This paper presents a new multi-stage adaptive search (MSAS) vector quantization algorithm for image compression. It permits improved solutions which approximate the exhaustive- search multistage solution. The standard multi-stage vector quantization algorithm had the advantage of simple structure and low complexity. However, the performance degrades rather rapidly when the number of stages increases. Our algorithm has the same advantage as the standard algorithm, but the performance is much higher and also higher than the tree/full search vector quantization. The adaptive search algorithm can also be applied to the tree-search vector quantization (VQ). Tree adaptive search algorithm is a very powerful algorithm. The larger the tree-structured codebook, the better the performance of tree adaptive search VQ algorithm. However, when the codebook size increases, the codebook generation complexity and the required codebook memory increases exponentially too. A multi-path algorithm is also presented in this work. It can improve the performance of multi-stage adaptive/non-adaptive search VQ. It only increases the complexity of encoder. For the decoder, it is identical to the multistage algorithm.

The cost of processing imagery that exhibits a large data burden can be ameliorated by compressive processing, which simulates an image-domain operation using an analogous operation over a given compressed image format. The output of the analogous operation, when decompressed, equals or approximates the output of the corresponding image operation. In previous research, we have shown that compressive processing can lead to sequential computational efficiencies that approach the compression ratio. This effect is due to the presence of fewer data in the compressed image, as well as to the occasional occurrence of an analogous operation whose cost per pixel is less than that of the corresponding image operation. We generally claim computational efficiencies that approach the compression ratio. A further advantage of compressive processing can occur in parallel computing paradigms, where a consequent reduction in the processor count may approach the compression ratio. This has important implications when the compressive operation requires less computing time than the corresponding image operation. That is, a reduction in the fundamental complexity may occur, which facilitates computation in nearly-constant time, given sufficient parallelism. Additionally, the degree of parallelism is reduced with respect to that required for image-domain computation by a factor that approaches the compression ratio. In this paper, we discuss fundamental theory that unifies compressive processing at a high level, as well as present and evaluate general formulations of the BTC and VPIC compression transforms, Analyses emphasize effects of information loss and computational error inherent in VPIC and BTC, as well as computational efficiency. Our algorithms are expressed in terms of image algebra, a rigorous, concise notation that unifies linear and nonlinear mathematics in the image domain. Since image algebra has been implemented on numerous sequential and parallel computers, our algorithms are feasible and widely portable.

The processing of compressed imagery can exhibit advantages of (1) reduced space requirement for image storage, due to fewer data; (2) computational speedup resulting from fewer operations on reduced data; and (3) increased data security due to an obscure encoding format. We call this technique compressive processing, which we have shown can simulate an image-domain operation using an analogous operation over a given compressed image format. The output of the analogous operation, when decompressed, equals or approximates the output of the corresponding image operation. In Part 1 of this three-part series, we show that compressive processing can lead to sequential computational efficiencies that approach the compression ratio. Additionally, we present unifying theory that portrays the derivation of compressive operations at a high level for image operations such as pointwise, global reduce (e.g., image summation or maximum), and image-template (e.g., linear convolution) operations. Further discussion and analysis concerned formulations of block truncation coding (BTC) and visual pattern image coding (VPIC) compressive transforms. In this paper, we analyze high-level formulations of the vector quantization (VQ) and JPEG compression transforms. Additionally, we illustrate the utility of our high-level derivational methods by demonstrating the derivation and operation of several pixel-level operations over VPIC- and BTC-compressed imagery. Such operations are extended to include VQ- and JPEG-compressed imagery. In Part 3, we consider the pixel- level operations of edge detection and smoothing, as well as higher-level operations such as target classification and connected component labeling. Analyses emphasize computational efficiency, as well as effects of information loss and computational error. Our algorithms are expressed in terms of image algebra, a rigorous, concise notation that unifies linear and nonlinear mathematics in the image domain. Since image algebra has been implemented on numerous sequential and parallel computers, our algorithms are feasible and widely portable.

