Recent development in vision and image understanding related study reveals that a signal decomposition before processing may provide enormous useful information about the signal. Various signal decomposition models, such as the Gabor and wavelet transforms have been proposed. While the Gabor signal expansion creates a fixed resolution space-frequency signal representation, the wavelet transform provides a multi-resolution signal space-scale decomposition. Digital implementation of these transforms are computationally intensive both because of the nature of the coordinate-doubling of the transforms and due to the large quantity of convolution/correlation operations to be performed. Optics with its inherent parallel processing capability has been applied to many useful linear signal and image transformations for feature analysis and extraction. This paper is intended to study the suitability of using optical processing techniques for the signal Gabor and wavelet analysis. Gabor and wavelet transforms of both one- and two-dimensional signals and images are discussed. System parameters and limitation are analyzed. Preliminary experimental results are presented.

An optical Harr wavelet is created using a magneto-optic spatial light modulator (MOSLM). Two methods of controlling wavelet dilation are explored: (1) an aperture positioned in front of a binary modulated MOSLM; (2) spatial filtering of a ternary modulated MOSLM. Segmentation is performed through Vander Lugt correlation of a binarized image with the optical wavelet. Frequency-plane masks for the correlation process are generated using thermal holography.

We consider two new techniques (wavelets and Gabor transforms) plus morphological methods for use in the detection, image enhancement, and feature extraction stages of scene analysis. We find the new wavelet and Gabor methods to be of potential use for textural/statistical measures that complement other established methods. The optical realization of all methods on a unified optical correlator architecture is noted. The dimensionality problems of the two new methods should be solved by use of linear combination filters (LCFs) of several wavelet or Gabor functions.

Labeling objects in an image is an important step in many application areas such as target tracking, circuit board and IC mask inspection, medical image analysis, environmental analysis, and character recognition. However, in a real time system the speed of the labeling operation is often hindered by bottlenecks. Our new, fast method to perform labeling first obtains the connectivity information for each pixel in the entire image by means of convolution. Then, it uses the connectivity information to assign a temporary label and generate a small equivalence table. Finally, the algorithm uses the equivalence table to obtain the result. The advantage of this algorithm is its ability to exploit high-speed convolutional processors such as HNC's Vision Processor (ViP). Using the ViP and HNC's Balboa 860 coprocessor board, it takes between 36 and 66 milliseconds for most 512 X 512 images of interest (the time taken for the second step is dependent on image content). This type of fast algorithm running in processors such as the ViP, will yield a new wave in imaging processing algorithm development.

The processing of compressed imagery exhibits computational advantages due to the processing of fewer data, as well as the advantage of low-level data security afforded by the encoding format. We have elsewhere discussed the general theory of compressive processing, and have presented complexity analyses which support the claim of computational speedup. In this introductory paper, general methods are described for the processing of signals and imagery encoded via transform-, block-, runlength-, and derivative-coding schemes. Operations of interest include unary and binary pointwise operations, the global reduce operation, and neighborhood operations, such as morphological erosion and dilation. Implementational analyses emphasize the relationship between the compression ratio and the time complexity of compressive computation.

Tactile information processing has received a relatively small amount of attention in the area of pattern recognition. However, while most of the attention has been given to image, acoustic, and electromagnetic signal processing, there are many applications where tactile information processing is applicable. For example, underwater salvage operations where the water may be opaque with debris and target objects may be covered with viscous substances like silt, sand, mud, or camouflaged by crustaceans. These environments may also be hazardous to humans, because of pressure, temperature, or contamination. Another example is industrial assembly where subcomponents, initially in random position and orientation, need to be assembled together. We assume that an object has been tactilly sampled into a polyhedron mesh. Our concentration in this writing is to identify this mesh as belonging to a target object independent of orientation and position. To solve this problem we present a fundamental approach we call Complex Linear Morphology (CLM). This technique involves non-linear architectures which rely on banks of linear correlation filter elements thus the term linear. These elements are comprised of complex weighted training images or solids, thus the term complex. These complex weights are used to approximate logical operations on the input images or solids which result in the discrimination of target objects from clutter objects, thus the term morphology. There are two architectures presented. The first architecture assumes the 3-D polyhedron mesh is converted to a 2-D image by projection. CLM is applied to these 2-D images which are rotation-variant. The second architecture uses CLM techniques to process 3-D information directly. Results are presented for the 2-D CLM approach and techniques are presented for the 3-D CLM approach.

