This paper describes an image compression method based on a hierarchical segmentation scheme into polygonal homogeneous regions. Adequate homogeneity criteria are selected through statistical discriminant analysis. Once segmentation is completed, the coding process consists in formingacompact representation of region shapes together with a suitable approximation of each region's content. Two types of coding are represented: a graphics-quality coding based on zero-order approximation of region content and a television-quality coding based on moment-preserving bi-level truncation.

TWo entropy coding schemes are investigated in this paper by estimating the entropies that specify the lower bounds of their coding rates. In the firSt scheme, we use a traditional combination of runlength and Huffman codes. Arithmetic codes are used in the second scheme. The results indicate that binary arithmetic codes outperform runlength codes by a factor of 34 % for low-rate coding of the zero-valued coefficients of the cosine transform of digital images. Hexadecimal truncated arithmetic codes provided a coding rate improvement as high as 28 % over truncated Huffman codes at low rates. The complexity of these arithmetic codes is suitable for practical implementation.

The Discrete Cosine Transform (DCT) has found wide applications in various fields, including image data compression, because it operates like the Karhunen-Loeve Transform for stationary random data. This paper presents a recursive algorithm for DCT whose structure allows the generation of the next higher-order DCT from two identical lower order DCT's. As a result, the method for implementing this recursive DCT requires fewer multipliers and adders than other DCT algorithms.

A surface distortion measure for a motion compensated image coding algorithm is described. This distortion measure provides a new approach to determining the rate of the residual quantizer in such an algorithm. The surface distortion measure (SDM) is a function of both the signal variance and the error variance. This functional relationship circumvents a major problem that is associated with using the ratio of these quantities as a quantizer design parameter. When the surface distortion measure is used as the fidelity criterion for the residual quantizer of the motion compensated image coding algorithm, superior data compression capabilities with little added distortion are achieved. A comparisqn of the results of using both SNR and the SDM as a fidelity criterion in a motion compensated image coding procedure is presented.

It is shown that the performance of image coding techniques can be enhanced via the utilization of a priori knowledge. Critical features of the image are first identified and then they are accounted for more favorably in the coding process. For satellite imagery, thin lines and point objects constitute critical features of interest whose preservation in the coding process is crucial. For a human visual system, the impact of the coding degradation at low rates is much more detrimental for these features than for the edges which constitute boundaries between regions of different contrasts. A highly non-linear, matched filter-based algorithm to detect such features has been developed. Pre-enhancement (highlighting) of the detected features within the image prior to coding is shown to noticeably reduce the severity of the coding degrada-tion. A yet more robust approach is the pre-enhancement of the slightly smoothed image. This operation gives rise to an image in which all critical thin lines and point ojects are crisp and well-defined at the cost of non-essential edges of the image being slightly rounded off. For the transform coding techniques, distortion parameter readjustment and variable-block size coding provide promising alternatives to the pre-enhancement ap-proaches. In the former, the sub-blocks containing any part of the detected critical features are kept within a low distortion bound via the local rate adjustment mechanism. The latter approach is similar to the former except that the image is partitioned into varying size sub-blocks based on the extracted feature map.

In some digital image processing systems, input images may be oversampled or only the low-pass content of the image will be retained due to some processing such as data compression. Therefore, input signals can be decimated to a lower sampling rate, processed at that low rate, and interpolated back to the same input rate or other desired rates. This approach allows performing the required processing on the size-reduced signals and may save significant computation time. In another application, decimation and interpolation are required to enlarge or reduce the size of the digital image to fit that of a display device. For high-ratio decimation and interpolation, the multistage approach is more computationally efficient than the one-stage approach. In this paper, the optimal multistage implementation of decimation and interpolation is derived for digital image processors of various arithmetic environments. The theoretically derived optimal multistage decomposition is then verified by comparing it against numerical results and is found to be valid for a wide range of decimation ratios.

Associative memories represent a major new artificial intelligence type of processor. We consider their use in pattern recognition, with particular attention to distortion-invariant and adaptive pattern recognition. New associative memory techniques and pattern recognition oriented architectures suitable for multi-class distortion-invariant pattern recognition (including systems that provide adaptive updating, forgetting, achieve reduced dynamic range and improved performance) are discussed and initial results presented. The first results of distortion-invariance, multi-class associative memories for pattern recognition are presented together with new architectures and algorithms for multi-stage associative processors, iterative processors for associative memory synthesis, and multi-class distortion-invariant associative processors. The issue of orthogonal projection vectors, associative memory capacity and new results and techniques to synthesize associative memories are included throughout.

