A system for determining the position, orientation and motion of a satellite with respect to a robotic spacecraft using video data is advanced. This system utilizes two levels of pose and motion estimation: an initial system which provides coarse estimates of pose and motion, and a second system which uses the coarse estimates and further processing to provide finer pose and motion estimates. The present paper emphasizes the initial coarse pose and motion estimation subsystem. This subsystem utilizes novelty detection and filtering for locating novel parts and a neural net tracker to track these parts over time. Results of using this system on a sequence of images of a spin stabilized satellite are presented.
Convergence mechanism of vectors in the Hopfield's neural network is studied in terms of both weights (i.e., inner products) and Hamming distance. It is shown that Hamming distance should not always be used in determining the convergence of vectors. Instead, weights (which in turn depend on the neuron representation) are found to play a more dominant role in the convergence mechanism. Consequently, a new binary neuron representation for associative memory is proposed. With the new neuron representation, the associative memory responds unambiguously to the partial input in retrieving the stored information.
A 2-D hybrid optical neural network using a high resolution video monitor as a programmable associative memory is proposed. Experiments and computer simulations of the system have been conducted. The superior resolution and the number of gray levels of the video monitor lead to the implementation of a larger number of neurons and a larger dynamic range. The system operates in a high speed asynchronous mode due to the parallel feedback loop. The programmability of the system permits the use of the Orthogonal Projection(OP) and the Multilevel Recognition (MR) algorithms to increase the error correction ability of the network. Using current integrated optics and electronic technology this optical neural system attains high learning and operational speed.
Neural networks are composed of three components: neurons, weights, and connections. Limitation on implementing these components in electronics and optics are discussed. We show that using electronically fabricated neurons and a combination of optics and electronics for the weights and connections efficiently utilizes the advantage of each technology. The choice of the technology used to implement the weights and connections depends on the type of neural network being built.
A new capacity measure is proposed for the class of dynamic autoassociative neural memories (ANMs). The proposed associative memory capacity measure is based on the memory's performance vs. the number of stored memory vectors, where performance is defined in terms of the memory's error-correcting ability and its fundamental memories' attraction volumes. This new method of measuring ANM capacity is very effective when ANMs are used in pattern recognition and error-tolerant content-addressable memories. The proposed capacity/performance measure has been tested for several ANMs having the same dynamic Hopfield memory-like architecture, each employing a different recording technique. Correlation, generalized inverse (orthogonal), and Ho-Kashyap memory recordings have been investigated. Monte Carlo analysis has been performed on ANMs recorded with randomly generated patterns, in order to determine and compare performance characteristics and dynamics.
An optical associative memory system suitable for both auto- and hetero-associative recall is demonstrated. This system utilizes Hamming distance as the similarity measure between a binary input and a memory image with the aid of a 2-D optical EXCLUSIVE OR (XOR) gate and a parallel electronics comparator module. Based on the Hamming distance measurement, this optical associative memory performs a nearest neighbor search and the result is displayed in the output plane in real-time. This optical associative memory is fast and noniterative and produces no output spurious states as compared with that of the Hopfield neural network model.
New optical architectures for the implementation of associative memories based on neural networks are proposed. The application of the electron trapping (ET) materials to the new architectures of associative memories are described. Preliminary experimental results of the ET materials to evaluate the feasibility of the proposed architecture is presented.
An adaptive dynamic artificial neural memory is proposed for pattern recognition applications. The proposed neural memory has a simple layered structure of neural processing units (neurons) with feedback which is ideal for parallel optical implementations. An adaptive version of our earlier-proposed high-performance neural memory recording algorithm (Ho-Kashyap recording algorithm) is utilized for the memory learning phase. This learning algorithm is computationaly inexpensive and leads to high-performance associative memory characteristics. The combination of this algorithm with a dynamic heteroassociative memory architecture gives rise to high associative memory capabilities which are suitable for adaptive and robust pattern classification applications. The state-space characteristics of dynamic heteroassociative memories (DAMs) utilizing various recording/synthesis algorithms are studied and the advantages of the proposed associative memory over the earlier proposed bidirectional associative memory (BAN) and generalized inverse-recorded heteroassociative memory are established and analyzed.
Due to the linear characteristics of conventional optical pattern recognition processors, many useful recognition operations cannot be easily performed. In this paper, the simplest higher order nonlinear recognition system, i.e., the bilinear system, will be introduced for optical pattern recognition applications.