The method of block truncation coding (BTC) was originally proposed by Delp and Mitchell and later extended to color images by others. The idea is to retain important visual features while discarding details which are not to be noticeable to human observers. In our paper, we further explore the block truncation coding method by minimizing the within group variance measure proposed by Otsu and the information distance suggested by Kullback to divide every 4 by 4 subimage into two classes, and an intuitive vector quantizer to further compress the coded output. As a result of the combined application of BTC and vector quantization methods, we get better bit rates (bits per pixel) for the test image used in experiments without significant perceivable errors in its appearance. First, we divide a color image into 4 by 4 small nonoverlapping blocks. The Otsu or Kullback thresholding technique is then used as an optimal method to minimize the mean square error of classifying each pixel in a block into two classes and encode it in one-bit adaptive vector quantizer. After classification, for each 4 by 4 block, there is a bitmap corresponding to one-bit adaptive vector quantizer and a six-dimensional mean vector corresponding to each of the two classes. In the second part, the vector quantizer proposed by Linde, Buzo and Gary (known as LBG) is used to compress the bit-map and mean vectors separately. This is a six- dimensional signal compression for the mean vectors and a binary compression for the bitmap. Vector quantization of these BTC output results in a reduction of the bit rate of the coder. By using BTC and vector quantization methods, we have obtained 1.0 bit/pixel compression result for a color image of size 512 by 480 given with 8 bits/pixel and R, G, B specifications. The mean square error was also measured as low as 0.07 without much deformation in the reconstructed image.

A recent trend in computer graphics and image processing has been to use iterated function system (IFS) to generate and describe images. Barnsley et al. presented the conception of fractal image compression and Jacquin was the first to propose a fully automatic gray scale still image coding algorithm. This paper introduces a generalization of basic IFS, leading to a conception of partial iterated function system (PIFS). A PIFS operator is contractive under certain conditions and when it is applied to generate an image, only part of it is actually iteratedly applied. PIFS provides us a flexible way to combine fractal coding with other image coding techniques and many specific algorithms can be derived from it. On the basis of PIFS, we implement a partial fractal block coding (PFBC) algorithm and compare it with basic IFS based fractal block coding algorithm. Experimental results show that coding efficiency is improved and computation time is reduced while image fidelity does not degrade very much.

In Parts 1 and 2 of this three-part series, we show that the processing of compressed imagery can yield computational efficiency, due to the presence of fewer data. In Part 1, we show that such compressive operations can simulate an image- domain operation such that the output of the compressive operation, when decompressed, equals or approximates the output of the corresponding image operation. Additionally, we present unifying theory that portrays the derivation of compressive operations at a high level for image operations such as pointwise, global reduce (e.g., image summation or maximum), and image-template (e.g., linear convolution) operations. Further discussion and analysis concerned formulations of block truncation coding (BTC) and visual pattern image coding (VPIC) compressive transforms. In Part 2, we analyze high-level formulations of the vector quantization (VQ) and JPEG compression transforms. We further illustrate the utility of our high-level derivational methods by demonstrating the derivation and operation of several pixel-level operations over VPIC- and BTC-compressed imagery. In this paper, we extend our previous derivations to include image processing operations such as edge detection and smoothing, as well as higher- level operations such as target classification and connected component labeling. Analyses emphasize computational efficiency, as well as effects of information loss and computational error. Our algorithms are expressed in terms of image algebra, a rigorous, concise notation that unifies linear and nonlinear mathematics in the image domain. Since image algebra has been implemented on numerous sequential and parallel computers, our algorithms are feasible and widely portable.

In this paper we analyze the performance of the nonlinear joint transform correlator in terms of output signal to noise ratio. The main assumption used is that the signal energy is small relative to that of the additive noise; this assumption is defensible in practice due to the generally small spatial extent of target images relative to scenes. This work is an extension of that in an earlier paper. The previous analysis was carried out under a restriction that the signal and noise spectra were similar (actually multiples of one another). In the current work there is no such constraint. Further, the current work assumes sampled data; this is mostly for the sake of variety, and serves as a complement to the spatially continuous formulation of the earlier paper.