A prototype digital image processor for enhancing photographic images has been built in the Research Laboratories at Kodak. This image processor implements a particular version of each of the following algorithms: photographic grain and noise removal, edge sharpening, multidimensional image-segmentation, image-tone reproduction adjustment, and image-color saturation adjustment. All processing, except for segmentation and analysis, is performed by massively parallel and pipelined special-purpose hardware. This hardware runs at 10 MHz and can be adjusted to handle any size digital image. The segmentation circuits run at 30 MHz. The segmentation data are used by three single-board computers for calculating the tonescale adjustment curves. The system, as a whole, has the capability of completely processing 10 million three-color pixels per second. The grain removal and edge enhancement algorithms represent the largest part of the pipelined hardware, operating at over 8 billion integer operations per second. The edge enhancement is performed by unsharp masking, and the grain removal is done using a collapsed Walsh-hadamard transform filtering technique (U.S. Patent No. 4549212). These two algo-rithms can be realized using four basic processing elements, some of which have been imple-mented as VLSI semicustom integrated circuits. These circuits implement the algorithms with a high degree of efficiency, modularity, and testability. The digital processor is controlled by a Digital Equipment Corporation (DEC) PDP 11 minicomputer and can be interfaced to electronic printing and/or electronic scanning de-vices. The processor has been used to process over a thousand diagnostic images.

A peripheral card has been designed to interface the Texas Instruments Professional Computer to a standard RS-170 video camera, providing real-time digitization and storage of a single video field as a 256 x 256, eight-bit pixel array. The digitizer card is bus-compatible with any Intel 8088 microprocessor based microcomputer e.g. TI-PC or IBM-PC. At present, the TI-PC based image processing system allows for a high resolution four-bit display of the digitized image as sixteen grey levels for the monochrome monitor, or sixteen color-mapped levels on the color monitor. The system is being extended to reconstruct the RS-170 standard monochrome monitor for eight-bit grey level resolution. Further system extensions include eight-bit digitization, storage, and display of true color video by front end interface of the system to a high resolution TI-CCD color camera. The use of lower level languages has allowed for the implementation of software features such as image display and image enhancement algorithms. FFT spatial frequency filtering and small kernel convolution are available for 128 x 128 and 256 x 256 image array sizes, respectively. In addition to providing a number of digital system hardware and software educational opportunities, the applicability of this system will be invaluable as a cost-effective, compact, yet high performance and quantitative image analyzing system.

A system for automated picture and parameter analysis is described. The system has been devised a) for routine evaluation of photographs taken from the output window of X-ray image intensifiers at the end of the production line, and b) as a versatile image analysis tool to support development of new image intensifier technology. The system set-up is discussed and some first results are given.

The quality of a photographically recorded image can be enhanced through digital image processing. It is first necessary, however, to convert the film densities into an accurate digital record, free of additional noise and artifacts. The rate at which photographic negatives can be digitized to provide high-quality input for digital image enhancement is severely limited by current technology, particularly for small formats (8 mm x 10 mm images). At high speeds, the shortcomings of light sources, optics, and detectors, and certain peculiarities of existing photographic negatives, reduce the signal-to-noise ratio of the digital information to an unacceptable extent. A new method for scanning photographic negatives which uses an area CCD imager to provide substantially improved performance is described. The method recognizes the architectural limitations of CCD technology and makes optimal use of the silicon surface area. It also optimizes use of the light provided by simple illumination systems. The optomechanical implementation of the method is very simple, but this simplicity is gained at the expense of electronic complexity. Substantial digital preprocessing is required to reconstruct and correct the recorded image. A breadboard system has been built to demonstrate the method and to determine the design parameters required to produce excellent images. It was found that the method imposes strict requirements on the CCD electronics, but that all systematic artifacts could be removed by suitable processing of the digital image data. High quality images were obtained at the rates required.