Three-dimensional adaptive morphological operators are developed for image sequence processing. The theoretical aspect of the operators has been studied in a previous contribution. In this paper, we develop fast algorithms for the implementation of the 3-D operators and investigate their applications. At first, a basic algorithm is described, which is simple but slow. Then a number of propositions are proved, which can be used to modify the basic algorithm to develop fast algorithms. It is shown through experimental results that the algorithms developed are fast and have the potential for real-time video signal processing. The performance of the operators in filtering impulsive noise is also evaluated through experiments. The results show that the operators are suitable for high quality video signal processing.

This paper presents research on decomposition of morphological grayscale structural functions so that grayscale morphological transforms can be implemented quickly. An investigation into some theoretical and methodological issues about decomposition of structural function is conducted based on the results of decomposition of binary structure element in our previous work. Then a hybrid decomposition into neighborhood configurations for a large class of structural functions, which includes almost all useful structural functions especially some nonconvex ones, is proposed by use of the decomposing techniques for structure element in binary morphology. This hybrid decomposition can be calculated efficiently. From this decomposition, a fast implementing algorithm for grayscale morphological transforms is provided. It matches the common parallel image processors directly and so does not need support from complicated and special hardware.

An optical/digital/mechanical six degree of freedom tracking system using an optical correlator for image processing is being constructed at NASA's Johnson Space Center. The degrees of freedom are expressed in sensor coordinates as azimuth, elevation, range, line-of-sight rotation, and the two out-of-plane object rotation angles. Hardware for an initial configuration has been assembled and various tracking algorithms and filtering techniques are being implemented and evaluated. The current correlator hardware is based on LCTV SLMs from a commercial television projector. Correlation peak detection and measurement are made using commercially available digital image processing boards. Out-of-plane object rotation, range, and line-of-sight rotation are tracked by various correlation filter techniques. Performance of the current system is presented, as are plans for future configurations.

Synthetic estimation filters, introduced by Juday and Monroe, have been shown to be very useful in the estimation of pose parameters of objects from their images. These filters are designed from a composite image made up of a linear combination of images which have undergone variations in their position by a known amount. Each filter is designed such that its response for each of the constituent images lies on a straight line. The peak response of the filter was chosen as the response of interest. Though these filters were designed to have an affine response with respect to the pose parameter, the resulting response in general is not affine and this causes considerable error in the estimate. On a detailed study of the SEF filter design, it is found that this discrepancy results because of the use of the maximum response of the filter rather than the response at the origin. Hence, in this paper, new types of synthetic estimation filters constructed on the basis of the filter response at the origin are proposed. These filters, except the phase-only filters, yield exactly the desired response for the constituent images. Three filters of this type -- matched, phase-only, and composite phase filters -- are considered in this paper. Simulation results conducted on these filters using a set of images are presented. The accuracy of estimation is compared with the previous two SEFs - - matched and phase-only filters. It is found that the new filters possess better estimation accuracy. Noise analysis of these filters were also carried out. Both analytical and simulation studies were made. The matched SEFs designed on the basis of the response at the origin were found to possess good noise resistance characteristics.

An optical terrain analysis based sensor system specifically designed for landing hazard detection as required for NASA's autonomous planetary landers is introduced. This optical hazard detection and avoidance (HDA) sensor utilizes an optoelectronic wedge-and-ting (WRD) filter for Fourier transformed feature extraction and an electronic neural network processor for pattern classification. A fully implemented optical HDA sensor would assure safe landing of the planetary landers. Computer simulation results of a successful feasibility study is reported. Future research for hardware system implementation is also provided.

We describe our planned use of optical correlation in landmark navigation associated with planetary landing. Standard correlation provides 'pointing-to' information, giving the vector to the landmark from the spacecraft in the spacecraft frame. The synthetic estimation filter (SEF) provides 'pointing-from' information, estimating the vector from the landmark to the spacecraft. Digital and optical SEFs were constructed and compared using a Martian-like 3D modelboard to provide test images. The digital SEF worked reasonably well, but the optical SEF did not perform as expected. The optical SEF was implemented with a liquid crystal television (LCTV) correlator that had been used successfully in previous SEF experiments using a spacecraft model. The results suggest the SLM model for the LCTV needs further refinements. Both the digital and optical filter provided good pointing-to results. We did not plot the correlation surface for the digital SEF response, and though it is less sharp than the POF, it had only a one-pixel variation in the peak location. Surface plots for the conventional optical phase-only filter produced a correlation peak that was sharp enough to be located within two pixels.

