The paper will describe a new and unique systems approach to the scanning, digitizing, vectorizing, and editing of Cartographic images as well as other graphic media. Key elements in the system are the Scan Graphics, scanning system and a vectorizing program known as RAVE^TM. The scanning system will be described in detail with attention to the document scanners use of flatbed plotter technology, also highlighted will be the RAVE Software's ability to convert the raster scanned data into a vector format for further information processing or editing at a cost effective price in a minimum amount of time. The system runs on a medium sized minicomputer with moderate storage requirements and requires no operator intervention or interaction during the Scanning process of a typical document which may contain up to 2.4 x 10⁹ information Pixels.

The texture, contrast and "noisiness" of meteorological data belongs to the class of visual images that requires unconventional and non-classical processing techniques. For instance, the digital Laplacian operator cannot be directly applied to these images without further modification. This is vividly portrayed in the application of classical edge, detection techniques to visual images that are very "busy", which tends to amplify the granularity of an image rather than generate useful edge detection. In this paper, a comparative study of classical edge detection techniques is described with actual atmospheric data obtained from geostationary satellite data. These results are compared to novel techniques developed at the Image Processing Center for Atmospheric Studies at Colorado State University.

This paper describes the development and performance demonstration of two charge-coupled device (CCD) chips for image processing. The aim of the work is to demonstrate the feasibility of developing custom CCD architectures that will enable the time-consuming, low-level processing functions (such as feature extraction) to be performed in real time. We describe the circuit concepts and device layout for six commonly used algorithms and include photographs of the raw and processed imagery.

The Sobel and Maximum Gradient edge detection operators are examined with respect to their response to Gaussian noise, and closed form solutions to their resulting probability density functions are presented. The results of million event Monte Carlo simulations are presented for various contrast to noise ratios (CNR) for an ideal edge and an expression is derived for edge detection performance as a function of CNR. It is shown that acceptable edge detection performance is achieved for CNR values greater than 4. For comparison purposes, the edge direction images for an aerial photograph are shown at four different signal to noise ratios.

Discontinuities in a noisy image can be detected by thresholding a differentiator adapted to the general spatial quality of the image. This paper develops a simple "blind" algorithm which can be used on the resulting binary map to locate edges. The data is first subjected to a thinning operation and the map is processed to yield clusters of points which are best represented by straight lines. It is not necessary to identify the subsets, and the algorithm is therefore highly independent of operator interaction. A mathematical basis for the technique and several examples are presented.

This paper is a sequel to a paper presented in 1977 regarding the development of an automated visual edge match system (AEMS). The AEMS is a system for measuring photographic image quality by means of edge sharpness and is particuarly useful for assessing focus and image motion in controlled tests. In latest developments, the system has been expanded to incorporate a focus control system which has virtually eliminated the operator as an error source. Other developments include the incorporation of a graphics terminal, a new sharpness algorithm, and greatly simplified operator controls. Data are presented which show the features of the machine and engineering details of the reengineered system.

The perception of surface luster in achromatic single view images seems to depend on the existence of regions with source-like properties, called here Light Source Effects. Light source effects are broken down into three categories according to gross aspects of the physical situation in which they occur, and criteria for detecting the regions they cause are suggested.

The specific problem addressed by this paper is the reading by computer of handprinted ocean depth soundings from "smooth sheets," records obtained from survey missions. The objective is to reduce the soundings and coordinates from the smooth sheets to digital tape for editing and preparation of standard oceanographic charts. In the system described the smooth sheet soundings are "read" by a video camera and recognized by software. Commercially available OCR systems are not suited to this task because the text is unconstrained in both font and format. In order to allow for differences in handwriting styles, the recognition software must work with those features that distinguish one character from another, while tolerating variation in the formation of a single character. For production purposes, a high through-put rate is needed. A practical recognition algorithm embodying these concepts will be described, along with the preprocessing steps found necessary. The performance of the system will be discussed, with examples.