As part of the Landsat Thematic Mapper (TM) engineering evaluation effort, there have been several studies to measure a modulation transfer function (MTF) for the TM. In contrast to work that measured or modeled only the optical and electronic components of the MTF, our goal was to measure the complete imaging system MTF that includes factors such as optics and electronics. Measurements of the TM MTF using targets-of-opportunity have been previously reported, however this paper describes new results using a specially designed target. To arrive at an estimate of the system MTF, the TM point-spread function (PSF) was measured using a two-dimensional array of black squares constructed at the White Sands Missile Range in New Mexico. The target provides relatively good spatial and radiometric control of the scene and thereby reduces the uncertainties introduced by targets-of-opportunity. The clever design of the target allows 1/4 pixel shifts of "point sources" throughout the instantaneous-field-of-view (TFOV) of the TM. The shifting allows exploitation of sample-scene phasing to effectively resample the PSF at the finer rate. A method for sorting the image data and "assembling" a PSF and a technique for correction of rotation of the target relative to the sensor scan direction are discussed. Also, a technique for partial elimination of variations in the scene background is presented. The resulting estimates of the TM imaging system PSF and MTF are presented and compared with earlier results.

This paper describes a workstation and a software development environment designed to facilitate the investigation of integrated signal/symbol processing in image understanding and graphics processing. The workstation is comprised of two complementary, high-performance processors connected via a high-speed interface: the Pixar 2D Image computer and the Symbolics 36xx Lisp machine. Some of the current and potential applications of the system include developing and evaluating new image analysis and feature extraction algorithms, investigating the role of the extracted features in perceptual aggregation, object detection/identification and scene analysis, and assisting in the creation of prototypical image understanding systems.

An image processing system based on a MC 68000 microprocessor has been developed for analysis of fundus photographs. This personal computer based system has specific image enhancement capabilities comparable to existing large scale systems. Basic enhancement of fundus images consists of histogram modification or kernel convolution techniques to determine regions of specific interest such as textural difference in the nerve fiber layer or cupping of the optic nerve head. Fast Fourier transforms and filtering techniques are then utilized for specific regions of the original image. Textural difference in the nerve fiber layers are further highlighted using either interactive histogram modification or pseudocolor mappings. Menu driven software allows review of the steps applied, creating a feedback mechanism for optimum display of the fundus image. A wider noise margin than that of digitized fundus photographs can be obtained by direct fundus imaging. The present fundus image processing system clearly provides us with quantitative and detailed techniques of assessing textural changes in the fundus photographs of glaucoma patients and suspects for better classification and early detection of glaucoma. The versatility and computing capability of the system make it also suitable for other applications such as multidimensional image processing and image analysis.

An automatic inspection machine, designed and manufactured by the Perkin-Elmer Corporation for the U.S. Bureau of Engraving and Printing, is capable of real-time inspection of currency at rates compatible with the output of modern high-speed printing presses. Inspection is accomplished by comparing test notes (in 32-per-sheet format) with reference notes stored in the memory of a digital computer. This paper describes the development of algorithms for detecting defective notes, one of the key problems solved during the development of the inspection system. Results achieved on an analytical model, used for predicting probability of false alarms and probability of detecting typically defective notes, are compared to those obtained by system simulation.

The Wigner-Ville Distribution (WVD) has been shown to be a valuable tool for the analysis of non-stationary signals such as speech and Electrocardiogram (ECG) data. The one-dimensional real data are first transformed into a complex analytic signal using the Hilbert Transform and then a 2-dimensional image is formed using the Wigner-Ville Transform. For speech signals, a contour plot is determined and used as a basic feature. for a pattern recognition algorithm. This method is compared with the classical Short Time Fourier Transform (STFT) and is shown, to be able to recognize isolated words better in a noisy environment. The same method together with the concept of instantaneous frequency of the signal is applied to the analysis of ECG signals. This technique allows one to classify diseased heart-beat signals. Examples are shown.

A method for calculating the three-dimensional structure of suspended chains of carbon black revealed by transmission electron microscopy is presented. The method is a general one involving the interpretation of inclined projections of thin transparent objects to determine the spatial characteristics of microscopic features. The reconstruction technique consists of five sections: preprocessing, segmentation, classification, correlation and spatial reconstruction. Preprocessing consists of sharpening the features in the electron micrographs which are inherently too diffuse for the application of standard image analysis techniques. Image segmentation involves dividing of the image into its basic components (feature records) which are later combined to reconstruct the object in three-dimensions. The feature records are classified using a clustering k-mean algorithm and the correlation between the images is accomplished using an algorithm which learns class correlation based on the known axis of rotation and the uniqueness of the correlation. Once the features have been correlated it is possible to locate the three-dimensional features in space based upon their relative disparity. The spatial information can also be used to identify the features which may not otherwise be a unique solution and the artificial intelligence methods used for this aspect of the solution will also be discussed.

