Vision for industry

What's new in the world of industrial vision, and what's halting new developments.

Although machine vision is still often considered a 'new' technology, it has been used in industry for more than 25 years. To put this into perspective, the Cognex Corporation has shipped over 400,000 systems worldwide since 1982, the year after it was founded.

Today, the machine vision industry divides broadly into 'application specific' products and 'general purpose' systems. The former tend to be high value devices aimed at spec-ific tasks, often in parti industries. For instance, the semiconductor industry relies on machine vision to monitor and control many operations in the manufacturing process, but you will never see these products at an exhibition devoted to vision such as IPOT/VTX in Birmingham or Vision in Stuttgart.

Surface Inspection, of Bristol, started over 20 years ago by doing just what its name suggests, but fairly soon specialised in inspecting nothing but ceramic tiles, and became a world leader.

In Israel, Orbotech dominates the world market for machines that can inspect bare printed circuit boards at very high speed. It has a less dominant position inspecting flat-panel glass, as used in displays and mobile phones.

In Germany, Parsytec has made an interesting transition from making high performance parallel processing computers, to producing machines which inspect rolled steel strip at very high speed, now complemented by ones which inspect paper at even higher speeds.

For traffic applications, the technology has become sufficiently reliable to operate in all kinds of weather conditions and is used for recognising number plates for the London congestion charge and the more complex German Toll Collect system, which must not only recognise number plates but check the size of the vehicle, as it applies only to relatively large trucks.

Smart camera

The 'general purpose' sector includes three generic types of system as classified by hardware, with a variety of software types not aligned to the hardware classes. In recent years the 'smart' or 'intelligent' camera, in which the processing power is built into the camera body, has become very popular. It has the advantage that once set up, it occupies very little space and does not need additional environmental protection for a display or processor.

Of course, if you need more than one camera at a particular place on the production line, the economics of using a smart camera at each point begin to be outweighed by the alternatives, both of which can usually cope with several cameras feeding into just one processor. The first alternative is the 'embedded' system - a dedicated system optimised for image processing and analysis, this contrasts with the third type, the PC-based system where an 'ordinary' personal computer is used as the processor, and can be used for other tasks including program development.

Software vision

This brings us to the software types available. The introduction of graphical user programming has opened up the application of vision to non-programming staff and requires very little training.

Rapid development software needs more training and allows constructs such as 'if, then' to be built into the program. At the more complex end, programming in a language such as 'C' will usually produce the most efficient result but does require programming skills.

What about the tools available to the programmer - at whatever level? There was an important advance in vision in the late 1980s and early 1990s, led by Cognex in 1986 when they developed a fast version of an algorithm that had been known for many years but had been too computationally expensive to implement in a production environment.

This was a technique known as 'normalised correlation', which could 'find' a known pattern in a scene without being influenced by the light level or degree of contrast present in the scene that could be very different to the 'golden' pattern. The technique also found the 'best fit' even if parts of the pattern were missing - an important feature when aligning successive layers of a semiconductor device, or placing components on a printed circuit board.

For a short time, Cognex had this as a unique capability, but advances in processor power and in competing algorithms eventually led to many suppliers offering comparable tools. One limitation of the technique was its inability to cope with changes in orientation or scale of the sought pattern, and more recently 'geometrical pattern matching' has been developed, again in different versions by many vendors, to overcome this limitation.

These are by no means the only advances in software tools, but they give the message that machine vision today is far more robust than the early systems which were prone to fail if ambient light levels changed or if there were slight changes in the appearance of the objects of interest.

Slow adoption

Machine vision has been relatively slow to adopt digital transmission of images from camera to processor, and analogue transmission is probably still the most prevalent, even though there are significant advantages in digital transmission. The latter was helped by the introduction of a standard about eight years ago called CameraLink allowing any compliant camera to feed any compliant framegrabber using a standard, relatively low cost, cable. Much more recently, a version of CameraLink has emerged allowing power to be supplied over the same cable, simplifying connectivity even further.

Some cameras use IEE1392 (Firewire) or even USB2 to convey the image direct to the processor memory, eliminating the need for a framegrabber but requiring more knowledge of how to control and exchange data with the camera.