Several functions of the Real Time Edge Processing Unit developed by the Northrop Research and Technology Center have been found to be directly applicable to certain visual automation problems. In the manufacturing of Advanced Composite Graphite assemblies, the following visual tasks are very time consuming and difficult for humans to accurately quantify: • epoxy impregnated graphite (prepreg) material quality evaluation as material is unrolled • accuracy and quality of the cut edge of prepreg • determination of direction of graphite fibers in cut plys • detection of correct ply (size and shape) for pick-up • accurate placement of each ply on a tool • inspection of assembly during lay-up • radiographic inspection of completed assemblies. This paper discusses the application of real time grey scale level slicing and gradient edge extraction to the above problems. Experimental results are also given.

This paper describes the design and engineering of an automated recognition system for detecting minute voids in lines and measuring the line width of unloaded printed circuit boards. An automated recognition system which utilizes a precision optical front-end, an electronic error-recognition system, an x-y table-positioning assembly and a printer was proposed, and a prototype model was con-structed to prove its feasibility. The described system is capable of detecting line flaws and errors in conductor registration and in network characteristic impedance. The line-width information detected can be used both for impedance measurement and for detecting line flaws due to overwidth and underwidth conductors. Based upon performance, results show that the recognition system has good error-detecting resolution for all serious defects. It also provides a fast, reliable, practical method for measuring and testing printed wire width, thus solving a serious problem in manufacturing circuit boards for computer use.

An approach is described for detecting and classifying tactical targets in FLIR imagery. The basic assumption used for segmenting objects from their background is that the objects to be detected differ from the background in grey level, edge, properties, or texture. Potential targets are selected from a large frame, by locating combinations of grey level, edge, value, and texture that occur infrequently over the entire frame. Once potential objects are obtained, they are segmented from their backgrounds using the identical process as above, except applied on a local level. The segmented objects are classified into three, types of vehicles or into false, alarms. The classification procedure uses features measured on projections made through the segmented objects. Results are shown for 32 test images.

With the increasing availability of three dimensional data, a need has arisen for three dimensional picture processing systems. One possible use for such systems is the simulation of surgery, which can make use of data collected via tomography and standard X-rays. Once the data is collected it needs to be reconstructed, organized, and displayed. To this end, we outline the design and focus of a three dimensional system, and suggest some key algorithms. Each phase of processing in this system may use a different data representation, and so the shift in representation between phases is highlighted. Finally, we present a scenario for the simulation of surgery using the described system.

This paper describes the ongoing research at Purdue University in the area of biomedical image processing and understanding. Problems associated with the above research together with the techniques used to solve them are discussed, and the research activities are also viewed in the context of the artificial intelligence (A.I.) field. Specific applications, like the automatic analysis of chest radiographs and the white blood cell neutrophils image segmentation are presented.

Templates in pattern recognition are typically derived by averaging a number of patterns from the same class. An average distance is usually computed using the Euclidean metric. In this paper we investigate the use of an angle measure when deriving templates for angle tests. Applications of angle tests to check waveform similarity in recognizing printed letters and in remote medical diagnosis are illustrated.

The cytocomputer, an image processing computer employing logical neighborhood operations similar to cellular automata transition functions has been developed. Various sequences of cytocomputer operations transform a primitive image configuration (a single dot) into various geometric shapes. A system has been developed which, when presented with an image containing some shape, produces a sequence of cytocomputer operations which generates that shape. The system uses a genetic algorithm to find the correct sequence of operations. In the genetic algorithm, strings of cytocomputer commands play the role of chromosomes, and undergo reproduction, crossover, and mutation in a process that mimics the scheme evolution uses to find increasingly fit biological organisms. The commands for generating a shape, once known, can be transformed for use in a pattern matching process which detects the presence of the shape in an image.

A new measure of scene content based on the concept of structural entropy is presented. The measure utilizes unary and binary relationships extracted from a relational representation of a scene. Bounds on the entropy measure are dependent only on the number of unary and binary relations. Experimental results illustrating the concepts developed in the paper are presented.

