Our study consists of three phases: mapping the algorithms on the three topologies, simulating the execution of these algorithms, and design of the array. Our mapping follows the standard methodology proposed for the design of systolic arrays, since our application domain is very specific and the selected algorithms very regular. After the mapping is done we simulate the algorithms using the SES/workbench simulation package which allows us to collect statistics on the execution time and efficiency of our mappings and evaluate the performance of the three topologies in our application domain using different array and problem sizes. For each algorithm and topology the range of scalability is determined as a function of image size. In the design phase we propose an SIMD array with 2-D torus interconnection topology as a cost- efficient solution to the scalable implementation of the selected algorithms. Considerations entering the design phase are performance as determined by simulations, cost of implementation, and ease of scaling the machine size.

This paper discusses the use of neural network's for handwritten signature verification using the Fourier and Haar wavelet transforms as methods of encoding signature images. Results will be presented that discuss a neural network's ability to generalize to unseen signatures using wavelet encoded training data. These results will be discussed with reference to both Backpropagation networks and Cascade-Correlation networks. Backpropagation and Cascade- Correlation networks are used to compare and contrast the generalization ability of Haar wavelet and Fourier transform encoded signature data.

Discrimination technique between objects with similar colors are presented including traditional color measurement such as Lab and spectral signature. A number of methods have been evaluated for discriminating between different ceramic tile pigments. Ceramic products come in a wide variety with similar colors. Labeling mixups often occur between similar pigments resulting in incorrect products being shipped to customers. Mislabeled products are sometimes installed, since subtle color differences are difficult to perceive, especially with marginal illumination. This will result in direct expenses associated with rectifying these errors as well as customer inconvenience and dissatisfaction.

It is impossible to collect more than a tiny proportion of all of the possible examples of a given hue, to form a training set for a machine that learns to discriminate colors. In view of this, it is argued that color generalization is essential. A mechanism of learning colors defined by a human being has been developed by A. P. Plummer and is implemented in a commercial device known as the Intelligent Camera. (This device is being used by the authors in association with software written in Prolog. This combination has been described in an earlier publication and is being used in a study of methods for declarative programming of machine vision systems for industrial applications.) The Intelligent Camera can learn the characteristics of colored scenes presented to it. This paper presents four procedures which allow the range of colors learned by the Intelligent Camera to be broadened, so that recognition is made more reliable and less prone to generating noisy images which are difficult to analyze. Three of the procedures can be used to improve color discrimination, while a fourth procedure is used when a single and general color concept has to be learned. Several experiments were devised in an attempt to demonstrate the effectiveness of color generalization. These have shown that it is indeed possible to achieve reliable color discrimination/recognition in situations where it would hitherto have been difficult.

A high-speed algorithm for extracting line segment features in an image is described. The extracted line segments are used as the geometric features for identifying and locating objects of interest, specifically in a Model Based Vision system. The algorithm is divided into two parts with each part being performed using separate computer hardware. The first part of the algorithm involves locating pixels in an image which correspond to edges. This part of the algorithm was implemented on the Datacube MaxVideo-20 pipelined image processing hardware executing in real time. The second part of the algorithm involves the extraction of the edge pixels in a connected manner so that line segments can be identified. This part of the algorithm was implemented in software on a Sun Sparc 2 workstation using a run-length encoded image and a chain-code mapped image generated by the Datacube MaxVideo-20 hardware.

Interest in the discrete wavelet transform has grown explosively in the last five years, even though the underlying concepts are decades old and nearly identical transform techniques were being applied in industry 10 years ago. The most important aspect of the new work is the development of the underlying theory. Most if not all of the current applications of wavelets are software based, implying either slow execution times or very expensive computers. This paper shows the feasibility of using moderately-priced commercially-available image processing boards to carry out multi-band 2-dimensional (2D) wavelet transforms at real-time (30 images/sec) or faster-than-real-time rates. Implementations for both real and complex wavelets are shown. Word length and kernel size limitations are discussed, along with methods to overcome them. One-dimensional wavelets are mentioned as a special case of 2D wavelets. Because of the high speed and moderate cost of these implementations, much wider application of wavelets to industrial problems is now possible.

The discrete wavelet transform (DWT) provides a new method for signal/image analysis where high frequency components are studied with finer time resolution and low frequency components with coarser time resolution. It decomposes a signal or an image into localized contributions for multiscale analysis. This paper presents a parallel pipelined array processor for 1-dimensional (1-D) DWT. Unlike other VLSI DWT processors which processes signal data sequentially in a pipeline, this array processor can process all data in a signal segment in parallel and successive segments in the same pipeline which computes the multiple levels (octaves) of DWT. The speedup is linearly proportional to the width of the array (or the size of a segment), and thus the architecture is scalable.

