“Simulated defects open the door to AI applications”

Prof Längle, AI is causing a lot of disruption Does that include your business unit?

Thomas Längle: Machine learning (ML) has played a big role in computer vision for quite some time now. It has direct applications for us in fields such as remote sensing – so, in analyzing and interpreting aerial images – and in sorting bulk goods. One exciting trend we’re seeing right now is in the area of inspecting surfaces and complex objects, where the advances in detecting and classifying defects that we might expect from ML approaches have often been stymied by lack of data. To train the AI models, they would need to be fed a large number of examples of good products, but also be supplied with precisely labeled bad ones. There just aren’t enough of those, if any at all. So we’ve turned the focus of our efforts on synthetic image generation.  

You’re planning to generate the data to train the AI with AI?

Well, we already do have experience with ML approaches such as generative adversarial networks, variational autoencoders, and diffusion models for generating new versions of bad examples on the basis of existing images. But primarily, we’re working with simulations. The idea is to simulate the entire testing and inspection environment – specimen geometry, material properties, lighting, sensor technology – to produce images that are synthetic, but still sufficiently realistic. And then we can take data on defects that we have gathered in the past to add virtual errors as well and vary them in a number of ways. With a scratch, for example, we can cycle through different lengths, depths, and shapes. On this basis, we can then calculate the images that the analysis AI would ultimately face in the chosen inspection setup.

What are the advantages of this simulative approach?

It might help us solve the “chicken or egg” problem. ML-based reproduction of images of defects requires that there be at least some images already on hand, so it still depends on the quantity and quality of the input data. We can also build in any kind of error we want, and, of course, the synthetic images created in this way are always labeled perfectly. That’s because we know which defects we simulated and where. We can also change other relevant parameters. In real life, defects aren’t the only things that vary. There’s a certain amount of spread to a specimen’s position and dimensions, the lighting varies somewhat, etc. Ultimately, the simulation gives us a digital twin of the inspection setup – which we can also vary virtually. So we can test and optimize the entire system and draw conclusions about how it will perform before translating it into hardware.

Prof. Längle is the spokesperson for the Inspection and Optronic Systems business unit and head of the Visual Inspection Systems (SPR) department.