This paper describes a method of active light ranging, which in principle is based on triangulation, but where the range information is deduced from at least two measurements of the luminance of the object. With this tech- nique it is possible in a simple manner to obtain range data for each pixel in a computer vision system. The sources of error inherent in the method are discussed and a number of examples of photometric range finders and 3-D vision systems are given.

Features &e extrzted from the angular specrwn of light scattered fnxn the sur(e of machined surfxes which are then used as inputs to a hrarcha1 neural net. The net is "U1nCd by aselected training set of feures from surfaces whose quality has already been independently establishet These samplessicrepeatedlypresented to the sensors and the network makes a deCISIOn about the surface roughness which is then compared to the rrect answa, and the TOf used to modify the conntion weights. Following this training pcziod, the net is be able to identify the quality of new surfes presented to IL Indeed, it is able to ck, so even in the presence of noise which is likely to be caused by poor illumination. Results, using this laser scattering technique, from a set of prepared sirfaces are dixussed with regard to fusion of different features in der to obtain an adequate measure of surfe roughness ung the harchkal newal nwor

The pixel-wise classification of CCD colour Images Into previously learned colour classes at video-rate is a demanding vision task, both with regard to the complicated cluster shapes encountered in natui- al scenes and to the required computing power for real-time operation. We discuss two classical solutions based on an algorithmic statistical classifier and on a Neural Network paradigm and propose an alternative simple and low-cost classifier based on approbriately trained look-up-tables. Two different learningrules for the supervised training of this LUT classifier are presented for the colour classification of both synthetic and natural blotechno1ojr scenes. The proposed LUT classifier shows all the positive features of a (simulated) 3-layer perceptron Neural Network, but performs 60.000 times faster with simple, commercially available components.

To make efficient practical use of the attractive properties of standard neural network models it seems reasonable to com bine a recognition network with appropiate conventional preprocessing. In this contribution we describe the current research at IPA concerning such a hybrid approach to acoustic pattern recognition. The inputs of the recognition network are feature vectors consisting of local frequency characteristics extracted from the Wigner representation of the patterns. Simulations show that the system is capable to recognize individually learned objects in a scene.

In the manufacturing of continuously formed products such as metal plates and plastic sheets, on-line monitoring of thickness is important for quality control. Most cases require noncontact measurements. The most widely used optical technique is "optical triangulation" in which a lateral effect position-sensing detector is used. However, the technique has a limited measurement range to resolution ratio making it not suitable for some on-line applications. In this paper we propose a new approach to measure thickness in which the position of a scanning laser beam is timed. It eliminates the restriction of measurement range by the dimension of the detector as in traditional optical triangulation. Also, because small area silicon photodiodes are used, signal-to-noise ratio is largely enhanced thus improving the resolution. The probe we built demonstrates a 47 % measurement range-to-standoff distance ratio and a 2.8x104 measurement range-to-resolution ratio. Also, its performance is not influenced by large changes in object reflectivity.

We developed pattern inspection techniques for simultaneously measuring both the width and the thickness of conductive patterns printed on ceramic green sheets. The measuring optics use a He-Ne laser scanner. Thickness is measured by beam splitting, and width, by light diffusion. The beam-splitting optics measure the height by determining the ratio of reflected to transmitted light intensity. By optimizing the arrangement of the optics, we attain a resolution of 5 im and a height measurement range of [70 pm. Pattern width measurement is based on the fact that a beam of light diffuses when it strikes the substrate, but does not diffuse when it strikes the conductive pattern. A spatial ifiter in the focal plane raises the S/N ratio to four regardless of the state of the conductor surface. We used these techniques to inspect the conductive patterns on green sheets. Our system inspects conductor patterns 3-dimensionally at 12 Mplxels per second with a 5 i.m resolution.

The inadequacy of the stuck-at fault model has been well aired and oced1 ,2All studies agree that this model does not reflect the physical failures of real devices,3 principally because such failures do not exhibit a 1:1 mapping onto the logic domain.2 ,' Circuit layouts which are based on stick diagrams do however reflect the physical domain in sufficient detail to enable both structural defects, together with shorts and opens in metallic and non-metallic domains, to be detected and located. The author has proposed the adoption of a novel method which processes information obtained from a scanfling laser beam reflected from a surface profile. Scanning may be of a raster nature over the surface, or follow a suitable path search along layouts. The latter search type has been simulated in PROLOG using breadth-first (BRFS), Euler (ES), and neighbour-first (NFS) searches. It is suggested that by creating and modifying an acquired-knowledge database (AKDB), according to defect occurrence, it is also possible to search those regions where defects may be present in order of decreasing probability. Thus a useful library of the distribution of defect density statistics would be created by virtue of this proposal. In this paper the experimental hardware results presented are based upon reflectance measurements obtained by raster scanning in three different optical modes. In addition results of search simulations and a discussion of the AKDB are included. Such a topological approach to the testing problem offers a test structure for exploitation which is technology independent, fast and is adaptable to parallel processing.

