There’s a three-story-high Wall of Fame at Cognex that tracks the company’s successes over 23 years of risk-taking. That lobby wall includes framed patents, article reprints, and champagne bottles. Cognex employees recognize every major accomplishment by popping a champagne bottle and having each project contributor sign the bottle.
That willingness to take risks has resulted in the creation of four major cutting-edge products, and helped usher in related breakthroughs in the machine-vision field:
DataMan: This first Cognex product proved to the doubting world that a machine could quickly read numbers, letters, and symbols printed, etched, stamped, or embossed on product surfaces. It changed manufacturing methods—products could now be manufactured based on identifying individual units.
Cognex 2000: A watershed product, it replaced a set of machine-vision boards with just one board, thus greatly reducing manufacturing costs for Cognex. The company used grayscale normalized correlation, a pattern-recognition algorithm called Search. It recognized patterns far faster than anyone thought possible. As a result, it changed the way objects were located in semiconductor manufacturing, electronic assembly, etc.
PatMax: Here, Cognex took advantage of Intel’s MMX instruction set and turned it into a software tool for machine vision with much higher vision accuracy. Far easier to train than earlier methods, it handles pattern and size variations previously deemed impossible. It uses geometric pattern matching to identify objects when they vary in angle or size, when appearance is degraded, or when partially hidden from view.
In-Sight: This tool, the company’s fastest growing product, lets engineers install a machine-vision system without getting involved in complex programming. It uses spreadsheet theories and a vision sensor that integrates all of the machine-vision components into a single package.
The "triple whammy"
When Cognex was created, the founders initially faced three challenges, or what Bill Silver, senior vice president of research and development and chief technology officer, calls the "triple whammy." Machine-vision products needed to be good at understanding what they see almost all the time, they needed to be fast, and they needed to be cheap.
"We had to figure out how to do it. There was no textbook on it. We had to invent it," says Silver. "The hardest thing with machine vision is to get the right answer." Machines are not good at judgment, humans are, but machines are good at calculating and measuring. Machine vision though, includes both.
The biggest initial design challenge for Silver’s team was speed. Machine-vision tools could not slow down a manufacturing line. Yet, with so many numbers to analyze, it took a long time. So Cognex developed special programming tricks. Software algorithms were matched to the then-popular DEC PDP-11 minicomputer from Digital Equipment Corp., which interpreted the information from the cameras.
Cognex engineers had been writing software code in assembly language, but it wasn’t fast enough. To get the computation speed needed, the engineers had to get down to the microcode level, says Silver. This was proprietary to DEC, though. Cognex worked with its DEC salesman, Justin Testa (now senior vice president of marketing at Cognex), to persuade DEC on Cognex’s need to write in microcode. Then, using DEC’s documentation, Cognex designed and manufactured the writable control store hardware and created the needed microcode.
Cognex thus met its goal in 1982 with its first product: DataMan, machine vision that was reliable, fast, and economically justifiable.
Single board brings major changes
The next watershed product was Cognex 2000, introduced in 1986. Cognex 2000 was the first machine-vision system built on a single pc board instead of a set of circuit boards. It used a patented circuit made with TTL (transistor transistor logic) medium-scale integration, the only alternative for small companies in the mid-1980s. It provided an inexpensive implementation of crucial image analysis operations, most significantly the ability to compute correlation without multipliers.
Cognex 2000 also used a "new" pattern-recognition algorithm based on grayscale normalized correlation (NC) named Search. NC provides a numerical measure of the similarity between a pre-trained object and a portion of the image that might contain the object. Search was able to use NC far faster than anyone thought possible at the time. The result was a more accurate, easier to train system, with far fewer errors than earlier machine-vision tools.
"Before, machines would examine every possible position in an image and compute a measure of similarity, then declare the position that is most similar the winner," says Silver. "The problem with this is so many possible positions existed and thus required so many mathematical operations. It would take hours." But Cognex took a different approach. "We realized you don’t have to look everywhere."
Silver compares the Search system to looking for a set of car keys lost in a kitchen. "You don’t grab a magnifying glass and pour over every inch of counter space," he says. Instead, you first scan the room for things that resemble car keys. When you find a slight similarity, you look closer. "You search the space more intelligently, so it can be done very quickly, but also so that there is no chance that you will miss it," he adds. "The key was planning how to search the space."