Yet another alternative is GigEVision, a standardised way of using Gigabit Ethernet for image transmission, with the advantage that distances up to 100m are feasible and the Ethernet switches and other devices are in mass use, hence relatively low cost.

The still-evolving Gen<I>Cam standard makes it very easy to control any compliant camera as it can inform the processor automatically of its capabilities and control requirements.

In development

Cameras, especially digital cameras, are getting faster and faster and have more and more pixel sites on them. A faster version of CameraLink (already the fastest option) - or an alternative to it - will probably be developed to cope with the latest cameras. It is possible that it will be 'leapfrogged' in speed by the use of 10 Gigabit Ethernet for image transmission.

It is probable that the majority of advances in camera technology will be for CMOS cameras rather than the longer established CCD cameras. This is partly because there is room for improvement in CMOS which still, in very general terms, lags behind CCD in image quality, but also because CMOS lends itself to development by small organisations without their own 'silicon foundries' but who can easily do the design work themselves and contract one of a number of foundries to actually produce the sensors for them to their exclusive design.

Earlier this year, the founders of Belgian Fillfactory broke away from the company (owned by Cypress Semiconductor) and set up Cmosis to 'return to their roots' as independent developers of CMOS sensors, as they had done originally when spinning off from IMEC, the Inter-University Microelectronics Centre in Leuven.

Blackballed vision

At the recent annual Business Conference of the European Machine Vision Association, held in Berlin in April, Professor Robert Massen, who managed to be simultaneously a Professor at Constance Polytechnic and found and run a vision company, looked at the contribution that German academic universities had made to the machine vision industry, and his conclusion was remarkably little technically, though obviously many of the founders had been trained at such universities. (He didn't consider the non-PhD awarding technical universities.)

Indeed, he felt that the academic universities had 'blackballed' machine vision as a topic 'not acceptable for PhD studies'. In recent years, it has been clear that the UK's EPSRC (Engineering and Physical Sciences Research Council) rarely funds research in machine vision, though that may be true for many other disciplines too. Members of the UKIVA (UK Industrial Vision Association) comment informally that UK universities are not producing the kind of 'rounded' engineer that they would like to employ - vision demands a knowledge of optics and illumination as well as computer processing of images. The Association does offer free membership to academic organisations concerned with industrial applications of machine vision, and about ten such organisations are listed on

Vital vision

There are some industries where machine vision is a sine qua non - most notably in the production of semiconductors and printed circuit boards - but increasingly in safety-critical industries such as the pharmaceutical and medical device industries, the manufacture of glass for flat panel displays, and for any component concerned with safety in the automotive and aerospace industries.

There are some less obvious industries where vision is considered essential - fruit juice cartons all have a plastic lining to the cardboard outer, and for many years the plastic film has been inspected for pinholes before use. Why? Well, consider a pallet of such cartons with a leaky one on the top layer - it will ruin the whole pallet at a considerable loss to the supplier. The manufacture of cigarette packets may not seem an important use of vision, but in fact the cigarette companies are extremely concerned about counterfeiting and they demand that every pack they produce is printed to the highest standards, making it that much harder to duplicate satisfactorily.

An important cross-industry reason for using vision systems is for traceability, particularly using the two-dimensional matrix codes which can hold much more information than a conventional barcode, and which contain much redundancy so that even if some of the code is obscured, its information content can still be read in full. The advantage of detailed traceability in the automotive industry is that, if a component failure leads to a recall notice for replacement, only those cars that have a component from the suspect batch need be recalled, instead of all cars manufactured between certain dates.

Non visionaries

Members of the Association say that one of the factors holding back the adoption of machine vision in the UK is the lack of suitably experienced system integrators who can provide a 'turnkey' solution for manufacturers who do not wish to engineer a solution for themselves, even with the availability of 'point and click' or 'drag and drop' programming.

They contrast this with Germany, where there appears to be no shortage of system integrators familiar with vision, and where the market for vision systems appears to be several times that of the UK.

Recent articles

Info Message

Our sites use cookies to support some functionality, and to collect anonymous user data.

Learn more about IET cookies and how to control them