Segmentation of an image often suffers from the lack of information contained in every pixel. The multisensorial camera produces for every pixel not a scalar attribute, but a complete feature vector with many, preferably uncorrelated components. We show how a 'color and height' line scan camera can be designed which generates a feature vector (intensity, hue, saturation, height) for every pixel at resolutions of typically 2048 pixels along the line of scan and with scanning frequencies up to several kHz. The processing of such a vectorial image starts with a LUT- based, trainable pixel classifier who transforms the vectorial image into a stack of binary class label images. This significant data reduction results in only little information loss and leads to further processing based on well-established binary image processing techniques.

An optical configuration for the detection of faults was developed and tested. The optical fourier transformation is the basic working principle of the system. When good fabric passes in front of the optical system the Fourier image, captured by the camera, shows well defined spots corresponding to the spatial frequencies of the tissue. If a defect occurs during production on the loom, the pattern changes significantly and a defect is easily detected in real time. A very simple electronic image processing based on thresholding and binary histograms allows to obtain very encouraging performance for its applicability to the looms. A compact device has been realized and tested in real working conditions on the loom.

One problem in the quality control of textiles is the measure of the pilling degree in worn fabrics. Wear and tear make some fibers separate from the woven yarns and get entangled in pills distributed on the web. Pilling degree is often evaluated by experts from a visual comparison with standard images. In this work, we use some techniques of digital image processing to evaluate the pilling degree. From the analysis of a set of standard images we establish and develop a sequential method for an objective measurement. The method involves operations in the frequency domain as well as in the spatial domain. In the final processed images we segment pills from the background fabric and we measure the total area of pilling for each image. We have verified that there is a logarithmic relationship between the total pilling area and the degree of pilling.

With the PVD (physical-vapor-deposition) technology various metallic and ceramic coatings, for example coatings like TiN and CrN, can be used on different materials as wear- resistant coatings. To verify the quality of the coating especially on components with a complex shape the nondestructive testing of the coating properties is very important. Established nondestructive testing techniques like ultrasonic and eddy current techniques often fail for thin coatings because of the large wavelength. In this article the possibilities and limits of the laser supported thermography for the characterization of such coatings are presented and discussed. Beside the thermal wave technique using a periodic laser heating and an infrared detection from the front side of the component the transient behavior of the temperature in the case of a step like heating is analyzed. By a systematic variation of the measurement parameters and comparison with calculations about the heat flow in coating systems the thermal properties near the surface of the component are analyzed. The results of the laser supported thermography are compared with the results of a metallography of the coatings.

Time-of-flight diffraction (TOFD) is a relatively new method of ultrasonic inspection and is well suited to semi- automation using methods such as robotic scanning, computer conditioned data acquisition and signal and image enhancement. However very little work has been documented on the full computer understanding of such scans. Instead, most work has been directed at aiding the manual interpretation process to determine defect characteristics. This paper describes the application of image processing and neural networks (ANNs) to the task of completely automating the decision making process involved in the interpretation of TOFD scans. Local area analysis is used to derive a feature vector which contains 2D information on defect/component and non-defect areas. These vectors are then classified using an ANN trained with the backpropagation algorithm. The labelled image is then further segmented using binary shape analysis to discriminate between component echoes, or defect signals. Time-of-flight correction techniques may be then used in order to determine the location of defects within a scanned weld.

This paper addresses the development and application of the next generation of JL Automation's industrial machine vision systems. An application which took advantage of new technical developments is discussed. Advances in system development are also discussed. The customer application was required to check for the presence and correct location of approximately 600 small holes in a plastic panel with a very complex surface. The solution involved four vision systems working together in a master/slave configuration connected via a serial link. It is worth noting that without the inclusion of machine vision in this process the component would not have been cost effective to manufacture.

Moire methods are optical methods that are based on the effect of superposition of grating lines and have been widely used in the context of industrial applications for shape analysis, for non-contact measurements, and for quality control of industrial components. In applications the following computations: image filtering, fringe skeletonizing and fringe numbering have to be performed for each test object, before comparison between the numerically reconstructed test object shape and its CAD model. In order to reduce the computing time required by the preceding computations, the inverse moire technique has been introduced by Harthong. Instead of using a grating made of parallel straight lines, the inverse moire technique uses a pre-computed specific gratin, that is formed of curved lines such that the moire pattern is composed of parallel straight fringes if the test object shape is conformed to its CAD model. Defects are then characterized by a deformation and a curvature of these parallel fringes. In this paper, we present examples showing that standard fringe extraction by automatic thresholding is not that easy. To overcome this difficulty, we propose a four stage process algorithmical approach that allows fringe detection in inverse moire images with high sensitivity and specificity. First we used the well-known image processing technique called unsharp masking, to enhance moire image and to emphasize low contrasted fringes. The second step is to extract bright fringes by image segmentation and constrained contour modeling. After detection of these bright fringes inside the zone of interest of the moire image, we get the thick skeleton of adjacent background and of dark fringes. The third step is to skeletonize this thick skeleton of adjacent background and of dark fringes, using morphological thinning of well-composed sets, that assures that each fringe skeleton will be one pixel thick, at the difference of standard thinning techniques. The fourth step is to apply a graph technique to isolate the individual dark fringes. When all these four steps have been followed, one is left with a binary image showing the dark fringe pattern skeleton. The experimental results that have been obtained have shown the robustness of this algorithmical approach, for the analysis of noisy inverse moire images.