In recent years, significant progress has been made in image segmentation and classification, but still no global theory of image segmentation exists. The wide variety of segmentation techniques used are basically ad hoc and are very dependent on the way the desired features are presented. The main local features used in various segmentation algorithms are image brightness, color and texture. One of the most important features for image segmentation by the human observer is texture, yet it has been difficult to measure and characterize. Actually, texture segmentation is at a very early stage of development at this time. This paper describes an experimental investigation into digital image segmentation using texture features. Texture features are extracted from a few widely different examples of image data. Several different feature sets are used, and the resulting files are input into an unsupervised clustering algorithm. Several variations on the clustering algorithm are explored: some partition the image into segments by using similarities only in the space of features, and others include spatial information such as the location of individual pixels. The various experimental results are also compared, and a new direction for investigation is described.

The object of this contribution is to explain the necessity of using Fuzzy Mathematics to recognize the pattern and employing the method of automatic recognition of the mechanical parts based on teh Membership's principle of Fuzzy sets. This article looks briefly at the recognition system and the recognition classifier designed by the micro computer Z-80. The authers endeavour to enumerate the recognition classification of screw and screw cap and draw up a recognition program. The results of recognition can be displayed on the fluorescent screen or may appear in print.

In computerized Chinese character printing, it is infeasible to use the fully-formed character approach since there are about 8,000 Chinese characters in common use. Therefore, dot-matrix printing with a large dictionary of binary images of Chinese characters stored in memory is commonly used. To generate these Chinese character patterns in the dot-matrix form by manual operation is tedious. A better approach is to apply image processing techniques to automatically convert the image of a character into its corresponding dot-matrix pattern. We developed a system that can automatically generate a Chinese character multifont. This system includes image processing and CAD subsystems. Each input picture, consisting of about 100 Chinese characters, is scanned by a scanner. The digitized line-scanned image is processed by the image processing subsystem to form the Chinese characters by a dot matrix. The modules of the image processing subsystem include noise reducer, text detector, adaptive threshold, slicer, and size corrector. Due to the effect of quantization error, there are some defects in these digitized Chinese characters. The CAD subsystem is used to trim these characters. The modules of the CAD subsystem include radical extractor, radical classifier , radical generator, radical copier, stroke extractor, and stroke trimmer. This system can automatically generate Chinese characters in a wide range of resolutions ( 24x24 to 240x240 ) and in any specified font, such as Sung style, Ming style, Formal style, Running style, and Script style of Chinese characters. Using the proposed system, we have generated about 160,000 Chinese characters, which consist of five styles in four dif-ferent resolutions. The advantages of this system are time saving, cost saving and high quality.

A problem of much interest is the estimation of the delay or shift between two signals or images. In this paper, we investigate the use of recently proposed [Yegnanarayana et.al., IEEE Trans. ASSP, Vol 32, 610, 1984] Group Delay Functions (GDFs) for estimating the delay between two signals. The two types of GDFs (one based on the Fourier phase and the other based on the Fourier magnitude) contain different types of information relevant to the delay estimation problem. The phase-based GDF contains the delay information whereas the magnitude based GDF contains the Signal-to-Noise Ratio (SNR) information. We propose an adaptive delay estimation technique that utilizes both types of information. This adaptive technique weights the delay estimates from the high-SNR regions more than the delay estimates from the low-SNR regions. Since the GDFs take on a continuum of values, the proposed technique does not distinguish between integer delays and fractional delays. Thus the GDF-based delay estimation procedure is inherently capable of subpixel delay estimation. Simulation results are also included to illustrate the capabilities of this procedure.

The hit-or-miss transform forms the basis of Serra's mathematical morphology in image analysis. It has many applications, including texture enhancement and object isolation. A resolution pyramid is a sequence of images of decreasing size and resolution derived from a single, relatively high-resolution image. The combination of hit-or-miss transforms with resolution pyramids is a useful technique for representing and ordering structural information in digital images. In this paper we define a mask-driven, mask-generating hit-or-miss transform that can isolate image features and order them hierarchically with respect to size and relative brightness. We show how this mask-driven transform, either alone or in conjunction with a resolution pyramid, can generate a sequence of nonlinear approximations to the original image. We show applications to blob detection, data compression, contour mapping, and image complexity measurement.

