Deep learning is rapidly becoming an indispensable element in machine vision solutions. Its application is proving to be particularly useful for identifying objects and features in images.
Deep learning is a subset of machine learning and is based on neural networks. The technology is unparalleled when used in cases where there are complex and varying imaging conditions. In manufacturing applications, deep learning tools complement traditional machine vision, as they are enhancing engineering capabilities and optimizing processes by performing impossible tasks.
Andrew Zosel, senior vice president and general manager, Zebra Technologies, characterizes deep learning tools used in computer vision as a reciprocal set of technologies to traditional machine vision. While machine vision systems use cameras and sensors to capture visual data, he said, AI algorithms process this data to make decisions or trigger actions.
READ MORE: At the Edge: Vision AI Software Improves Quality Inspections
Computer vision is an application of machine learning that processes information from digital images and videos into meaningful information for decision-making. Manufacturers achieve unprecedented accuracy when using the overlapping technologies of both machine vision and computer vision. These vision technologies are transforming quality control, automated inspections and defect detection in a range of vertical markets, said Zosel.
Zebra Technologies, for example, developed deep-learning solutions for automotive, semiconductor and electronics applications, and inspection-type applications for cosmetic defects.
“Think of body panels on a vehicle and whether there are scratches on a cosmetics case or a cell phone,” Zosel explained. “There are also classification applications, such as what type of wheels are installed on a vehicle, for example. If you have 100 different wheel sets then one can much more easily classify those through a deep-learning classifier than through a traditional machine vision.”
Zosel said his team is seeing a multitude of OCR (optical character recognition) use cases leveraging deep learning. “We are trained from early on, as humans, to read text, and there can be all kinds of written text that we easily interpret as readable,” said Zosel. With traditional machine vision tools, OCR has been a challenging area to solve. The text had to be specific, and it required specified lighting. To date, since systems are globally read by humans, there is a lot of text on parts and products.”
Zebra Technologies is exploring how deep learning-based OCR can enable the machine or the camera systems to read text in a manner that compares favorably to the way humans read without developing specialized text or processing for that task, said Zosel. This has unlocked applications across many industries.
“Think of date and lot codes on pharmaceuticals and cosmetics and those type of things,” pointed out Zosel. “We’ve done applications where we’re looking at those for returns processing and for some of those things we’ve looked at OCR. For instance, characters on a stamped assembly in automotive, where they’ve embossed characters and that may not be highly legible with a standard machine vision system.”
Zebra Technologies uses deep learning OCR to read such applications with ease, Zosel said. “It really allows us to deploy and help our customers with camera systems and in ways that they traditionally would have found difficult to automate or digitize.”
Justifying the Cost of Vision System Implementations
The cost of vision solutions can range from simple vision sensors to high-end, high-resolution cameras—but it all depends on the application.
“The cost of a system is, of course, the product and technology that you deploy, but it’s also the integration of that technology and the setup and the validation and the systems around it,” Zosel said. “And when the machine vision system is used in parallel with a robot or some other handling system, a lot of times that system can be a much higher cost. So, it is all relative.”
Fundamentally, as vision systems specialists continue to drive ease of use and ease of deployment and make the user interfaces better to implement, the cost of deployment will be more feasible, Zosel said.
Watch additional parts of this interview series with Andrew Zosel:
Part 1: How Zebra Technologies Uses Machine Vision to Transform Production Automation
Part 2: Deep Learning Algorithms Help Vision Inspection Systems See Better