Car identification is often necessary in traffic control, transport planning, origin-destination inquiries and so on. It can be reached by reading the license plate contents. Segmenting the plate is a critical preprocessing step for this. We present a multi-resolution plate segmentation algorithm, looking for constrained texture regions, and giving very high confidence to the final plate candidate, with no reading attempt. As real-time or near real-time execution are required in such traffic-based applications, a pyramidal approach is often necessary to reduce the amount of raw input data. Since vertical edges are critical in the plate detection, we created a new construction method emphasizing edges on the top level of the pyramid. The plate candidates are searched in between the textured regions on the top level of the pyramid and they are accepted or rejected, in function of various geometric criteria and some constrained texture regions features. Experiments, led on 405 outdoor scene images of static as well as moving cars under uncontrolled illumination, showed high accuracy and high scores in detecting the plate. Moreover our algorithm is dealing very well with cases with many character strings present in the image, and not only the one corresponding to the plate.

The identification of impurities in seed lots is required before their commercialization. Artificial vision can therefore be used for this goal. The aim of the present study was to identify rumex and wild oat in lots of lucerne seeds by color image analysis. A set of 58 morphometrical and textural parameters were assessed from the images of seeds. Among them, 3 relevant parameters were selected by stepwise discriminant analysis. The main purpose of this paper was to investigate the application of condensed nearest neighbor (CNN) rule for the recognition of seeds. The results were compared with those obtained by the well- known k-nearest neighbor rule, which requires a lot of computer time. CNN was found to substantially reduce the size of the training set. It was surprising that with only 9 observations selected by the CNN from the sample collection, it was possible to correctly classify all the 1194 observations of the training set. It also appeared that the seize of the consistent subset increased in relation to the number of required k-neighbors. CNN presented good classification performances despite the heterogeneity of wild oat seeds. Moreover, the algorithm converged fairly quickly and the number of iterations did not exceed 4.

In this paper we outline an approach to a fast low-cost system for the inspection and evaluation of surface-mount assemblies and solder joints, giving particular focus to the need for robustness in such a system. We describe the software modules involved in extracting features, and shape analysis for the purpose of inspection. A common problem during the early stages of image processing is the loss of information about the edges of shape because of noise, weak contrast, or occlusion, which often increases the amount of incorrect interpretations of a given image when presented to higher levels of processing. An algorithm is described which significantly reduces the effect of such problems before information is passed to later stages of the vision process. The algorithm, which uses information about contour to correct dropouts is generic, employing no application specific information and is therefore applicable to vision systems in general.

This paper deals with the problem of segmenting a 3D scene obtained by range imaging. It assumes scenes of arbitrary complexity in which the objects to be recognized are newly added or removed and investigates how the methods of change detection and image difference used in classical image processing can be used in range imaging. In a first step, we consider the case of ideal range images and conduct an analysis of segmentation by range image difference that shows the direct applicability of this principle. In a second step, we consider the case of the wide class of range sensors that suffer from shadowing effects which leads to missing data in the range image. An interpretation of this ambiguity in difference calculation and means to remove it will be given. Additional rules for the practical segmentation of 3D scenes by range image change detection are described. The presented methods lead to the possibility to segment a scene by isolating newly added or removed objects. They are tested using range images from two distinct range imagers of the light stripping type. Results indicate the success of this approach and the practical possibility to use it in the frame of an assembly task.

One of the tasks of a navigation autonoumous vision system is to determine his motion from the environment in a sequence of images. This information can be obtained from the focus of expansion (FOE) or the focus of contraction (FOC) due to the translational egomotion. We propose a simple method which allows to extract the FOE (or FOC) from time varying images. This algorithm uses neither the computation of the optical flow, nor image features tracking. It is based on a simplified Hough transform processing which is applied to the horizontal and vertical edges of the epipolar image captured by a translating camera, the optical axis of which is near from the motion direction. The epipolar image is obtained by a recursive operation between two images of a no dense sequence of a structuring or weakly structuring scene. This image stores the trajectory of the horizontal and vertical segments of the scene objects, because these edges have the particularity to have a size which linearly decreases until they are going out near the FOE. Then, by a simple analysis of the HOugh space in the horizontal and vertical direction, we deduce the FOE location. The FOE (or FOC) coordinate correspond to the Hough histogram minima. This methods can be implemented in hardware by a parallel pipelined image processing system.

