General Defect Detection

Fewer than most people expect for a first working model, and more than most expect to reach production confidence. A useful starting point is a few dozen good parts and a similar number of examples per defect class you care about. The limiting factor is almost never the count, it is the variety: twenty images that cover the real range of lighting, position, and acceptable finish beat two hundred near-identical ones. Rare defects are the hard case, which is what synthetic defect generation exists to solve.

That is the normal situation, and it is the central problem in applied defect detection. A well-run line produces very few defects, so the classes you most want to catch are the ones you have the least data for. Waiting for them to accumulate can take months. The alternative is to generate them: start from real images of your own good parts and render the defect onto them, which gives a model something to learn from without waiting for the failure to happen.

Some, and it is worth being precise about what to expect. A model trained on specific defect classes recognises those classes. A model trained to recognise a good part will flag anything that departs from it, which catches novel defects but also flags harmless novelty and needs a human to sort out early on. Most deployments start with the second approach to find out what actually comes off the line, then tighten into named classes as the real failure modes reveal themselves.