The underlying mathematical models employed in reflec-tion and transmission computed tomography using diffracting wavefields (called diffraction tomography) are reviewed and shown to have a rigourous basis in inverse scattering theory. In transmission diffraction tomography the underlying wave model is shown to be the Rytov approximation to the complex phase of the wavefield transmitted by the object being probed while in reflection diffraction tomography the under-lying wave model is shown to be the Born approximation to the backscattered wavefield from the object. In both cases the goal of the reconstruction process is the determination of the object's complex index of refraction as a function of position F and, possibly, the frequency w of the probing wavefield. By use of these approximations the reconstruction problem for both transmission and reflection diffraction tomography can be cast into the simple and elegant form of linearized in-verse scattering theory. Linearized inverse scattering theory is shown to lead directly to generalized projection-slice theo-rems for both reflection and transmission diffraction tomography that provide a simple mathematical relationship between the object's complex index of refraction (the unknown) and the data (the complex phase of the transmitted wave or the complex amplitude of the reflected wave). The conven-tional projection-slice theorem of X-ray CT is shown to result from the generalized projection-slice theorem for transmission diffraction tomography in the limit of vanishing wavelength (in the absence of wave effects). Fourier based and back-projection type reconstruction algorithms are shown to be directly derivable from the generalized projection-slice theorems.

This paper addresses some of the fundamental issues involved in the development of mathematical models for the generation of ultrasound pulse-echo data in the human body. These models provide a rational basis for the development of algorithms used for medical diagnosis. Models for transducers, wave propagation and scattering, and image formation are considered. Wave propagation models are based on a linear, weak scattering assumption. Examples of state-space models for parameter estimation and tissue classification are given.

Ultrasonic imaging diagnosis has become more sophisticated as instrument performance has advanced and as the educational base of its practitioners has extended. By and large, the method is used qualitatively to answer specific medical questions, usually those involving anatomic features delineated by fluid spaces or involving local changes in tissue macrostructure. Some physiologic information is inferred from the anatomic appearance, from target dynamics, and from ancillary forms of data manipulation and display, such as the Doppler frequency shift.

The physical and statistical properties of backscattered signals in medical ultrasonic imaging are reviewed in terms of: 1) the radiofrequency signal; 2) the envelope (video or magnitude) signal; and 3) the density of samples in simple and in compounded images. I INTRODUCTION There is a wealth of physical information in backscattered signals in medical ultrasound. This information is contained in the radiofrequency spectrum--which is not typically displayed to the viewer--as well as in the higher statistical moments of the envelope or video signal--which are not readily accessed by the human viewer of typical B-scans. This information may be extracted from the detected backscattered signals by straightforward signal processing techniques at low resolution. II THE RADIOFREQUENCY SIGNAL In the weak scattering regime of medical ultrasonic imaging (1), (2) the Born approximation is valid and yields the frequency dependence of back-scattered radiofrequency signals from a diffuse cloud of similar scatterers. The result in the acoustic application (3) is similar to that found in electro-magnetic wave scattering and the frequency dependence of the backscattered intensity is given by I cc f⁴ F ² (1) where the frequency to the fourth power is the dependence associated with Rayleigh scattering, and I F I 2 is the square of the scattering form factor: the Fourier transform of the distribution of the scattering potential, here the change in compressibility or acoustic impedance of the medium. F I 2 has the form of a low pass filter. Some examples of Eq.(1) are given in Figure la for Gaussian form factors with diameters (four standard deviations) as shown, and in reference (4) this simplified treatment is shown to agree well with an exact calculation after the method of Faran when the product of the wave number k and particle radius a is less than unity. In Figure lb the same curves are presented after spatial attenuation by a factor exp(-,2.cx(2x) f), a reasonable approximation to the round trip attenuation at a depth x and at fre-quency f. The attenuation has appreciable effect only at the higher frequencies, as in many other applications.

Some of the hasic ideas and approaches of medical ultrasound tissue characterization are overviewed in critical vein.