Space-invariant lateral interconnects and winner-take-all concepts can be combined to produce optical neural networks to process the output plane of a conventional optical pattern recognition system. Advantages in clutter rejection, sensor fusion, preference for smoothly moving targets, etc. are readily obtainable. The output plane energy distribution will be vastly simplified as well. In the limit, the output plane will be one or more isolated points representing target locations.
We describe a binary joint transform correlator (JTC) that uses a synthetic discriminant function (SDF) at the input plane to store the reference images. The SDF is displayed at the input plane side by side the input scene. The input scene, the SDF, and their joint power spectrum are thresholded to only two values. Thus, binary spatial light modulators (SLMs) can be used to implement the correlator. The correlation performance of the SDF-based binary JTC is compared to that of the classical SDF-based correlator in the areas of correlation peak intensity, peak to sidelobe ratio, signal-to-noise ratio, and correlation width. We show that the SDF-based binary JTC outperforms the classical correlator in all these areas. Furthermore, a single binary SLM can be used to read in sequentially the binarized input signals and the binarized joint power spectrum which results in a significant reduction in cost, size, and complexity of the system.
The use of a hybrid (optical-digital) system for image processing is presented. The implementation of a magneto-optic spatial light, modulator in the Fourier plane for real-time generation of spatial filters is investigated. The application of this system to document analysis and fingerprint, classification is examined.
We describe a method for bandwidth-efficient processing of video imagery to be viewed by the teleoperator of a remotely-operated vehicle on which the camera is mounted. The method comprises image coding, transmission, and reconstruction. We make certain initial assumptions: first, that transmission bandpass is the limiting factor rather than encoding/decoding schemata; second, that image coding and reconstruction will be done within the general abilities of the NASA/TI Programmable Remapper; and third, that the ratio of retained local detail to the operator's visual resolution is held constant throughout the large-field image he sees. Novel features include that the compression and reconstruction address certain characteristics of the human visual system (HVS), that two-way communication controls a moving "fovea" in the transformation, and that resolution varies over the image. Conventional motivations accommodated include the Cartesian raster-scan nature of available imagers and display devices and a need for low bandwidth in the image transmission. Unique image processing hardware, NASA's Programmable Remapper, allows demonstration of the method. Once refined, the technology could be adapted to special purpose imagers and display devices, or otherwise to dedicated image processing hardware.
We are investigating image coordinate transformations possibly to be used in a low vision aid for human patients. These patients typically have field defects with localized retinal dysfunction predominately central (age related maculopathy) or peripheral (retinitis pigmentosa). Previously we have shown simple eccentricity-only remappings which do not maintain conformality. In this report we present our initial attempts on developing images which hold quasi-conformality after remapping. Although the quasi-conformal images may have less local distortion, there are discontinuities in the image which may counterindicate this type of transformation for the low vision application.
We propose an innovative hybrid real-time vision system concept for autonomous landing on planetary bodies. It contains digital and optical processing methods. It includes high speed digital image warping to preprocess a video image for optical correlation, position estimation by synthetic estimation filtering, image stabilization by joint transform optical correlation, and hazard identification from measurements of optical flow by sequential image subtraction.
New optical correlation filter results are reviewed for use in symbolic and inference correlators for ATR scene analysis. MACE filter synthesis has attractive properties that include: easily detectable peaks, distortion-invariance, simplified training set selection, solutions to input bias effects, performance in high noise, performance in real background clutter, region-of-interest MACE filters, filter quantization tests, and less clutter with its reduced number of training set images.
A multi-channel holographic correlator has been constructed which can identify and track objects of a given shape across the input field independent of their in-plane rotation. This system, derived from the classic Vander Lugt correlator, incorporates a hololens to store an array of matched spatial filters (MSFs) on thermoplastic film. Each member of the MSF array is generated from a different incrementally rotated version of the training object. Rotational invariant tracking is achieved through superposition of the corresponding array of the correlations in the output plane. Real time tracking is accomplished by utilizing a liquid crystal light valve (LCLV) illuminated with a CRT to process video input signals. The system can be programmed to recognize different objects by recording the MSF array on re-usable thermoplastic film. Discussion of the system architecture and laboratory results are presented.
