Described herein is intelligent camera technology suitable for a wide range of inspection applications including webs, widgets, gauging, etc. The system is modular whereby it can sue from one to fifteen intelligent cameras networked together to a personal computer, which in turn may be networked. The inspection system software is multi-threaded, autoconfigurable, and offers: friendly GUI, expert system illumination control, and interchangeable components accommodating numerous post-processing algorithms within the cameras. Camera algorithms allow for the processing of image data at speeds in excess of 400 Mbytes/sec using one frame grabber. Various camera technologies are available including TDI, allowing for high gray scale resolution at high data rates. Multi camera systems can have different network topologies, according to the user's requirements. Each camera can be individually addressed/instructed further increasing system functionality. Diagnostics are incorporated both within the camera and the inspection software to enhance the serviceability of the system.

We have developed a high speed automatic inspection system which verifies the validity of electron beam exposure data used for fabrication of VLSI photomasks. By employing neural network accelerator board and adopting a flexible verification scheme based on parallel processing, the system performs 100 times faster than the previous system running on a general purpose workstation. In this paper, the architecture of this system and some salient techniques implemented in the system are presented.

The results of a program to develop a low-cost, real-time approach for 100 percent inspection of fast-moving process lines for textures, paper, and steel are reported. Initial results will be described of a laser based scanning system that uses a neural network for training and the implementation of a neural network with an optical processor for real-time defect inspection.

A system for measuring barriers in a color AC plasma panel has been developed. Barriers are used in this type of display to prevent phosphors in cells adjacent to lit cells from being excited which adversely affects color purity. The geometry of the barriers is a significant factor for successful operation of color plasma panels and must be measured to verify that the barriers are within specifications. Barrier height is on the order of several mils with a pitch on the order of about 10 mils. A system developed for spacer measurements was available for this application. However, it did not have sufficient light sensitivity because the barriers reflect light much less efficiently than traditional panels. The original system employed a light section microscope for height measurement. The video amplifier gain was boosted significantly in the frame grabber and frame integration was provided to reduce noise. Finally, background subtraction was provided to remove shading variations associated with the normally insignificant dark current of the CCD sensor. Once a good image had been obtained, morphological processing was performed to reduce noise and centroid calculations were performed to provide an accurate measure of the barrier surface height.

This paper presents an overview of work underway at Industrial Research Limited directed ultimately at developing an automated grading system for pickled sheep pelts. The wide variety of defects and indistinct nature of some of them illustrate the difficulties associated with the automatic inspection of natural and varying products which poser significant technical challenges. A novel imaging approach has been taken to highlight the features of interest and thus simplify the inspection task. A laser scanner has been developed which provides simultaneous acquisition of three image types representing transmission, reflectance and fluorescent properties of the pelts. Of particular interest is the fluorescence image which highlights pelt defects not normally apparent with either the naked eye or a camera. The transmission and reflection images can be used in conjunction with the fluorescence image for defect detection as well as for calculating pelt area and for recognition of pelt identification marks. Various types of pattern recognition algorithms are under investigation to assess their potential for automating the grading process using as inputs the three image types. The approach taken is based on supervised learning using feature vectors derived in various ways for the pelt images.

This paper focuses on producing a state-of-the-art technique for designing an image recognition system for machine vision applications. The motivation behind the new system design is to provide a unique methodology, using strategic design techniques, to implement a system that addresses real-world image recognition applications. The introduction of application-specific, massively parallel array of processors, where low-level processing is accomplished on reconfigurable hardware structures, highlights the scheme. The system was built and simulated on a VLSI chip and results were verified using Electric Rules Check and Harris Timing Analysis examination tools. The system is composed of there functional layers and a main control unit. The top two layers are used for image loading and manipulation and the third layer is used for processing the pixel values. Each layer has a local control unit, while the main control unit oversees the operations of the whole system and synchronizes the processes. A CAD designer tool was implemented to facilitate the design and reconfiguration f the low-level processing elements. This tool has a modular library of processing elements and a logic verification algorithm. The current architecture of the chip was built to accommodate a 64 by 64 array with unlimited stackable characteristics to handle larger images. This paper will present the top level design layers, the distributed control, and demonstrate that the proposed system is image complexity invariant. The massively parallel approach of transferring data and control signals and the processing of image data is presented. The system takes into consideration cost, size, speed,and reliability needs of today's applications.