"The world knew about it (normalized correlation), but everyone thought it would take hours to run the algorithm when an answer was needed in a tenth of a second. What we invented was a hardware and software system capable of doing it at the speed and cost industry needed. This completely stunned people. People thought it couldn’t be done. We changed the way an object is located in semiconductor manufacturing, electronic assembly, and other significant areas of manufacturing. To this day, that change is a crucial part of the way these devices are manufactured," says Silver.
For the first time, normalized correlation was practical. The NC formula makes the measure of similarity independent of the overall brightness and contrast, and that makes it immune to significant changes in lighting and object reflectivity. It’s what made Search so greatly accepted by end users.
Cognex 2000 also represented a new business approach for the company. The first Cognex products were sold to end users, but this one went to original equipment manufacturers (OEMs). Cognex sold them a toolkit that includes the NC software plus a software library that ran on the Cognex hardware platform. This let others with specialized knowledge and technical expertise design the Cognex machine-vision tools into their products.
At the same time, Cognex marketing focused on industries where machine vision is vital to maintaining a competitive edge, such as semiconductor manufacturing. That meant selling to Japan instead of buying from them. "This was mostly Shillman’s doing (Dr. Robert Shillman, now president, CEO, and chairman). He had the courage and smarts to succeed in Japan when Americans were not selling to Japan," says Silver. Many companies, Silver explained, hesitated to sell to Japan for fear the products would be copied. Cognex avoided that by developing "new stuff faster then they could copy old stuff," he says.
The total onslaught paid off. "This business strategy of a toolkit for OEMs to solve problems and a worldwide sales force was unique in the industry then, and every bit as responsible as the technology in making us a dominant player in the world. It really established Cognex and made us what we are today," says Silver.
Grabbing advances elsewhere
Machine-vision technology has always relied on the technological state of standard computer components developed for high-volume applications unrelated to vision. That meant using minicomputers in the early 1980s, switching to microprocessors in the mid-1980s, and employing ASICs starting around 1990. By the early 1990s, the high-volume applications had switched to desktop PCs.
"But PCs at that time were lousy at image analysis," says Silver. That changed when Intel introduced its MMX instruction set in 1997. "Now desktop PCs became fabulous at image analysis. MMX was intended for graphics, but we turned it into something useful for machine vision. That was a key technological change," says Silver.
By this time, normalized correlation was 10 years old, and Cognex wanted to address its limitations. For example, existing NC systems couldn’t handle variations in size or angle. "MMX enabled new inventions," says Silver. The result was PatMax, a software tool with greater accuracy and ease of use that recognizes changes in an object’s size, angle, or appearance.
"We like it a lot, " says software solutions developer Adil Shafi, president of SHAFI Inc. (Brighton, Mich.). "We needed a machine-vision tool to find a hole or notch on a product and guide a robot to that spot. PatMax is an extremely good pattern-finding tool. It can find an object in an image despite variations in color and contrast, orientation changes, and focus changes."
SHAFI serves Daimler Chrysler, Ford Motor Co., and other major automotive/plastics companies/suppliers. SHAFI software solutions support 11 major robotic platforms with Cognex vision products, including ABB, Kawasaki, Kuka, Nachi, Panasonic, and Staubli.
The key engineering challenge in creating PatMax was finding a way to turn brightness values into a shape-based description of an object independent of those brightness values. To accomplish this, Silver created calculations to do a search in tens of milliseconds to get a rough idea of where an object is likely to be within an image. Then Cognex engineers used assembly-language programming with MMX to make those algorithms operate at a speed and price point that industry would accept. Finally, Silver invented a method to refine the rough search results to get a highly accurate position of where the object is located within the image.
The result was a new way to conduct industrial pattern recognition: geometric pattern matching. It describes an object not by brightness value, but by computing measures of similarity in an object’s shape. This could not be done before.
"The first industry reaction was, ‘you can’t make anything that accurate,’" says Silver. Customers asked, "Can it really work?" But time, says Silver, has proven its benefit and value.
"In the past, present, and future, our added value is figuring out things no one has figured out. It really does depend on technical uniqueness," observes Silver.