This paper discusses signal processing for the detection of defects in automated inspection using machine vision. A generalized methodology is presented applicable to a wide range of tasks. A method is presented for design of a system to achieve quantitative performance requirements in terms of probability of defect detection versus probability of false alarms. Some modifications which can improve performance of the basic scheme are described. Applications of the method are mentioned briefly.

This paper describes a vision system designed for automatic inspection of galvanized metallic strips in real-time. First, we present the image acquisition system whose two main components are a linear camera and a specific lighting device. In the second part, an original procedure is proposed in order to segment the line images at the line image acquisition rate. It consists in a fast adaptive thresholding scheme which determines global thresholds for each line image. In order to achieve an exhaustive control of the whole production, these tasks are run in real-time on a specific hardware architecture. The prototype described in the last section of this paper has been integrated on a production line to evaluate its efficacy.

A general architecture for detecting and analyzing surface defects in aluminum strip is described. Information concerning visual information from the aluminum surface, surface temperature and strip dimensions--profile thickness-is processed jointly by means of an expert system in order to determine the quality level of each aluminum coil produced; control actions over the casting process, derived from this information, are also suggested by an expert system. This paper shows in deep work related to surface image analysis. The data volume to be processed, up to 20 Mbytes/s, has forced up the development of a high parallel architecture for high-speed image processing. A specially suitable lighting system has been developed for enhancing matrical image acquisition from metallic surfaces that includes reflect avoidance as well as uniform incident angle of light along the scanned portion of surface--about 120000 square mm.

Developers and users of industrial inspection systems can choose from a variety of CCD- cameras with large differences in performance and price. Many industrial inspection tasks require the extraction of metric information and ever more frequently of accurate 3D metric information. The performance of CCD-cameras for metric inspection tasks is influenced by a large number of factors such as illumination, sensor characteristics, signal transmission and frame grabbing, calibration techniques, and algorithms. Recent advances in characterization and calibration of CCD-cameras are presented. Measurement accuracies exceeding 1/50^th of the pixel spacing were attained with off-the-shelf equipment in controlled laboratory tests. Selection criteria of cameras as a factor of resolution and accuracy requirements are discussed.

In 1972, Barnea and Silverman presented a new approach to the wide field of template matching, the SSD-algorithm. Further work has been done to adapt the method to gain subpixel accuracy. Intense investigation of the proposed algorithms led to our new approach: by interpolating the template instead of the reference image, and by applying sort of an error- correction to the resulting subpixel-value, both computation time and accuracy can be improved. Exhaustive experiments with a CCD-camera and various kinds of reference images showed that a maximum error of 10% of the pixel period can be expected. Depending on the kind of image, mean square errors range from 0.4% to 4%.

Machine vision systems are often used in gaging applications. Analysis of the measurement system is important to determine if the measurement process is adequate to measure the part- to-part variability in the manufacturing process and if the system can be implemented as a reject station. The paper provides the principles of measurement system analysis and the principles of Three-Sigma and Six-Sigma quality control. This paper introduces formulas for evaluating the additional yield loss due to rejection of good parts in Three-Sigma and Six- Sigma process design as function of the measurement system resolution. At the end of the paper we will discuss gage resolution requirements for Three-Sigma and Six-Sigma quality control.

In this paper we propose a step-by-step design procedure for image acquisition arrangement planning. An example design for a steel sheet inspection application is presented to illustrate the different phases of the procedure. In the example, the inspection aspects of three different defect types, a dent, a scratch and rough local coating are handled. This design procedure is as follows. Firstly, some important optical parameters of the surfaces to be inspected are measured, for example, scattering, surface roughness, spectrum and polarization properties. Secondly, the optical behavior of the surface is modelled, e.g., by using reflectance models. The third step is to specify the components of the imaging system and to simulate the whole imaging process in a computer by using optical analysis tools. Finally, the simulation results and the real world are compared and the usefulness of the computer based design is evaluated.

Most industrial vision applications use black and white sensors. Specific lighting or optical filters are often used to analyze colored parts. Nevertheless, this approach can be inadequate for applications in particular domains (foods, printing, etc.). In some cases, it can be necessary to analyze the hue of the parts in addition to their intensity. The availability of low- cost CCD color cameras allows the development of this kind of solution. The purpose of the project presented here is to define and develop a color vision system corresponding to industrial requirements, especially in the range of low to medium performance systems. Such a system must respect the following constraints: (1) wide variety of problems and small amount of equipment for one kind of application, (2) cadences: many applications need to control about 10 parts per second, and (3) cost.

