Efficient high-speed algorithms are in great demand for applications in which the geometrical configuration of the environment must be assessed before a subsequent move can be performed. The knowledge of the spatial configuration of the object distribution is either a priori or obtained from fused and converted sensor data. The new method can: (1) readily implement known and new sensor data inputs (2) process the resulting geometry in three-dimensional space for location and intersect and (3) permit a system response with a best path in less than a second. Due to its simple architecture the system can treat threats targets terrain and moving objects in the same hardware. 1.
Unlike MSS LANDSAT imagery and other photography the specific characteristics of the intensity of water and shadow in an SAlt image make the task of discriminating them extremely difficult. In this paper we present a scene analysis system for automatically identifying the water regions and shadow regions whose differences appear as subtle differences in tone and texture. In the preprocessing a region dependent Multi-threshold Adaptive Filter (MTAF) is proposed for texture preserving noise removal. In the low-level labeling a probabilistic relaxation algorithm with dynamic adaptive compatibility coefficients is provided to extract the initial object regions. In the high-level interpretation a relational graph model based on the knowledge of water and shadow regions on SAR imagery is constructed. Then spatial reasoning is carried out according to this relational model. The contextual information from different processing modules either at the pixel level or region level is consistently combined to reduce the labeling ambiguities. The experiments shown that more than 85 of shadow regions and water regions on a set of four SAR images can be identified correctly by this system. 1.
The problem of fusion of mnultisensor data for multitarget tracking is considered. A fusion model is presented for fusion of numeric and nonnumeric data from multiple sensors and a clustering technique is introduced. A modified version of Dempster theory is used to combine evidence and to determine object type. An example is presented to illustrate the effectiveness and methodology. 1.
This paper describes the implementation of a test-bed for the real-time evaluation of target detection and tracking algorithms for infrared (IR) imagery. It contains a description of the test-bed algorithms their mapping into real-time digital hardware and an analysis of the system performance. 2.
In the present device, which consists of a SAW sensor and a specialized IC processor, the former generates a set of voltages whose pattern represents the delay or phase difference of the input signals, while the latter produces a parallel digital readout of delays or phases on the bases of voltage comparisons and unclocked logic operations. If equipped with a delay on one input branch, this SAW sensor can measure frequency using the same processor. Six- and seven-bit devices are envisioned with response times of the order of 100 nsec.
Inverse SAR (ISAR) and target-motion resolution (TMR), both of which are based on the phase of the carrier signal rather than the reradiated pulse''s time-of-arrival, entail the resolution of phase variations due to translation. A novel method is presented for such motion compensation which employs a spline function to both smooth the moving target trajectory and complete vernier correction on the basis of phase compensation. Simulation results are presented which demonstrate an ISAR and TMR rms range-error reduction from 7 m to only 0.001 m.
A parametric analysis is conducted for device and system design issues associated with the miniaturization of a hybrid optical correlator that incorporates an electronically addressed liquid-crystal spatial light modulator (SLM). Attention is given to the requirements of drive and readout electronics, as well as the associated optics. The parametrics resolve around the SLM, which is the correlator size-limiting element; emphasis is accordingly placed on the importance of small pixel pitch and minimization of ''dead space'' in order to maximize the miniaturized correlator''s performance.
System design considerations are presented for a rugged optical-processing system for real-time pattern recognition in a military environment, with a view to the criteria of significant system-capabilities demonstrations. The solid-optics correlator presented employs an electronically addressable magnetooptic spatial light modulator at the input to allow easy interconnection with a wide variety of sensor systems; the output device used is a commercially available CCD array. An ideal demonstration of the correlator would involve an antiarmor/antihelicopter beyond-line-of-sight missile that is controlled via bidirectional fiber-optics link.
In this paper, the application of Joint Transform Correlator (JTC) techniques for use in automatic target recognition using actual sensor data is addressed. The problem of interest is the detection and classification of objects in forward looking infrared (FLIR) images. A Joint Transform Correlator architecture using a single magnetooptical spatial light modulator (MOSLM) and a charge coupled device (CCD) camera is designed and tested. A unique technique for binarizing the fringe pattern, commonly called the joint transform power spectrum (JTPS), that enhances the application to actual sensor images is presented. The modification of the architecture to allow for scale and rotation invariant target recognition is also presented.
De8cription of boundary curves (shapes) is an important problem in image processing and pattern recognition. During the last two decades there have been a variety of approaches to the problem. Among these approaches the Fourier description (FD) techniques seem to be the most promising in extracting the features of an object. But the problem of the FD techniques that have been practised is the difficulty in describing local information. A modified FD technique is suggested in this paper which use a combined I requencyposition space as the shape descriptor domain. This new set of position-dependent Fourier descriptors provides the best spectral information along the boundary curve. Experimental results are presented in this paper. 1.