This paper introduces a novel technique called photometric stereo. The idea of photometric stereo is to vary the direction of the incident illumination between successive views while holding the viewing direction constant. This provides enough information to determine surface orientation at each picture element. Traditional stereo techniques determine range by relating two images of an object viewed from different directions. If the correspondence between picture elements is known, then distance to the object can be calculated by triangulation. Unfortunately, it is difficult to determine this correspondence. In photometric stereo, the imaging geometry does not change. Therefore, the correspondence between picture elements is known a priori. This stereo technique is photometric because it uses the intensity values recorded at a single picture element, in successive views, rather than the relative positions of features.

We present a general scheme for the representation and use of procedural knowledge in an image understanding system. Each procedure has associated with it a procedure descriptor, which characterizes the procedure's capabilities in terms of its input/output behavior and its performance characteristics. A general control paradigm, based on descriptor-directed invocation of procedures, is presented and shown to allow for: intelligent, goal-directed invocation; dynamic planning using local context; modular addition and change of procedural knowledge; and flexible allocation of computing resources. The representation scheme and control paradigm are being applied in the domains Of medical imagery (ultrasonograms and radiographs) and aerial imagery (of industrial sites).

This paper describes a scene analysis system (SAS) which is a general-purpose software package for processing, manipulating, and analyzing digital pictures. The SAS package is structured to allow a sequential operation of image processing operators on pictures stored in core. There are three categories of operators used in SAS: basic operators, general operators and user-developed operators. This paper aims to illustrate the versatility and usefulness of SAS in developing and analyzing image processing techniques.

Digital image processing techniques have been applied to the images from the meteorological satellite (NOAA) in order to utilize the information more effectively, whereas they are observed and analyzed as analog images. Processings such as contrast enhancement, extraction of sea surface, observation of sea surface temperature, Mercator and polar stereo mapping, enlargement by interpolation of picture elements, and enhancement of details have been studied. And it is shown that these image images can be applied to not only meteorology but also environmental measurement and fishery.

A computed tomography system designed to function in an industrial environment is being constructed. The modular design maintains the flexibility necessary in an environment where source strength and energy, as well as detector type, could change with each application. The static nature of industrial objects also allow economies of design not possible where dynamic objects are to be analyzed. Simulation of the device has been completed and expected image qualities obtained.

A computer control algorithm utilizing the modern control theory and the heuristic artificial intelligence approach has been developed to control the MIT Scheinman electric arm. The control algorithm is task-oriented with the capabilities of accepting visual input coordination and discrete word voice commands in addition to the linguistic commands typed in by an operator at the remote terminal. The motion of the arm is considered to be composed of motion of the wrist and orientation of the hand. A human operator, always present in the control loop and interacting with the system, generates linguistic task-directed commands at the remote terminal. A task-directed command is recognized, interpreted, and decoded into sequence of subtasks. The first subtask corresponds to the motion of the wrist which is controlled by the suboptimal feedback controller. The remaining four subtasks are further broken down into combinations of six primitive movements which govern the positon/orientation of the hand. The objective of the visual recognition algorithm is to identify objects and their locations surrounding the arm from its environmental library or model. The library is then updated to initiate the arm to complete its execution of task. Areas and circumferences of the objects are the two features chosen for recognition. The recognition of an object is then based on the threshold value of the weighted sum of area and circumference of the object. The set of weights is trained again when new object is introduced to the arm and incorporated into the library. The real time implementation of the algorithm on AARL MIT arm connected to a PDP 11/45 computer shows that it can recognize the objects surrounding the arm within 50 seconds.

The acquisition of analog video data, its conversion to digital data, and subsequent storage, transmission and reconversion to visual form is a well-practiced art. The specific application discussed herein is to record, measure and calculate the size distribution of a large number (several thousand) of air bubbles in a water flow. Previously, experimenters had acquired these data from 8 x 10 photographs, employing mechanical measurement techniques. In this technique data were taken from individual frames of high speed 16 mm motion pictures of the two-phase flow. Each frame was scanned and stored digitally as light intensity levels in a 256 x 256 array. The data array was processed using an on-line PDP-11 minicomputer. Using an input discrimination level, the PDP-11 examined each pixel sequentially until one was found which was part of a bubble. The program then followed the contour of the bubble, measuring perimeter, maximum X and Y excursion and integrating to determine area. At the completion of the contour program, control dropped into a routine which calculated desired bubble statistics. Control then reverted to the routine examining pixels. The techniques described herein were developed to solve a specific problem. These specific methods, if generalized, could be used to extract numerical data from digitized video data in a wide variety of fields.