Development of vision systems which are easy to use is the `holy grail' of vision system engineering. This paper describes the architecture of our so-called `SmartCamera' hardware platform and describes the type of modules and functions which can be provided in such a system. The aim of the paper is to describe how this approach can simplify vision system application in some cases. By providing an integrated software and hardware package which has been configured for a highly specific application, the system integrator or end-user will find the applications engineering much easier and most importantly, less time-consuming.

A binocular stereo vision system has been used to demonstrate freespace determination for local navigation of a laboratory robot vehicle, and object recognition, location and tracking for docking of the vehicle with known objects. Experiments on the performance of the object recognition, location and tracking of the system implemented on a multi-transputer architecture hosted by a SUN workstation are reported and discussed. It is shown that the choice of focus feature groups critically affect the success of the object recognition, but that once recognized, objects can be located consistently to an accuracy of a few mm and approximately 2 degree(s) in the working volume of 2 - 3 m. Small, relative displacements of an object by a few cm can be detected to a similar accuracy whilst larger displacements of 0.5 - 1.0 m can be measured to within a few %. Tracking of objects in 3D at 2 - 3 Hz using a small set of focus features to limit processing of the incoming image data to a small number of regions of interest is shown to be only slightly less accurate.

Two systems are described and results presented demonstrating aspects of real-time visual guidance of autonomous mobile platforms. The first approach incorporates prior knowledge in the form of rigid geometrical models linking visual references within the environment. The second approach is based on a continuous synthesis of information extracted from image tokens to generate a coarse-grained world model, from which potential obstacles are inferred. The use of these techniques in workplace applications is discussed.

Obstacle detection and avoidance is an important issue for autonomous robots that work in a dynamic environment. In order to solve such a problem, we propose a method based on passive stereo-vision. Our implementation is a fast algorithm that achieves directly a rough estimation of the 3D position of obstacles by avoiding the stereo correspondence problems. This obstacle detection method has been experienced and working in various types of environments: from 2 m to 6 m in an indoor environment and from 5 m to 100 m in an outdoor environment.

For ten years, unstructured road following has been the subject of many studies. Road following must support the automatic navigation, at reasonable speed, of mobile robots on irregular paths and roads, with heterogeneous surfaces and under variable lighting conditions. Civil and military applications of this technology include transportation, logistics, security and engineering. The definition of our lane following system requires an evaluation of the existing technologies. Although the various operational systems converge on a color perception and a region segmentation optimizing discrimination and stability respectively, the treatments and performances vary. In this paper, the robustness of four operational systems and two connected techniques are compared according to common evaluation criteria. We present the identification of typical situations that have been used to create the image data base, the experimental protocol designs to compare the performances of the techniques and the analysis of the experimental results.

Dynamic vision is a powerful and effective sensing modality for a broad variety of numerous industrial applications. Two of them will be introduced as examples. One, relates to a novel implementation of motion stereo allowing precise distance measurements by a moving robot in real time with an uncalibrated camera. The other, behavior-based navigation, allows a robot to accomplish a desired navigation task by activation of an appropriate sequence of system- immanent behavior patterns.

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 3D 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.

Ultrasonic non-destructive testing of carbon fiber composite (CFC) aircraft panels has, in the past, been a time-consuming and laborious process. Data acquisition (using C-scan techniques) takes of order 1 hour per m², and the decision as to whether the panel meets the testing standard (technically known as sentencing) is an unexciting and repetitive visual task for a human operator. This paper introduces a new system for automated sentencing of CFC panels of solid or matrix (honeycomb) construction. It begins with a brief description of a new parallel-scanning ultrasound rig which greatly reduces the time required for data acquisition. A detailed description is then given of the design and implementation of a computer vision system which processes the resulting ultrasound images.

Smooth surfaces are widely applied in modern technology. Therefore a large variety of approaches for the determination of microtopographic surface descriptors has been developed. But there is a lack in adapted techniques for in-process surface assessment and defect detection. This paper deals with the development of an in-process sensor based on fiber optics and multivariate statistical signal processing.