The probability of correct object recognition is calculated assuming a maximum-likelihood algorithm, accounting for the modulation transfer function (MTF) of the atmosphere-imaging- system combination, and for the presence of clutter. The MTF is determined by the spatial bandwidth of the adaptive distortion-compensation filter. The algorithm recognizes an object based on minimum distance in feature space. Each feature has a corresponding scale size, with a corresponding blur from the MTF. It is demonstrated that the Vander Lugt filter is a maximum-likelihood estimator for a broad range of clutter and distortion statistics. Specific probability distributions are assumed for the compensated distortions (log normal), and the clutter strength (beta and Gaussian). The results presented assume an informationless distribution of targets in feature space, and show the variation of performance with the number of features, the quality of the compensation filter, and the strength of clutter.

Infrared imagery of 512 x 512 pixels were processed with 128 x 128 arrays by computer simulation of an optical correlator using various correlation filters. Pyramidal processing using binary phase-only filters (BPOFs), synthetic discriminant function (SDF) filters, and feature-based filters was used to process an entire image in parallel at different resolutions. Results showed that both SDF and feature-based filters were more robust to the effects of thresholding input imagery than BPOFs. The feature-based filters offered a range of performance by setting a parameter to different values. As the value of the parameter was changed, correlation peaks within the training set became more consistent and broader. The feature-based filters were more useful than both the SDF and simple BPOFs for recognizing objects outside the training set. Furthermore, the feature-based filter was more easily calculated and trained than an SDF filter.

The function of any processor is to map input data to output data. Multi-layer processing systems can implement mappings not feasible in single-layer systems. A layered architecture not only facilitates the implementation of non-linear operations, but also provides successive stages for linear processing. We describe the use of layered architectures in optical processing.

Our numerical results demonstrate that both Fraunhofer and Fresnel diffractions provide good approximations to forward scattering by elastic spherical shells and rigid/soft spheres, for nondimensional frequency values 20 < ka < 80, and scattering angles 0 degree(s) < 0 < 10 degree(s). There are only small differences among the Fraunhofer, Fresnel, and rigid/soft predictions for differential scattering cross sections. They arise mainly from the slightly varying degrees of compression (in angular space) of one pattern relative to another. The magnitudes of the maximums predicted by the four methods are in good agreement, but the minimums do not agree. Calculations for 1%, 5%, and 10% spherical steel shells indicate that variation of shell thickness has only a small effect on forward scattering.

We modeled an optical system for estimation of the fractal dimension to provide a measure of surface roughness for an entire image and for image segmentation. Although the simulated optical result was similar to that calculated by digital techniques, both suffered from problems known to occur with estimating fractal dimension. Furthermore, the optical estimation did not have as good a resolution as that obtained with digital estimates due primarily to the limited dynamic range of the detector.

A hybrid optical processor for extracting the invariant moments of images is presented in this paper. The processor is composed of two parts: the optical one consisting of a holographic mask and two Fourier lenses for computing geometric moments of the input images in parallel, and the optoelectric one consisting of a charge-coupled device detector and a microcomputer for calculating invariant moments from the measured geometric moments. The capital letters, C, F, G, H, J, and L, as the input images, are tested to extract their moments by the processor. The produced invariant moments show that the moment values of a letter are approximately independent of shift and rotation, and the moments of different letters are distinct enough for one to recognize them.

A fiber optic based wide-word personality ROM capable of data rates in excess of 125 Mbs has been constructed and tested. Because of the special fan-out characteristics of optical devices, the optoelectronic ROM is capable of operating in a wide-word regime that is inaccessible to all-electronic ROMs. The principle of operation permits n electrical control bits to select among 2ⁿ groups of p optical output bits. In theory, each output word of width p can consist of thousands of bits. The device constitutes a high-speed electro-optic control system permitting a small number of electrical bits to control a geometrically larger number of optical output bits.

A potential application of the photorefractive time-integrating correlator is the real-time radar jamming interference rejection system, using the adaptive filter method; a fast photorefractive crystal is needed for adapting a rapidly changing jamming signal. An effort is presently made to demonstrate and characterize a GaAs-based photorefractive time-integrating correlator, since GaAs crystals are 2-3 orders of magnitude faster than most other alternatives.