A description of research work on the automatic visual inspection of printed circuit boards is presented as an example of a practical industrial automation problem. The major goal of this research is to develop a programmable visual inspection system applicable to printed circuit boards and other electronic assemblies. Described methods are the dimensional verification technique and the pattern matching technique. In dimensional verification, the inspection is accomplished by verifying the dimensional accuracy of certain features of the board. In pattern matching, standard features of the board are extracted interactively. The inspection is accomplished by matching these standard features with patterns of the unit under test.

Techniques for the automatic detection of defects, such as cracks and cavities in radiographs of artillery shells, have been developed and demonstrated. Because the defect indications are subtle and ride on an intensity trend that varies greatly across the field, it is necessary to precede the detection algorithms by pre-processing steps which "flatten" the trend and emphasize the defects. The algorithms are described and examples of their performance are shown.

Holographic interferometry has been demonstrated to be a useful nondestructive technique with applications, for example, in the inspection of small stainless steel pressure vessels. However, holographic interference patterns can be very complex and difficult to evaluate manually. We have developed an automated technique for interferogram interpretation using a PPP-12 minicomputer, COHU television camera and Hughes scan converter. A digitized image of the interferogram is stored on disc, and a small area is read into central memory. The fringe density in that region is estimated based on the number of peaks found in several line scans across the area under study. This calculation is repeated for successive small areas until a map of fringe density covering the entire part is compiled. If the fringe density map falls within an acceptance profile, the part is accepted. Experimental results demonstrate that this technique works well on interferograms having substantial variations in intensity and fringe contrast.

Recent work in coded aperture imaging has shown that the uniformly redundant array (URA) can image distant planar radioactive sources with no artifacts. This paper investigates the performance of two URA apertures when used in a close-up tomographic imaging system. It is shown that a URA based on m sequences is superior to one based on quadratic residues. The m sequence array not only produces less obnoxious artifacts in tomographic imaging, but is also more resilient to some described detrimental effects of close-up imaging. It is shown that in spite of these close-up effects, tomographic depth resolution increases as the source is moved closer to the detector.

This paper describes techniques being developed to characterize the features found during ultrasonic examination of stainless steel welds which are indicative of defects. Ultrasonic waveforms obtained from both defects and grain boundaries have similar time-domain characteristics. This phenomenon, together with variable signal attenuation and dispersion, is commonly encountered. The problem is to develop feature extraction techniques which will enable the examiner to discriminate reliably between weld defect signals and the other noise. Techniques presented use both time and frequency domain algorithms. The use of these techniques has demonstrated significantly better discrimination than conventional ultrasonic methods.

The low cost and simplicity of microprocessors will enable the image-processing field to expand widely. Data acquisition, conversion, storage, and display are becoming similarly inexpensive. Described is a low-cost image-processing system that illustrates the use of microprocessors. Also, a set of high-level image-processing primitive functions that are machine independent are proposed. These primitive functions may be used to write high-level languages in less space for shorter execution time. Their machine-independency should encourage widespread exchange of programs among members of the image-processing community.

As a step to automate assembly of various industrial parts, this paper describes a machine vision system that can recognize a variety of complex industrial parts and measure the necessary parameters for assembly, such as the locations of screw holes. Emphasis is given to a method for extracting useful features from the scene data for complex industrial parts so that accurate recognition of them is possible. The proposed method has the following features: 1) simple features are detected first in the scene and more complex features are examined later, using the locations of the previously found features; 2) the system is provided with a high-level supervisor that analyzes the current information obtained from the scene and structural models of various objects, and proposes the most promising feature to be examined next for recognizing the objects in the scene; 3) several sophisticated feature extractors are used to detect the complex features. The system has been tested on several sets of parts of small industrial gasoline engines and the results were satisfactory.