The problem of fusing identity declarations emanating from different sources is explored and decision makers are offered a quantitative analysis based on statistical methodology that can enhance their decision making processes regarding the identity of detected objects. The context is naval warfare where commanders and their staff require access to a wide range of information to carry out their duties. This information provides them with the knowledge necessary to determine the position, identity and behavior of the enemy. Statistical analysis rooted in the Dempster-Shafer theory of evidence is used to further the fusion of identity declarations.

We extend the decentralized signal detection problem to the case where the observations available at the sensors display measurement inaccuracy. The imprecise measurements at each sensor are modeled as fuzzy information systems. The fuzzy information system model consists of fuzzy partitions defined on the corresponding crisp (nonfuzzy) observation space. A Bayesian decision criterion is employed in the design of the signal detection schemes at the local sensors and the decision fusion center. Compression of the input fuzzy systems at the sensors into a fewer number of levels is studied. Both optimal and suboptimal methods are employed to partition the fuzzy observation space for signal detection applications.

A new approach to fusion of target identity and state vector information acquired from multiple sensors, based on combining approximate reasoning concepts derived from fuzzy logic and evidential reasoning fields, is presented. The primary emphasis is on the fusion of target identity information, assumed to be available in the form of fuzzy class membership estimates, to derive more robust target identification. The fusion of track state vector information is viewed as an essential ingredient of the identity fusion process and is carried out with only minimal sophistication to minimize computational demands that enable real-time implementation. The methodology, although designed for fusion of fuzzy identity information, can be adapted to cases wherein the information provided is in the form of crisp labels from unreliable or imperfect sensors. This is accomplished with a learning phase wherein the imperfectness of the sensor inputs are modeled using the available ground truth data to derive fuzzy membership estimates for the various sensoridentity reports. The methodology takes on a sensor-to-track fusion (rather than sensor-to-sensor fusion) approach to enable a variable number of sensors and minimize dependence on congruency of sensor reports along the time line. The fuzzy identity membership estimates of the tracks along with track state vectors are updated through appropriate data association (based on near-enough-a-neighbor principles) of the incoming sensor reports. The estimates also include a measure of ignorance that monotonically decreases as more and more sensor reports are fused with the tracks. Details of the algorithmic approach are presented followed by results of application of the methodology to some real-world data. These results clearly show the efficacy of the process, which synergistically combines fuzzy logic and evidential reasoning concepts.

A multisensor fusion algorithm that integrates model-based target identification (ID) and multiple-hypothesis tracking (MHT) is described. The algorithm augments the target state representation to include ID information and to manipulate the new target state. The major innovation of the algorithm is using Bayesian networks to modify the hypothesis-evaluation process by taking into account target ID. This research was conducted as part of a larger effort to design a decisiontheoretic sensor management system. In the system, two types of sensors, electronically scanned radar (ESA) and IR search and track (IRST) are assumed to be available. The ESA radar is modeled to have search and update capabilities as well as three radar identification modes: ultrahigh- resolution radar (UHRR), radar signal modulation (RSM), and radar electronic support mode (RESM). A Bayesian network is used to model the detection and observation processes for these ID techniques and compute the association likelihoods between measurements and tracks.

The ability to efficiently fuse information of different forms to facilitate intelligent decision making is one of the major capabilities of trained multilayer neural networks that is now being recognized. While development of innovative adaptive control algorithms for nonlinear dynamical plants that attempt to exploit these capabilities seems to be more popular, a corresponding development of nonlinear estimation algorithms using these approaches, particularly for application in target surveillance and guidance operations, has not received similar attention. We describe the capabilities and functionality of neural network algorithms for data fusion and implementation of tracking filters. To discuss details and to serve as a vehicle for quantitative performance evaluations, the illustrative case of estimating the position and velocity of surveillance targets is considered. Efficient target-tracking algorithms that can utilize data from a host of sensing modalities and are capable of reliably tracking even uncooperative targets executing fast and complex maneuvers are of interest in a number of applications. The primary motivation for employing neural networks in these applications comes from the efficiency with which more features extracted from different sensor measurements can be utilized as inputs for estimating target maneuvers. A system architecture that efficiently integrates the fusion capabilities of a trained multilayer neural net with the tracking performance of a Kalman filter is described. The innovation lies in the way the fusion of multisensor data is accomplished to facilitate improved estimation without increasing the computational complexity of the dynamical state estimator itself.