N-diniensional filters are employed to alleviate the bottleneck in optical correlators caused by the output detector array. These filters are implemented by time integrating the response of a number of filters on the detector array. Ndimensional filters provide superior clutter rejection and can effectively accelerate the speed of the sequential filtering operations that are needed for robust optical pattern recognition. 1.
A goal-driven approach to the design of SDF correlation filters is described. It is shown that a variety of filters, each designed for a specific goal or result, are derived from a common generic form which is a multi-stage extension of the conventional SDF filter. The SDF construction affords many degrees of freedom which can be exploited to effect a specialized result.
Synthetic Discriminant Functions have had several different names over the last ten years of their development from the earliest type called linear combinatorial filters to one of the latest versions called the minimum average correlation energy filter. The ten years of development produced many different variations and efforts toward significant advancement over two dimensional matched filters. Most of this filter development was oriented toward optical implementations however the test results presented here are selected from many filters tested digitally and are considered exemplary of the major types for use in both optical and digital implementations. 2. SDF Development Stages We categorize three major stages in SDF development each of which is considered an improvement over the Standa SDF using N training Images d long previous one. The original or first SDF where d " N. method is a linearly combined reference N set'' using the technique outlined In h . a. . SDF filter Figure 12. liii The limitation of this original T v approach Is that the correlation surface is 1 not guaranteed to be anything specific or N T N defined except at the registered position. In 1 a X1 X V1 a A j123 - - - N other words there is no control over the Ri i1 output correlation surface except at one point. This result is of course in general . much different than matched filter would a R V produce and was not very useful until phase encoding schemes were introduced. 3 The values of v1 t The are the training Images. The phase only encoding shown in ______________________________ Figure 2 and the binary phase encoding Figure 1 produce dramatically improved results. 110 / SPIE Vol 1297 Hybrid Image and Signal Processing /1(1990)
In this paper we propose a digital optoelectronic computer capable ofperforming a rich variety of full text search operations. Most of these operations require searching a significant number of documents to retrieve only a few pages. In the proposed configuration textual data are stored on optical disks retrieved in parallel and processed onthefly by the optical computer. Only data that satisfy a particular query are converted into electronic signals and transferred to the enduser. A combination of twodimensional spatial light modulators and photodetectors are used for array processing. In this way the system takes full advantage of the speed and parallelism of digital optical technology. Finally the implementation of a set of text processing operations is shown. 1.
The present comparative study of fundamental differences between electrical and optical interconnects proceeds from an analysis based on Maxwell''s equations. Attention is given to the combinational properties of the electrical interconnects for N gates, within the constraints of microelectronics technology. A gate-level comparison is conducted between electronic and optoelectronic fan-out vs switching time and fan-out vs power. Signal distortion, electromigration, and thermal transfer problems are associated with interconnect fan-out; in general, smaller feature sizes lead to greater electromigration effects than signal distortion effects at all frequencies.
A bidirection matrixvector raultiplication scheiie leads to faster convergence as well as guaranteed convergence in a relaxation processor for parallel solution of linear algebraic equations when used in a Bimodal Optical Computer. 1.
We present a new hybrid optoelectronic method for nonlinear image processing and demonstrate its application to the enhancement of linear features (e. g. striaghtline segments) in noisy low contrast images. An input image is convolved with a long narrow 2D kernel which is rotated through 360 either continuously or discretely in a large number of steps. The convolution output is measured and the maximum [ Max(x and minimum [ Min(x values at each point (x y) are stored. The output image is then given by some applicationdependent function of Max(x and Min(x We refer to the method to as a rotating kernel mmmax transformation (RKMT). In the enhancement of straightline features two types of kernels are especially useful: (1) a long narrow rectangular profile and (2) a long narrow triangular profile. Calculating the function [Max(x produces significantly enhanced linear features while nondesirable features are suppressed. Better results can be obtained by a cascade system combining a Max(x operation with [Max(x - Min(x Comparison is made with a conventional filtering method. 2.
We present a nonlinear joint transform processor that can perform image enhancement. Image enhancement results are obtained for different degrees of nonlinearity applied to the joint power spectrum. The first order harmonic term at the output plane produces the enhanced image which has the exact Fourier phase information of the input image and a Fourier magnitude of the input modified by the nonlinearity. For compression types of nonlinearities the thresholding will redistribute the energy in the Fourier magnitude of the image by increasing the magnitude of the higher spatial frequencies. Thus the fine details of the image that are contained in the high spatial frequencies of the joint power spectrum are enhanced. We investigate the effects of various types of nonlinearities on the enhanced images. Analytical expressions for the enhanced images obtained by the nonlinear techniques will be provided. Computer simulations of the nonlinear processor for image enhancement are presented to study the performance of the system. We show that the nonlinear technique produces reasonably good results. II.
