Updated July 2026 · 7 min read · Adente Vision Engineering Team
What is a golden sample in visual inspection?
A golden sample is the reference standard that defines what a correct part looks like, so an inspection system has something to compare every incoming part against. In classical machine vision the golden sample is often one physical master part or a template image, used to set thresholds for dimensions, position and presence. Parts that match the reference pass; parts that drift too far from it get flagged.
The idea is old and useful: fix a known-good reference, then measure deviation from it. What changes with AI inspection is what the reference actually is. For an anomaly-detection model the reference is not a single perfect part but a description of the full range of good, learned from a set of example images. That shift, from one master part to a golden image set, is the point of this article.
Why is a single golden part not enough for AI inspection?
A single golden part describes perfection, not the normal range a good part is allowed to occupy. Real production varies: casting texture, surface finish, colour batch, print position and fixturing all move within tolerance without making a part defective. A reference built from one pristine unit teaches the system that only that exact appearance is acceptable, and everything else looks wrong.
Anomaly detection avoids that trap by learning from many good parts at once. The model builds a picture of the normal distribution, the spread of acceptable appearance, from a set of good images, then scores each new part by how far it deviates. 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 trains this way: it starts from about 20 good reference images and needs no defect samples to begin.
Because the model trains on good parts only, the golden set does a different job than a single master part. It is not there to define one ideal geometry; it is there to teach the boundary between acceptable variation and a real defect. Getting that boundary right is what separates a clean inspection from one that over-rejects.
How do you choose the golden images that define normal?
Choose good parts that represent variation, not perfection. The golden image set has to cover the natural spread the line actually produces, so the model treats that spread as normal instead of flagging it. Twenty near-identical shots of one flawless part are weaker than 20 images that span the real range of good output.
- Include the natural variation your process allows: colour batches, surface texture, and minor position or fixturing differences that are all still good.
- Keep imaging consistent with the line: the same framing, focus, working distance and lighting the running inspection will use, so the model does not learn accidental differences.
- Exclude defects and contamination: a scratch, a smudge or a dirty lens in a good image teaches the model that the flaw is normal.
- Capture and check on the unit: on-device preview lets you frame, light and confirm the reference set at the machine, with no laptop and no separate PC.
Model training on that set completes in under 48 hours, so building or rebuilding a golden set is a task measured in a day or two, not a multi-month data project.
What happens when the golden set is wrong?
A weak golden set shows up as false rejects. When the reference images are too clean or too few, the model learns an artificially narrow definition of normal, and ordinary variation, a slightly different colour lot or a part sitting a millimetre off, reads as a deviation. The line then rejects good product, an operator loses trust in the check, and someone widens the threshold until real defects start slipping through too.
The opposite failure is a golden set that accidentally contains defects. When a scratched or contaminated part is included as good, the model folds that flaw into its idea of normal and stops flagging it. Both failures trace back to the same root cause: the reference did not describe the true boundary of a good part. The comparison below contrasts a strong and a weak golden set.
| Golden set attribute | Strong golden set | Weak golden set |
|---|---|---|
| Coverage of normal variation | Spans lighting, colour, texture and position spread | One pristine part, single appearance |
| Image consistency | Same framing, focus and lighting as the line | Mixed cameras, angles and lighting |
| Defect content | Good parts only, nothing contaminated | Accidental scratch, smudge or dirt included |
| Sample count | About 20 representative good parts | Too few, or many near-identical shots |
| Effect on false rejects | Normal spread passes, real deviations flag | Normal variation reads as a defect, over-rejection |
How does the golden image set fit your inspection line?
Treat the golden set as an asset you maintain, not a one-off. When the process shifts, a new material lot, a tooling change, or a lighting update, the definition of normal moves with it, and the golden set is what you recapture and retrain from. Because training runs on good parts only and finishes under 48 hours, keeping the reference current is a routine operation rather than a re-engineering project.
For the full method behind training from about 20 good images, see the sibling post on few-shot and anomaly inspection. For the wider picture of how AI visual inspection fits a line, see the pillar guide, and to match a golden-set workflow to your enclosure and mounting, see the system page.