Ordered dither is a popular method for printing/displaying continuous tone images on bi-level devices such as laser printers, ink-jet printers and liquid crystal flat panels; the success of this algorithm comes from its simplicity and its non recursive nature. A major contention about this technique is its poor performance for rendering sharp edges. To alleviate this problem, we propose a new half-toning method called the dynamic ordered dither algorithm; it is an adaptive scheme based on the standard ordered dither operation preceded by an adaptive high pass filter. The dynamic ordered dither algorithm yields a far better edge rendition while keeping the good properties of the standard dithering process, namely the preservation of the object location, the good color reproduction and the possibility to use parallel archi-tectures for processing pictures in real time.

Quadtree decomposition is one of the methods of representing binary image data at multiple resolutions. Such representations have found application in many areas of image processing. This paper describes a novel implementation of quadtree generation using histograms. The histogram based implementation is fast and easy to implement on any histogram generating hardware. The number of histograms to be computed depends upon the size of the original image, the size of the smallest block into which the image is to be resolved, and on the number of gray levels handled by the histogram generating hardware. The paper also describes a fast algorithm to reconstruct binary images from quadtree decomposition using video rate lookup table processing.

We treat the linear estimation problem with two simultaneous, competing objectives: minimum mean-squared error and minimum error-signal correlation. The latter objective minimizes the signal component in the error and maximizes the correlation of the estimator with the signal. The problem is solved, both for the scalar and stationary random process cases, as an optimal trade-off which produces a slightly higher mean-squared error and a much larger reduction in error-signal correlation over that of the minimum mean-squared error single objective solution. The optimal trade-off solution, which we call the mini-mum-error, minimum correlation (MEMC) filter is then applied to the problem of recovering space-invariant, blurred images with additive noise. As the theory predicts, the images restored through the MEMC filters are sharper and clearer than their minimum mean-squared error (Wiener) filter counterparts, but slightly noisier in appearance. Most viewers prefer the MEMC restorations to the Wiener ones, despite the noisier appearance.

A method for detecting motion based on the statistical properties of a video image will be discussed. This technique computes the mean and fourth moment about the mean, called kurtosis, for an image and compares the result with a previously stored value. The method involves observing a change in the shape of a video images' distribution and requires the storage of only a 16-bit digital word. A security intrusion system is described which incorporates a two level approach to motion detection: (1) motion detection based on the real-time pixel-by-pixel comparison between four selectable video lines and previously stored reference data; and (2) a higher order, near real-time verification of motion, using all of the images' pixels and based on the statistical properties of the image. The system is implemented in hardware using a TMS320 signal processor and charge coupled device (CCD) cameras.

The need arises in certain vehicular surveillance applications for an electronic imager that can perform without human interaction and can transmit definitive images over very low bandwidth channels. The investigators have found that line scan imagers offer particular advantages toward these ends. Specifically, the use of a line scan imager facilitates the following processing steps: 1) Segmentation of the vehicle from the background, 2) Auto-matic exposure control, 3) Light level equalization prior to quantization, and 4) Implementation of an adaptive sampling scheme. These processing steps together with the source encoder may be implemented on a relatively low throughput processor and achieve near real time operation. The specific encoding method used here is an extended differential pulse code modulation (DPCM). A prototype system has been developed, producing medium resolution images at less than 10K bits per frame.

A rule-based expert system is being developed to segment and label the major elements in digitized images of simple road scenes. The system acts on a set of uniform regions, generated by the split-and-merge algorithm, which may be regarded dS a non-standard primal sketch. Simple descriptors attached to regions and to boundaries between regions are referenced by the rule base, which incorporates general Knowledge of the typical benavior of the split-and-merge algorithm, as well as specific knowledge of the scene domain. The rules merge oversegmented regions to form an essentially correct description of the scene. The system is intended to investigate tne limits of applicability of methods which do riot introduce an explicit tnree-dimensional (3D) representation

Machine recognition and generation of Chinese characters has been a challenging research subject due to the structural complexity of the characters. There are about 40,000 Chinese characters in total and about 5,000 in daily use. An exhaustive approach to recognizing or generating all Chinese characters is almost infeasible in practice. Consequently, most techniques try to segment the characters into suhpatterns, called roots, which form a basis used to compose Chinese characters. The number of roots to be dealt with is much smaller than the number of Chinese characters. Therefore, the character-gap, is a useful and natural feature to segment the characters.

