Anomaly detection: inspecting with only good samples

The most common reason an inspection project stalls has nothing to do with the technology. It is that nobody can produce the defect images. The failure happens once a month, the parts were scrapped without being photographed, and the quality team cannot describe the next one because they have not seen it yet.

Anomaly detection inverts the problem. Instead of learning what defects look like, the model learns what a good part looks like and flags anything that departs from it. The product description is exactly this: train using only good samples, and the model learns the normal appearance of your parts and flags anything that deviates as a potential defect.

Supervised and unsupervised, in one sentence each

A supervised model is taught the difference between named categories. You show it scratches, you show it clean surfaces, you label both, and it learns a boundary between them. It can then tell you which defect it found, because you told it the names.

An unsupervised model is taught only what normal looks like. It builds a representation of good parts and measures how far each new part sits from that. It cannot tell you the defect is a scratch rather than a smear, because nobody ever gave it those words. It can only tell you this region does not belong.

Three model types, and how to choose

Anomaly detection is not one algorithm. On the OV80i the unsupervised block offers three, and they trade sensitivity against speed and against how much data they need. The product describes them plainly:

The image count is worth pausing on. Twenty good parts is a coffee break, not a data collection project. The reason anomaly detection changes project timelines is not that it is more accurate than a trained model, because often it is not. It is that the data you need already exists on the line right now.

What the model actually outputs

A supervised model returns a class and a confidence. An anomaly model returns something different, and understanding the shape of it is what stops the sensitivity slider becoming a source of argument.

The minimum blob area filter is the most important setting on the block, and it is the one most often left alone. Every real surface deviates a little everywhere. If the rule is any deviation at all, the station will fail good parts constantly. If the rule is a deviating region larger than a defined area, it behaves. Setting that number is a conversation with the quality team about what size of mark is actually a reject, which is a conversation worth having anyway.

When anomaly detection is the right call

Four situations where it is clearly the better starting point.

1. You have no defect images and cannot manufacture them. The most common case, and the honest reason most inspections never get built.
2. The defect catalogue is open ended. Contamination, foreign objects, and cosmetic damage do not come from a finite list. A model trained on the eight defects you know about will miss the ninth by construction.
3. The surface is textured or patterned. Where no two good parts look the same, describing defects by appearance is close to impossible, but describing the boundary of normal is tractable.
4. Low volume or high mix. If a part number runs a few hundred times a year, you will never accumulate a labelled defect set for it. Twenty good parts you can get on the first run.

And the honest inverse. Use a supervised model when the disposition depends on which defect it is, when you need the defect name for process feedback rather than just a reject signal, or when you already have a good labelled set, in which case a trained model will usually be both more accurate and easier to defend in a capability study.

These are not exclusive. A recipe on the OV80i can run several models over one image, so a common and sensible pattern is a supervised model handling the known defects with named classes, and an unsupervised model running alongside it as a safety net for everything nobody predicted.

The failure mode to plan for

Anomaly detection has one characteristic weakness, and it is better to name it than discover it. The model defines normal from the good parts you gave it. If your twenty good samples do not include the full range of acceptable variation, everything outside that range reads as an anomaly.

In practice that means a station commissioned on parts from one material lot will flag the next lot when the finish shifts slightly. It is not a model defect, it is an accurate report that the part looks different, which is precisely what you asked for. The fix is to include real variation in the good set: different lots, different shifts, the acceptable end of the tolerance band as well as the middle of it.

This is also why a retraining path matters more for anomaly detection than for supervised work. When production shifts, you need to feed genuinely-good-but-flagged parts back into the good set. Being able to find those images in production history and add them to the training set without a data science exercise is the difference between a station that survives a year and one that gets switched off in month three. Our guide to machine vision systems covers where this sits in the wider station design, and the lighting guide covers the other half, because an anomaly model is even more sensitive to illumination drift than a supervised one: a lighting change looks exactly like an abnormal part.