Machine vision is proving ideal in helping humans perform tedious but crucial manufacturing tasks. That is why it is poised to grow significantly in the next few years. It is powered by artificial intelligence (AI) in both its deep learning and more traditional rule-based methods, and the question today is how to employ even more of it.
Many see software as the key to its current success and greater adoption. “[We are] an AI company that offers machine vision technology,” said David Michael, director of core vision tool development for Cognex Corp. It is a surprising statement about a company that has a strategic position in machine vision hardware, but it illustrates that from its roots in MIT’s AI lab, the company saw processing power as key.
Engineers can apply two flavors of complementary AI. One is the more traditional rule-based machine vision, where logic is programmed by an experienced engineer. The other is the hot emerging field of deep learning neural nets, which statistically create their own rules by training on a data set of tagged images. Academic research can employ tens of thousands of training images, but most industrial applications yield good algorithms with just a few hundred, Michael said.
No panacea, machine vision is better at some tasks than others.
Michael said there are four applications especially well suited for machine vision: vision guided automation and robotics; metrology; visual inspection; and reading various identification codes.
In vision-guided robotics, machine vision is used to identify an object and determine its location so that a robot can operate on it, usually in a pick-and-place operation. “This area is really growing fast,” he said.
Metrology is precise measurement of dimensions. Machine vision is growing in this field through increasing collection of scan data in 3D. Rule-based AI is best in this application because dimensional tolerancing is precise and well-defined.
Inspection, another fast-growing area, uses machine vision to determine surface defects, color, or presence/absence. It does not require strict measurements, like metrology, but compares what is being measured with an ideal object or condition, such as a clear surface compared with one that is scratched and pitted. Because the problem is less mathematically precise, deep learning neural net AI is ideal and creating new opportunities, Michael said.
Identification is another growth field in both manufacturing and logistics. To identify a part or a package, machine vision can read 1D or 2D bar codes, text using optical character recognition, and even color or shape. It uses a mix of rule-based and deep learning neural nets, depending on the means for identification.
Machine vision is certainly at least poised to penetrate the industrial market.
A good example is Marposs, a provider of a wide range of high-precision measurement, inspection and process control equipment.
Marposs executives see machine vision eventually improving productivity, quality and costs. While manual operations are flexible and adaptable, these pros are not paramount in high-volume productions. “The down sides of manual inspection are well known. Human operations cannot be faster than a machine vision system when it comes to long, repetitive tasks,” said Matteo Zoin, senior manager and head of new market development for the company.
“That’s why automatic visual inspection for surface quality assessment, final checks for assembly completeness and robotic guidance for part-handling are some of the main applications we see for machine vision in manufacturing,” he said. “Artificial intelligence techniques are widely investigated in machine vision applications. They are not yet mature for industrial applications. But it’s just a matter of time.
“From my perspective, AI is currently one of the best solutions to analyze the growing amount of data generated by modern industrial processes and achieve ambitious goals like predictive maintenance,” Zoin continued. “It will be a key enabling technology not just for image data, but [also] for the whole industry process.”
Another company viewing machine vision as important is Jenoptik Automotive North America, a provider of metrology, gaging, and inspection equipment. Its products range from air gages to systems using cameras and telecentric lenses dedicated to measuring shafts.
Machine vision is becoming another tool for the firm’s customers, said Darren Dawes, a metrology product specialist at Jenoptik. Why? Machine vision systems have advanced to the point where convincing customers on cost alone to replace human inspectors is increasingly easy.
One of the company’s most popular applications, using its IPS B10, is looking for porosities, inclusions and scratches in bores. “Machine vision today is much more reliable than a person with a flashlight staring at bores for eight hours—especially at the end of the shift,” he said.
Jenoptik has gotten good results with rule-based algorithms. “We are experimenting with deep learning software,” Dawes said. “Deep learning will be useful as a complement to rule-based methods, especially where you are looking for large errors, such as cracks and porosities in castings, where we do not know quite how to define those problems or what they will look like.”
With the right training set, a deep learning algorithm will flag “suspicious” as well as known errors, improving inspection robustness.
“[Deep learning] is going to open up a lot of other opportunities for parts where a customer cannot really define what the problem is, but they know that it is out of the ordinary and they want to be alerted to it.”
An excellent example of where rule-based algorithms shine is in Jenoptik’s IPS GAP device, designed to check that valve seats are properly pressed into aluminum heads.
Using a four-point check on an image and some mathematics, it can check against the largest allowable gap along the circumference of the ring and flag errors.
“The industry was really asking for this,” Dawes said.
Big growth expected
Looking at the broader market, AI is growing at a blistering pace. The market for AI in manufacturing is expected to grow from $1.03 billion last year to $17.22 billion by 2025, said Lisa Whalen, global VP for automotive & transportation at a research firm called MarketsandMarkets.
“Quality control and machine inspection are the major applications where machine vision in combination with AI is expected to grow at a significant rate,” she said. “Within the AI market, machine vision is expected to grow at the highest CAGR of 54.8 percent between 2018 and 2025.”
AI in combination with machine vision can be used for visual inspection in order to identify imperfection in products, she acknowledged.
“Machine vision tools, along with AI algorithms, can be used to inspect minute components and identify microscopic defects as they are more sensitive than human vision,” she said. “Visual inspection using machine learning reduces inspection time significantly by detecting a variety of defects as compared to conventional techniques.”
Automotive, energy and power, pharmaceutical, and semiconductor and electronics are some of the major industries adopting machine vision combined with AI for quality control and machine inspection, Whalen said.
She has thoughts about the future, as well. It lies in robotics.
“Industrial robots are already being used in manufacturing … to speed up [the]production process and increase productivity,” Whalen said. However, there is a need for collaborative robots that can work with humans and respond to their environment—and AI and machine vision has a role.
“AI is also being used to power robotic inspections and train robots or drones to inspect areas that are hazardous and difficult for humans to access,” she added.