Through our experiment, a hybrid optical/electric method has been developed to combine two color (3 - 5 micrometers and 8 - 12 micrometers ) thermal images together. During the fusion process, we have adopted an equivalent optical time delay and integration method which can add dual waveband images directly on the focal planes. Meanwhile, a technique of dynamic range normalization of thermal image has been used to limit the dynamic range of thermal image to match the T.V. monitor. The results of the experiment show that this method can provide a real time (30 frame) single enhanced dual waveband thermal image with a better system performance than the single waveband image.

We report on two new computer generated hologram (CGH) elements for optical processing. They are a one-to-many optical interconnection element (that allows analog weights, high efficiency, and is not restricted to a regularly spaced grid) and an element to provide separate 1-D collimation of the laser diodes in an array (with high efficiency). Error diffusion encoding and multilevel phase CGHs are used to achieve high accuracy and high efficiency. Simulations are used to show the advantage of error diffusion (ED) encoding. Optical laboratory data are included to show the feasibility of the elements and the validity of our simulator.

The authors have achieved a 6.5-to-one image compression technique for thermal images (640 X 480, 1024 colors deep). Using a combination of new and more traditional techniques, the combined algorithm is computationally simple, enabling `on-the-fly' compression and storage of an image in less time than it takes to transcribe the original image to or from a magnetic medium. Similar compression has been achieved on visual images by virtue of the feature that all optical devices possess a modulation transfer function. As a consequence of this property, the difference in color between adjacent pixels is a usually small number, often between -1 and +1 graduations for a meaningful color scheme. By differentiating adjacent rows and columns, the original image can be expressed in terms of these small numbers. A simple compression algorithm for these small numbers achieves a four to one image compression. By piggy-backing this technique with a LZW compression or a fixed Huffman coding, an additional 35% image compression is obtained, resulting in a 6.5-to-one lossless image compression. Because traditional noise-removal operators tend to minimize the color graduations between adjacent pixels, an additional 20% reduction can be obtained by preprocessing the image with a noise-removal operator. Although noise removal operators are not lossless, their application may prove crucial in applications requiring high compression, such as the storage or transmission of a large number or images. The authors are working with the Air Force Photonics Technology Application Program Management office to apply this technique to transmission of optical images from satellites.

A feasibility study was conducted to investigate the advantages of data compression techniques on multispectral imagery data acquired from airborne scanners maintained and operated by NASA at the Stennis Space Center. The technique used was spectral vector quantization. The vector is defined in the multispectral imagery context as an array of pixels from the same location from each channel. The error obtained in substituting the reconstructed images for the original set is compared for different compression ratios. Also, the eigenvalues of the covariance matrix obtained from the reconstructed data set are compared with the eigenvalues of the original set. The effects of varying the size of the vector codebook on the quality of the compression and on subsequent classification are also presented. The rate of compression is programmable. However, the higher the compression ratio, the greater is the degradation between the original and the reconstructed images. The analysis for 6 channels of data acquired by the thermal infrared multispectral scanner (TIMS) resulted in compression ratios varying from 24:1 (RMS error of 8.8 pixels) to 7:1 (RMS error of 1.9 pixels). The analysis for 7 channels of data acquired by the calibrated airborne multispectral scanner (CAMS) resulted in compression ratios varying from 28:1 (RMS error of 15.2 pixels) to 8:1 (RMS error of 3.6 pixels). The technique of vector quantization can also be used to interpret the main features in the image, since those features are the ones that make up the codebook. Hence, vector quantization not only compresses the data, but also classifies it. The original and reconstructed images were not only analyzed for their RMS error but also for the similarity in their covariance matrices. Using the principal components analysis the eigenvalues of the covariance matrix of the original multispectral data-set were found to be highly correlated with those of the reconstructed data-set. The algorithms were implemented in software and interfaced with the help of dedicated image processing boards to an 80386 PC compatible computer. Modules were developed for the task of image compression and image analysis. These modules are very general in nature and are thus capable of analyzing any sets or types of images or voluminous data sets. Also, supporting software to perform image processing for visual display and interpretation of the compressed/classified images was developed.