A new shape-from-focus method is described which is based on a new concept named Focused Image Surface (FIS). FIS of an object is defined as the surface formed by the set of points at which the object points are focused by a camera lens. According to paraxial-geometric optics, there is a one-to-one correspondence between the shape of an object and the shape of its FIS. Therefore, the problem of shape recovery can be posed as the problem of determining the shape of the FIS. From the shape of FIS the shape of the object is easily obtained. In this paper the shape of the FIS is determined by searching for a shape which maximizes a focus measure. In contrast with previous literature where the focus measure is computed over the planar image detector of the camera, here the focus measure is computed over the FIS. This results in more accurate shape recovery than the traditional methods. Also, using FIS, a more accurate focused image can be reconstructed from a sequence of images than is possible with traditional methods. The new method has been implemented on an actual camera system, and the results of shape recovery and focused image reconstruction are presented.

This paper introduces the main issues involved in modelling the motion of a target that evolves in a 3D environment and presents a new approach to this sort of problem. The results are compared to those obtained through a whole family of polynomial curve-fitting algorithms based on statistical optimization and those called SMOP and MOCPA.

While morphology use is gaining in popularity, it can be a computationally intensive process. This is particularly true for applications requiring large structuring elements. In this paper we will describe the use of field programmable gate arrays (Xilinx 4000 FPGA family) to implement a systolic architecture for a morphology engine. The engine has a 14 bit data-path, a reconfigurable structuring element size and a 512 X 512 image size. We will describe the architecture, the FPGA implementation of the engine and the interface with the host (datacube).

The Euclidean distance transform (EDT) converts a binary image into one where each pixel has a value equal to the Euclidean distance to the nearest foreground pixel. It has important uses in image analysis, computer vision and robotics, and so its VLSI implementation is very useful. In this paper, a sequential algorithm which does not require global operations is first presented. We then present a square and a triangular shaped systolic arrays to realize the algorithm. For a n X n image on an equal size systolic array, the computing time is 5n- 5.

The prospects for building practical automated inspection machines, capable of detecting the following faults in ordinary, everyday loaves are reviewed: (1) foreign bodies, using X-rays, (2) texture changes, using glancing illumination, mathematical morphology and Neural Net learning techniques, and (3) shape deformations, using structured lighting and simple geometry.

A study was conducted to evaluate differences in bread crumb grain using digital image analysis. The objective of this study was to develop an image texture-based model which recognized the effect of different technological factors (formulation, mixing time and water absorption) on bread crumb grain. Co-occurrence matrix evaluations of 18 image texture features from 64*64 pixel subimages within the bread slice images produced the data based used for image texture feature analysis. Shape and size features of each slice were also included in this study.

This paper describes the development of a modular inspection system for the on-line inspection of printed-web (a continuous sheet of paper on a roll) material. Real-time inspection during the high-speed printing of high-quality, multicolor web materials can determine process deviations that could degrade product quality in time to take corrective action. This action can effectively and directly reduce spoilage within the current process and eliminate losses in subsequent processes. The system was initially developed for the U.S. Bureau of Engraving and Printing (BEP) for postage stamp production but can be adapted to almost any printed-web quality-monitoring applications.

Inspection systems for wide web materials have been unable to effectively image fine defects as they are detected. The amount of data produced by highly parallel video inspection cameras can exceed 400 MBytes/sec. The system described in this paper is capable of analyzing and displaying a detected image within seconds of the event using a single frame grabber and a 386 computer. The system can operate at processing speeds of greater than 400 MBytes/sec since it makes use of a novel post processing algorithm within the camera itself. The video cameras are based on TDI (Time Delay and Integration) technology to provide high grey scale resolution at high data rates and low light levels. The system has an adjustable resolution ranging from 256 to 24,000 pixels per line scanned. The scanning rate is adjustable to a maximum of 20,000 lines scans per second.

Industrial automation, robotics and automatic inspections include the processing of image sequences. The data rate can be very high in dynamic image systems and often real-time processing requires the use of special hardware. As an example for a typical image processing problem the following paper presents a visual motion detection systems which allows non- tactile registration of process parameters in a continuous rolling process. During an image sequence analysis the motion and the dimension of hot tubes will be measured on-line by a real-time image processing system. The motion is detected and measured by observing the translation of specific templates on the surface within a sequence of grey-scale images. This motion is analyzed by using a correlation procedure. The introduction of a state space model for the process parameters which includes a-priori knowledge and dynamic knowledge of the rolling process allows a compact motion description.

