Updated July 2026 · 6 min read · Adente Vision Engineering Team
What is the difference between supervised and unsupervised inspection?
Supervised inspection learns from labeled examples of both classes: this image is good, that image is defective. The model is trained to reproduce those labels, so it can only recognise defect types it was shown during training. Unsupervised inspection learns from one class only, the good part, and treats anything that deviates far enough from that learned normal as a defect, whether or not it has seen that failure before.
The distinction is not academic; it decides how much data you have to collect before the line can run. Supervised needs a labeled catalogue of defects. Unsupervised, also called one-class or anomaly detection, needs only a set of good parts. Which one your line needs comes down to a simple question: do you actually have the defect examples that supervised learning demands?
Why does labeled defect data cost so much on a real line?
Labeled defect data is expensive because defects are rare by design. A line running at healthy quality produces very few bad parts, so collecting hundreds or thousands of labeled defect images, enough for a supervised model to generalise, can take weeks or months of running, sorting and hand-labeling. Every new defect type restarts the collection.
There is a second, structural problem: you can only label the defects you already know about. A supervised model trained on scratches and dents has no representation of the short-shot, the contamination or the first-seen flaw nobody catalogued. When that novel defect appears, the model has no class for it and tends to pass it as good. So the very cost of supervised data buys you coverage of known failures and leaves the unknown ones uncovered.
Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and it takes the unsupervised path: it trains on good parts only, from about 20 reference images, so a line does not have to accumulate a defect library before inspection can start.
When does unsupervised inspection win?
Unsupervised inspection wins whenever defects are rare, variable or unknown in advance. Because the model learns the boundary of normal rather than a fixed list of defect classes, it flags deviation itself, which is exactly what a first-seen or one-off flaw is. That covers most surface-defect, contamination and anomaly work, where the catalogue of possible failures is open and no one can enumerate it up front.
It also collapses the timeline. Training on good parts only removes the defect-collection phase entirely, so setup is a matter of capturing about 20 representative good images and training the model, rather than running the line for weeks to harvest bad parts. For a plant that changes parts often or cannot afford to stockpile defects, that is the deciding factor.
When does supervised classification still fit?
Supervised classification still fits when you have a small, closed set of well-defined classes and enough examples of each. Sorting a part into one of several known variants, grading a defect into named severity classes, or reading a fixed set of categories are jobs where the classes are stable and repeatable, so labeling is finite and worthwhile. If the same handful of defect types recurs constantly and you can collect them, a supervised model can classify them precisely.
The two approaches are not mutually exclusive. A practical line often runs unsupervised anomaly detection to catch anything abnormal, then applies a supervised or classical step to sort or grade the specific cases that matter. The table below frames the choice by the data you have, not by the technology label.
| Dimension | Supervised inspection | Unsupervised inspection |
|---|---|---|
| What it learns | Labeled good and defective examples | Only what a good part looks like |
| Data needed | Many labeled defects, hundreds or more per class | About 20 good images, no defects |
| Rare or first-seen defects | Misses classes absent from training | Flags any deviation from normal |
| Best for | Known, repeatable defect or variant classes | Rare, variable or unknown defects |
| Time to start | Long defect collection and labeling | Good parts only, training under 48 hours |
How do you choose for your line?
Choose by your data reality, not by which method sounds more advanced. If your defects are rare, changeable or not fully known, and you cannot realistically collect a labeled defect library, unsupervised inspection is the fit: it starts from about 20 good parts and catches deviations you never catalogued. If your task is sorting a small set of well-defined, repeatable classes and you can gather examples of each, supervised classification earns its place.
Most lines land on unsupervised as the base layer because the rare, first-seen defect is the one that hurts, and supervised as a targeted add-on where a closed set of classes justifies the labeling. For the mechanics of training from about 20 good images, see the sibling post on few-shot and anomaly inspection. For the broader method, see the pillar guide on AI visual inspection, and to match the approach to a unit and mounting, see the system page.