Static and dynamic distributed interacting multiple model (IMM) fusion algorithms for multiplatform multisensor tracking are developed. Each platform contains a model set, which may or may not be the same as that of other platforms. An interacting multiple model filtering is performed on each platform. An equivalent platform model and an equivalent global model are constructed. To save the bandwidth of an interplatform communication datalink, only combined IMM tracks are allowed to communicate. Taking advantage of the equivalent models, both static and dynamic fusion algorithms have very decent and comparable results. But simulations show that the computation complexity for the dynamic fusion algorithm is far lower than that of the static fusion algorithm. Both algorithms benefit from multiple models and distributed tracking.

We consider the problem of over-the-horizon multiradar track association. The operation center receives the registration of tracks from three different over-the-horizon radars. The fusion of these tracks presents a large operator workload. We investigate automating this process. We propose two methods for feature extraction. The first uses the Hough transform and the second uses some track affinity measures. For every case, we consider two systems, one using the multilayer perceptron and the other assuming independence of the features and combining the probabilities. For the Hough transform, we consider the range-time image of the tracks and transform all track data to a unified coordinate system. In the second method, we consider a set of affinity measures and study their relative and cumulative frequencies for associated and nonassociated pairs of tracks. Error rates of less than 7% were achieved with the Hough transform method and less than 1% using track affinity measures.

Two multisensor multitarget track initiation techniques are proposed. One is the multisensor Hough transform and the other is the multisensor logic-based method. Both are extensions of their single sensor versions. Since a track initiator is basically a detector, the receiver operating characteristic (ROC), which is a measure of target track detection probability versus false track probability, is used to evaluate the performance of these multisensor track initiation techniques. Simulated and real life multiple radar data are used to evaluate performance. Results show that both multisensor track initiators perform well in heavy clutter environments, and the multisensor logic-based track initiator has better performance than the multisensor Hough transform method for maneuvering targets.

Image fusion is a difficult task due to the large amount of data that needs to be processed and the intricate dispersal of the pertinent information that are within many channels. The pulse-coupled neural network (PCNN) has previously been shown to be a very powerful image processing tool for single channel images. A multichannel PCNN for processing of multichannel images is introduced. Fusion of the pulse images produced by the multichannel PCNN is easily performed by a single channel fractional power filter.

The qualitative and quantitative results of three image data fusion algorithms and their target signature variation are shown. We describe these methods for comparison along with a new and easier to understand process developed for their implementation. We also present an illustration of the effects of variation of the target signature on each of the fusion methods along with measurement of these effects using a new distortion metric.

Although most naturally occurring objects move in a nonrigid manner, the vast majority of the motion research in computer vision has focused on rigid objects due to mathematical elegance and legacy. Driven recently by new applications in telecommunications, medicine, and industry, a new body of research is beginning to emerge that is focused on the recovery of structure and motion for nonrigid objects. Most of these methods, however, have abandoned the large body of work in feature-based rigid body motion recovery in favor of modelbased techniques, which unfortunately limits their abilities in completely unknown environments. Four token-based, rigid body structure from motion (SFM) algorithms are modified for the recovery of structure from nonrigid motion by replacing the rigidity constraint with the temporal fusion of an image sequence. The results of numerous experiments using simulated imagery are also presented to demonstrate and characterize the performance of the four modified techniques.

The problem of dynamic distributed sensor fusion, where time critical sensor readings are fused concurrently by physically isolated modules, is considered. A large number of critical applications including target tracking, remote sensing, and autonomous vehicle navigation can depend on interpreting sensor readings in real time. Independent sensor modules receive partially contradictory data, but their answers must agree within computed accuracy bounds. Real-time constraints cause additional uncertainty factors. The dynamic nature of this problem requires that the problem be viewed in four dimensions and processed efficiently. We present a new method that robustly tolerates a limited number of component failures. This method has complexity O(f 4N log N), where N is the total number of modules and f is the number of faulty modules tolerated. Experimental results are given showing the merits of the proposed method.

The problem of registration for multiple mobile sensors in a fusion system is considered. We develop a registration model for mobile sensors and implement the registration based on the nonlinear Kalman filter theory. A parallelized extended Kalman filter (EKF) structure is also derived in which the algorithm is parallelized and processed at each sensor. The parallelized EKF is shown to have the performance of a centralized system. It does not rely on a single central facility to perform and thus is less sensitive to sensor defects than the centralized system. Finally, computer simulations are used to show the effectiveness of the proposed registration algorithm.