In this paper we present a new approach to reduce the computations involved in recognition applications. Fourier descriptors are treated as a occurrence of a complex random variable. Statistical function measures are then used to characterize the behavior of the complex variable. A study of pattern regeneration based on these statistical features was carried out. Some of these statistical measures were found to comprehend most of the object global features. Thus, they could be used for classification and recognition purposes.

The automated video surveillance market is booming with new players, new systems, new hardware and software, and an extended range of applications. This paper reviews available technology, and describes the features required for a good automated surveillance system. Both hardware and software are discussed. An overview of typical applications is also given. A shift towards PC-based hybrid systems, use of parallel processing, neural networks, and exploitation of modern telecomms are introduced, highlighting the evolution modern video surveillance systems.

An improved technique is described for segmentation of digitized grey scale image of fingerprints. Conventional approaches to the segmentation problem are apt to yield binarized images of acceptable visual clarity, but tend to emphasize noise, which may lead to inefficiency in data compression and storage. They may be computationally intensive, which has implications on time taken for the processing of current archives. The advantages of the technique described are high visual acceptability of the resulting binary images, and speed of execution using simple computing architectures. The method makes use of a matched spatial domain high pass filter, combined with a statistical operator that attenuates noise and effects due to paper grain. Examples are presented of the use of the technique with fingerprint images of widely varying quality, and it is shown that visually acceptable results are obtained even when the original images is heavily degraded. Results after compression are described showing merits of the statistical operator.

This paper presents a neural network approach to classify three-band RGB color images for automatic visual inspection of seed maize. A back-propagation neural network classifier was developed and tested. The effectiveness of the neural network classification was evaluated by comparing with two conventional statistical methods, minimum distance (MD) and maximum likelihood (ML) classifications. Experimental results showed that the BP neural network classifier, the MD classifier and ML classifier provided the overall accuracies of 93.4%, 91.5% and 96%, respectively. The neural network classifier showed a better performance than the MD and ML classifiers in classifying the shady boundaries or blurred edges of an interested class when the training samples were selected from the boundary areas of the class. This research indicated that neural networks are suitable for the pattern classification of RGB color images.

In this paper, we describe an application of neural network for the classification of natural materials textures. We developed this solution in the context of leather quality control. This leather is used in car sits manufacturing. The aim of this control is to make sure of the compatibility of every visual aspect in the whole car. This job is currently processed by human experts that cannot inspect every sits with the same attention. The automation of such a process is very complicated because it is necessary to build a model of human vision in order to take into account how a texture is interpreted as an aspect.

The rapidly-growing interest for neural networks has been emphasizing their capabilities for object classification. Nevertheless, few applications have been turning out to be really industrially significant. In this paper, we present an application of neural network to 2D object recognition. The RCE model is used to retrieve the reference of an aluminum profile among 1600 possible classes. Although the learning phase is time-consuming, the recognition is carried out on-line.

We describe here a knowledge-based system, NEXSYS (Nextwork EXtraction SYStem) which was designed for the recognition of communication networks in SPOT satellite images. NEXSYS is a frame-based system and uses a co-operative and distributed structure based on a blackboard architecture. Communication networks in SPOT images are composed of thin linear segments. Segments are extracted using mathematical morphology and a Hough transform. An intermediate image representation composed of geometric primitives is obtained. Then an expert module is able to process the segments at the symbolic level trying to recognize networks.

In this paper we describe the computer vision system TOROS developed at INETI to automate the volumetric evaluation procedure as the loads of logs enter the paper mill. The method consists of on-line analysis of the images of all the visible sides of the load of wood, assuming that the logs are always stacked the same way, namely across the width of the truck. The lateral images captured by the system contain information about the sectional area of the logs, and the rear image provides the data related to their length. Once a calibration procedure has been applied, the measurements of the gross volume presented to the vision system are evaluated, based on the automatic detection of the contours of regions of wood in all the images considered. The most important feature of this computer vision system is the possibility to provide a measure of the percentage of wood that is really present in the consignment, through the identification of all the spaces between the ends of logs in the lateral images.

