For higher-order transient solutions the time-domain moment method is very promising in comparison to frequency domain approaches where phase wrapping has been proven to be a fundamental problem. A preliminary numerical study has verified the Cavicchi's expansion of the moment method equation duplicating the close agreement between the numerical approximation and the exact solution. The primary purposes of this paper is to introduce a numerical implementation of the exact scattered fields by using a Bessel function series and the discrete Fourier transform, and to show that algorithm artifacts occur due to circular convolution aliasing and time domain aliasing. Computer simulations have reconstructed the aliasings individually and shown that a great improvement in numerical verification can be obtained by using alias-free estimation data.

An experiment is reported which had as its main requirement the gathering of a wide range of information about the scene under investigation. It took place at the area of the Nurnberg airport (Bavaria, West Germany) in the summer of 1991 and the spring of 1992. The time frame of the experiment was planned to get satellite images in addition to airborne images at nearly the same time. The airborne scanner provided images in eleven spectral channels from the visible to the thermal infrared with geometric resolutions between 0.75 m and 4.50 m. In parallel we performed several radiometric measurements of reference targets at the ground as well as meteorological measurements to take care of atmospheric effects and lighting conditions. To get a better knowledge of the whole area we gathered additional information of the Nurnberg airport and surroundings. In particular, we investigated the functionality and task of certain areas (airport, agricultural, and industrial) and the type of surface materials of objects (airplane, taxi way, park way, shed or hangar roof, natural surfaces) within the scene of investigation. A combination of vegetation filters, unsupervised classification, and gray scale morphological image processing was applied to the data. As a result we obtained pixel by pixel and morphological classification of buildings, roads, and vegetation.

The analytical method for calculation of image correlation to object detection from a real scene using a description of objects and scenes by the fragments of random fields is described. The discrimination capability of the algorithm and the effect of random and geometrical distortions on the performance of cross-correlation technique are considered.

Some results of textured objects recognition experiments using spatial matched filtering technique are presented. These experiments demonstrate the fact that the recognition results depend greatly on the parameters of image and this technique does not allow recognition of naturally textured objects reliably.

The authors consider the methods of detecting edge based on the Fourier transform. They show connection between this method and well-known enhancement/thresholding edge detectors. The quantitative and qualitative estimations of different methods of detecting edge are given.

The solution to the problem of constructing the system capable to recognize objects and to determine their coordinates via images is presented. The theoretical approach based on a set of features invariant to such affine plane transforms as the scale changes and rotations is developed. The algorithm for computation of the invariant features of the objects and the optimization procedure for determining its coordinates are suggested. The algorithm involves the Hough-transforms (HT), threshold processing in the HP parameter plane, logarithmic transform of the coordinates, and direct calculation of the invariants with respect to the above groups of the affine transforms. The block-diagram of the iconic system is presented. The algorithm application to aircraft images processing is provided.

The problem of the relation of the iconic system's recognition probability (under known a priori statistics) with the system parameters and the necessary image scale is considered. It is shown that the resolution as a system quality criterion results in the essential errors. The new expression to correlate the recognition probability, the image scale, and the system parameters for the aerial photographic system by using as a criterion the information estimate C₁ is proposed. The information estimate C₁ is connected with the signal-to-noise ratio of the system. The original signal model is given. The distinctions between the estimate C₁ and the information capacity C₀ are shown. On the basis of the proposed method the application program package for the numerical optimization of the air photographic systems and their modules is realized. The mathematical model of the system implemented in the package takes into account the conditions of the real aerial shooting. The package can be used to optimize the parameters of the black-and-white photographic films and their photochemical treatment.

Image registration measurement systems (MSs) often possess the following specific properties: the measurement results of these systems demand for further complex processing; these MSs deal with large amounts of data; and they possess a high level of invariance. In this paper a general description of measurement systems invariant with respect to a given group of transformations (e.g., translations, reflections, rotations of field of view) is considered. Then the problem of an optimum measurement computer system (MCS) synthesis for a given invariant MS is stated. It consists in obtaining a mapping (describing processing algorithm), delivering an optimum to the MCS as a whole. It is shown that proper allowance for invariance can significantly simplify MCS's synthesis problem solution and, as a consequence, appreciably cut down the computational costs involved both for constructing an optimum computational component of MCS and for its subsequent functioning.