The use of joint transform correlators for pattern recognition has the advantages such as low resolution requirement of the spatial light modulators and easy optical implementation. Based on a programmable JTC architecture, we propose two methods to detect objects under rotation or scale change. Rotation invariant pattern recognition can be achieved by writing the real-value circular harmonic expansion of an image function as reference. By using multiple images, each rotated by a certain angle or at different sizes as composite reference in a JTC, objects of different orientations or sizes can be detected.
The effectiveness of multichannel optical correlators for distortion invariant pattern recognition is examined. The geometric notion of a decision surface is used to assess the performance of two possible systems.
We describe a joint Fourier transform image correlator that employs thresholding at both the input plane and the Fourier plane. This suggests using a single binary spatial light modulator (SLM) to read in sequentially the binarized input signal and the binarized Fourier transform interference intensity. The performance of the single SLM joint Fourier transform correlator (JTC) is compared to that of the classical JTC in the areas of correlation peak intensity, peak to sidelobe ratio, signal-to-noise ratio (SNR), and correlation width. We show that the single SLM JTC outperforms the classical JTC in all such areas. Using a single binary SLM results in a significant reduction in cost, size, and complexity of the system.
A Joint Transform Correlator which uses inexpensive liquid crystal televisions as spatial light modulators has been constructed. The advantages that this architecture offers over the Vander Lugt matched filter type correlator will be examined. The disadvantages shared by both architectures, as well as some constraints unique to the Joint Transform configuration, will also be discussed.
We report initial experimental investigation of the Texas Instruments deformable mirror device (DMD) in a joint optical transform correlator. We used the inverted cloverleaf version of the DMD, in which form the DMD is phase-mostly but of limited phase range. Binarized joint Fourier transforms were calculated for similar and dissimilar objects and written onto the DMD. Inverse Fourier transform was done in a diffraction order for which the DMD shows phase-mostly modulation. Matched test objects produced sharp correlation, distinct objects did not. Further studies are warranted and they are outlined.
A new type of composite filters called the Linear Phase Coefficient Composite Filters (LPCCF) is introduced. Unlike previous approaches to composite filter design, this method considers the training set selection and the filter design simultaneously. The LPCCF is obtained by summing the training images weighted by complex weights of unit magnitude and linear phase. This paper also presents a way of combining the outputs of a bank of LPCCFs. Computer simulations are also included to quantify the performance of this approach.
Efficient methods for determining the passband of the Binary Phase-Only Filter (BPOF) such that we get the maximum possible output Signal-to-Noise Ratio (SNR) are identified. Computer simulation results are presented to illustrate the advantages of including this optimal passband.
Optical correlator systems using the binary phase-only filter (BPOF) have shown excellent performance in a variety of both theoretical and experimental studies. The output of the system represents the correlation of the input function g(x,y) with the impulse response function of the BPOF designed for a given target pattern f(x,y). Unfortunately since this impulse response function contains both the desired pattern f(x,y) and and its inverted image f(-x,-y), the correlator output contains two signals representing the correlation of g(x,y) with both f(x,y) and f(-x,-y). The phase relationship between these two contributions can be varied depending on the algorithm used to make the BPOF. This paper discusses factors which influence the performance of the BPOF when the input pattern is degraded with background noise including the location of the object used to make the filter, the extent of the background noise, the algorithm used to make the BPOF, and object features which affect the filter response. In addition we will present a Fresnel lens-encoded BPOF which very greatly reduces the correlation with the f(-x,-y) term.
Recent developments in spatial light modulators have led to devices which are capable of continuous phase modulation, even if only over a limited range. We used one of these devices, the Texas Instruments deformable mirror device, to compare the relative merits of binary and partially-continuous phase filters in a specific problem of pattern recognition by optical correlation. Each filter was physically limited to only about a radian of modulation. Researchers have predicted that for low input noise levels, continuous phase-only filters should have a higher absolute correlator peak output than the corresponding binary filters, as well as having a larger signal-to-noise ratio. When continuous and binary filters were implemented on the DMD and they exhibited the same performance, an ad hoc filter optimization procedure was developed for use in the laboratory. The optimized continuous filter gave higher correlation peaks than did an independently optimized binary filter. Background behavior in the correlation plane was similar for the two filters, and thus the signal-to-noise ratio showed the same improvement for the continuous filter. A phasor diagram analysis and computer simulation have explained part of the optimization procedure's success.