The objective of this paper is to utilize the vector space methods to illustrate the fundamental structure of the reconstruction algorithms for linear acoustical imaging systems. In this paper, continuous, discrete, and partial discrete models and the asso-ciated optimal algorithms are presented. In addition, the commonalities and differences of these algorithms and the sensitivity and performance parameters are discussed.

This paper gives an overview of pattern recoanition techniques (P.R.) used in biomedical image processing and problems related to the different P.R. solutions. Also the use of knowledge based systems to overcome P.R. difficulties, is described. This is illustrated by a common example ofabiomedical image processing application.

Image processing and pattern recognition techniques are used in many aspects of seismic data processing and interpretation. These tools can effectively complement and simplify the task of the geophysicist and geologist in detecting and mapping potential hydrocarbon reservoirs. In the past, seismic interpretation was concerned mainly with mapping subsurface structures. With the advent of digital field recording and improved signal processing techniques, it became clear that the seismic section contains more than just structural information. The data contain details about attenuation and other attributes of the seismic signal that can be used to indirectly determine rock properties. The geophysicist now has the capability not only to map structure but to determine subsurface lithology. Image processing and pattern recognition techniques are being used to classify and sort the myriad of data available into an easily understood depiction of subsurface structure and lithology. This paper consists of an overview of the seismic technique and recent applications of pattern recognition to seismic exploration. A case history is presented showing how pattern recognition techniques were used to locate the termination of a gas-bearing sandstone between a producing well and a dry well.

Standard performance measures and statistical tests must be altered for research on animal sonar. The narrowband range-Doppler ambiguity function must be redefined to analyze wideband signals. A new range, cross-range ambiguity function is needed to represent angle estimation and spatial resolution properties of animal sonar systems. Echoes are transformed into time-frequency (spectrogram-like) representations by the peripheral auditory system. Detection, estimation, and pattern recognition capabilities of animals should thus be analyzed in terms of operations on spectrograms. The methods developed for bioacoustic research yield new insights into the design of man-made imaging and pattern recognition systems. The range, cross-range ambiguity function can be used to improve imaging performance. Important features for echo pattern recognition are illustrated by time-frequency plots showing (i) principal components for spectrograms and (ii) templates for optimum discrimination between data classes.

A technique to display digitally sampled RF A-lines using false color is described. The backscattered 5 MHz ultrasonic signal obtained from the constant laminar flow of a blood-mimicking substance is sampled at 50 MHz with an 8 bit A/D converter. The RF samples of each backscattered A-line are displayed vertically and consecutive A-lines are displayed side-by-side horizontally to form a 2D false color image. The 2D image of a typical blood flow phantom data set is used to graphically demonstrate the manner in which conventional Doppler processing and newer correlation processing concepts are related to patterns in the 2D M-mode RF display.

The study reported here investigates the use of a time-varying filter to compensate for the spreading of ultrasonic pulses due to the frequency dependence of attenuation by tissues. The effect of this pulse spreading is to degrade progressively the axial resolution with increasing depth. The form of compensation required to correct for this effect is impossible to realize exactly. A novel time-varying filter utilizing a bank of bandpass filters is proposed as a realizable approximation of the required compensation. The performance of this filter is evaluated by means of a computer simulation. The limits of its application are discussed. Apart from improving the axial resolution, and hence the accuracy of axial measurements, the compensating filter could be used in implementing tissue characterization algorithms based on attenuation data.

A new digital beam-forming technique for the ultrasound imaging is proposed. This technique is based on the holographic beam-forming and the signal-decomposition using uncorrelated pulses, and it permits high speed data acquisition with simple hardware. Some algorithm for the decomposition and beam-forming is described. The ability of the method is discussed through the analysis of the point spread function simulated by the computer, and an experimental system using an airborne sensor array is shown.

