IEN: How do you define sensor vs machine vision system?
Dawson: I think of a "sensor" as something that translates energy patterns into data, like a CMOS sensor for a camera or a temperature sensor. A machine vision system includes one or more sensors, usually in cameras or part-in-place sensors. Some vendors call their machine vision systems "sensors," perhaps because the factory floor is comfortable with the term.
IEN: What issues confront machine vision? How can they be resolved?
Dawson: Machine vision issues can be divided into fundamental and ease-of-use. Fundamental issues include how to reliably find objects of interest in an image, how to separate these objects from their background, and how to make reliable and accurate measurements of the individual objects. The fundamental issues surrounding machine vision are complex, so the question is how to package what has been learned in laboratories and real-world environments during the past 30 years so that someone who is not an expert in machine vision can use it to solve a problem. Vision Appliances from ipd resolve this problem by embedding the knowledge needed to solve problems in an easy-to-use human interface. Unlike other machine vision systems, Vision Appliances do not require users to program, understand the details of how to find objects of interest, separate them, etc. In contrast to traditional vision systems, a Vision Appliance does one or a few tasks well and is not intended to be a general, and therefore complex, tool to develop and run a machine vision solution. (Vision Appliances like ipd''s iLabel, pictured, feature a graphical, easy-to-use interface so that users with little to no machine vision experience can get applications up and running within hours.)
IEN: Has the use of vision systems spread beyond traditional end-use markets, and if so, why?
Dawson: Actually, it is somewhat the reverse. The traditional markets have been mostly OEMs, who embed vision systems into their products, such as robot guidance, web inspection, and optical metrology equipment. The end user then has a vision system tuned to a specific need and might not even know there is a vision system in that equipment. As vision systems get smaller, faster, cheaper and -- most importantly -- easier to use, more and more end users are purchasing vision systems for specific tasks, yielding more reliable and repeatable results than human inspectors can provide, and enabling substantial manufacturing cost savings.
A second way to view this question is, will there be markets other than the "traditional" markets for machine vision? Traditional markets include electronic parts (wafers, connectors, PCBs, etc.) inspection, robot guidance, high-end optical metrology and OCR (optical character recognition). Applications in these markets generally have controlled imaging conditions and limited variation in the objects being imaged. As fundamental knowledge increases and machine vision systems can handle greater environmental and parts variation, machine vision is expanding beyond these traditional markets into areas like face recognition.
IEN: What innovations are in store for users in the areas of: color, intelligent sensing, and high-end edge detection?
Dawson: Color is an important feature of most products and can provide valuable information about the product (is a tomato ripe? rotten?). Human color vision is more robust than machine vision, but it is also more subjective. I expect that improvements in color cameras and multi-spectral imaging (more than 3 bands), as well as improvements in color image processing algorithms, will significantly increase the use of color machine vision systems.
I think of intelligent sensors as having some intelligence right on the sensor, but others call their combination of a sensor and a separate processor an "intelligent sensor." Putting the intelligence right on the sensor has been both disappointing and fascinating. It has been disappointing because processing capabilities and flexibility have been limited by what we can do and what we know. For example, attempts to build a "silicon retina" have been interesting but of little practical value for machine vision. One fascinating aspect of intelligent sensing is building a vision sensor that is designed for a specific task; for example, detecting movement to activate a video camera or open a door, or for collision detection. In this case, the sensor can be simple and cheap while doing a good job. One might argue that this doesn''t seem like a high level of "intelligence," but the point is that the sensor performs a required function at a very low price and with little effort.
Edge detection is a fundamental issue and one that is constantly being investigated. In optical metrology, for example, the limitations of resolution and repeatability are set by the physical spacing of the sensor elements (pixels) and the knowledge or assumptions we can make about the edge sensor structure. With little knowledge and few assumptions, a resolution of roughly 1/4 pixel is easy to attain. As we know more about the edge and the microstructure of the sensor, we can get better resolution, perhaps as small as 1/40 of a pixel under ideal conditions. To improve resolution, sensors are increasing the number of pixels, and pixel sensor structure is becoming more uniform (a higher "fill factor"). Algorithms that include additional knowledge of the edge shape, the length of the edge and so on can help boost resolution in less than ideal conditions.
IEN: Will wireless play in machine vision?
Dawson: Multiple machine vision systems -- one to inspect the product, one to look at the packaging, another to read the bar code, etc. -- are often used on assembly lines. These systems have to be integrated with each other, into the manufacturing cell and higher levels of control and, sometimes, into enterprise business networks. Serial and digital data lines and, now, Ethernet, are currently the preferred methods for these connections. Many manufacturing lines present a difficult environment for wireless, due to the quantity of metal that can block the signals and the quantity of electrical noise (think: relay closings and arc welders) that can corrupt the signals. However, due to the potential advantages of wireless on the plant floor -- including the ability to move equipment and systems around or add new equipment and systems to a network more easily and seamlessly -- manufacturers are showing more interest in wireless plant floor networks in manufacturing. I''m taking a wait-and-see approach regarding how this will play out.
IEN: How are software and equipment being improved?
Dawson: Machine vision systems are smaller, faster, smarter, and easier to use. The fundamental issues in machine vision continue to be researched, leading to software that is more flexible and precise.
IEN: Can sensing be part of a lean/flexible manufacturing solution? How about machine vision?
Dawson: Since machine vision systems are becoming smaller, faster, smarter and simpler to use, less development and implementation time is required. Instead of requiring months, or even years, of development by a vision expert, ipd''s iLabel Vision Appliance, for example, can be brought online in a matter of hours to solve specific machine vision problems, and it can be reprogrammed on the fly by a line operator, because the product automatically learns examples of good labels.