A novel methodology offering the potential for resolving two of the significant problems of implementing multisensor target recognition systems, i.e., the rational selection of a specific sensor suite and optimal allocation of requirements among sensors, is presented. Based on a sequence of conjectures (and their supporting arguments) concerning the relationship of extractable information content to recognition performance of a sensor system, a set of heuristics (essentially a reformulation of Johnson’s criteria applicable to all sensor and data types) is developed. An approach to quantifying the information content of sensor data is described. Coupling this approach with the widely accepted Johnson’s criteria for target recognition capabilities results in a quantitative method for comparing the target recognition ability of diverse sensors (imagers, nonimagers, active, passive, electromagnetic, acoustic, etc.). Extension to describing the performance of multiple sensors is straightforward. The application of the technique to sensor selection and requirements allocation is discussed.

The self-organizing neural network Dignet is used in the design of a two-stage parallel data fusion system. The data fusion is applied to target detection problems in a multichannel moving target indication (MTI) system. Features of the received data from three different radar channels are extracted via digital signal processing techniques. Pulse compression, clutter canceling, and the fast Fourier transform (FFT) are used to transform data from the time-range domain to the range-Doppler domain for feature processing. A first-stage Dignet module on each channel is used for feature clustering and prefiltering. The clustering results from each channel Dignet module are passed on to the fusion Dignet for the second-stage clustering. Map regularization, circular metric, and contrast enhancement are used to resolve the misalignment problem of features on the range-Doppler maps from sensors on different channels. Each first-stage Dignet module performs feature extraction and clustering as a filter-upon-demand module that generates adaptiveness as it is required by the data clustering. Experimental results on field data have shown that a Dignet-based multiradar data fusion system successfully detects a moving target embedded in clutter.

This paper introduces the single-lens joint transform correlator (SLJTC). The SLJTC is a two-stage processor. The first stage computes the chirp-modulated joint power spectrum. The chirp-encoded joint power spectrum may be viewed as containing an amplitude-encoded lens function or Fresnel zone plate. The correlation output is in the focal plane of this zone plate. As in the chirp-modulated joint transform correlator (JTC), the output contains a single correlation peak. The SLJTC has been experimentally demonstrated with a peak-to-noise ratio (PNR) of 17.1 dB. The same chirp-encoded joint power spectrum used in the second stage of a binary chirp-modulated JTC resulted in 6.5-dB-poorer PNR performance.

We design, implement, and test an optical joint transform correlator, using a nonlinear optically addressed ferroelectric liquid-crystal spatial light modulator in the Fourier plane. This compact correlator performs real-time and on-board recognition tasks. Results of in-car experiments for road sign recognition in real environments are presented.

When using the single-point diamond-turning (SPDT) technique to produce oxygen-free high-conductance (OFHC) copper reflector, it is found that surface quality, in terms of microroughness, reflectivity, laser damage threshold, residual stress, etc., has a close relation with the microstructure and the processing sequence of the material. In addition, the uniformity of the crystal grains and the heat-treatment processes largely affect the microroughness and the residual stress of the finished surface. Finally, an optimized processing method before diamond turning is given to obtain the best surface quality.

It is demonstrated experimentally at the output of an optical fiber excited by a low-coherence source that the mutual interference of two linearly polarized (LP) modes can be resolved in the spectral domain when the optical path difference between both LP modes exceeds the source coherence length. The spectral interference between two LP modes, which shows up as a periodic modulation of the source spectrum, serves as an illustration of the feasibility of a novel measuring technique utilizing a high-resolution spectrometer in the evaluation of the group delay time difference between both LP modes. Moreover, it is shown that the spectral analysis performed across the output modal pattern of the two-mode optical fiber leads to spectral modulations in which the effects of both polarization properties and overlaps of modes are inscribed.

The performance of a self-referenced fiber sensor is analyzed. The theoretical limits of its performance are identified.

Based on possible optoelectronic realization of adaptive filters and equalizers using fiber optic tapped delay lines and spatial light modulators we describe the design of a single-layer fiber optic Adaline neural network that can be used as a bit pattern classifier. In our design, we employ as few electronic devices as possible and use optical computation to utilize the advantages of optics in processing speed, parallelism, and interconnection. The described new optical neural network design is for optical processing of guided light wave signals, not electronic signals. We analyze the convergence or learning characteristics of the optoelectronic Adaline in the presence of errors in the hardware. We show that with such an optoelectronic Adaline it is possible to detect a desired code word/token/header with good accuracy.