Lens offset, an image combining prism and a shadow graph are used to meet the performance objectives of an on-line precision dimensional gauge for transparent containers using focal plane array sensors and digital image processing.

Analysis of shattered automotive glass for determining particle size, shape and count is legally required to demonstrate that safety criteria have been met. Manual methods are labor intensive, subjective and prone to errors. This paper presents an approach for automating this analysis. Problems associated with missing boundary segments, can cause serious errors unless corrective measures are applied. Techniques for efficient extraction of particle boundaries and restoration of missing segments are presented.

A machine vision system was developed by BIRL to provide verification and identification of tire treads on the fabrication line. This paper discusses technical aspects of this successful application, from the original design to the final implementation. Specific emphases are given to the hardware and software implementation of the image acquisition and processing of the system. Aspects regarding the imaging of highly absorptive material (black rubber) and rapidly moving target are also discussed. Problems encountered during the development and integration of the system are described, and solutions are presented.

This paper describes a computer vision system for the high-precision inspection of bearing shells. We have developed algorithms to solve the problem of inspecting the wearing surfaces of sputter-coated metal shells for surface defects (high spots, cavities, blisters, grooves, and pores). The quality goal to be achieved was 0.3 m2/h, which for a typical 90 mm bearing shell being measured would mean about 0.5 minutes/shell. The resolution to be achieved was of each pixel covering an area of 24 micrometers by 24 micrometers . The analysis method was based on gray-scale rather than a binary algorithm. The quality standards were those defined by the Motoren and Turbinen- Union GmbH, Germany, and Daimler-Benz AG.

This paper reports on some recent developments in a software product that recognizes printed music notation. There are a number of computer systems available which assist in the task of printing music; however the full potential of these systems cannot be realized until the musical text has been entered into the computer. It is this problem that we address in this paper. The software we describe, which uses computationally inexpensive methods, is designed to analyze a music score, previously read by a flat bed scanner, and to extract the musical information that it contains. The paper discusses the methods used to recognize the musical text: these involve sampling the image at strategic points and using this information to estimate the musical symbol. It then discusses some hard problems that have been encountered during the course of the research; for example the recognition of chords and note clusters. It also reports on the progress that has been made in solving these problems and concludes with a discussion of work that needs to be undertaken over the next five years in order to transform this research prototype into a commercial product.

The paper presents a new approach to inspect the objects containing structure at different scales. The approach obtains correct descriptions of objects at a specified scale by processing it only at that scale. This feature of the technique is used to segment and inspect the coarser shape first, the results of which are then used to decide whether or not to segment and inspect the finer parts. This reduces the computational effort, and it also eliminates the need for a sophisticated segmentation technique, which may otherwise be necessary for such objects. The results on real 3-D data of a telephone instrument and a loaf of bread are presented.

An interactive vision system fully integrated with the VGA display of a PC controller has been developed for a robotic vehicle operating in hazardous environments. The system combines operator control information and real-time 3D video feedback on a standard PC display, thus obviating the necessity for separate displays and fully integrating the functions of observation and control.

The Machine Vision Testbed (MVT) is an integrated hardware and software environment for prototyping machine vision algorithms. The system incorporates special purpose image processing hardware (Datacube MaxVideo20^TM,) multiple high performance general purpose computers (SKYbolt^TM) and a Sun SPARCengine 1E^TM. A software environment provides the Sun processor with the capability to thread and control tasks amongst various processing elements. To support video processing requirements, special purpose hardware has been designed which provides an image broadcast bus from the Maxvideo20 into multiple SKYbolt processors. The underlying programming model for the system is to use the special purpose image processing hardware to perform computationally intensive image processing tasks and to have the local (or neighborhood) type operations run on the general purpose processors. This partitioning provides a path to migrate large portions of an application, which would typically run on special purpose hardware, onto fast general purpose processors for the purposes of software reuse and portability. This paper is intended to discuss the motivation for choosing such a hybrid architecture and to discuss the details of how such a heterogeneous collection of components was integrated into a system capable of processing video data.