The increasing integration of semiconductor circuits requires absolutely defect-free masks. Before their exposure to an c-beam writer the unstructured masks (blanks) have to be inspected in a clean room environment for particles on surfaces, pin-holes in the chromium layer and inclusions or scratches in the glass. Based on a HeNe laser scanner and by comparing relative directions and intensities of the scattered light, the position, size and type of defects were automatically classified. The criteria for classification will be explained. The smallest detectable defect diameter was 0.3 p.m. For presentation under a TV-microscope each defect could be located within 100 m.

The use of CCD sensors in optical metrology requires synchronous sampling of the image with a good signalnoise performance. . A system has been developed to digitize optimally the signals from high-resolution CCD sensors. The data acquisition system is split into two parts -the first is a storage unit for the IBM PC/AT family of computers with a fast, digital, input-output interface, with 8-bit transmission speed DC-4OMHz and 16 bit operation DC-2OMHs. The digitization of the analogue signal is performed on separate units, up to 2m from the computer. Separating the analogue processing from the computer and using a separate power supply not only reduces the electrical noise from the digital electronics to a minimum but also allows greater flexibility in designing custom 'front ends' for a wide range of sensors. The storage card has two 1M byte banks of memory. These are normally used to provide double buffering of 1M pixel images, but can also be used to store 2M byte images without double buffering. Practical experience, using 8 and 10 bit video front ends, indicates that the geometrical resolution possible with modern CCD sensors is approaching 1/100 of the pixel period. The digital signal processing required for this performance does not depend on the CCD camera's PSF and it is insensitive to variations offocus and orientation.

Non contact video inspection systems measuring geometric elements of workpieces. These measurements should be done with high accuracy and speed. During the design period of the UMS432 of Wild Leitz the parameters that could reduce the expected measuring accuracy and allowed speed were investigated. In this paper especially the influence of the imaging part and the image analysis is analysed.

Surface parameterisation using phase measurement of interferometric fringe patterns is supported by a strong theoretical basis. However, little work has been published with regard to the practical implementation of the available methods as tools for industrial inspection. This paper attempts to justify the choice of Fourier Fringe Analysis for this purpose and addresses some aspects of its implementation.

This paper discusses the construction and capabilities of a time-of-flight laser rangefmder system for profiling measurements on hot surf,es (< 1400C), especially measurement of the thickness profile of the lining of a converter in iron works. Thebkgmtmdradiation is the main source ofthe noise in the receiverin this kind ofmeasurement. The inherentlylow signalto- noise ratio, typically 10 - 40,can be improved by the construction, adjustment ofthe parallel optics and optimization of the gain of the receiver avalanche photodiode. A stability of +1- 5 mm and a resolution of a few mm can be achieved in real measurement sitoations with a measurement time ofless than 1 second per poinL The non-linearity of the device at the near end ofthe measurementrange due to the paralleloptics and optical fibres is reduced to 5 mm by modifying the optics and using a mode mixer. The poor temperature stability, typically about -2 mm/'C, ofa device with different receiver channels for start and stop pulses can be improved by stabilizing the temperature ofthe electronics, orbeuer still, by using one electrical channel, in which case the temperature drift is almost zero in the temperature range -10C -+35C.

The feasibility of an opto-electronic inspection system for the sorting and grading of apples with respect to ripeness and quality is studied. This study comprises a detailed spectral analysis of the healthy skin as well as a variety of visual defects. Three different apple species were studied: Cox orange, Golden delicious and Golden rennet. Spectral wavebands of interest appear to be the chlorophyll peak around 570nm, the ripeness peak around 64Onrrt and the near infra-red beyond 75Onxn. Stem and calix of apples pose a separate problem because their spectral reflection cannot be distinguished from that of most visual defects. A solution for this problem is given.

An optical/digital approach to the classification of rough surfaces using Fourier spectrum sampling is described. The sampling of the 2-D Fourier spectrum is achieved with wedge ring detector which reduces an infinitely dimensioned spectrum image into a set of 64 measurements. To discriminate three metal milled samples in this reduced subspace we employ the Karhunen-Loève transformation. The classification procedure then selects automatically the best subspace from the K-L feature vectors.

A special application case of multiresolutlon image processing is pointed out. From the imager's original picture. a set of image copies is derived. Each of these copies is generated with different demagnification factors along both image coordi- nates. Simple compact object structures are demagnifled from picture to picture and appear in a specific reduced image as single pixels. For a real-time application, a tailored hardware solution was constructed, which involves a special filter design to minimize accumulation of anti-allasing errors. The application for rough defect classification during strip material fabrication is discussed. In this case, flaws of a given length/width ratio can easily be detected by simple, minimum templates.

Image processing has been used to analyze TV -holographic fringe patterns of deformations and vibrations. The processed pictures give improved detection and information for the operator of the interferometer. A lower detection limit of about ?./50 for deformation analysis has been achieved, while amplitudes down to V500 and phase values with an accuracy about 3° have been measured on vibrating objects. Examples of applications of the techniques are shown.