When acoustic waves propagate through a nonlinear reactive medium, they can generate a variety of secondary waves. This generation depends on the magnitude of the nonlinear parameter β of the medium. Large variations of β in biological tissue (from 3.5 to 6.5) are reported in the literature. Tomographic imaging of the nonlinear parameter β in human organs is currently being proposed as a potentially powerful tool in medical diagnosis. The magnitude of the generated secondary waves can be described by a line integral of the distribution of along the propaga-tion path. We analyze two types of tomographic image reconstruction of the nonlinear parameter. In one case, two primary waves of different frequencies are used to parametrically excite a difference-frequency secondary wave, which is then detected. In the other case, a single-frequency primary wave produces a second-harmonic wave, which is then detected. In this paper, we present the underlying mathematics and physical assumptions which describe the generation of the secondary waves in the above two cases. The approximations and simplifications used in image reconstruction are justified. The experimental results we have thus far obtained are stated, and possible modifications in the reconstruction algorithm and imaging system are considered.

A new class of algorithms is given for the precise computation of general multidimensional recursive discrete transformations from data that is unevenly-spaced in time and/or space.

In this paper we introduce a new application for ultrasonic imaging that has not received much publicity: ultrasonic imag-ing for underwater robots. After a brief survey of some of the applications we consider the conceptual design of an underwater ultrasonic imaging system. The goal of the system is to provide 3-dimensional images about objects that are distances of 10's of meters away at real time video rates. In order to achieve this goal, we suggest that a principle of spatially variant insonification (SVI) be employed. This technique propagates multiple sonar beams into the medium simultaneously and upon reception, the individual beams are separated. In order to accomplish this, each sonar beam uses a specific code. The mathematical properties of the specific code that are em-ployed insure that this is possible. The design of these codes is identical to that encountered in multi-access, spread spec-trum, communication systems. Therefore, the time-bandwidth properties of the system limit the simultaneous requirement of decorrelatable beams and good range resolution. As a specific implementation of the technique a frequency hopping code is presented. This code is then combined with a sparse receiving array. The resultant system satisfies the above criteria.

Frequency-domain techniques for characterizing ocular and abdominal tissue are being investigated in our laboratories. Results of tissue-characterization processing, often displayed in graphical or numerical format, are also being displayed in image format. This allows for viewing the spatial distribution of the particular parameter being studied. This capability may prove to be important clinically, since it provides a means for visualizing and comparing different pathologies, monitoring changes due to treatment, and imaging the extent of disease processes. The displayed parameters are derived from clinical data bases of independently verified pathologies, as well as an analytic model that relates spectral parameters to physical properties.

A brief review is provided of methods available for speckle reduction. Relatively new techniques for reducing speckle by "intelligent" (i.e. speckle discriminating) image filtering are introduced. These utilize methods of echographic texture analysis, and the multivariate instrument signature of speckle, to recognize and largely eliminate speckle from the display without substantially destroying the structural information in the image. Further developments from a simple filter described previously²⁰, in the form of a non-linear transform applied to one feature and the use of a two-feature discriminator, are shown (qualitatively) to improve filter performance.

When image texture analysis methods are used for ultrasonic tissue characterization, the discrimination results obtained by statistical pattern discrimination methods must be interpreted carefully in order to avoid pseudo-discriminations due to differences in exa-mination procedures and system settings. This study examines the dependence of popular texture analysis methods on transducer-specific diffraction characteristics, B-mode image reconstruction and sampling factors, i.e. size and position of the selected Region-of-Interest. It is shown that image analysis should always be based on diffraction-corrected ultra-sound signals. In large-organ applications, e.g. liver, polar reconstruction yielded more stable results than cartesian reconstruction, especially when texture measures from the greylevel runlength matrices or power spectrum are used. Analyzing clinical and synthesized ultrasound images, we found that the first-order greylevel statistics: Mean greylevel, skewness and excess as well as the second-order sta-tistics: Correlation of greylevel cooccurrences proved to be stable with respect to tissue-independent factors as well as sufficiently sensitive to tissue differences.

In hyperthermia treatment, ability to predict complete tissue temperature fields from a limited sampled temperatures or non invasively is greatly desirable to assess treatment success. Non uniform heating of the lesion causes less pronounced cell kill in inadequately heated regions. Tissues were suspended in a temperature controlled water bath, the temperature of the bath is increased at different high rates to ensure no tissue variations. Two thermistors were inserted at the upper and lower edges of the region of interest (ROI) with-in the tissue sample. A real time sector ultrasound image along with the radio frequency signal are accessed in real time at different temperatures of the (ROI). First and second order grey level statistics were calculated for fresh liver kidney and brain (cow) tissues in the temperature range of 32-47°C. The average temperature in the (ROI) is correlated with the calculated parameters. Entropy is a sensitive parameter to assess the uniformity of heating. Possibility of in vivo study, and to predict long term hyperthermic effects via the calculation of backscattering coefficient is further investigated. A multi-layered model is also presented to aid in spatially resolving temperature dependent acoustic parameters.

