Distillery Tests Vision System to Improve Manual Processes

Dairy Distillery (Almonte, ON, Canada; www. www.dairydistillery.com) produces vodka and cream liquor from milk permeate, a byproduct leftover after cream, fat, and proteins have been removed from whole milk during an ultrafiltration manufacturing process. With such an unusual product, the distillery, founded in 2018, created a distinctive bottle and label to help tell the brand’s story. The bottle is styled after a traditional milk bottle, and the three-part label includes a round emblem featuring a cow and silo.

These brand elements help the 15-employee Dairy Distillery’s brand, Vodkow, stand out on shelves filled with many similar products. “Consumers often judge what’s inside a bottle based on its appearance,” Neal McCarten, co-founder and director of marketing at Dairy Distillery, explains in a news release.

They’re looking for a visually pleasing presentation. If a package is damaged or a label is askew, consumers may choose another product in a more pristine package, says David Geros, chief scientific officer at the distillery. “We have a high value product here,” he adds, noting that the company’s products sell for $30 to $35 a bottle.

In addition to vodka, the distillery makes four lactose-free cream liquors—plain, chocolate, maple, and coffee. It sells the products primarily in Canada and San Diego, California. “We’re hoping to push more in the next year into the into the U.S. market,” explains Geros, says. “It’s been tough, with the pandemic, getting out and around.”

While Dairy Distillery’s unique package is a marketing asset, it presents operational challenges. Two parts of the label are applied by machine, and then employees apply third element—the emblem—manually by lining it up with the others—a large label on the bottom of the bottle and a small label that is affixed to the cap and top of the bottle. For bottles to look good, all three labels need to be lined up properly.

Role for Vision in Manual Process

That is not easy for employees to manage consistently, particularly in a production environment in which the distillery produces 1,000 bottles in an eight-hour shift.

“There’s a lot of subjectivity around it for the operators, and they actually found it quite stressful because they were worried that they were putting it in the wrong location,” Geros says. “What was happening was this one label would be too far to left or too far to the right. It just looked funny.”

In an earlier iteration of the label-application process, the distillery applied the third label using a machine but ended up with error rates of 50%. The machine “had an optical sensor on it, and it just wasn’t sensitive enough to pick up the labels in the way that we needed it to,” says Geros. “The tolerances were just a little bit wider than we could accept.”

That’s when the distillery went back to applying the third label manually, but it also began testing a decision-support solution to assist employees with the process. The product is a turnkey visual inspection system based on artificial intelligence (AI) from Pleora Technologies (Ottawa, ON, Canada; www.pleora.com). The camera-based edge-processing system has an AI app that can be trained on a specific task. In this case, the distillery trained the app to recognize what the properly applied labels should look like. To do so, Dairy Distillery provided the AI app with “a golden reference,” or one image of each of the company’s products with labels applied correctly.

“What we really aim to do there is provide a level of repeatability and intelligence to augment that manual operator. We are not trying to displace the operator,” says John Butler, vice president of sales and marketing at Pleora. The goal is to “reduce the risk of error counts and drive up the overall efficiency and productivity of those folks who are doing that manual task,” he adds.

Vision Components

The system includes the software, NVIDIA (Santa Clara, CA; USA; www.nvidia.com) Jetson TX2 industrial edge processor, keyboard, monitor and document-scanning 4K camera from IPEVO (Taipei City, Taiwan; www.ipevo.com). The lens and sensor are included as part of the camera. There is a USB connection from the camera to the edge processor and an HDMI connection from the processor to the monitor.

“Really the main requirement for the initial system is high frames per second, so there’s minimal latency between the camera and display,” explains Ed Goffin, marketing manager at Pleora Technologies.

Goffin says a light is part of the camera, although the camera-based illumination hasn’t been necessary at the distillery.

When employees are ready to apply a label, they place the bottle under the camera in a holder designed to fit the bottle. The camera, which is mounted on a stand, captures a live image, which is fed to the software, running on the NVIDIA processor. The software then overlays a real-time image of a crosshair, or grid, on the monitor. As the employee places the third label on the bottle, an image of the label appears on the monitor, allowing the employee to see if the label is placed correctly within the crosshair.

The distillery began working with Pleora in fall of 2021 on a prototype version of the system and was scheduled to begin testing a second version in May 2022.

The newest version is designed to overcome an issue Dairy Distillery had with the earlier version of the system. With that version, the operator had to line up the bottle precisely under the camera, which slowed down the production line. If they were off by even 2 or 3 mm, the image with the crosshair would be fuzzy, making it difficult to apply the label. “It was a bit fiddly,” Geros says.

In the new version, the software adjusts automatically to “lock” on the image of the bottle—rather than the other way around. “They just have to put it under the camera and now it is recognized,” Goffin says.

While the earlier version of the product is not currently in use in the production environment to assist with the labels, the distillery has been using it as a training tool. There are two full-time staff members in the packaging area, but Geros hires three or four contract workers annually for the summer and fall seasons. The distillery needs the extra staff to help because 60% of the company’s sales occur in November and December. In addition, Geros crosstrains employees on various functions, including applying the labels.

Using Vision to Improve Human Inspection

With the new version of the system, Dairy Distillery plans to not only focus on the labeling process but also manual inspection. The goal is to ship bottles with labels that are within 1 mm of the ideal placement. In this process, employees manually spot-check bottles to determine if they meet the distillery’s bottling standards.

In this use case, the AI app will visually highlight the differences that it sees in the labels on the screen, compared with the golden reference, giving the operator the option of accepting or rejecting those bottles.

Geros says the addition of the tool at the end of the production process is designed to help remove subjective decision-making. Human inspectors get “really good at finding minor, minor defects,” including those within the tolerance range for proper placement of the labels, he says, making it somewhat stressful for them to decide if they should pull a bottle to be relabeled. He hopes the tool will give “them that confidence that what they’re putting out is a quality product.”

As is the case with other AI tools, Butler explains that the software learns and improves as the human inspectors use it, making it possible to build more of an automated inspection process over time. As human operators react to the differences the software notes during the comparison process, “what they are actually doing in the background is building up a repository, or library, of reference images,” he says. Over time, the software makes pass/fail decisions accurately without human input.

However, inspectors would still be required to place the bottles under the camera. “To automate that, you would essentially need an automated feed mechanism, and now you are into more of a traditional machine vision inspection application. This is really intended to be more of an offline system,” Butler says. “But the dependence on the operator to make the decision is gradually taken away over time.”

The next step, Geros says, is to train the staff on the new version of the product and then assess how to apply it to the work at the distillery. He believes the product has potential as a tool for training, applying labels, and inspection.