An optimal distortion-invariant filter for detecting a distorted target in input noise is designed. The input noise consists of two sources of noise, overlapping additive noise and non-overlapping background noise. We obtain the filter function by statistically maximizing the peak-to-energy ratio criterion, which is defined as the ratio of the expected value squared of the output signal at the target location to the expected value of the average output energy. This results in a filter output with a well defined output peak at the target location and low output noise floor. The filter is designed to take into account the effects of both overlapping additive noise and non-overlapping background noise, the finite size of the input data, and the target distortion. Computer simulation results are provided to show the performance of the filter.

We introduce an optical classification method that allows the simultaneous comparison of an input object with multiple stored objects. The stored objects are compressed into one- dimensional vectors by means of an optical integrating ring detector that focuses the incident energy incident upon each ring to a point. The neural weights corresponding to the vectors are orthogonalized by means of a delta rule algorithm, and the classification site is carried out with the help of a joint transform correlator.

Because massive amounts of various information from embedded fiber sensors in composite are required to intelligently monitor the integrity and health of the structure, smart structures face the challenge of signal processing in future applications. In this paper, we give our studies on the processing of signals from polarimetric fiber sensors to assess strain distribution by an improved optical Hopfield neural network.

A method of extracting improved features for object identification by correlating with a wavelet filter is described. The wavelet filter used is a linear combination of Gabor wavelets, which is designed by a neural network algorithm to extract features that are useful for discriminating different classes of objects. The neural network algorithm achieves this by iteratively adapting the filter parameters and linear combination weights of the wavelet filter so that the features extracted maximize the Fisher ratio between the classes. The algorithm thus provides an automated technique of designing a filter which extracts improved features for identification. Results are presented which show the ability of these improved features to increase the classification performance of a pattern recognition system.

Traditional wavelet edge detection and encoding schemes preserve shape features of objects effectively at a variety of spatial scales and also allow an efficient means of image and video compression. However these schemes also remove texture from imagery and thus image reproduction quality suffers. Fractal encoding techniques on the other hand generate high compression ratios but tend to introduce blocky artifacts in imagery. Thus we describe an encoding method that combines the shape preserving capability of wavelets with the compression qualities of fractal compression for a hybrid multiresolution technique that achieves high compression, selective and accurate feature preservation, and is computationally efficient.

Fluctuations are always present in underwater sound propagation, and are generally viewed as a complication in signal detection and identification. However, in some cases where the signals fluctuate less than the noise, it is possible to take advantage of the different magnitudes of fluctuations of signal and noise to improve detection. Wagstaff's integration silencing processor (WISPR) is an example of such a processor. The original version of the WISPR processor utilized power values derived from complex pressures in a given frequency bin, but ignored the phases of these complex pressures. An improved processor that takes advantage of the phase as well as the amplitude is described below. Its performance is verified using measured data, where detection has been accomplished by a margin of 4 decibels. Simulations using synthetic data show that the new processor can be effective for signal-to-noise ratios greater than minus 20 decibels.

The accuracy of digital and optical Hough transform (HT) processors is analyzed. New correction techniques to achieve improved accuracy are addressed. A new output format optical HT system using computer generated holograms (CGHs) is described and analyzed and its accuracy is found to be superior to that of digital HT processors. Its speed is much faster; the CGH space bandwidth product (SBWP) requirements are much less than for other methods; CGH error sources are addressed; and simple multiple binary exposure CGH fabrication is found to be sufficient.

The watershed is a complex but very powerful morphological operation that is often used in image segmentation. It involves the iterative use of the skeleton by influence zones (SKIZ). In this paper a parallel algorithm for the determination of the SKIZ is presented. The proposed algorithm can be efficiently implemented on an optoelectronic array processor. The optoelectronic processor is constructed via multilevel, high-speed, fiber-optic programmable logic arrays.