The ever-increasing demand for quality goods call for more and more efficient inspection systems in order to reduce the amount of scrap. Especially for natural products like wooden boards it is necessary to have an early classification and fault detection in order to manufacture an optimal product. An optimal solution was found using the, at that time, newly developed SBIP/L (Single Board Image Processor for Line Cameras). The problem was to inspect the boards from all four sides with an image processing system and detect and classify the faults on the lateral surfaces. This year RMV E&S presented a new system based on the SBIP/L combined with a color line camera from DALSA thus significantly enlarging the range of its wood inspection systems. This system allows the detection of color faults on wood (e.g. blue stains). Of course, the color fault analysis can be used for the inspection of other continuous materials as well.

This paper describes the development of a vision system at the heart of a machine capable of manufacturing cork stoppers automatically. The vision system, which uses three monochrome CCD cameras and a solid-state laser, is capable of generating an optimized punching pattern for producing cork stoppers from strips of raw cork material introduced by an unskilled operator. The vision system allows objective quality control constraints to be applied and decreases the amount of waste produced when compared with the manual method. A fully functioning prototype of the machine has been built and successfully tested.

A potential on-line automatic inspection system for grading seed maizes using machine vision was proposed. A number of samples of the maize images can be acquired as the maize cobs passing through the inspection system. The samples represent the quality of inspected maizes at different layers of unloading maizes from a truck. Machine vision algorithms were developed to measure the amount of residues mixing up with maize cobs and the loss of kernels on cobs. The methodology will be presented and discussed. Two parameters, residue mixture ratio and kernel loss ratio are introduced as indicators for quantitative measurement of the amount of residues mixed with cobs and kernel lost on the cobs.

Fully standardized machine vision systems won't require task specific hard- or software development. This allows short project realization times at minimized cost. This paper describes two very different applications which were realized only by menu-guided configuration of the QueCheck standard machine vision system. The first is an in-line survey of oilpump castings necessary to protect the following working machine from being damaged by castings not according to the specified geometrical measures. The second application shows the replacement of time consuming manual particle size analysis of fertilizer pellets, by a continuous analysis with a vision system. At the same time the data of the vision system can be used to optimize particle size during production.

Two VIKON-3D optical inspection systems assure the correct assembly of piston rings and guard rings in a new Volkswagen piston/rod assembly line. Both systems use laser light stripe measurements to locate and identify the relevant parts with high accuracy. The piston ring assembly is checked dynamically in video real time using laser light stripe and parallel projection techniques. In addition structured light is used to verify the correct piston/rod assembly. Both inspection systems are fully integrated into the manufacturing line. All types of pistons assembled can be checked without any mechanical changes to the measurement setup.

In this paper a number of practical applications of one sensor type the `Seampilot OPS' made in the Netherlands will be presented and developments like adaptive fill-in properties, and multi-layer possibilities will be illustrated. The working principles and the communication of the system with various controllers has been described briefly. With the implementation of the optical sensor system in industrial daily practice, experience was made regarding the many industrial demands. For these demands new software developments had to be included in the total package. By means of some practical examples a number of such software developments will be explained. The Seampilot OPS system is sufficiently flexible to be used with many different controllers.

This paper describes a non-contact system for surface shape measurement of manufactured parts. The technique consists of a coherent optical system combined with powerful parallel processing hardware, thus allowing rapid inspection rates to be realized. A `structured lighting' system is employed involving the projection of a multi-stripe fringe pattern onto the object surface thus enabling 3D data to be obtained. Phase measuring techniques are applied in order to increase accuracy and resolution. A comparison is made between two different phase measuring methodologies. These techniques require computationally intensive algorithms and processing of large amounts of image data. Image processing hardware should be fast enough to achieve results within a reasonable time scale. A parallel image processing system has been designed for such applications based on the Texas Instruments digital signal processor type TMS320C40. The system is based on MIMD message passing architecture and uses an EISA bus for the host interface. The issues of speed of data acquisition and processing, reliability and robustness of the technique and accuracy are discussed.

This paper presents a developed system for digitizing 3D objects in the footwear industry (e.g. mould, soles, heels) and their introduction in a CAD system for further manipulation and production of rapid prototypes. The system is based on the acquisition of the sequence of images of the projection of a laser line onto the 3D object when this is moving in front of the laser beam and the camera. This beam projection lights a 3D curve on the surface of the object, whose image is processed in order to obtain the 3D coordinates of every point of mentioned curve according to a previous calibration of the system. These coordinates of points in all the curves are analyzed and combined in order to make up a 3D wire-frame model of the object, which is introduced in a CAD station for further design and connection to the machinery for rapid prototyping.