An investigation is made of means to the standardization of methods for parameter characterization in the cases of spatial light modulators of optically addressed, electrically addressed, amplitude-modulating, and phase-modulating types, with a view to recommendations for future practice. The speed, resolution, and visibility parameters are noted to compare well irrespective of modulator type; phase and amplitude modulating devices cannot be directly compared, however, due to the fundamentally different ways that parameters are measured for each. To be meaningful, the MTF for an optically-addressed device must be accompanied by speed and illumination data. Lifetimes should be specified for all devices.
A recently introduced algorithm called the Successive Forcing Algorithm (SFA) is generalized for designing Binary Phase-Only Synthetic Discriminant Functions (BPOSDFs). Preliminary simulation results obtained using this filter are presented. 1
Ternary Phase-amplitude Filters (TPAFs) used in a real-time hybrid (optical/electronic) correlator system coniprise a promising pattern recognition approach with potential for iiear-term j)ractical applications. Range images obtained from LADAR sensors present unique problems due to their particular signal and noise characteristics. We report computer simulation results of the application of both known and new TPAF formulations to the problem of target recognition on actu al LA DAR. images. Binary input im ages suitable for input using magneto-optic spatial light modulators were gemierated by a simple preprocessing step which seems particularly suitable for these images. Experimental results verifying the simulated smart filter performance are presented. 1.
The inputs to optical correlators have in general always been an amplitude modulated input either in a binary or in an analog fashion. Little work has been done on the performance of phase modulated inputs to optical correlators. Simulations are performed using phase inputs and amplitude inputs binary and analog and the Homer efficiency is used as the metric of comparison. An ideal phase filter and classical matched filter cases were compared. Results show phase inputs are a viable means of encoding input information. 1 .
Graylevel patterns are used to identify the correlation discrimination performance of the phase-only filter (POF) and amplitude-modulated inverse filter (AMIF) . The correlation tests were repeated with Gaussian random noise added to the input signal. The computer simulation results show that the ANIF based correlator has better discrimination capability than that of the POF based correlator in the absence of noise as well as in the presence of Gaussion random noise over most of the practical range of noise variance. 1.
With the advent of high speed spatial light modulators, it is possible to write binary phase-only filters (BPOFs) faster than standard video rates. However, an optical correlation system which uses a camera operating at standard video rates to capture output correlation responses is limited to a maximum filter rate of 60 Hertz. This paper examines the use of a charge injection device (CID) camera to speed up the processing time of a BPOF-based rotation-invariant optical correlation system. Having integrating capabilities, the CID camera can integrate several output correlation plane responses during the frame time of the camera. In this paper, the performance of an integrating correlator is presented. The influence of factors such as the input image noise density, the number of integrated correlation responses, and the size of ternary low-frequency blocks on correlation responses are explored.
This paper describes the design of a very large scale integration (VLSI) application specific integrated circuit (ASIC) for use in pattern recognition. The pattern recognition scheme uses Hu1 and Mailra''s2 algorithms for moment invariants. A prototype design was generated that resolved the long delay time of the multiplier by custom designing adder cells based on the Manchester carry chain. Use of the Manchester carry chain effectively incorporated the lookahead carry function into the adder cells. The prototype ASIC is currently being fabricated in 2. 0-mm compiled simulator for metal oxide semiconductor (CMOS) technology (simulated at 20 MHz). The prototype consisted of a 4x8 multiplier and a 12-bit accumulator stage. The present ASIC design consists of a 9x26 multiplier (maximum propagation time of 50 ns) and a 48-bit accumulator stage. The final ASICs will be used in parallel at the board level to achieve the 56 MegaPixels/s [230 million operations per second (MOPs)] necessary to perform the moment invariant algorithms in real time on 512x512 pixel images with 256 grey scales. 2.
This paper presents a new type of adaptive smoothing technique for range images based on intrinsic properties of differential geometry. The concept of intrinsic surface distance and the parallel transport theorem were used to design a filter that is invariant to the viewpoint and is capable of preserving depth and orientation discontinuities. Experimental results are presented. 1
We present a multiscale edge- and region-based segmentation scheme for range images which leads to a rich representation in terms of focused edges and constant-sign curvature regions. At a given scale the edge detection scheme extracts the discontinuities in surface depth and orientation. The extraction is accomplished by first detecting the presence of significant edges at that scale and then determining their precise location by tracking them over decreasing scale down to the finest scale at which they are most accurately positioned. At the same scale the region-based segmentation scheme consists of applying anisotropic filtering to the regions delimited by the previously detected edges and segmenting these regions into surface primitives of constant Gaussian and mean curvature signs. Experimental results are presented for both synthetic and real images and a comparison is done with some recently published techniques. 1.