The methods of statistical pattern recognition are well suited to the problems of in vivo ultrasonic tissue characterization. This paper describes supervised pattern recognition methods for selecting features for tissue classification, calculating decision boundaries within the selected feature space, and evaluating the performance. We address the considerations of dimensionality and feature size which are important in classification problems where the underlying probability distributions are not completely known. Examples are given for the detection of diffuse liver disease in the clinical environment.

This paper describes a procedure for classifying tissue types from unlabeled acoustic measurements (data type unknown) using unsupervised cluster analysis. These techniques are being applied to unsupervised ultrasonic image segmentation and tissue characteriza-tion. The performance of a new clustering technique is measured and compared with supervised methods, such as a linear Bayes classifier. In these comparisons two objectives are sought: a) How well does the clustering method group the data? b) Do the clusters correspond to known tissue classes? The first question is investigated by a measure of cluster similarity and dispersion. The second question involves a comparison with a supervised technique using labeled data.

The information about the condition of parenchymal tissues is obscured by the performance characteristics of echographic equipment. The authors investigated by realistic 3-D simulations the so-called beam diffraction effects on two echographic imaging modalities: amplitude modulated (AM) and phase derivative (PD) echograms. Furthermore the modification of the image texture by attenuation was quantified. In order to assess the potentials of statistical analysis of texture for medical diagnostics the effects caused by varying the density of scattering particles in a homogeneous medium were studied. It is concluded, that unless beam diffraction effects are either prevented, or adequately corrected for, quantitative texture analysis is not meaningful. In addition, the data have to be corrected for the non-linear and time dependent amplifier characteristics. Data-acquisition and preprocessing equipment performing these tasks has been developed at the authors' laboratory.

In this work the effects of averaging (compounding) the intensities from partially correlated phased array antennas are studied as a means of speckle reduction. Compounding intensities from subarrays is often referred to as mixed-integration in the radar literature. Correlated subarrays are obtained from a large coherent phased array by division into subarrays after coherent signal detection. Overlapping subarrays share common elements. Model calculations of the covariance matrix are made for both synthetic aperture and fixed phased array systems. These results are then used to calculate the speckle contrast reductions for multiple compounded partially correlated arrays for these different configurations. The results are applicable to a variety of coherent systems such as microwave radars, laser radars, sonars, and ultrasound.

A new approach towards the recognition and removal of the speckle artefact in ultrasound pulse-echo images is presented. The technique utilises the behaviour of the signal's instantaneous frequency, which is the time derivative of the temporal phase. In contrast to all other approaches, the proposed method allows the speckle artefact to be recognised on a deterministic and local basis. Although presented in the context of medical ultrasound imaging, the technique is of much wider generality.

Speckle is believed to be a significant degrading factor in ultrasonic images, and likely impairs the detectability of low contrast tumors and clinical scans. One method of reducing speckle, termed frequency diversity processing, uses multiple, narrow-band filters applied to the received RF signal. For this study, several images of targets from a contrast detail phantom were acquired and stored on floppy disks, by digitizing the RF from a mechanically translated transducer. Digital filtering techniques were used to process the images by breaking the RF spectrum into several overlapping bands. Each processed image was computed as both the incoherent and coherent average of the individual narrow-band images. The filters were designed with individually variable weights to broaden and whiten the original transducer spectrum. The resulting processed images showed a degradation of spatial resolution, but a significant increase in grey scale information, and a substantial improvement in the signal-to-noise ratio. The results indicate little difference between coherent and incoherent processing of the images.