A scene matching approach which utilizes histograms of edge directions is presented. This approach is expected to have utility in matching imagery taken by different sensors and matching images taken at different times of the day or night. Examples of using this technique to match white hot to black hot FLIR imagery and match synthetic to FLIR imagery are shown.

The extraction of a binary image from a gray level image is a common image processing operation particularly for document image analysis and optical character recognition. Various methods for this task are described in the literature including global and adaptive binarization. This paper evaluates three adaptive binarization techniques viz., a contrast measure approach, a weighted running average approach and a second derivative approach, and compares them to global binarization methods. Experiments with noisy document (postal letter mail) images lead to the following conclusions. Image contrast binarization often yields nearly the same results as the edge operator, with considerably less computation and is less sensitive to parameter settings. In addition, the edge operator is more sensitive to image resolution than the contrast operator. The weighted running-average approach is highly sensitive to the parameters involved in the calculation of the average but produces a quick binarization.

Much of the work done in digital enhancement and image restoration has been limited in application to black and white images. The extension to color image processing is not trivial. Most workers agree that processing should take place in a three-dimensional space whose dimensions reflect the visual attributes of luminance, hue, and saturation. Different schemes have been developed over the years for expressing these attributes in terms of the well known red, green, and blue primaries (RGB) components. In this paper we first discuss our choice of color space for computer image enhancement. Our strategy for processing is to display each attribute as a black and white digital image; each attribute is then processed independently to achieve the desired enhancement. Contrast and sharpness enhancement techniques are discussed. The computer processing algorithms are restricted to those which preserve the natural appearance of the scene.

A computerized stereotactic measurement system for evaluating rat brain metabolism was developed to utilize the large amount of data generated by quantitative autoradiography. Conventional methods of measurement only analyze a small percent of this data, because these methods are limited by instrument design and the subjectiveness of the investigator. However, a computerized system allows digital images to be analyzed by placing data at its appropriate three-dimensional stereotactic coordinates. The System automatically registers experimental data to a standard 3-dimensional image using alignment, scaling, and matching operations. Metabolic activity in different neuronal structures is then measured by generating digital masks and superimposing them on to experimental data. Several experimental data sets were evaluated and it was noticed that the structures measured by the computerized system had, in general, lower metabolic activity than manual measurements had indicated. This was expected because the computerized system measured the structure over its volume while the manual readings were taken from the most active metabolic area of a particular structure.

We treat the problem of estimating 3-dimensional motion from 2-dimensional projections. This problem is of primary interest in the field of image sequence analysis. The paper discusses a mathematical exact method, a conventionally used approximative method and a modified approximative method studied by the authors. It is verified that the later method performs well for relatively large values of the motion parameters while being computationally comparable to the conventionally used approximative method.

Traditionally, image segmentation algorithms work by either making point amplitude measurements or by scanning a small computational window over the scene to discover local texture statistics. Where these measurements significantly change, the boundary of an object is said to exist. In this new image segmentation algorithm the entire scene is first transformed with a global transformation such as with a Fourier or Hadamard transform. The consequence of this transformation is that coherently linked structures within the scene, such as texture fields, condense into one, or a few, distinctive peaks. These peaks may then be selectively extracted (or rejected) by a variety of supplementary algorithms. The result is a modified coherent spectrum of the original scene. Through inverse transformation of this modified spectrum back to the image domain, the coherently linked structures are extracted. With this technique structures of related texture may be selectively, and globally, extracted even if they are not contiguous in the original image -and even in the presence of very substantial noise.

Multiaperture optics deals with image formation employing a large number of optical elements, the insect eye being an example. If the apposition design is used, each optical element creates one pixel only, although due to FOV overlap, there may be some additional information collected. The consequence is that multiaperture optics devices produce a relatively small number of pixels. This presents a challenge to any pattern recognition procedure by requiring it to make assumptions on the information content concerning the spaces between the pixels. Another challenge for pattern recognition, common for all systems, is speed. Therefore, time should not be wasted in examining dead zones in the field. The algorithm presented here meets the first challenge by transforming the presented image into a combination of rectangles of varying aspect ratios. The second challenge is met by selectively examining only those data locations known to be responsive and eliminating any blank space above and below the pattern. This process consists of three stages: 1) an essentially random distribution of data points is converted into rectangular form; 2) these rectangles are then converted into a series of "code elements" which are actually equal to the value of their aspect ratios; 3) this pattern of code elements is then compared to numbers representing some known patterns to achieve identification. The algorithm relies heavily on the concept of a-priori knowl-edge as well as recognizing a fairly small universe of patterns.