Although there are many edge detection operators that are used in Automatic Target Recognition (ATR) Systems we believe that the required performance can be accomplished using simple operators such as the Sobel operator with minor modification through pre and postprocessing. In this paper we describe two methods that enhance the Sobel performance. The first which can be considered pre processing increases the effective size of the operator to 5*5 The second which is postprocessing modifies the edge magnitude output based on the consistency of the local edge direction. Both methods can be easily implemented on SIMD machines and they are effective in deleting isolated edge points which usually are not part of interesting targets. We will describe the implementation of these techniques on a SIMD machine and study their effect on the performance of an ATR system. 1.
The ability to classify texture regions in images is considered to be an important aspect of scene analysis. The information gained from such classification can be used by a computer vision system to assist in image segmentation as well as object identification. In this paper, the use of a neural network model in performing classification of images containing regular textures is investigated. The texture features used in the classification process are Hough transform-based descriptors. The performance and capabilities of the neural network approach are then compared to classical technique utilizing a linear associative memory.
Invariant pattern recognition is achieved by providing the binary codes of moment invariants to the input end of Hopfield network. Combination of moment invariants with neural computing allows us to recognize image patterns which have been subjected to various distortions such as noise translation rotation and scale variation. 1.
An optical neural network architecture and algorithm based on a Hopfield optimization network are presented for multitarget tracking. This tracker utilizes a neuron for every possible target track and a quadratic energy function of neural activities which is minimized using gradient descent neural evolution. The neural net tracker is demonstrated as part of a system for determining position and orientation (pose) of spinning sateffites with respect to a robotic spacecraft. The input to the system is time sequence video from a single camera. Novelty detection and filtering are utilized to locate and segment novel regions from the input images. The neural net multitarget tracker determines the correspondences (or tracks) of the novel regions as a function of time and hence the paths of object (sateffite) parts. The path traced out by a given part or region is approximately elliptical in image space and the position shape and orientation of the ellipse are functions of the satellite geometry and its pose. Having a geometric model of the satellite and the effiptical path of a part in image space the 3-D pose of the satellite is determined. Digital simulation results using this algorithm are presented for various sateffite poses and lighting conditions. 1
We present a neural network algorithm for the detection and classification of transients in noise. The algorithm is a feed-forward three layered network capable of training and testing in a single pass. It benefits from the use of a time-frequency representation that conveys temporal information about the signal. 1.
A user friendly flexible system for assembly line part inspection which learns good and bad parts is described. The system detects missing rivets and springs in clutch drivers. The system extracts features in a circular region of interest from a video image processes these using a Fast Fourier Transform for rotation invariance and uses this as inputs to a neural network trained with back-propagation. The advantage of a learning system is that expensive reprogramming and delays are avoided when a part is modified. Two cases were considered. The first one could use back lighting in that surface effects could be ignored. The second case required front lighting because the part had a cover which prevented light from passing through the parts. 100 percent classification of good and bad parts was achieved for both back-lit and front-lit cases with a limited number of training parts available. 1. BACKGROUND A vision system to inspect clutch drivers for missing rivets and springs at the Harrison Radiator Plant of General Motors (GM) works only on parts without covers Fig. 1 and is expensive. The system does not work when there are cover plates Fig. 2 that rule out back light passing through the area of missing rivets and springs. Also the system like all such systems must be reprogrammed at significant time and cost when the system needs to classify a different fault or a
A single-layer locally-connected neural network that uses a simple second-order neural architecture is presented. This network demonstrates surprisingly good performance in the detection and isolation of traveling pulses superposed on a randomly fluctuating background. The processin topology and dynamics can be traced to High Order eral Networks (HONNs) and motion sensitive cellular automata '' . Similar processing strategies have also been found in biological processing systems. 1.
The basic operations of mathematical morphology are quite useful for a broad area of image processing and analysis tasks. All morphological operations can be built from erosions and dilations. In this paper we develop a single chip VLSI architecture of an erosion/dilation algorithm for real-time image processing. The new architecture allows sequential inputs and performs parallel processing with 100 percent efficiency. The erosions (mm of differences) and dilations (max of additions) operate on 2-D gray-level image signals and use a 5 x 5-pixel gray-level structuring element. Two chips can be cascaded for a 7 x 7 structuring element. The overall chip design also provides the user with sufficient flexibility to optionally offset the output images by adding a constant level and/or scale their dynamic range. 1.