The processing of compressed and encrypted imagery can exhibit advantages of computational efficiency as well as data security. Due to the data reduction inherent in compression, the computational speedup achieved via compressive processing can equal or exceed the compression ratio. Encryptive transformations which yield compressed ciphertext can similarly facilitate computational speedup, and are well known. However, due to analytical difficulties inherent in the derivation of operations which compute over the range space of commonly employed compressive transforms, reports of such processing paradigms are not evident in the open literature. In this introductory paper, we describe compressive and encryptive transformations in terms of functional mappings derived from abstract mathematics and image algebra (IA). An emerging technology, IA is a rigorous, concise notation which unifies linear and nonlinear mathematics in the image domain, and has been implemented on a variety of serial and parallel computers. Additionally, we derive a taxonomy of image transformations. Each taxonomic class is analyzed in terms of computational complexity and applicability to image and signal processing. We further present decompositions specific to each transformational class, which facilitate the design of operations over a given transform's range space. Examples and analysis are given for several image operations.

An adaptive discrete cosine transform (DCT) method for image compression is presented. Clustering technique based on vector quantization (VQ) is used to reconstruct binary patterns representing the location maps of selected DCT coefficients. The DCT of image blocks are classified into broad categories and after quantization, the transform coefficients are encoded using a variable word length (VWL) coding scheme. It is demonstrated that this approach is well-suited for real-time visual communications systems, and performs efficiently at very low bit rates.

In the realm of flexible automation, viewpoint independent object recognition will become increasingly important. An algorithm for the recognition of plane shapes under pseudo- perspective projection is presented. The method is based on the comparison of features that remain invariant under 2-D affine transformations. In particular, the work presented is part of the wider framework of semi-differential invariants. It was successfully tested on a database containing 42 objects.

Many airborne survey and reconnaissance systems require very precise location or position fix information in order to correlate mineral, agricultural, oil exploration, or highway construction survey data with fixed geodetic information in a geographic information database. This paper describes work in progress that: (1) compares the correlation performance of an existing scene matching system with that of an artificial neural network based system; (2) determines the performance on scaled, rotated images; and (3) compensates for temporal variations in image gray scale and noise levels. The goal of this research effort is to demonstrate with artificial neural networks performance improvements in robustness, flexibility of use, and speed compared to the current digital correlation system.

An accuracy requirement of +/- 0.011 degrees in the declination measurement of a remotely imaged munition cannot easily be satisfied using conventional imaging instrumentation. A mixed vendor computer architecture is implemented to accommodate the design and execution of a dedicated image analysis software system. Based on internal testing, the developed system is expected to meet design goals during a formal certification process.

Image processing operations involve replacing the color hue at a given pixel by computations involving the color values of surrounding pixels. Traditional image enhancement computations employ a 3 by 3 matrix manipulation involving a pixel and its eight neighboring pixels. Although some information loss results from any image processing technique, the use of a 3 by 3 operator typically results in a 50% loss of detail. This paper investigates unbalanced operators which do not seem to possess this deficiency. In particular, 2 by 2 operators can eliminate as much noise as 3 by 3 operators with less loss of detail. The enhanced image can be interpreted as a `balanced' operation with a 1/2 pixel shift of the original image for each operation performed. In four successive operations, a border 2 pixel band is lost on 2 edges of the image. In our applications, the loss of a few borer raster lines is insignificant when compared to a 50% loss of detail.

The median operator has been demonstrated to be a very effective method for restoring recognizable images from very noisy image data. The power of the median operator stems from its non-algebraic formulation, which prevents erroneous data corrupting the final color computation. A principal drawback is that the median operator replaces all data, erroneous or not, the result being a net loss of information. This paper presents alternative statistical outlier techniques by which erroneous data is readily recognized, but valid data usually remains unchanged. The result is an effective noise removal algorithm with reduced loss of information.

This paper addresses the automatic interpretation of digital image of three-dimensional scenes, especially automatic recognition of three-dimensional aircraft types from digital images. First, an efficient coordinate transform from a series of two-dimensional aircraft posture silhouette images to invariant matrices is developed. The invariant matrix is independent of its translation, scaling, and rotation. Next, on the basis of the invariant matrix, an effective algebraic feature extraction method is proposed. The method is based on singular value decomposition (SVD) of the matrix. To compress the dimensionality of the singular value vector, an optimal discriminant transform for a small number of samples is introduced to transform an original feature space of singular value vector into a new feature space in which its dimensionality is very low. Finally, our method is used to recognize three-dimensional aircraft types Experimental results show that our algebraic method as a high recognition rate, and it is insensitive to translation, scaling, rotation, and noise.