Short answer
A golden sample is a part judged to meet specification and kept as a physical or visual reference. Golden-sample and known-good images anchor anomaly-detection methods, which fit on defect-free images only, and serve as alignment and audit references. They show one acceptable appearance, not the full range of acceptable variation, so they cannot define normal on their own.
What "golden" means in practice
On a production line, a golden sample, master sample or reference part is a unit approved by quality as conforming. Operators and machine-vision systems compare production parts against it. Some plants keep several: one at nominal, boundary samples at the accept limit, and defect samples for operator training. Images of these units are the visual record.
The label has more than one meaning in data sourcing. It can be a single reference image per part and view, a curated set of known-good images from a stable period, or a template used for registration. Buyers need to say which they want. This page treats the first two; boundary and defective samples are covered under physical defect samples and borderline samples. See the inspection cluster and the humanoid data section for scope.
Why anomaly methods depend on known-good images
Cold-start anomaly detection requires no defect labels: it models what normal looks like and flags deviation. PatchCore builds a memory bank of nominal patch features from defect-free images, subsampled with a greedy coreset, and reported image-level AUROC up to 99.6% on MVTec AD (Roth et al., 2021). The MVTec AD training set contains defect-free images only, while the test set carries pixel-precise ground truth for defects (Bergmann et al., 2021).
That number comes from a curated benchmark with controlled capture. It tells you the method class works under benchmark conditions, not that a given plant's images will give the same result. For the supervised alternative and when each approach fits, see anomaly detection versus supervised defect data.
What reference images can and cannot support
They can support training and fitting of nominal-only models (see sourcing defect image datasets for the defective side), registration templates, drift monitoring against a stable baseline, and audit sets where a labeler confirms that a "known-good" claim holds. They are also a useful control: if a detector flags a golden image, the fault lies with the model or the imaging, not the part.
They cannot define the acceptable range. A part can be good and still differ from the golden sample in finish, color lot or texture. A set built from one pristine unit teaches a detector that normal is narrow, producing false rejects. Reference imagery also contains nothing about handling, so it does not tell a robot how to position or pick the part. And "known good" often means "passed final inspection" rather than "verified by measurement", which is a weaker guarantee.
Illustrative reference-image package
This package is Illustrative; field values would be synthetic placeholders.
| Component | Content | Notes |
|---|---|---|
| Golden units | Images per unit, all required views | Record unit serial, approval date, approver role |
| Nominal pool | Images of production units dispositioned accept | Spread across lots, shifts and tooling |
| Boundary units | Images of units at the accept limit | Tagged with the characteristic that is near limit |
| Capture record | Camera, lens, lighting, exposure, working distance | Needed to compare across lines |
| Calibration | Target images, scale factor, color reference | Ties pixels to physical units |
| Lineage | Revision of drawing, tooling and process at capture | Reference goes stale with design change |
| Disposition link | Outcome for every image | Excludes late rejects and returns |
Imaging detail is covered in imaging conditions and metadata. Multi-camera setups add alignment concerns; see multi-view inspection imagery.
How to audit a reference set
Check that normal is varied. Compute simple statistics across the nominal pool: brightness distribution, pose spread, background and fixture variation. A pool with near-identical frames is one sample repeated.
Check that normal is clean. Run a sample past a qualified inspector. Hidden defects in the nominal pool contaminate nominal-only models silently, and the effect is hard to detect after fitting. Keep a held-out test set with known defects so contamination shows up as missed detections.
Check temporal validity. Compare the reference's revision with the revision of the parts being inspected. A tooling or finish change makes the old golden sample wrong.
Check cross-line transfer. Fit on one line's nominal images and score another's. Large drops indicate that your reference is tied to the capture setup; see domain shift across lines.
Check that evaluation data stays separate. Reference images used to fit a model must not appear in the evaluation split. See defect detection evaluation sets.
Supplier types and sourcing routes
Manufacturers with machine-vision systems already store golden templates and a stream of accepted-part images, often in the vision vendor's software with its own export format. Quality labs and calibration labs keep master samples with written approvals. Contract inspection firms hold reference images on behalf of clients; rights often rest with the client.
Existing archives are fastest, but only if capture parameters were logged. If they were not, commissioned capture of known-good units may be the better route, with the supplier photographing approved parts under documented conditions. That is a new-collection discussion and depends on site access and scope.
The acceptance rule that makes a sample "golden" is a label definition in its own right; see acceptance standards as label definitions.
Rights and sensitivity
A golden sample is a finished representation of the supplier's design, which raises product-design and customer-part confidentiality. Reference images may reveal tolerances, markings or geometry that a customer owns. Ask who owns the design, whether customer approval is needed, and whether the license covers derived models. Blurring logos does not resolve ownership of the design itself. This is not legal advice.
Limitations and extra preparation
Known-good images are only as reliable as the decision that labeled them good. If "golden" means "passed final inspection", some units with subtle defects will be in the set, and an anomaly model fit on them learns to accept those defects as normal. A reference set captured in one season, on one line or with one lighting recipe defines "normal" narrowly; production variation outside that window will be flagged as anomalous even when the part is acceptable. Reference images say nothing about how a part feels, deforms or should be gripped, so they support inspection and recognition, not manipulation.
Plan for extra preparation: confirming how each unit was verified, removing images taken before a process or design change, recording camera, lens and lighting per capture period, and holding out reference images by lot or by date for evaluation. Where the supplier cannot document verification, treat the set as "probably good" and budget for a manual review of a sample before relying on it.
Questions to ask suppliers
- Which units are designated golden, who approved them and when?
- Are images of production-accepted units screened for hidden defects, and by whom?
- What drawing and process revision does each image correspond to?
- How many lots, shifts and tooling sets does the nominal pool span?
- Which capture parameters were logged for each image?
- Do any images show customer-owned parts or markings, and has the customer approved licensing?
- Can you supply boundary or known-defective units for an evaluation set?
- How are images exported, and what do the file names encode?
Discuss your requirements
If you need known-good reference imagery from US manufacturers, describe the parts, views and imaging conditions to SourceX. SourceX reviews the requirement and follows up about potential sources, and checks the data and the supplier's licensing permissions. A request does not guarantee a match.
Tell SourceX what your evaluation set needs to cover
Sources
- Towards Total Recall in Industrial Anomaly Detection — PatchCore fits a memory bank of nominal patch features from defect-free images only, reaching image-level AUROC up to 99.6% on MVTec AD.
- The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — MVTec AD training set contains only defect-free images with pixel-precise ground truth on the test set.