Pattern recognition and image processing techniques have been developed and applied to automated visual measurement and inspection. These techniques are used to detect objects and to determine their locations, sizes and shapes. The system uses an Apple IIe microcomputer to process data transmitted from a video processing unit and sends commands to robot manipulator to perform the required tasks.

Digital image restoration requires some knowledge of the degradation phenomena in order to attempt an inversion of that degradation. Typically, degradations which are included in the restoration process are those resulting from the optics and electronics of the imaging device. Occasionally, blurring caused by an intervening atmosphere, uniform motion or defocused optics is also included. Recently it has been shown that sampling, the conversion of the continuous output of an imaging system to a discrete array, further degrades or blurs the image. Thus, incorporating sampling effects into the restoration should improve the quality of the restored image. The system transfer function (the Fourier transform of the point spread function), was derived for the Landsat Multi-Spectral Scanner (MSS) and Thematic Mapper (TM) systems. Sampling effects were included, along with the relevant optical, instantaneous-field-of-view and electronic filter data, in the system analysis. Using the system transfer function, a least-squares (Wiener) filter was derived. A Wiener filter requires the ratio of the power spectra of the scene and noise, which is often, for simplicity, assumed to be a constant over frequency. Our restoration includes models for the power spectra which are based on the study of several different types of Landsat scenes. The Wiener filter is then inverse Fourier transformed to find a restoration filter which is spatially windowed to suppress ringing. Visual evaluations are made of the restored imagery.

The laser scanning technology was developed rapidly in recent years to fulfill the great needs of high speed, high quality image and documentation transmission and retrieval especially for facsimile and computer input/output applications. The main components of a laser scanner include a laser source, modulator, and a deflector. Recently,semiconductor lasers which can be directly modulated become popular. The modulator used externally thus can be eliminated. Among several different deflection devices, the rotating polygon mirror is used most popularly because it has high speed, large deflection angle, and low cost. Ideally, each facet of a polygon mirror should be parallel to the rotating axis to produce uniform scanning lines. However due to the fabrication limitation, it is very difficult to have perfect parallelism of each facet. The tilt of the facet will displace the scanning lines to yield a periodical pattern to which the human eyes are very sensitive. Many researchers reported several methods such as tightening the tolerance of the polygon mirror, using cylindrical lens, and employing optical-acoustic device to eliminate the wobble error. This paper will report a new method which employs a piezo-electric device to achieve the above mentioned purpose. We use an optical sensor to detect the displacement of the dislocated scanning line and feed the information to a piezo-electric device with a mirror to compensate the wobble error. This paper will describe the method briefly and present the theoretical calculation and experimental results.

Within the scope of the Fresnel diffraction theory the threedimensional transfer function OTF of the microscope was obtained. Inverse filtering with OTF-1 can be used for the reconstruction of 3-D light-microscopic images. Comparison with Stokseth's approximation for the optical transfer function used in other works on 3-D reconstruction shows a rather large relative error for this approximation. Transforming OTF-1 back into the spatial domain a closed form for the inverse point spread function IPSF (with respect to convolution of the point spread function PSF) was obtained. Direct deconvolution of 3-D images in the spatial domain proofs to be advantageous compared to deconvolution in the frequency domain. Closer examination of the reconstruction procedure shows that the spatial resolution defined by the Rayleigh criterion can be improved by a factor 1.8 perpendicular and by 1.5 parallel to the optical axis.

We introduce a simplified technique for calculating the maximum entropy solution for restoration of degraded images. We show that the maximum entropy solution can be formulated as an inverse filter estimate plus a correction term. The estimate starts with the inverse filter solution and an iterative algorithm produces the correction term. At each iteration we use an entropy gradient and an analytically calculated step size. The algorithm uses two Fourier transforms per iteration. Examples show that, after only a few iterations, the technique reduces the artifacts often found in the inverse filter estimate.