The degree of speckle reduction achievable by spatial and frequency compounding is a function of the rates of speckle pattern change induced by varying the imaging system's illumination angle and acoustic frequency, respectively. We have measured these rates under a variety of conditions and derived a method of maximizing speckle reduction using the average of partially correlated speckle patterns. Our experimental results agree well with theoretical predictions of these phenomena and indicate that, under limited conditions, improved target detection is possible using spatial compounding. Frequency compounding appears to be counterproductive in improving target detectability.

A model of echogenic structures based on previous conventional inhomogeneous continuum models is presented for generating cardiac muscle textures in two-dimensional echocardiograms. It is designed to study the dynamic aspects of echographic texture in relation to cardiac dynamics. The theoretical basis and assumptions made in the simulation of myocardial textures are discussed. Finally, preliminary results showing average gray level variations with different orientations and states of contraction of myocardial fibers are presented.

An autoregressive model is generated for a two dimensional pulsed ultrasound data. The autoregressive parameters are used to differentiate between various liver textures,potential for pathology characterization is discussed. The same model could be used to seperate regions of normal and diseased tissues,subtracting original image out of the developed texture. The theory of 2-D autoregression is developed as well as discussing the overall applicability and drawbacks of the technique.

Ultrasonic speckle in B-scan images is usually modeled with simplifying assumptions about the propagation phenomenon. Moreover the resolution cell is supposed to contain a large number of scatterers. This assumption leads to a Rayleigh distribution for the image amplitude and more generally to speckle statistics which do not carry information about the scatterers distribution (Rayleigh limit). To take into account any propagation and diffraction effects we introduce a complex 3-dimensionnal point spread function which is not constrained to be space invariant. To deal with situations in which few scatterers are present in the resolution cell (low scatterer density or near-focus measurements) we establish the expressions of the first and second order statistical moments of the image intensity without the assumption of large scatterer density, i.e. below the Rayleigh limit. As a result, it appears an explicit dependence of the image contrast and of the speckle pattern autocovariance on the volumic scatterer density: the speckle carries information about the scattering medium. Of course, classical results are obtained by letting n tend towards infinity. The so-called SNR is shown to be inferior to its Rayleigh limit, depending on the product of scatterer density by the "effective volume": a defintion introduced to give an unambiguous (quantitative) measure of the volume of the resolution cell. The speckle spot size in both axial and transversal directions is also simply expressed as a function of the scatterer density.

Motion parameter estimation has been one of the most important objectives for target tracking. And holographic systems have significant advantage over conventional motion detection devices because of the capability of detecting the range information. Motion estimation is usually performed in the image domain after the image reconstruction process. Typically, motion estimation operates in the discrete mode given a finite number of selected point features with correspondence. These constraints and requirements reduce the feasibility of motion estimation for real-time target tracking. Recently, a statistical motion estimation method has been developed for holographic acoustical systems. This methods utilizes the mean and covariance matrix to identify the translation vector and rotation matrix operation associated with the target motion. This method is capable of performing motion estimation in both space and spatial-frequency domains and it does not require point feature selection and correspondence identification, and therefore, it signi-ficantly enhances the potential for real-time tracking. In this paper, we first introduce the development and formulation of this statistical motion estimation method. To separate the translation component of the motion from the rotation operator, we perform the motion estimation in the spatial-frequency domain using the power spectra of the received wave-field sequences for the identification of the rotation operator and the phase variation for the estimation of the translation vector. This method can be also directly applied to navigation and position identification for radar systems, and displacement estimation in tomographic reconstruction.

In this paper we describe a pulse-echo Beam Tracking technique which can be used for precision estimations of the speed of sound in biological tissues. The theoretical trade offs among the estimation parameters are described, followed by experimental results from phantoms and from tissues in vitro. Finally, a computer simulation is described which demonstrates further improvement in the precision of the technique.

A tone burst of ultrasonic waves in air is scattered by small pads of cotton of different composition. Spectrum analysis of the transmitted ultrasonic tone burst is found to produce a spectrum which can be related to the physical properties of the scatterer. Use of well-characterized cotton samples has made possible a correlation with data from other studies made with the same samples. The role of pattern recognition in a future grading system is described.