For many vision-based inspection tasks, clear measurable features inherent in an image are sufficient to allow classification of the image content. Sometimes, however, it is difficult to select suitable feature sets, as the classification can only be made on the basis of subtle, diffuse relationships within the image. It has previously been shown that it is possible to automatically select sets of 'feature values' in such applications, using a procedure based on a modified version of the Karhunen-Loeve Transform (KLT), applied to window (imagelets) within images. This paper discusses the extension of that work in three directions. It describes the possibilities for using this data- reduction procedure in conjunction with more traditional and better understood classification methods for the decision-making stage. It discusses the potential for application of these ideas by combining the statistical transform coding stage with a range of image pre-processing operations. It also examines some of the issues of industrial integration of this procedure.

This paper describes the design and construction of an open, automated, solder bond verification machine for the electronics manufacturing industry. The application domain is the higher end assembly technologies, with an emphasis on fine pitch surface mount components. The system serves a measurement function, quantifying the solder bonds. It interfaces with the manufacturing process to close the manufacturing loop. A geometric model of the solder in a joint, coupled with a finite element analysis of the physical properties of solder, lead to objective measurement of the solder. Principle illumination systems are laser, X-ray and noncoherent lighting. Open, Objected Oriented design and implementation practices enable a forward looking system to be developed.

In the modern computer industry the Laptop computer has become an exciting and revolutionary product. While some engineers work to make computers bigger, faster and more powerful, others are striving to make them possible. Due to the obvious size and power constraints it is necessary to use Liquid Crystal Devices (LCD) in Laptop Computers. These displays consist of matrices of individual LCD cells. These individual cells form the pixels of the screen. It is thus possible to display a complex pattern on the screen by controlling the state of these individual pixels. In the Apple Computer PowerBook^TM Laptop computer each pixel is 0.5 mm square and on the display there is a total of 640*480 pixels.

An inspection system has been developed to improve both the reliability and the yield of surface treated components. Digitized data streams from photo detectors are fed into a hybrid DSP/multi-processor transputer system. The photo detectors collect scattered light from twin laser scanners (one scanner for each side of the relay contact). For real time operation the system inspects up to 10 contacts per second, finding defects as small as 5 (mu) in raw combined data streams totaling approximately 20 Mbyte/sec. The inspection system is currently inspecting the surfaces of plated relay contacts on the factory floor. This paper describes the inspection system with emphasis on the constraints imposed by the requirement for real time processing of large volumes of data.

The European Community has, since 1984, funded ESPRIT--the European Strategic Program for Research and Development in Information Technology. Amongst the many projects supported are a number relating to the development of machine vision for applications in manufacturing industry. Other vision related projects address applications outside the strict field of manufacturing, whilst yet others impinge upon vision but are not limited to it. This paper identifies these industrial-vision-relevant projects, completed and current, within the ESPRIT program, and indicates their current status and summarizes reported results in the public domain, but in no way sets out to report in depth on any project. The motivation for the preparation of this paper has been the author's involvement as a review for one such project, MOVE, (Microprocessor Open Vision Environment) which by its nature necessarily interacts with a number of the others. This paper is written for a mainly North American audience and is intended to give a very brief introduction to the ESPRIT program for those not familiar with it; it should not be seen as an authoritative statement of how ESPRIT is administered.

Opto-electronic sensing methods are increasingly being used for the inspection of manufactured goods. Such systems have the advantage of being non-contact methods, and are fast and flexible. One particular aspect of inspection is the measurement of three-dimensional shape. Even in shape measurement, there is a wide range of requirements. Some components, such as mechanical parts, may have exacting geometric requirements if they are to function correctly. Here, inspection is concerned with checking geometric details. Correctness of other objects such as loaves of bread is much more difficult to pin down, and symbolic methods are more appropriate than high-precision numerical methods. This paper considers a range of inspection tasks, and appropriate methods for carrying them out using visual sensing techniques.

In order to improve flexibility when changing between different laser processing workheads we have developed an automatic tool changer for laser machining centers. This tool system was designed for large multi axis machines such as gantries suitable for three-dimensional processing, but can also be used for other types of laser operations like robots for example. The system also offers the possibility to combine laser processing with deburring and milling on the same machine.

This paper describes the development of a visual programming environment for the Max Video 20, called MAPE. It is a tool for the mapping of image processing algorithms on Max Video hardware.

This paper is concerned with the modelling of safety requirements using Petri nets as a tool to model and simulate a Flexible Manufacturing System (FMS). The FMS cell described comprises a pick and place robot, a multi-head drilling machine together with a vision system and illustrates how the hierarchical structure of Petri nets can be used to ensure that all fail- safe requirements are satisfied; block diagrams together with fully detailed example Petri nets are given. The work demonstrates the use of cell and robot control Petro nets together with robot subnets for the x, y and z axes and associated output nets; the control and output nets are linked together with a safety net. Individual machines are linked with the control and safety nets of an FMS at cell level. The paper also illustrates how a Petri net can act as a decision maker during image inspection and identifies the unsafe conditions that can arise within an FMS.

