When a Tier 1 automotive parts supplier needed to cover 8,000+ product models across 6 production lines running at 11,520 units per day, the team started with golden-sample matching. Within three months, the reference library had drifted: new SKU variants, a relighting after a ceiling repair, and a tooling revision on two lines had made roughly 15% of the golden images unrepresentative. Inspection confidence fell; the quality team faced a choice between re-photographing hundreds of references or rethinking the approach. They chose to retrain. The retrain took 30 minutes of setup and one additional training session on the affected SKUs. The re-photography would have taken two days.
That outcome shaped how we deploy on high-mix lines. Both golden-sample matching and learned defect models work. The question is which one holds up on your specific line — and which one imposes costs that do not show up in a six-week pilot.
What golden-sample matching actually does
Golden-sample matching — also called reference-image comparison or template-based inspection — compares each incoming part image against a stored reference image of a known-good part. If the incoming image deviates from the reference by more than a configured threshold, the system flags a reject.
The operational appeal is clear: no training data, no labelling pipeline, no machine-learning workflow. The golden image is the specification, made visual and machine-readable. For high-volume, single-SKU lines with controlled lighting and a narrow defect taxonomy, it deploys fast and validates straightforwardly.
HyperQ AI Vision includes a Pattern Inspection module built on this principle. Setup runs in approximately 30 minutes and produces live results immediately. For the right application — a stable line with a limited product range and well-understood failure modes — that speed is a genuine advantage.
The problem is not with the method. It is with what happens when line conditions change, and they always do.
Three ways golden samples lose validity over time
SKU proliferation. Every new product variant needs its own reference image. A line running 50 SKUs today may run 200 in 18 months. Each new reference requires capture under production lighting, review against the engineering drawing, sign-off by quality, and entry into the library management system. The Tier 1 auto-parts supplier above eventually built a library of several thousand references. Maintaining it (auditing for drift, archiving retired variants, updating changed SKUs) became a full-time quality task.
Lighting drift. LED arrays age over thousands of hours, replacement bulbs vary slightly in colour temperature, and seasonal shifts in ambient light bleed through factory skylights. A golden image taken in Q1 under one effective illumination is compared in Q4 under a slightly different one. The part has not changed; the comparison has. False positives climb without any corresponding change to reject quality.
Tooling and process change. When a supplier modifies die tooling, adjusts a surface treatment process, or changes a component geometry within engineering tolerance, the surface finish shifts. The golden image now depicts a surface condition the current process no longer produces exactly. The matcher interprets a legally conforming part as a deviation. On a line running first-article inspection (FAI) protocols, this forces a re-photograph and re-validation cycle every time the process is touched. A quality engineer managing multiple FAI lines described the pattern: a single engineering change order that covered four part numbers required 11 reference re-photographs before the inspection system was back in a validated state.
Each vector is manageable for a small number of SKUs in a stable environment. Each compounds on high-mix lines with active engineering change orders.
How learned defect models handle the same conditions
Learned defect models train on labelled images of good parts and defective parts, then generalise across visual variation rather than comparing against a single reference. HyperQ AI Vision initialises a model on 1,000 images — roughly one-tenth of the labelled dataset that older hardware-locked vision platforms specify. The model learns which visual features correlate with defects, not which pixel values differ from a reference.
When process conditions change within the variation the model has seen during training, it continues producing accurate decisions without any intervention. When conditions shift outside training distribution — a new defect class, a new surface treatment, a significantly different lighting setup — targeted retraining on a few dozen boundary images brings performance back to spec. The previous model version stays on record for rollback.
On the Tier 1 auto-parts line, the transition to learned models reduced the reference-maintenance overhead because model updates are triggered by detected drift, not by calendar or change-order cadence. Retraining is event-driven, not scheduled.
Where learned models lose to reference matching
Before the decision table, the honest countercase.
A learned model only catches what it has seen. If a defect class is absent from the training set, the model will not flag it. Golden-sample matching, by contrast, flags any deviation from the reference — including failure modes the system has never encountered. For a line producing bespoke surface treatments in low volumes, where novel defect types are a genuine operational risk, the "unknown unknown" detection property of reference matching is a real advantage.
The practical threshold is whether the line generates enough reject samples per defect class to train a reliable model. Defect rates below 0.5% on short-run production may not accumulate sufficient labelled images in an acceptable timeframe. In those cases, reference matching holds the line while training data accumulates.
A Busan-based manufacturer running low-volume, high-variety precision components used reference matching for initial deployment, then transitioned specific high-volume SKUs to learned models as training data matured. The hybrid approach is an honest allocation of the two methods to the problems each solves best.
Decision table: matching the method to the line
| Criterion | Golden-sample matching | Learned defect model |
|---|---|---|
| SKU count | Low (under 30) | High (30+, growing) |
| SKU change frequency | Rare (less than quarterly) | Regular (monthly or faster) |
| Defect variety | Narrow (1–3 defect classes) | Wide (5+ defect classes) |
| Lighting stability | Controlled and validated | Variable or subject to change |
| Training data availability | Minimal — good parts only | Requires labelled defect samples |
| Unknown defect type risk | Detected — any deviation flagged | Not detected — trained classes only |
| Audit record | Reference library + version log | Model version + retrain audit trail |
| Initial setup time | ~30 minutes | ~60 minutes (setup + initial training) |
| Change event response | Re-photograph affected SKUs | Targeted sample addition, shadow validation |
| 12-month maintenance estimate | High on high-mix lines | Lower once initial training is complete |
Lines rarely sit cleanly in one column. A narrow SKU set with complex defect taxonomy, or a high-mix line with unusually stable lighting, falls in the middle. For mixed cases, a two-pass approach (reference matching as the first-pass anomaly detector, learned model as the adjudicator on flagged parts) reduces false positives from the reference matcher while preserving its broad-net sensitivity.