This paper introduces a novel networking strategy for RAM- based Neurons which significantly improves the training and recognition performance of such networks while maintaining the generalization capabilities achieved in previous network configurations. A number of different architectures are introduced each using the same underlying principles. Initially, features which are common to all architectures are described illustrating the basis of the underlying paradigm. Three architectures are then introduced illustrating different techniques for employing the paradigm to meet differing performance specifications. The architectures are described in terms of the structure of the neurons they employ. Greater detail of the various training and recognition algorithms employed by the architectures may be found in the referenced papers.

The purpose of this paper is to demonstrate how a structural matching approach can be used to perfonn effective rotational invariant fingerprint identification. In this approach, each of the exiracted features is correlated with Live of its nearest neighbouring features to form a local feature gmup for a first-stage matching. After that, the feature with the highest match is used as a central feature whereby all the other features are correlated to form a global feature group for a second.stage matching. The correlation between the features is in terms of distance and relative angle. This approach actually make the matching method rotational invariant A substantial amount of testing was carried out and it shows that this matching technique is capable of matching the four basic fingerprint patterns with an average matching time of4 seconds on a 66Mhz, 486 DX personal computer.

The purpose of this paper is to demonstrate how a structural matching approach can be used to perfonn effective rotational invariant fingerprint identification. In this approach, each of the exiracted features is correlated with Live of its nearest neighbouring features to form a local feature gmup for a first-stage matching. After that, the feature with the highest match is used as a central feature whereby all the other features are correlated to form a global feature group for a second.stage matching. The correlation between the features is in terms of distance and relative angle. This approach actually make the matching method rotational invariant A substantial amount of testing was carried out and it shows that this matching technique is capable of matching the four basic fingerprint patterns with an average matching time of4 seconds on a 66Mhz, 486 DX personal computer.

A variety of time delay and integration (TDI) arrays have been developed. The TDI image sensor offers significant improvement in performance over a linear CCD-sensor with respect to sensitivity. This is particularly significant in low level operations since the exposure time is increased by a factor which is equal to the number of the TDI stages in the sensor. Typically, the use of a TDI camera is restricted to cases where the surface is viewed from the direction of the surface normal. This is because the TDI-camera sensor has to be parallel to the viewed surface plane, in order to avoid a decrease of image quality due to varying magnification. However, in certain visual inspection applications it would be advantageous to use non-zero viewing angles. Three different solutions were tested and analyzes: a commercial shifting lens, standard lenses with an extension structure to support the lens in the shifting position and a commercial shifting and tilting lens. The results indicate that the TDI-camera can be used with viewing angles up to 30 degrees from the surface normal,which has been found to be the optimal viewing angle in some visual inspection applications.

A system for 3D gauging of small fibers has been developed for process monitoring. The basic hardware consists of a pair of 2048 linear cameras orthogonally positioned, an IBM PC-compatible Pentium computer with frame grabber, a stepper motor and associated hardware for translating the fiber, a bright-field light source and special optics. The fiber is moved vertically past the two cameras as they scan. the computer acquires each scan line, processes it and then issues control signals to the stepper motor. Several different image processing operations are used to minimize the effects of illumination nonuniformity since fibers will sometimes have low contrast due to their small size. There are two sources of illumination variations, spatial and temporal which are processed independently. Image analysis is performed to provide 3D fiber shape characteristics.

With this paper we present a modular built multi sensor for visual inspection around the coin. Due to the application of the most modern sensor and calculating techniques, a sight check of highest reliability is realized. The sight check can be extended by modules, e.g. detection of metals or control of alloys.

The paper describes some early results obtained by applying a multiscale approach to texture analysis for effect detection on leather surfaces. Texture properties have proved to be a valuable help in characterizing the chaotic structures normally present on leather and the oriented structures typical of defects in spite of the wide range of associated visual appearances. Texture based defect detection can be done effectively by using the coefficients of the projection of a flow field into a vector space spanned by an appropriate basis function set. Being texture strongly related to scale we propose to extend the work of Rao and Schunck by employing a bank of filters, tuned at different scales, in order to perform the flow field estimation. Measures based on the coherence of outputs are used to determine at each point the scale providing the strongest oriented texture. Results showing the effect of synthesizing multiscale data into a single flow field will be provided.

