AI in automated scratch-dig inspection (Conference Presentation)

The tricky part in deploying AI models to production is often training, with two important prerequisites: One, training data must be representative of the data that the algorithm will see in later use. And two, training data must be properly labeled manually. In algorithms for automated optical inspection, there is a further problem: What if there are only a few examples of specific defect types? We tackled these problems with different strategies when developing our ARGOS system for scratch-dig inspection. We will present real-world examples of how AI algorithms can be used for defect detection and classification without large training databases.