The maintenance cost vendors do not show in a pilot
A hardware-locked vision platform vendor running a six-week pilot shows you setup speed and detection performance. The pilot rarely shows you what happens in month seven when three SKUs change and two lighting arrays are replaced.
On a line with 200 active SKUs and quarterly engineering changes affecting 30% of the product range, the re-photography burden runs to roughly 60 reference images per quarter. Each requires the part staged under production conditions, the capture validated against current engineering drawings, the new reference approved by quality, and the outgoing version archived with an effective date. That is a recurring cost denominated in hours, not in capital.
The total-cost model over four years tells a different story than the pilot cost model. Reference-matching platforms appear economical when the pilot uses a stable, narrow SKU set under controlled conditions — the exact conditions under which the maintenance burden is invisible.
This is not an argument against reference matching. It is an argument for running the right cost model before committing to either approach.
How to audit your current golden-sample library
If your line is already running reference matching, a structured audit every six months reduces drift-related false positives.
For each reference in the library:
- Pull a current good part off the line under live production conditions.
- Compare the live part to the stored reference using the same illumination and focal settings.
- Record the pixel-deviation score under the current comparison threshold.
- If the deviation exceeds 5% of the alert threshold, flag for re-photograph.
- Log the audit date, part number, version, and auditor.
Libraries above 200 references benefit from a rolling audit protocol — segment the library into four groups and audit one group per quarter, so the entire library turns over annually. For each re-photographed reference, retain the previous version with its retirement date for audit-trail purposes.
This protocol maps directly to ISO 9001 clause 7.1.5 (monitoring and measurement resources) and to first-article inspection documentation requirements common in automotive and electronics supply chains.
If the audit reveals drift that has been accumulating undetected, the appropriate response is not an emergency re-photograph of the entire library. Start with the highest-volume SKUs and the defect classes that have generated customer escapes or internal quality holds in the past 12 months. A targeted 20-reference update corrects the most consequential drift without burning a week of quality engineer time on stable references that have not changed.
Separate the audit cadence from the engineering change order process. ECO-triggered re-photographs should happen within 5 working days of the ECO taking effect. The six-month audit is for drift that accumulates without a discrete change event. Conflating the two creates both gaps (ECO changes not actioned promptly) and redundancy (stable references audited more often than necessary).
Frequently asked questions
What is golden-sample machine vision?
Golden-sample machine vision compares each incoming part image against a stored reference image of a known-good part. Any deviation above a configured pixel threshold triggers a reject flag. The method requires no training data beyond the reference image itself and can be deployed within 30 minutes for applications with a stable product range and controlled lighting.
When does a golden sample become invalid?
A golden sample loses validity when lighting conditions shift beyond the tolerance used during capture, when the production process or tooling is updated within engineering specification, when a new SKU variant is introduced, or when normal process drift moves surface characteristics outside the original reference band. There is no fixed expiry interval; regular validation against current production parts is the accepted practice.
How many images does a learned defect model require?
HyperQ AI Vision initialises a learned model on 1,000 images — a representative mix of good parts and labelled defect examples, with boundary cases at the edges of tolerance adding the most value per image. Legacy hardware-locked vision platforms typically specify 10,000 or more labelled images before deployment. The 10x reduction in required training data shortens the time from sample collection to live deployment.
Can a learned model detect defect types it has never seen?
No. A learned model classifies based on patterns in its training data. Defect classes absent from training are not detected. This is the primary argument for retaining reference matching as a first-pass check on lines where novel failure modes are an operational risk. The two methods are complementary, not mutually exclusive.
How does retraining work when a process changes?
The standard procedure is shadow-mode validation: the candidate new model version runs in parallel with the production version, logging its decisions without affecting line output. After 24–72 hours of shadow operation, the two versions' decisions are compared on a held-out sample set. If the new version meets the acceptance criteria (typically parity or improvement on catch rate and a defined ceiling on false-positive rate), it is promoted. The previous version is retained for rollback.
What change-control record should a model update generate?
The minimum record includes: the trigger for the update, the date and batch range of images added to training, the shadow-mode duration and comparison outcomes, the acceptance criteria applied, and the name and role of the approving engineer. This maps to ISO 9001 clause 8.5.6 (control of changes) and, for medical-device supply chains, to 13485 design-control documentation requirements.
For guidance on building the sample set that feeds into either method, see what goes in the sample box. For the change-control procedures that keep production vision models audit-ready, see change control for production vision models. For the HyperQ AI Vision product overview, see the AI Vision solution page.
Send us a sample box — 50 good parts plus examples of the 3 most common reject types from your current line, captured under production lighting conditions. We will run both reference matching and a learned model on the same samples, document the detection and false-reject rates for each method, and return a written comparison within 10 working days. No contract until the recommended approach passes your own acceptance criteria on your own parts.