This paper describes research into 3D line-scan systems for applications covering inspection and both autonomous and semi-autonomous control. Image generation using line-scan devices requires relative motion between camera and object. The type of motion investigated has concentrated on lateral displacement and the rotation of the line-scan device in relation to a static object or scene. Results are presented on the accuracy of the lateral displacement arrangement in resolving the position of points in the object space in all three co-ordinate axes.

Programmable optical grids can be projected directly on the surface of unpolished test-pieces to realize an absolute measurement method called `coded light approach' to obtain a full depth- image, requiring only a small number of projections. In combination with the phase shifting method subpixel accuracy is achieved. Geometry-checking and controlling the surface quality of objects with mirror-like surfaces is done with indirect illumination. Using a line-projector, stripes are projected to a diffusely reflecting surface, e.g. a wall. The test piece is positioned so that the lines are seen indirectly by the camera, reflected in the surface.

The Offshore Oil and Gas Industry in the North Sea has many requirements for 3D measurements in air and underwater. A market audit found that the use of film based photogrammetry was being rejected for many applications because the information was not available fast enough. A development project was set up to replace the photographic cameras with a choice of video of high resolution digital electronic cameras, and the analysis system with a personal computer based image processing system. This product has been in operation with Remotely Controlled Underwater Vehicles since September 1992. The paper deals with the ongoing development of the system, including the automation of the measurement process. It introduces the application of the system as a closed-loop control system for underwater manipulators.

Photogrammetry is well known in geodetic surveys as aerial photogrammetry or close range applications as architectural photogrammetry. The photogrammetric methods and algorithms combined with digital cameras and digital image processing methods are now introduced for industrial applications as automation and quality control. The presented paper will describe the photogrammetric and digital image processing algorithms and the calibration methods. These algorithms and methods were demonstrated with application examples. These applications are a digital photogrammetric workstation as a mobil multi purpose 3D measuring tool and a tube measuring system as an example for a single purpose tool.

A photogrammetric station is a configuration of several video cameras which are used for automated 3D measuring of industrial objects. All measurements are performed from the image observations only, and the targets are natural features or projected light patterns. The first `on-line' applications are for car body orientation within a robotic sealing cell and for 3D verification of ship screws within a robotic machining step. The `off-line' applications are for object reconstructions for reverse engineering purposes.

We will present a new system for knowledge-based image analysis which exploits the benefits of object oriented programming. The knowledge base is built using a formal language for semantic nets, which are already successfully applied in an industrial project. The knowledge base is compiled into an image analysis program. Time consuming search in the knowledge base can thus be avoided. The system provides a general interface to pattern analysis techniques which are included in the generated program as specified by the user. The approach combines the advantages of a compiled program for a special purpose with the flexibility of a general knowledge base tool. The resulting program is used for the reconstruction of the 3D shape of industrial objects using stereo techniques. Images are taken from several viewpoints. Models of 3D objects are then created by an integration of the segmentation data.

MOVE is one of the ESPRIT III OMI feasibility studies projects is Vision/Robotics for industrial applications. The objective of MOVE is to study a modular environment specifically designed for the field of Computer Vision, which permits the integration of heterogeneous processors and of specific software. This environment is open -- it allows integration of existing and future processors, actuators and visual sensors -- and it facilitates communication with the external world. It is used directly in the design of a set of industrial vision applications which correspond to real requirements and therefore it facilitates product development and make it cost effective. There are many areas of industry and commerce where application of machine vision would make an important impact on productivity and product quality, but the progress has been slow because of the lack of a suitable high-performance low-cost hardware. This paper gives some elements on how MOVE intends to provide the needed solution.

Using an incoherent optical correlator and a liquid crystal light valve, some binary image morphological transformations such as dilation, erosion and edge detection are implemented all optically. The experimental results are given.

In this paper, we describe a hierarchical framework for the computation of displacement fields from a sequence of images acquired by a moving camera and containing several moving objects. We also describe a matching algorithm which is consistent with the proposed framework and we present some experimental results demonstrating the robustness of our approach.