The use of ultrasound pulse-echo signals to produce quantitative parameter mappings (or images) of an inhomogeneous medium is complicated by the presence of echo interference and transducer field effects. In this paper it is briefly discussed how the use of appropriate measurement techniques can counter the latter problem, leaving the speckle due to interference as the major source of artefact. It is shown in more detail how fluctuations in the instantaneous frequency of the received signals may be used to pin-point the location of the destructive interference in the pulse-echo signals. When assessing bulk attenuation values by a backscattering technique, this allows the data to be selectively edited to exclude interference-corrupted segments, considerably reducing the need for gross averaging over large amounts of data. The approach also points towards a technique for reducing the (artefactual) speckle-like appearance of attenuation images obtained via a backscattering technique.

Heart sounds are generally believed to be caused by acceleration and deceleration of blood in the cardiovascular system. Some theories proposed attribute the sounds directly to the opening and closing of the cardiac valves themselves. Regardless, some components of the heart sounds are ascribed to certain events in the cardiac cycle, such as the aortic and pulmonary components of the second sound. We are studying the possibilities of analyzing multichannel phonocardiogram (PCG) signals using passive sonar signal processing techniques, with the aim of locating the sources of the heart sound components in three dimensions. So far, we have been able to obtain only two-channel recordings of the PCG with an ECG channel with available equipment. Digital signal processing techniques have been developed for comparison of PCG segments by correlation methods. These techniques permit detection of similar PCG segments in the different channels which may be related to the same cardiac event. The time differences between such segments obtained from a number of PCG channels, along with the known locations of the PCG transducers, should aid source localization and ranging.

Diffraction effects are important in acoustic imaging and tissue characterization because of the relatively large wavelengths used and the fact that applications are frequently used in the near-field of the source. It is difficult to intuitively anticipate the shape of the field there, yet the description of the field's spatial acoustic potential or pressure distribution is necessary. This problem is more complicated when focused transdu-cers or phased arrays are used. Using the spatial frequency, domain it is possible to model propagation in lossless and lossy media as a transfer function. The sources are represented as planar sources with separable arbitrary time excitation and arbitrary spatial excitation. Transfer functions can be obtained for lossless media, media with a linear frequency dependence of attenuation coefficient, and media with a quadratic dependence of attenuation co-efficient. The transfer functions are shown to be simply related to the two-dimensional spatial transform of the Green's function of the wave equation for propagation in the medium of interest with the assumed boundary conditions. The transfer functions of the lossy and lossless propagation models are shown to be interdependent. For any given observation plane, these transfer functions are time-varying spatial filters that attenuate higher spatial frequencies with increasing effectiveness as time proceeds. The effects of source excitation and apodization, source boundary conditions, assumed media properties, and receiver aperture effects are easily incorporated in this model. Several numerical simulations of computed acoustic potentials and pressure distributions are shown.

In this paper, a description of the scanning acoustic microscope (SAM) is first given to be followed by a brief and selective review of the recent applications of the SAM in nondestructive study of solid-state materials and devices. Some of the recent progress on the imaging and the methodologies for quantitative characterization of thin-and thick-solid state electronic materials and devices that are being made at the authors' institution is stressed.

Acoustic Microscopy is an important branch of non-destructive evaluation which provides high resolution for imaging the detailed structure of a small object. When an acoustic microscope operates in the transmission mode, the micrograph is simply a shadowgraph of all the structures encountered by the acoustic wave passing through the object. Because of diffraction and overlapping, the resultant images are difficult to comprehend in the case of specimens of substantial thickness and structural complexity. We used the principles of diffraction tomography and acoustical holography along with digital calcuations of wavefield propagation to overcome this problem. We have described in previously-published work how a scanning laser acoustic microscope (SLAM) can be modified to obtain data for subsurface tomographic imaging. In this paper, we review the principles of scanning tomographic acoustic microscopy (STAM). The required modification of SLAM to obtain STAM and the reconstruction process are described. We show how we are able to accurately acquire the complex-amplitude information necessary for image reconstruction. We demonstrate the power of this technique by comparing digitally-computed images thus obtained with analogue images of a conventional SLAM. The results show that high-quality, high-resolution subsurface images can be obtained from experimentally acquired data. We also describe techniques to obtain projection data from different angles of wave incidence enabling us to tomographically reconstruct different planes of a complex specimen in microscopic detail. With these modifications in place, STAM should shortly become a powerful tool in non-destructive evaluation.

