With segmentation models, customers also have the advantage of being able to specify exactly what should and should not be considered as a defect. This can be particularly important in surface inspection, for example. Adrien explains, “Surface inspection using segmentation is much more efficient than anomaly detection in most of our cases. You don’t know what you will get with anomaly detection. Some images are very complex and not easy at all to detect defects and sometimes it’s a bit borderline. So some patterns might appear anomalous, but they are not. Anomaly detection models can be very, very sensitive and detect many, many anomalous patterns because the variability in the patterns is just too high.” For a segmentation model, on the other hand, you can specify exactly what to look for. This significantly reduces the risk for pseudo scrap. Adrian tells me, “For example, we have this case right now: one of our customers regularly changes the definition of a defect type. We can re-annotate accordingly each time and re-train the models. With anomaly detection, you don’t have that control or the flexibility to do that.”

This precise specification of error types and what to look for also offers the advantage that segmentation models can detect very small errors. In training, the models can be trained on exactly these small errors by focusing on them through right annotations. For anomaly detection, it’s a different story, Adrien tells me: “For anomaly detection, detecting small errors is rather hard, because especially on larger surfaces, the anomalies are too small to affect the basic appearance of the objects strongly enough.” However, this detection of small defects on surfaces is often important in industrial quality control: “Often it really comes down to finding small, localized defects on large surfaces. Hence, anomaly detection will not be sensitive enough to find the defects that we want whilst segmentation is.”