A digital image smoothing procedure is described that meets two requirements inferred from a recent model of biological vision. First, the smoothed image is a linear combination of basis functions formed by convolving a Gaussian function with each pixel. Second, the linear coefficients are evaluated by requiring that the integral of the smoothed image over each pixel equal the product of the gray value and area of that pixel. These requirement are in accordance with a model of visual hyperacuity that explains the ability of biological vision systems to resolve some image details that are much smaller than system photoreceptors. The procedure is demonstrated and compared with standard Gaussian convolution smoothing for both a simple one- dimensional example and a practical corner-of-an-eye test image.

We introduce a three-dimensional target attitude estimation method based on the linear optimal filtering and the signal processing prediction algorithms. By estimating the three- dimensional attitude and the scale of the target, we are able to predict the temporal trajectory of the target and to perform a robust tracking. Most cases of parameter estimation with the help of correlation deal with the two- dimensional estimation of position, scale and object rotation. Treating a three-dimensional parameter estimation problem leads to an increase in the number of correlation filters. We propose a strategy to reduce the number of correlations and thus reduce the computational burden and the processing time. We demonstrate the efficiency of the strategy in the case of optimal trade-off filters.

The ImSyn^TM Processor is an optoelectronic device developed by Essex Corporation to accelerate the SAR imaging process. This paper focuses on the application of the ImSyn Processor to SAR/ISAR imaging of moving targets from a moving platform. The targets may have non-uniform translational and rotational motion during the data collection process. The ImSyn processor operates directly on the polar frequency data permitting interaction with the SAR phase history without data interpolation. The process produces high quality sub-aperture images for determining the phase errors, which enables the implementation of motion compensation and autofocus algorithms. Simulations and examples of SAR/ISAR images from the prototype system for both stationary and moving targets are presented.

Recently, the theory of the synthetic aperture microscope (SAM) was presented. A SAM is a three dimensional imaging system that makes use of the principles of synthetic aperture radar to obtain a high resolution, complex valued image at a large working distance. Theoretically, a SAM can achieve resolution of approximately (lambda) /4 in all three dimensions. A typical system consists of a holographic sensor head and a reconstruction processor. This implementation will use the Essex ImSyn^TM optoelectronic discrete Fourier transform (DFT) processor to reconstruct the synthetic aperture image. Over the past year Essex has constructed a breadboard of the system and obtained initial results consisting of a single digital hologram and its computer-reconstructed image. The ability to collect complex valued image data opens the door to image processing and pattern recognition algorithms that are not applicable to intensity images, such as holographic interferometry for mapping strain fields. Applications include industrial inspection, robotics, and biological imaging.

We present a single operation mode binary image algebra (SOMBIA) based on a basic operation that is a dilation- complement operation followed by an intersection operation. We also propose the parallel architecture of a compact optical morphological image processor (COMIP) with only one operation mode to efficiently realize the parallel algorithm of SOMBIA. The optical implementation of the basic operation of SOMBIA is finally discussed.

A novel scheme for implementing gray scale image hit-or-miss transform (HMT) in one optical step with only one optical correlator is presented. This scheme uses area- complementary-encoding to represent each pixel of gray scale images. Using an incoherent optical correlator to optically perform correlation and a microcomputer to perform nonlinear thresholding electrically, the area-complementary-encoding gray scale image HMT is realized for auto target recognition.

The methods and software of the fast-running processes' investigation in various media (solids, liquids, gas, plasma) and technical objects were developed. The mathematical models of the objects under research were created in the conditions of exposure to the fields possessing different physical nature. The hardware and programming complex of the film-phototelerecording and data processing combining the computer equipment, television- and film technique, was created, permitting the use of the advantages of every type of the technical systems for the investigation of certain processes, objects and their characteristics. The hardware and programming complex of the film-phototelerecording and data processing is used both in the development and designing periods, and during tests, as well as under the long-term operation of technical objects in the ordinary and extremal situations. The fast-running processes' investigations in various media and technical objects were performed, and by their results the data bank was made.