A central problem in automatic visual inspection and computer vision is determining the extend to which one shape differs from another. This is the key element of any inspection algorithm. Pattern recognition operations such as correlation, template matching and model based vision methods can all be viewed as techniques for determining the difference between shapes. One of the methods for measuring the 'difference' between two shapes such as a 'model' and an 'image' is using distance transforms. In this paper the problem of visual inspection of deformable materials in general and lace in particular is considered. A mechatronic approach based on correlation and distance transforms using a line-scan CCD camera is presented. The advantages and disadvantages of each procedure compared with other approaches are also discussed.

A new high performance CCD camera family is presented. The camera incorporates a micro-controller/PLD combination to provide users with computer control of image acquisition, image processing and analysis. User control of image acquisition includes adjustable gain and offset, data rate and timing.Image processing and analysis algorithms are implemented within PLDs and regulated via the microcontroller. A variety of image processing algorithms are discussed including gray scale thresholding, RLE, edge detection and gauging. Parameters that govern the nature of the image processing algorithm can be computer controlled and randomly accessed. The camera transmits a combination of high speed digital video for frame grabber acquisition and low speed serial analysis and status information to a PC serial prot. Pixel rates at up to 20Mbyte/sec/channel for up to 8 channels combine for data throughputs of up to 160Mbytes/sec/camera. Camera configurations include single and multi-tap, linescan, TDI and area array formats. Up to 15 cameras can be easily integrated together to form a single network through the use of a programmable communication HUB unit. An even larger number of cameras can be linked together by combining several networks into a web. Applications for the programmable camera/HUB technology include web and parts inspection, template matching and gauging.

A method for modeling a hazardous environment automatically, for real time task planning, using laser range images of multiple partial views of a single work space scene, is presented. Viewpoint invariant properties of differential- geometric shape descriptors like the mean curvature and the Gaussian curvature are utilized to classify a pre-smoothed laser range image into one of eight basic surface types. Connected components of these classified pixels, that satisfy specific planarity constraints, are clustered into planar regions. Selected image processing techniques are applied to the planar regions in order to extract their critical features, and to synthesize those polygons, with normals approximately orthogonal to the sensor view-axis. Detailed shape of the objects in the scene develop through view integration of multiple partial views of the objects in the scene.

Considerable work has been reported in recent years on the utilization of hierarchical architectures for efficient classification of image data typically encountered in task domains relevant to automated inspection, part sorting, quality monitoring, and so on. Such work has opened up the possibility of further enhancements through the more effective use of multiple-experts in such structures, but a principal difficulty encountered is to formulate an efficient way to combine decisions of individual experts to form a consensus. The approach proposed here can be envisaged as a structure with multiple layers of filters to separate an input object/image stream. In an n-way classification problem, the primary layer channels the input stream into n different streams, with subsequent further processing dependent on the form of decision taken at the earlier stages. The decision about combining the initially filtered streams is taken based on the degree of confusion among the classes present. The filter battery effectively creates two separate types of output. One is the relatively well-behaved filtered stream corresponding to the defined target classes, while the other contains the patterns which are rejected by different filters as not belonging to the target stream. Subsequently, more specialized classifiers are trained to recognize the intended target classes only, while the rejected patterns from all the second layer filters are collected and presented to a reject recovery classifier which is trained on all the n input classes. Thus, progressively more focusing of the decision making occurs as the processing path is traversed, with the resultant increase in the overall classification capability of the overall system. In this paper, classification results are presented to illustrate the relative performance levels achieved with single expert classifiers in comparison with this type of multi-expert configuration where these single experts are integrated within the processing framework outlined above. A number of conclusions are drawn in relation to the value and potential of hierarchical/multi- expert systems in general and, more importantly, some guidelines are offered about optimizing classifier structures for particular application domains such as automated inspection processing.