The robot application of arc welding relies upon the robots welding gun being kept in a known position relative to a workpieces joint to be welded. The use of sensor generated position and geometric information is critical if the robots nominal path is to be altered to compensate for the workpiece position and geometric variations. This paper describes a multistripe structured light sensor which establishes the saddle type joint formed by two small diameter intersecting tubes. The CCD based sensor derives geometric information which is transformed into an arc welding robots base coordinate system by using a calibration feature in the image. With a robot following a path based on the sensor generated positions each unique workpiece is successfully welded. The surface texture of the steel tubing and the curved surfaces involved mean that image enhancement and noise reduction methods are critical to the success of this method. The positional information extracted is accurate to +/- 0.1 mm, well within the welding process tolerance of approximately +/- 0.4 mm. Industrial collaboration has proven that the sensor can be integrated with an industrial robot and is capable of analyzing eight joints every 2 minutes from a PC host machine.

A pilot-scale instrumental inspection system for poultry carcasses was assembled and is being tested for its robustness and real time operation. Two different techniques were implemented for the separation of wholesome and unwholesome carcasses on-line in real time. The first technique utilizes the visible-near-infrared spectroscopy of the carcasses and the other technique is based on multiple spectral imaging of the carcasses. They system acquired spectral images of the whole carcass and visible-near- infrared reflectance of its breast on a moving shackle and processes these spectral data for classification in real time. This paper presents the set-up of the pilot-scale instrumental inspection system and describes briefly the hardware and software modules implemented. Some preliminary test results are also presented.

Optical scanning systems can be used for measuring the 3D position of objects. In the case of large objects large errors can occur depending from the area to be covered by the laser scanning system and from the accuracy obtainable on angular position of driven rotating mirrors. A numerical simulation has been performed to evaluate errors in a case of large objects. This is a typical situation that is encountered when a scanning laser system is applied to aid the fabrication process of composite materials for aircraft components. Criteria used to perform the numerical simulation are described and results are reported for a number of cases could be met in areal applications.

Optical 3D measurements can provide a fast and non-contact inspection of engineered surfaces. This paper presents the development of a system based on holographic methods and image processing to inspect and classify metallic surfaces for the evaluation of the quality of special machining processes. An ideal example for the investigation of the performance of the system is the surface of cylinder bores. A successful inspection of the complex microstructure of bores can be used to adapt this system to other machining processes.

The paper presents the application of image analysis and mathematical morphology for the investigation of fibrous structures. The directional analysis of fibers is widely applied in biology, geology, metallurgy etc. With respect to the purpose of the analysis the investigations were divided into two groups. Each fiber was distinguished and recognized individually or the surface areas containing fibers of the same orientation were segmented. Based on the methods of mathematical morphology, the algorithms of directional detection of fibrous structures were developed. The investigations were conducted utilizing the binary source images. These images were subjected to preliminary filtering and thinning. The directional erosion by a pair of points oriented in 6 or 9 directions was performed. The corresponding grey level was assigned to the result of each erosion. A grey image was obtained, encoded in such a way, that each grey level corresponded to a different structure orientation. The properties of the computer image and the structure of fibers did not allow to obtain fibers coded uniformly along their entire length. The coding of the particular fibers was unified to enable the description of each of them through one variable (a single grey level). Each fiber was a collection of pixels of various grey levels. The adjoining pixels of identical grey levels formed segments of various length. Distance functions for these segments were calculated. The global maximum of the distance function contained in the particular fibers corresponded to the central point of the longest segment. An algorithm assigned the grey level consistent with the grey level of this segment to the entire object. In order to connect again the fibers which might have been broken earlier the propagation method was used. The main advantage of the presented algorithm is the possibility of a simultaneous analysis of all fibers within the boundaries of the analyzed image. THe method has been used for analysis of various fibrous structures. The other group of applications of fiber analysis is connected with the detection of area in which the fibers are of the same orientation. The above algorithm also helped to solve this problem. The obtained results was used to calculate the mosaic image included areas of different grey levels. Each of these levels was analyzed separately. In the next step of the algorithm the watershed was used to separate objects connected by narrow strips, and next the opening by reconstruction in order to eliminate the negligible objects. The algorithm enable the segmentation of areas of the same direction of fibers. The minimal surface area was determined by the size of the opening by reconstruction. The algorithm is illustrated with some examples of metallurgical structures.

The goal of this work is to allow a precise measurement of bar turning pieces. The dimensions of the pieces under study were smaller than 50mm. We have developed an illumination system which is specifically designed for a telecentric objective. Furthermore a special positioning system, which allows an integration of the measurement in a fabrication process, was added to the system. Thanks to this arrangement, it is possible to follow and master the fabrication process with a relative tolerance of the dimension of about 0.07 percent for the process and about 0.014 percent for the final control.