Films and coatings are integral elements of devices used for a variety of purposes in engineering, electronic, optical and other applications. The performance, functional characteristics, and structural properties of the devices all depend on the adhesion between the film and substrate. Adhesion measurement techniques that are currently available are not comparable from one laboratory to another, nor are there any techniques that are noncontacting, quantitative and reliable. In our work, surface acoustic waves (SAWs) are used to introduce stresses between the films and their substrates. Regions of the film having weak bonding forces will produce a different response than the regions having strong bonding forces. The response is measured as a change in the velocity of the waves. The waves are excited on the substrate with a high numerical aperture cylindrical lens. The V(z) response of the film system is then obtained and processed to extract the SAW velocity. The results of these measurements on a film system with known regions of strong and weak adhesion will be presented. These measurements will be compared to simulations of the change in surface wave velocity as the bond stiffness be the film and substrate is changed.

Pattern recognition methods are described for classifying acoustic emission (AE) signals according to their source types. Simple time and frequency domain features of the AE waveforms are used in the classification to distinguish one type from another. Methods for classification using labeled waveforms, and clustering using unlabeled waveforms have been developed and applied to the detection of a fatigue crack growing from a fastener hole in a simulated aircraft structure. Sources of AE in this monitoring application are crack growth, crack face rubbing, fastener fretting, mechanical impacts, electrical transients, and hydraulic noise. Classification of labeled data to separate crack-related AE from the other types produced a 96-100% accuracy, and clustering of unlabeled data pro-duced an 82-94% accuracy. A system calibration method needs to be developed before the pattern recognition algorithms can reliably accommodate specimen geometry changes.

A new method for generating three dimensional images of flaws in elastic solids from backscattered data at a finite number of look angles (usually 2 to 3) is presented. The procedure is based on the physical elastodynamics aporoximation which leads to a relationship between the backscattered ramp response to an incident ultrasonic ramp pulse and the area of cross section of the object along the "line of sight."

A low cost high range resolution ultrasonic imaging system has been developed to nondestructively evaluate composite material as well as biological tissue and agricultural products. The system which is capable of dual displays, B and C scans, utilizes high frequency sampling and phase processing of the r.f. signal in order to achieve range resolutions of .5 mm and range accuracies of .1 mm. A pseudorandom sequence signal processing scheme is under development to enhance the signal-to-noise ratio for low level signals.

In this paper, a pattern recognition approach to the ultrasonic nondestructive evaluation of materials is examined. Emphasis is placed on cientifying effective features from time and frequency domains, correlation functions and impulse responses to classify aluminum plate specimens into three major defect geometry categories: flat, angular cut and circular hole defects. A multi-stage classification procedure is developed which can further determine the angles and sizes for defect characterization and classification. The research clearly demonstrates that the pattern recognition approach can significantly improve the nondestructive material evaluation capability of the ultrasonic methods without resorting to the solution of highly complex mathematical inverse problems.

After about twenty-five years of growth, image analysis technology has advanced to the point where the design of machines that model real scenes in real time is at the technological frontier. Consequently, we see the beginning of a merging of the technologies of image analysis and automatic control. This merger is leading to the development of complex distributed systems in which decisions and actions are based on realistic models of possibly time-varying scenes. We refer to this merged technology as vision automation. In the past decade, reflecting the impact of artificial intelligence, image analysis has expanded to scene understanding, and automatic control has evolved into intelligent control. Thus, vision automation consists of two major components: scene understanding and intelligent control. (We prefer the term "scene understanding" to the more commonly used "image understanding", since we want to understand the scene, not the image.)