A rich theoretical background to the problems that occur in the automation of material handling can be found in operations research, production engineering, systems engineering and automation, more specifically machine vision, literature. This work has contributed towards the design of intelligent handling systems. This paper will review the application of these automated material handling and packing techniques to industrial problems. The discussion will also highlight the systems integration issues involved in these applications. An outline of one such industrial application, the automated placement of shape templates on to leather hides, is also discussed. The purpose of this system is to arrange shape templates on a leather hide in an efficient manner, so as to minimize the leather waste, before they are automatically cut from the hide. These pieces are used in the furniture and car manufacturing industries for the upholstery of high quality leather chairs and car seats. Currently this type of operation is semi-automated. The paper will outline the problems involved in the full automation of such a procedure.

The field of automated optical inspection (AOI) is continuing to grow rapidly as manufacturing industries begin to recognize the potential of this area. However the AOI industry is rapidly approaching a bottle-neck in the cost of these systems. Computer hardware costs have dropped substantially while precision mechanical hardware costs have remained relatively fixed (and costly) over the last decade. While cost is a major contributing factor to the exclusiveness of AOI, another major factor is the lack of flexibility and general integration problems. This paper describes an AOI system prototype designed to counter some of the above problems regarding the perception and integration of AOI systems.

For several years, inspection integrators have had the use of fiberscopes which are employed as borescopes for use in remote inspection applications. However, these are limited to RS170 performance in terms of throughput, sensitivity, and resolution. This paper reports on a new machine vision camera which is small in size (2 inches by 3/4 inches in the image plane) yet offers the resolution and throughput of a variable scan camera. This miniature light weight camera for machine vision can be placed in spaces where physical size and weight are limitations, such as at the end of remote manipulator arms used in automated assembly. A miniature line scan camera using a 1024 pixel line scanner is discussed which has the added feature that video data from multiple remote camera heads can be electronically combined and digitized at a central controller. This permits multiple views of the same object from cameras which can be placed in a space less than 3 cubic inches. This flexibility permits the ease of placement of cameras on the assembly line.

Recently, unmanned operation in machining is requested. But the tool chipping is an important problem which prevents this factory automation. Many methods such as acoustic emission, motor current monitoring, motor torque fluctuation monitoring have been developed for tool monitoring. But, these methods have their own problems applying to actual machine. Authors have developed the inprocess monitoring system which detects the wear or chipping of a tool of lathe machining. This system consists of laser diode, CCD camera and image processing systems. This system shows the shape of tool tip and the roughness of cutting surface.

This paper describes the development of a practical object recognition framework for industrial inspection applications. The framework is being used to develop a prototype automatic inspection system for Surface Mounted Electronic Assemblies. The recognition technique uses the object oriented paradigm to both process and model the image information. Three inter-linked information structures, that allow both image and object modelling information to be represented at multiple resolutions, are described in this paper. The first structure is a hierarchical feature based description constructed from the recursive grouping of perceptual features. The second structure provides a containment tree to describe the connections between sub-component parts of a complex object. The final structure is the inheritance tree that provides the capability to model multiple variations of the same object type. The inspection module begins by processing image information from a newly developed non-linear shape abstraction technique. The information is delivered in the form of a multi- resolution scale space description of the captured image.

CAD-model based recognition and localization is one of the most important research areas where active vision principle can fully demonstrate its power and effectiveness. In this paper, we propose to use sparse range images to recognize and localize common mechanical parts that are isolated on the worktable. We emphasize the extraction of features that are suitable for the recognition and localization of objects and are easy to be detected by a light stripe range finder. Features of the mechanical parts are first carefully analyzed, which are then selected automatically from the corresponding CAD models. They are classified into two types according to types of the surfaces, that is, features for planar surfaces and features for curved surfaces, and are organized into a hierarchical feature base which is used for feature matching during the process of object recognition and localization. We explain our strategy in detail and show several experimental results of recognition and localization using a simulation system.

An optoelectronic moving object linear dimension inspection method is studied. The method is based on the outline image registration with the help of photodiode arrays operating in the signal accumulation mode. A theoretical analysis and experimental test of this method are carried out. Measurement errors of such objects as slit and roller are determined. The degree of object displacement effect on the accuracy of measurement in the measurement region is estimated.