Deep learning: a powerful tool for the right machine vision applications
As a certified vision integrator specializing in machine vision systems, LEONI Engineering Products & Services, Inc. (LEPS) is using deep learning to help industrial customers resolve challenges that cannot be solved with traditional machine vision solutions. But that doesn’t mean deep learning is a silver bullet for every application.
“Deep learning allows a machine vision system to ‘learn’ what a good or bad part is based on the statistical analysis of images that have been expertly tagged,” says Jim Reed, Technical Sales, Vision Solutions at LEPS. “It’s similar to how a human learns. However, if an application can be solved with traditional machine vision techniques, it will likely run faster and cheaper than a deep learning solution.”
There is a misconception among some that deep learning means the machine vision system trains itself, but that’s not the case. Andrew Meyer, LEPS Sales Engineer, explains that deep learning has two phases: training and inference. Training happens when deep learning software uses massive amounts of computational power to analyze images of objects that a third-party expert has tagged as “good” or “bad.” This is often done in the cloud or by using large PCs available to system designers.
Then, after training, the deep learning program can be run on local hosts as part of the inference phase. This is when the deep learning system makes judgements about the quality of new parts — that is, the inspection step.
“Deep learning is great at identifying defects that are hard to define, such as a random scratch on a cell phone cover,” Meyer says. “But it’s not good for traditional machine vision applications, such as metrology and measurements.”
When an application looks like a good candidate for deep learning, LEPS helps its customer develop a tagged image set to evaluate the application.
“If there’s a machine vision system in place, we can have the operator tag the images as they’re produced,” Reed says. “If there isn’t a system in place, we may install a camera to acquire sufficient sample images. While this may seem easier than programming using traditional machine vision algorithms, it’s still work that needs to happen.”
One recent application that benefited from deep learning was the inspection of molds that create automotive seat foam pads. Each mold, of which there are thousands of variants, contains any number of wires, clips, and other components. These molds are notoriously hard to rework. A single defective pad can erase the profits from the next series of pads. Traditional machine vision systems have been used to inspect all the parts of a mold, determine if a part is wrong or missing, and, in the case of missing parts, flag the mold for rework. However, constantly changing pad designs means the machine vision system needs a full-time employee to program each new mold into the system.
“The customer found that they can now use anyone on staff who recognizes a clip or component to add new pad models to the deep learning inspection system,” Meyer says. The system then trains itself with new models to continue the inspections, without the need of full-time staff support.
With many new technologies emerging in automation, LEPS focuses on choosing the best technology for each application to allow customers to make the most of advanced technology to improve manufacturing.