This paper introduces a texture descriptor that is invariant to rotation. The new texture descriptor utilizes the property of the magnitudes of Fourier transform coefficients that do not change with spatial shift of input elements. Since rotating an image by an arbitrary angle does not change pixel intensities in an image but shift them in circular motion, the notion of producing textural features invariant to rotation using 1D Fourier transform coefficients can be realized if the relationship between circular motion and spatial shift can be established. By analyzing pixels in a circular neighborhood in an image, a number of FOurier transform coefficients can be generated to describe local properties of the neighborhood. From the magnitudes of these coefficients, several rotation invariant features are obtained to represent each texture class. Based on these features, an unknown image is assigned to one of the known classes using a nearest neighbor classifier. All of the feature samples for the classifier are extracted from unrotated texture images only. The new texture descriptor outperformed the circular simultaneous autoregressive model in classifying rotated texture images taken from 30 texture classes.

As an inspection technology, x-rays have been used in food product inspection for a number of years. Despite this, in contrast with the use of image processing techniques in medical applications of x-rays, food inspection systems have remained relatively rudimentary. SOme of our earlier work in this area has been stimulate by specific x-ray inspection tasks, where we have been required to locate contaminants in batches of particular packaged products. We have developed techniques for contaminant detection in both canned and bagged products. This paper gives an overview of work undertaken by Industrial Research Limited on the development of automated machine vision techniques for the inspection of food products for contaminants. Our recent work has concentrated on the development of more generic techniques for detecting contaminants in a wide range of continuously produced products, with no requirement for product singulation. A particular emphasis in our work has been the real-time detection of contaminants appearing indistinctly in x-ray images in the presence of noise and major product variability.

In a previous paper, the authors described a real time approach for tracking of parametric objects. The system proposed used coordinate trees that enabled the accurate and efficient calculation of the projection calculus in real time. Frame to frame coherence constraints were provided for by ensuring incremental movements between scenes, that is, the change in the perceived view of projected object boundaries from one instance of motion to another. This requirement can be directly translated into geometric constraints that can be enforced in parallel. This paper provides necessary implementation details for the parallel aspects of the algorithm and further extends the algorithm to apply to the cases of a moving viewer.

Theoretical and experimental results of investigations which are leading to the digital TV-based systems for objects in motion against a non-steady-state background detectors design are presented. Specifically, the peculiarities of proposed adaptive threshold-based method of detection was explored for wide range of meteorological and observing conditions. The effect of multiplication objects marks for the spatially distributed moving objects was analyzed and several approaches to consolidation of them are proposed. They are based on threshold value back-feed varying and nonlinear cluster analysis of structures and images in question. The suboptimal of threshold estimation is developed and discussed.

Rapid agile metrology for manufacturing (RAMM) using multiple non-contact sensors is likely to remain a growing trend in manufacturing. High speed inspecting systems for manufacturing is characterized by multitasks implemented in parallel and real-time events which occur simultaneously. In this paper, we introduce a real-time operating system into RAMM research. A general task model of a class-based object- oriented technology is proposed. A general multitask frame of a typical RAMM system using OPNet is discussed. Finally, an application example of a machine which inspects parts held on a carrier strip is described. With RTOS and OPNet, this machine can measure two dimensions of the contacts at 300 parts/second.

It is very important to evaluate mechanical parameters of material by testing in mechanical manufacture, especially the tensile strength of material. The results of the testing showed that its accuracy is relevant to many factors, such as the manufacture technology of testing samples, load growing in number with speed in the test processing etc. Then a key to get a high accuracy (sigma) _b in testing is to measure the traverse area of the sample precisely since the precision of the traverse area measurement affected seriously on the precision of measuring (sigma) _b. Conventionally, the area of the sample is commonly measured by measuring its diameter in two-direction in one plane with a sliding gage used. This measuring method not only causes a high system error in experimental testing, but also measure the area only in 2D. In fact, the tensile distribution of the sample is in 3D and related with the tensile strength (sigma) _b clearly in 3D. For this reason, the relationship between the traverse area of the sample and its tensile strength is studied theoretically and experimentally in 3D. In this paper, it is emphasized that the traverse area is in 3D and then a method of measuring the are of one is designed in experiment.