A full scale model of an aircraft engine inlet was tested at NASA Lewis Research Center's Icing Research Tunnel. Simulated natural ice sheds from the engine inlet lip were studied using high speed digital image acquisition and image analysis. Strategic camera placement integrated at the model design phase allowed the study of ice accretion on the inlet lip and the resulting shed ice particles at the aerodynamic interface plane at the rear of the inlet prior to engine ingestion. The resulting digital images were analyzed using commercial and proprietary software to determine the size of the ice particles that could potentially be ingested by the engine during a natural shedding event. A methodology was developed to calibrate the imaging system and insure consistent and accurate measurements of the ice particles for a wide range of icing conditions.

The problem of shape recognition is studied through the use of relational models based on the hypergraph representation and the context similarity measure. Formal definitions are introduced and graph properties are calculated important to the matching process. A conflict is shown to exist between the interclass distance and the semantical distance between the vertices within a model. The representation is extended with the notion of vertex neighborhood, which increases the semantical distance and makes the processing of complex scenes feasible.

Image reconstruction from the measurements of the image Fourier transform modula is an important and difficult problem. Among all the approaches developed to solve this problem, the iterative algorithms remain the most efficient. However, these algorithms suffer from a major drawback that limits their practical application. In this paper we introduce a wavelet adaptation of one of the iterative algorithms that can significantly improve the performance of these algorithms while dramatically reducing their computational complexity.

Based on some important edge detection algorithms, this paper presents a compound edge detection method. In this method, the nonlinear Laplacian operator and the edge strength detector are combined to improve the edge location accuracy, the infinite impulse response (IIR) exponential smoothing filter and the length denoising operator are used to reduce the noise effect, and a fast algorithm is developed to make this method more efficient. To evaluate the performance of this method, we apply this new edge detection method on different kinds of images. The experimental results show that this method has high edge location accuracy and can decrease the influence of noise effectively.

The objective of data encoding is to transform a data array into a statistically uncorrelated set. This step is typically considered a 'decorrelation' step because in the case of unitary transformations, the resulting transform coefficients are relatively uncorrelated. Most unitary transforms have the tendency to compact the signal energy into relatively few coefficients. The compaction of energy thus achieved permits a prioritzation of the spectral coefficients with the most energetic ones receiving a greater allocation of encoding bits. There are various transforms such as Karhunen-Loeve, discrete cosine transforms etc., but the choice depends on the particular application. In this paper, we apply an approximate Fourier expansion (AFE) to sampled one-dimensional signals and images, and investigate some mathematical properties of the expansion. Additionally, we extend the expansion to an approximate cosine expansion (ACE) and show that for purposes of data compression with minimum error reconstruction of images, the performance of ACE is better than AFE. For comparison purposes, the results also are compared with discrete cosine transform (DCT).

Staring infrared detectors often produce low resolution images because the technology does not exist to produce higher resolution arrays with sufficient spatial sampling intervals. A proven approach to combat this difficulty involves recording multiple frames that have been optically shifted onto a high-resolution grid pattern and then combined together into a single high resolution image. This process is known as microscanning. In fact, if the infrared (IR) imaging system is mounted on a moving platform, the normal vibrations associated with the platform's movement can be exploited to generate shifts in the acquired images. We present an algorithm that can register this temporal image sequence at the sub-pixel level and then reconstruct a high resolution image from the shifted frames. The proposed algorithm uses a gradient based shift estimator which provides shift information for each of the recorded frames. The reconstruction algorithm is based on a technique of high resolution image reconstruction by solving a series of linear equations in the frequency domain. In this paper, we review the theory behind the registration and reconstruction algorithms and their limitations. We demonstrate that the registration is a viable real-time algorithm that is suitable for applications involving small image shifts (i.e. less than one detector element). We also show that the reconstruction program gives dramatic improvements in the image's resolution and does well in handling the aliased information.