• By Tim Heston
• August 28, 2025

Consider all the activities in metal fabrication that require visual inspection and counting. A person inspects a blank for distortion and edge defects and, if needed, sends it to a deburring machine followed by a part leveler. A press brake operator inspects a formed enclosure for dimensional accuracy, looking out for excessive bulging at the weld notches in the bottom corners. A worker inspects a blasted metal part, then makes a judgement call before sending the piece to powder coating.

Computer vision has come a long way in recent years, especially at large companies with high production volumes. Visit the typical small or medium-sized metal fabricator, though, and you won’t see vision used everywhere, at least not yet.

This might be starting to change. To uncover the potential of AI and vision technology, The Fabricator recently spoke with Eric Sutton, president of Datavision, a growth consultancy, as well as Suresh Yalamanchili, founder and CEO of Amniscient, a firm that builds computer-vision AI models. Sutton, who also sells for Spira Systems, has experience in industrial automation, while Yalamanchili and his company have been working on implementing AI models that aim to help simplify and streamline the use of vision across industries, including manufacturing.

Their message: Implementing vision technology is becoming simpler, less costly, and more flexible. And from a fabricator’s perspective, new vision technology, rooted in AI and machine learning, could have a profound effect on part flow and QA—essentially, any application where workpieces are counted and inspected.

Collect, Train, and Deploy

The current conundrum in AI, at least when it comes to company-specific industrial applications, is that effective implementations need a large data set. Consider the typical custom or contract fabricator. The organization might make hundreds of different products at different times. They’re not producing millions of an identical product. So much for that large data set.

But as Yalamanchili described, AI technology is evolving quickly, and this includes the arena of vision and object recognition.

“Today, when building ML models, they use synthetic data to offset the human effort to try to collect enough data for that object to be accurate. That synthetic data is augmented with computer graphics. The problem is that, by using CG, the data is false. It’s not based on reality. They’re using it as a noise substitute. So when you implement this technology into the real world, it doesn’t work very well. At Amniscient, we’re not using CG. We’re using a proprietary form of synthetic data augmented from real-world pixels that the user is collecting. This is how we ensure high accuracy and reduce false detections.”

No longer do computer-vision AI models need a massive number of images of a specific object. “Traditionally in the market, when you’re building AI models, you need thousands of images per object. We’ve reduced that to fewer than 30 frames per object,” Yalamanchili said, adding that this soon could be reduced to just one object per frame.

Cost and development time have been additional stumbling blocks. As Yalamanchili explained, “The goal always has been, ‘Can you build AI models without any human involvement, and what would that AI model look like?’ That’s what we’re solving.”

Amniscient boils down its AI model building process (which it calls AmniSphere) into three steps: collect, train, and deploy. The idea is to deploy new computer vision systems quickly, be they around QA, inventory management, defect detection after specific processes, or overall throughput monitoring.

Vision in Process Manufacturing

For a perspective on how far vision systems in general have come, Sutton described the current state of meat processing.

“Here, you need to identify large cuts of meat, and if you make a cut in the wrong place, you have a multibillion-dollar mistake. If you make a one-vertebra cut change, the value of the meat decreases. Here, the vision system quickly identifies where the cut needs to be.

“Now, let’s take that to the logical next step. If you’re a manufacturer, you need to step back and ask where your production bottlenecks are, and where the quality control issues are. You start by asking, ‘How can we use a vision system to solve these problems?’”

He added that it’s important to start asking these questions now, especially since applying computer vision in the factory is becoming easier than ever before. “We’re now to the stage where it doesn’t require so much manpower in order to implement a SaaS [software-as-a-service] application that’s never been done before.”

Consider the quality function. “If you have a single person taking measurements, and that’s all they do, think about how you feel sitting for an hour staring at a computer screen,” Sutton said. “You get mesmerized, and you start making bad decisions. We need to help those people. Is there a vision system that can help make those decisions for you, and faster?”

Sutton added that as vision systems become more intelligent, they can help with more nuanced inspections. “You might not have a simple pass/fail metric,” he said. “There could be a scale, a spectrum. Shot peening is an example.

“This is where the AI comes in. You want a system that’s not simply binary, and you can’t show a system examples of every possible defect. You need the system to make intelligent decisions very quickly.”

Vision for Metal Fab’s Future

Consider the hardware inventory for the insertion presses. There might be colored bins, each with bar codes and perhaps a picture of the hardware the bin is holding. Even with all of these safeguards, some incorrect hardware slips through and ends up being inserted into a sheet metal enclosure or other product where it doesn’t belong. The product ships and customer complaints ensue. It’s now off to the races to process the rework and ship replacements.

Now consider QA techs using coordinate measuring machines and scanning arms. A lot of parts flow through the quality department every day, and a few pieces receive signoffs when they should have been rejected.

Finally, consider in-process checks, like the sheet metal blanks associated with a specific traveler. For tracking, an operator denesting the blanks places a bar code printout on each stack, ensuring optimal traceability. Problem is, some similar-looking blanks are sorted to the wrong pile and have the wrong label. They reach the press brake, where the operator pulls up the wrong part program. The troubles only snowball from there.

Now, imagine if a computer vision system—one that doesn’t read a bar code but instead “sees” the actual hardware, component, or workpiece in front of it—caught these errors before they snowballed into something worse?

At the same time, could vision help fabricators track the velocity of all materials through an operation, from raw stock to the shipping dock, and ensure the correct pieces of metal, hardware, and purchased components are where they’re supposed to be?

The devil will be in the details, of course, but as Yalamanchili described, the computational horsepower is there to meet the challenge. “Think of medical research. It’s important to understand what the cell data is, its membrane. For cancer research, they’re looking to build computer vision models to detect objects in the billions.”

Such work puts challenges in metal fabrication—blank denesting, inventory management, and the rest—in a new perspective.

The Positive Snowball

Sutton grouped vision applications in manufacturing into several broad categories, including counting items, label presence, quality control, and sorting. Once a potential application is identified, “you identify the inputs, and you identify what information you want to make decisions on.

“If you want to keep it simple, you have one camera that makes one decision, you have a positive snowball effect. You solve the problem for one application, then you realize, ‘If it can solve this, can it solve this other problem?’ Increasingly, we’re finding the answer to be yes.”