A relaxation algorithm for removing impulse noise from highly corrupted images is proposed, where a nonlinear probabilistic model is used to reduce the ambiguity of each pixel based on the contextual information. Each pixel is given three labels which stand for ‘‘positive corruption,’’ ‘‘no corruption,’’ and ‘‘negative corruption.’’ The initial probabilities of a pixel whose gray level lies in the middle intensity range of the whole image are simply given some fixed values. Those of a pixel whose gray level lies in the upper or lower part of the intensity range are determined by its gray level and the difference with the median value in a 3×3 window. To display the image after each iteration, the gray level of a pixel with high no-corruption probability is set to its original value approximately, while the gray level of a pixel with high corruption probability will be replaced by the average value of its neighbor pixels with high no-corruption probabilities. Experimental results show that this algorithm can effectively remove both positive and negative impulse noise with very high probability and is superior in performance to some other methods for highly corrupted images.

This paper presents the design and VLSI implementation of a new ASIC that performs in real time the morphological operations of dilation and erosion. The ASIC’s architecture is based on the extension of the majority-gate algorithm for morphological operations. The ASIC was implemented using a DLM, 0.7-?m, CMOS, N-well process, and it occupies a silicon area of 14.78 mm2. Its maximum speed of operation is 92.5 MHz. Targeted applications include machine vision, where the need for short processing times is crucial.

The mutual relation and the observed correlation between reflective and thermal IR images of ground terrain are presented and discussed. The joint distribution density function of thermal and reflective images and its diurnal behavior are presented and analyzed. An approximated analytic expression of the correlation coefficient between thermal and reflective images derived from the explicit form of the joint distribution function is calculated and presented, but it fails to predict the observed correlation between images in these two bands. A modification of the distribution function to consider the influence of local ground topography in the scene is introduced. The prediction of the modified joint density function and correlation coefficient shows a much better agreement with the experimental data. The meaning and the practical applications of the model and the experimental data are discussed.

In industry, mechanical stylus profilometry is the prevailing technique for assessing surface roughness and related parameters of engineering surfaces. Though it is a conceptually direct method, it suffers from several drawbacks: the measurement is not area-covering, while many important features of engineering surfaces are related to three-dimensional topography; the technique is slow; and it involves physical contact with the surface under test, with the danger of damaging it. We have used infrared scattering at 10.6-?m wavelength to provide area-covering, noncontact surface characterization of engineering surfaces. The surface power spectral density is calculated from the scattering distribution by Rayleigh-Rice vector perturbation theory. Measurements of power spectral density functions and bandwidth-limited surface roughness are demonstrated for ground and polished surfaces in the root-mean-square roughness range of 0.06 to 1.7 ?m. Good correlation with contact stylus measurements is achieved for ground and lapped surfaces. Measurement of scratch and lay characteristics is also demonstrated

Reticle seekers temporally modulate target location onto the incoming spatial signal. When large or multiple targets are present in the field of view (FOV), however, it is hard to precisely modulate the incoming target signal by the reticle. To solve this loss-of-modulation-depth problem, we present an improved reticle seeker using the segmented focal plane array (FPA). The new reticle system uses the normalized difference as well as the modulated signal of each detector output in the segmented FPA. In simulation, we have ascertained the proposed system can perform effective analysis and tracking for multiple or large targets.

In a three-guide coupler with multimode central waveguide, more than two modes of the entire structure participate in the coupling between the outer waveguides. Using a three-mode approximation we found simple conditions for complete power transfer between the outer waveguides: the device length has to match certain multiples of the conventionally defined coupling length. The specific form of the relevant modes allows us to design a magneto-optic isolator or circulator with significantly reduced device length (as compared to the conventional nonreciprocal coupler). The performance of the proposed devices is simulated by propagating-mode calculations. Estimates for admissible fabrication tolerances for the layer thicknesses are presented.

`A lateral-shearing interferometer specially devised for production-line inspection of lenses is presented. The interferometer is configured with four square prisms whose relative sliding motions readily provide lateral shearing and phase shifting necessary for evaluating the beam wavefront collimated by the lens under test. The prisms are attached face-to-face using immersion oil so that undesirable disturbances from external mechanical vibration and atmospheric turbulence can be minimized in the measurement. Further, a special phase-measuring least-squares algorithm is adopted to compensate the phase-shifting error caused by the thickness variation in the immersion oil holding the prisms.