We describe a compact optic correlator architecture which does not require a CCD camera to input the image to be interrogated. The white light illuminated image can be polarized and imaged directly onto a thin bacteriorhodopsin film which modulates the film's birefringence. Read out of this written information can be achieved using a low power diode or HeNe laser in order to put the image information onto a coherent wavefront. Architectures with the bR film in the input and Fourier plane are considered using performance measures such as the fidelity of feature identification and speed.

The paper starts with a review of integration techniques for calculating height maps from dense gradient fields. There exist a few proposals of local integration methods, and two proposals for global optimization. Several experimental evaluations of such integration techniques are discussed in this paper. The examined algorithms received high markings on curved objects but low markings on polyhedral shapes. Locally adaptive approaches are suggested to cope with complex shapes in general.

A machine vision system used for quantitative analysis of the uniformity of powder blending has been built at the CAIP center of Rutgers University. A wide variety of instruments are used in the system in order to perform the image processing algorithms required. This paper will introduce the system, focusing on the high speed image processing hardware configurations and discussing the scheme for developing the software to manage this comprehensive system.

A system for high speed capture of image sequences and real- time filed flattening has been designed and implemented. This system is useful for applications where the objects of interest in a scene are moving or changing, since it provides a means for obtaining the images as the objects move, as well as compensating for lighting variations. The system is implemented on a Datacube MV200 hosted by a SUN Sparc 5 workstation. The system has been successfully used to capture image sequences as part of a system to develop algorithms for dynamic face recognition.

A high frame rate display system has been developed to present image sequences at 114.4 frames/sec as opposed to the video frame rate of 30 frames/sec. In order to do this, we are using a video display monitor in stereo mode and a Datacube MV-200 image processing hardware system to display images at rates approximately twice as fast as would otherwise be possible. We propose to use this system to assess the elderly diver's ability to locate and distinguish computer generated images of vehicles and to determine their direction of motion in a variety of simulated conditions.

Prolog image processing (PIP) is a multi-media prototyping tool, intended to assist designers of intelligent industrial machine vision systems. This is the latest in a series of prolog-based systems that have been implemented at Cardiff, specifically for this purpose. The software package provides fully integrated facilities for both interactive and programmed image processing, 'smart' documentation, guidance about which lighting/viewing set-up to use, speech/natural language input and speech output. It can also be used to control a range of electro-mechanical devices, such as lamps, cameras, lenses, pneumatic positioning mechanisms, robots, etc., via a low-cost hardware interfacing module. The software runs on a standard computer, with no predecessors in that the image processing is carried out entirely in software. This article concentrates on the design and implementation of the PIP system, and presents programs for two demonstration applications: (a) recognizing a non-picture playing card; (b) recognizing a well laid table place setting.

In a series of eleven previous papers a radically different method of implementing a wide range of neighborhood image processing operations has been presented, under the acronym SKIPSM (separated-kernel image processing using finite state machines). Simply by changing the contents of two lookup tables, one can use the same software code or the same hardware configuration can carry out a long list of operations, including binary morphology with multiple large structuring elements, multiple simultaneous steps of the grassfire transform, 'fuzzy' binary morphological operations, grey-level morphology, binary skeletonization, binary correlation, binary openings and closings in one pass, and certain global image processing operations. The execution time is very fast, and is totally independent of the size of the neighborhood or of the number of simultaneous operations being performed. This paper gives a detailed description of the steps for creating lookup tables for binary morphology.

PIP (prolog image processing) is a prototyping tool, intended to assists designers of intelligent industrial machine vision systems. This article concentrates on the multi-media extensions to PIP, including: 1) on-line HELP, which allows the user to satisfy PIP goals from within the HELP facility, 2) lighting advisor, which gives advice to a vision engineer about which lighting/viewing arrangement is appropriate to use in a given situation, 3) device control, for operating a robot work cell, 4) speech input and (simple) natural language understanding, 5) speech synthesis, 6) remote operation of PIP via a local area network, and 7) remote operation of PIP via a local area network. At the time of writing, on-line access to PIP, via the Internet, is being developed.