Spud smarts: AI program learns to sort potatoes

KENNEWICK — Brian Greaves spends his days teaching a computer how to sort potatoes.

A software engineer for California-based Exeter Engineering, Greaves has been teaching an artificial intelligence algorithm how to spot defects in potatoes.

“So with AI, you’re training the computer to think more like a human or train them like you would a human,” Greaves said as he sat in the midst of Exeter sorting equipment on display at the annual Washington Oregon Potato Conference.

“I’m going to show you the potato, and you’re going to determine whether it’s good or bad. And you’re going to learn based on that,” he said.

Exeter Engineering makes sorting, handling and packing equipment for all sorts of vegetable and fruit processors — the company’s website presents case studies of its machines handling watermelons, sweet potatoes, bell peppers and, of course, potatoes — and combined software and imaging systems designed to sort potatoes as well.

“I’ve been doing this now for 25 years, and Brian nearly as long as I have,” said Boomer Batchman, Exeter Engineering president. “AI, generally speaking, is far and away the most effective tool we found for effectively quality sorting potatoes.”

Greaves said teaching an AI how to recognize defects in potatoes is much easier than trying to write a program to do the same thing. But it’s a time consuming process that involves showing an AI pictures — thousands and thousands of pictures — of potatoes, many with identified defects of various kinds, from bruises to rot.

“You need somebody who’s an expert at potatoes, who knows this is a growth crack, this is good, this is bad, this is green,” he said. “You have to tag each one of those objects as you separate them out. And then you can use that to train a data training set for the AI.”

That gives the software a chance to learn what a cut looks like, and eventually identify a damaged potato all on its own, Greaves said.

“AI is kind of a black box. You don’t know really how it’s doing it,” he said.

Scott Parrott handles business development for Smart Vision works, an Orem, Utah-based company that produces artificial intelligence applications for food sorting, including an AI system designed to sort potatoes by size and grade.

He sits underneath a big flat screen monitor running an endless loop of his company’s system scanning potatoes in a sorting line. The goal, Parrott said, is to train the AI to recognize a bruise or a cut in the same way humans recognize an animal is a cat, for example, even if they’ve never seen that particular cat before.

“AI works the same way. We train it to identify the characteristics of cats. And then when it sees a cat, it knows it’s a cat,” he said.

Both Greaves and Parrott said their companies are constantly working with images from their customers — if the customers opt to have those images shared — to further train their respective AIs, and then distribute that work to their other customers.

“We’re continuing to collect images from all our locations,” Greaves said. “And then we’ll bring those back to our office, we will classify those images, and take a couple of days to produce a new model based on those images. And then we just remotely update those models.”

Greaves said Exeter uses Nvidia’s artificial intelligence, though he noted “everybody kind of uses the same core AI” whether it’s Google’s or Amazon’s.” While the engineers at Exeter are constantly working to improve the AI, there comes a point when how it actually works becomes something of a mystery, Greaves said.

Batchman described AIs as a “very stubborn technology,” in that his company’s AI has proven very good at learning a very particular task but not very flexible at adapting to other tasks. This means it can be difficult to distribute the same AI widely to customers who are having the software do slightly different tasks.

“Not all customers want to do the same thing,” Greaves explained.

The AI is good at yes or no type decisions, Batchman said — the potato has a cut or not — but dealing with USDA standards that allow a certain number or percentage of defective potatoes in a batch, or varying that depending upon the quality of the incoming crop, is more difficult for an AI to manage.

“You might want to move some bad stuff into your number one raid. Okay, just to help move things along,” Batchman said. “Everyone does it. With AI, it’s by itself difficult to do those types of judgments.”

“It’s not a quick or easy process to train AI to work precisely. It can be tricky,” Parrott said.

Parrott said Smart Vision Works’ system is primarily used to spot foreign matter and defects, and is not only used in several Idaho potato sheds, but is also employed to spot defects in dates, inspect frozen food, and detect foreign material in shrimp and meat for a major U.S. meat processor.

“We use it in the food industry for one or two things. One is defect detection, and the other is for foreign material detection,” he said. “We stick to those two.”

Batchman said Exeter’s system is deployed in produce packing houses across the United States, Canada and Mexico.

He also believes that while AI is already proving itself very useful in agriculture, the technology is only going to get better and cheaper so even the smallest family business can use it to help become more efficient.

“The AI itself is really good,” Batchman said. “It will evolve, certainly over the next few years, but it will be the great equalizer. Everyone’s going to be doing it.”

Charles H. Featherstone can be reached at cfeatherstone@columbiabasinherald.com.