Traditional machine vision is supervised: to catch a defect, you collect and label many examples of it, train a model, then repeat for every new defect type. Aletis flips this. You show it about 25 good parts, it learns what a correct part looks like, and it flags anything that deviates, including defects it was never shown. No defect libraries, no labeling, no data-science team. And Aletis is more than a stock anomaly detector: it is a proprietary inspection pipeline that combines multiple methods, generates synthetic training data, and is tuned to your specific parts and line.
Supervised inspection has a data problem built in. It needs a large volume of labeled defect images, yet "optical inspection tasks often lack defective samples or it is unclear what kinds of defects may appear" (MVTec AD, Bergmann et al., CVPR 2019). Labeling them is expensive: these models need "a large amount of data ... often requiring high-precision labels," so "the cost of the solutions would significantly increase due to the annotation requirements" (Božič et al., Computers in Industry, 2021). You are asked to collect many examples of the very thing you are trying to prevent.
Aletis learns from good parts only. It models what a correct part looks like from a small set of examples and flags any deviation as a potential defect, so it catches defect types it was never shown. Where a basic anomaly detector stops there, Aletis adds a proprietary pipeline: it generates synthetic training data to strengthen the model, combines multiple detection methods rather than leaning on one, and is configured to each line. That is the difference between a research technique and a production-ready inspection system.
| Supervised machine vision | Aletis | |
|---|---|---|
| Training data | Many labeled images per defect type | About 25 good parts |
| Defect samples needed | Yes, for every defect type | None |
| Catches unseen defects | No, only what it was trained on | Yes |
| New defect type | Collect, label, retrain | Already covered |
| Who sets it up | Vision / data-science engineers | Your IT or ops manager |
| Setup | A data-collection and labeling project | Under 30 minutes |
With supervised vision, a defect you have not seen before means a fresh data-collection and labeling cycle, then a retrain. With Aletis, a new defect is just another deviation from normal, caught without new data or retraining.
The cost of poor quality in a typical organization runs about 15 to 20% of sales revenue (ASQ, Cost of Quality). A defect that reaches a customer costs far more than one caught on the line.
Supervised vision is not obsolete. If you inspect for a single, rigidly defined defect at extremely high volume and already own a large labeled dataset for it, a model tuned to that one defect can be the right tool. Aletis wins when defects are varied, rare, hard to collect, or not fully known in advance, which describes most real production lines.