The application of image processing technology in production situations is now practicable. This is due to the availability of low cost computing and sensing devices with which to exploit the considerable body of techniques which have been demonstrated. The most obvious and probably profitable use of such technology is in providing reliable inspection and visual monitoring devices where existing manual methods are unreliable and expensive. The cost-of-quality in manufacturing industry is UK alone runs into hundreds of millions of pounds and novel methods are required to augment existing methods of quality control in order to reduce this cost. Image processing techniques offer new tools for use in quality assurance with properties not available with more conventional techniques. These include tolerance of misalignment and no mechanical contact with objects under inspection, and an ability to inspect several parts per second making economically possible 100% inspection of parts. This is significant if such parts are to be assembled automatically.
Image processing in a strict sense is the transformation of one image into another (better) one. We will include the automatic image analysis and classification or pattern recognition, just like computer graphics (the input is not an image but other data). Image processing in a strict sense is used a.o. for image restoration (compensation for deformation, blurring, degradation), enhancement, addition of pseudo colour, for interactive processing. Computer graphics may expand the ability of humans e.g. in decision making and designing. A human analyses the images in all these cases. Automatic analysis and classification makes is possible to analyse and classify automatically. Some illustrative fields of the application of image processing in industry are given. They concern: materials science for R & D and production; inspection of moving surfaces; inspection of internal structures including radiography, ultrasonics and NMR; inspection and handling of objects; measuring from a distance; administration and design. A few remarks concerning the relation with non-industrial applications are made. Finally some expectations for the future are discussed. The difficulties and bottlenecks of image processing are mentioned. Image processing is not always an ideal solution. One has to try to avoid image processing by a redesign of the problem or the system (for example when the objects remain ordered it is not necessary to sort and orient them). But when this is not possible image processing systems may be very helpful. Technological advances are to be expected in the field of hardware, the architecture of image processing systems, and in the field of the theory of images.
The information contained in an image may be represented, analysed and processed in a wide variety of ways. In particular, the image may be transformed mathematically so as to re-structure the image information into a form better suited to immediate needs. The paper discusses three such transforms (Fourier, Hotelling, Walsh) from an intuitive standpoint. Useful information in an image is almost invariably corrupted by unwanted noise. When this noise is non-deterministic, a special conceptual framework is needed for its characterisation and subsequent analysis and processing. This framework is briefly discussed.
An outline of the principles of digital image processing is presented. The considerations necessary in on-line, real time processing of images for industrial applications are then examined. Image restoration, characterisation, the solution and efficient encoding of image features, and the effects of processing on a grid of cells are discussed.
XPLORE is an exploratory image processing system, namely a collection of computational tools which have been found useful in the design of image processing operations. The system is one of (mainly) ANSI FORTRAN subroutines to accept, process and display images on the PDP-11/40, a general purpose mini-computer. XPLORE can be used as it stands to carry out various display and transform operations, and extra facilities can be readily inserted by any user willing to learn about a very small number of utility subroutines. The sub-routines of the system can be used independently if the user wishes. The facilities available include window selection, histogramming, image transformations, creation of marginal totals and manipulation of image information expressed in curvilinear form. The system is portable, convenient to use and as fast as can reasonably be expected of one which relies on disc storage and calculation at mini-computer speeds.
Image analysis is a problem in data reduction. The only practical approach is to adopt a strategy where the reduction is effected in several stages. Structural analysis is probably the most useful method although difficult to implement properly. The crucial phase in which objects are delineated is the most problematic. Hardware to perform automatic image analysis can utilise structural analysis but to avoid undesirable restrictions a completely programmable system is required. The MAGISCAN is a high performance software based image analysis system for which high level problem solving languages have been developed.
Automatic or interactive image processing with the aim to detect, localize and classify objects in images is a very important task for automation and inspection in biomedical, industrial, remote sensing and military applications. One can distinguish several levels of complexity: The detection or tracking of objects with high contrast, fixed shape, variable shape after selection by an operator, or objects in general, the classification of objects into categories, and the separation of objects from an image. This paper describes some of the techniques developed for general and special applications and demonstrates some results.
The configuration of an industrial on-line inspection system is highly dependent on specific specifications imposed on its performance. Frequently, one of the more important specifications is speed, and the division between software and hardware is often based largely on this consideration. In this paper, various types of pattern classification software are discussed using specific examples from industrial inspection problems. Software is classified as performing a supervisory role or an execution role for data processing. Techniques for achieving the required performance in these various roles are indicated.
Conventional light sources in combination with mechanical scanning optical systems are severely limited in performance by significant aperture required to concentrate useful amounts of light into the scanning spot. The use of the laser source raises the figure of merit by several orders of magnitude which can be distributed between the various parameters of spot size, width covered and scan speed which multiply up to give very large bandwidth systems. The design of systems is described avoiding optical elements beyond the scanning polygon in order to obtain considerable flexibility in choice of scanned width and resolution from the one basic scanner design. The consequences of this are discussed along with the options on signal processing which include simple a.c. equalising of signal amplitude across the scan, to full digitally stored receiver/material reference profiles for long term d.c. stability required for example for on-line optical density monitoring. Multi channel receiver techniques for acquiring optical information for defect recognition systems are discussed and solutions to format accuracy in high resolution scanners.
The growing interest of image processing techniques in on-line industrial applications increases the demand for fast image digitalization units that offer sufficient quality at a low price. In this paper the use of television scanners is described as a peripheral in-put device for digital computers. Especially the design of an appropriate TV-sensor and scanner is emphasized by commenting the properties, the possibilities and the restrictions from users point of view. Several suggestions are proposed to solve typical designproblems like sample grid, data rate, shading and grey-value calibration.
This paper discusses methods of solving on-line inspection problems which have been developed and applied in two particular areas of industry. The cases studied are the automatic inspection of tinplate strip for visible defects in the presence of regular coding patterns on the surface, and the automatic checking of packaged products for dimensions and integrity of labelling. Both systems comprise electro-optical image scanning equipment and real-time, computer-based signal processing systems. The design philosophy of the image processing systems was in each case influenced heavily by the different functions to be performed, the different characteristics of the optical scanning heads, the presentation and condition of the products to be inspected, and the requirements and normal working practices of factory floor personnel. The designs were also constrained by processing speed and cost considerations.
A brief descriptive presentation is given of the various projects completed since work was initiated in sensory feedback some seven years ago. One of the first experimental intelligent robot systems to be completed in the United Kingdom is described together with subsequent developments. The underlying principles guiding the research philosophy is computational simplicity coupled with speed of data processing. Microcomputers and solid-state imaging arrays provide the essential ingredients for recent work.