In this paper we focus our attention on modeling nons-tationary texture behavior, thus resulting in accurate models for texture analysis and synthesis. We present a new class of nonstationary signals, the class of semi-Markov random fields. The likelihood function, which uniquely describes the statistical behavior of these random fields, is derived. We examine the validity of our two-dimensional semi-Markov random field models in synthesizing and analyzing textures. The appropriateness of the semi-Markov random field models for synthesizing images similar to real textures is studied and different models are optimally fitted to real textures by the use of a maximum-likelihood procedure.

Applications of fuzzy set theory to various disciplines like pattern recognition, image processing etc. have gained paramount importance in recent years. In many fuzzy pattern recognition systems identification of a set of independent fuzzy vectors in fuzzy matrices are required to be performed in parallel or pipe line fashion for enhancement of speed of the system. For example, in fuzzy syntactic analysis of patterns it is required to identify a set of fuzzy independent basis vectors for inference of pattern grammars from a set of positive samples of pattern strings. Due to the iterative nature of these operations, these jobs become computationally demanding ones. A semi systolic array of processors have been designed for parallel identification of such a set of independent vectors.

In order to reduce the classification errors when labeling texture pixels, the contextual information can be used to reduce the ambiguity attached to pixel classification. A post-processing approach is introduced whereby the context is modeled according to a Markov-Gibbs assumption to describe the local characteritics of a pixel label. Results of some pathological cases are presented.

In this paper we describe techniques for classifying a bandlimited signal based on properties computed from the signal's level crossings. We show that a Gaussian random process with a Gaussian shaped spectrum can be completely characterized using level crossings. For a non-Gaussian process we first construct a Markov model for the hard limited signal. We then show how this leads to the design of an efficient and useful classifier. The technique was applied to in vivo ultrasound liver data.

In this paper we shall consider the intrinsic complexity of patterns and the extent to which this notion is well defined. Past work in the computer science and artificial intelligence literature on the decomposition of pictures into their basic building blocks is consistent with our motivation, but this paper will look at the simplest description over all possible methods. Thus, the program complexity of Kolmogorov given the domain is mini-mum in finite length such that a Turing Machine can compute functions in finite time. So, the string approach is well used as Kolmogorov process. The Circular Layer Chart for Recognition (CLCR) is created with total number of elements (29,-1)*4, where L is the number of layers counted from innermost cell in even 2.1 number. A sequence of string arrays, called Circular Layer String Arrays (CLSA), are achieved for each layer. In this paper there are twenty layers used for the recognition of human face pictures. The distance between both arrays of corresponding layers of test pattern and sample is calculated. The inner five layers are applied for the normalization of orientation. That is, if any of these inner layers does not match, the arrays of test pattern are rotated one position and tried again until they are rotated one circle. If any array of inner layers does not match through the normalization of orientation, the sample is rejected. Thus it can save much time for matching.

Images are divided into equal area cells. The contour encountered in these cells are given WORD names, which describe the type of contour. Each cell has 9 pixels in a 3x3 matrix format and herutthere are a possible 512 different WORDs. However, for a class of images where the outline is sufficient for recognition purposes, a much smaller number actually exists. These images include lower case English handwriting and blood cell images. It is shown in this work that there are only 34 BASIC WORDs and all the others encountered in these images are rotations of these BASIC WORDs. Using this fact,features are defined for this class of images, standardized in size to 16 cells described above. Three types of features are defined namely, Independent, Dependent and Related to aid in application of syntactic methods of recognition from sentence structures defined using the above features. It is also shown that matrix algebra can be used for developing data reduction and recognition algorithms.

Two dimensional Median filtering has been used in picture processing for improved noise performance. Salt and pepper noise and spikes have been successfully removed using median filters. In this work the application of median filtering for finding hidden contours in an image is presented. The principle used is to obtain the correlation of pixels in a line of the image after passage through a median filter of suitable window width and determine the trend changes exhibited by this. Since those samples which show trend changeover are not passed by the median filter the output correlation must indicate changes whenever the part of the image exhibits trend changeover. This is most likely to occur when there are hidden contours in the image. The study also includes comparison of the DFTs of the image lines after passage through the median filter (standard scanning is assumed and the image intensity is available in a digitized form) with the original. This experiment also shows interesting results. It is shown that a theoretical basis can be established for this performance.