Deep learning, a subset of machine learning, is often the topic of conversation among machine vision professionals.
That’s because deep learning uses neural networks to mimic human intelligence, allowing it to automate tasks that traditional machine vision methods struggle to accomplish. For example, integrators and developers could create machine vision applications incorporating deep learning to inspect products in which there are subtle differences between acceptable and unacceptable variations in products, requiring subjective decisions. It can also inspect surfaces of products that are difficult to image clearly, such as those made of metal or silicon, or where the product orientation cannot be controlled. It also can help automate inspection in environments with insufficient or variable lighting or where lighting sources degrade over time.
In these situations, creating inspection algorithms based on rules is hard to do. In contrast, deep learning algorithms evaluate complicated inspection scenarios based on training with labeled images showing acceptable and unacceptable product variations.
In this issue, we have numerous articles in which the authors explore deep learning and its applicability to solving machine vision challenges. Here’s what you’ll find inside this issue of Vision Systems Design:
- An industry solution profile focused on using a deep learning-enabled software product to inspect fastening hardware on medical carts.
- A product focus section that explores the current and future impact of deep learning on embedded vision applications.
- An integration insights piece that discusses the merits of deep learning versus edge AI—an approach in which pretrained deep learning algorithms are embedded directly on an edge device such as a camera.
Despite its advantages, deep learning won’t work for all machine vision challenges. For example, traditional machine vision works better in situations where precise measurements are required. As David Wyatt, Founder of Automation Doctor and Vision System Doctor, explained to me, “Humans can guess the sizes of objects in an image, but when we need to know what it really is, we use a tool like a ruler.”
Since large clusters of neural networks are designed to mimic humans, “deep learning should use a conventional imaging metrology application as an input,” he adds.
In the end, I suspect that deep learning’s impact on machine vision will become more obvious over time.
On a completely different topic, I’d like to introduce you to Jim Tatum. He is Vision Systems Design’s new senior editor. He is a veteran journalist with experience in both newspapers and B2B publishing. He’s won many awards throughout his career as a writer and editor. We’re thrilled to have him on staff.