The wavefront sensor is a key component in applications such as adaptive optics, atmospheric turbulence analysis, and flow visualization. A prototype achromatic three-wave lateral-shearing interferometer is described. It is adapted to low-light-level conditions and enables high spatial resolution of the analyzed wavefront. It is essentially based on the use of a specific three-level phase grating.

For a wavefront tested by shearing interferometer, the Zernike polynomial coefficients of the wavefront are found in an analytic form expressed by the Zernike polynomial coefficients of the shearing interferograms. The validity of the new algorithm is tested by numerical simulations and real interferograms.

The cotton industry needs a method to identify the type of trash [nonlint material (NLM)] in cotton samples; learning vector quantization (LVQ) is evaluated as that method. LVQ is a classification technique that defines reference vectors (group prototypes) in an N-dimensional feature space (RN). Normalized trash object features extracted from images of compressed cotton samples define RN. An unknown NLM object is given the label of the closest reference vector (as defined by Euclidean distance). Different normalized feature spaces and NLM classifications are evaluated and accuracies reported for correctly identifying the NLM type. LVQ is used to partition cotton trash into: (1) bark (B), leaf (L), pepper (P), or stick (S); (2) bark and nonbark (N); or (3) bark, combined leaf and pepper (LP), or stick. Percentage accuracy for correctly identifying 139 pieces of test trash placed on laboratory prepared samples for the three scenarios are (B:95, L:87, P:100, S:88), (B:100, N:97), and (B:95, LP:99, S:88), respectively. Also, LVQ results are compared to previous work using backpropagating neural networks.

A rotation-invariant algorithm based on binary circular filters is developed for optical pattern recognition. The features of the genetic algorithm provide a highly efficient and rapid learning process. During training, the parameters of a circular filter are selected to maximize the distinction between the target and other expected objects in the image. These iteratively designed filters are good discriminators because they utilize all the spatial visual information about the target. Filters that optimize the trade-off between noise robustness and sharpness of the correlation peak can be determined. Binary circular filters when combined with spatial light modulators are appropriate for real-time applications.

We investigate an all-phase-encoding technique in the input plane for security verification. In this technique, a primary amplitude image is first phase encoded and then multiplied by a random phase code. We investigate the performance of this technique by evaluating the discrimination between different biometrics images and between different random codes using Fourier-plane nonlinear filtering such as nonlinear joint transform correlators. The amplitude images can be any biometrics, and in this study we use fingerprints. We show that we can properly choose the nonlinearity of the filter and the coding weight for the random code to ensure good discrimination against unauthorized random codes and fingerprints. We also study the robustness of the all-phase-encoding technique to input image distortions such as rotation.

Analytical geometry, ray tracing, and the Mueller-matrix calculus are used to determine the calibration matrix of the four-detector photopolarimeter (FDP) when the incident light beam deviates from the optical axis of the instrument. Detectors of the FDP are represented by their surface equations in Cartesian coordinates. The instrument matrices corresponding to angular deviations of ±0.5 and ±1.0 deg in reference horizontal and vertical planes, and their differences from the unperturbed matrix are given for two FDP configurations. Output Stokes parameters that correspond to several typical input states of polarization are calculated to demonstrate the effects of light-beam deviation.

While a single ferroelectric liquid-crystal (FLC) cell with a retardation of half a wave can be used as a switchable polarization rotator with monochromatic light, the variation of its retardation with wavelength results in a significant deterioration in its performance with white light. We show how a broadband response that is almost independent of the wavelength can be obtained by using one or two static birefringent elements in conjunction with an FLC cell. The influence of material characteristics on the results obtained with the two designs is discussed.

An original combination of acousto-optic modulators and electro-optic cells for stereoscopic laser videoprojection is presented. It is based on the separation by orthogonal polarizations of the two left and right images of a stereoscopic pair. As several video lines are scanned simultaneously, the acousto-optic modulators operate in multifrequency mode. The polarization states are controlled by the electro-optical devices. A practical bifrequency acousto-optic modulator using the slow shear acoustic wave in paratellurite has been designed, and an electrooptic device with four ammonium dihydrogen phosphate crystals has been implemented. Experimental results are reported.

The book provides an excellent survey of nonlinear optical materials and macroscopic phenomena, even covering such topics as periodically poled waveguides and cascaded second-order effects. Important equations are highlighted throughout and often are conveniently cataloged into tables.

This PDF file contains the editorial “Optical Storage and Retrieval: Memory, Neural Networks, and Fractals” for OE Vol. 36 Issue 03