Inspection Images Linked to Dispositions and Outcomes

From image folders to labeled data

A folder of images from an inspection station is not labeled data until you know what was decided about each part. Public benchmarks arrive with curated labels: MVTec AD, for example, provides pixel-precise ground truth for all anomalies across more than 70 defect types (Bergmann et al., 2021). A plant's labels are by-products of a workflow: a threshold crossed, a code chosen at a station, a form closed by an engineer. Each is a different kind of label, made by a different actor at a different time.

This page is about attaching those decisions and knowing what they mean. For what the camera archive itself holds, see in-line machine vision reject archives; for general sourcing and specification, see the guide to sourcing defect image datasets.

The records that have to be joined

Quality-system records are covered in nonconformance reports, and late outcomes in warranty and field failure data. A plant may run its quality system to ISO 9001, whose requirements are generic, cover documented information, operation, performance evaluation and improvement, and whose certification is optional (ISO, 2026). A quality system tells you records exist in some form, not that they contain the fields you need.

What a disposition label means, and what it does not

A disposition is a business decision, not a description of the image. A part with a visible scratch can be accepted because the customer specification allows it, and a part with no visible flaw can be scrapped for a dimensional or functional reason that the image never shows. Training on the decision teaches a model the plant's acceptance rules; training on what an expert sees in the image teaches appearance. Which you want depends on the job, so ask for both as separate fields: defect_observed, decision and reason_basis (cosmetic, dimensional, functional).

Boundary cases such as use-as-is are worth keeping and flagging; see defect severity and disposition labels and borderline samples.

Weak labels at scale, precise labels where it counts

Dispositions are image-level labels. They rarely come with masks. Božič et al. found that mixing a few pixel-labeled samples with weaker labels performed comparably to full supervision at much lower annotation cost (Božič et al., 2021). The practical reading for a purchase: buy dispositions across the whole archive, then add masks to a subset chosen for class coverage and borderline cases. Mask specification is in pixel-level defect segmentation, and old labels may need relabeling first.

Joins that fail quietly

• Unreadable serials. Parts without a readable serial were inspected but cannot be joined. Ask the unmatched rate and whether unmatched images are a random subset or a particular line or defect type.
• Clock offsets. When a join uses time windows, a small offset between station and MES clocks attaches a decision to the neighboring part.
• Re-inspection. One part can produce several images. The final disposition belongs to the part, not to each image, so keep part-level and image-level labels distinct.
• Selection. Dispositions are usually opened for rejects, so passes have none. Absence of a record is not evidence of a good part, particularly when lots are accepted by sampling: ISO 2859-1 describes acceptance sampling by attributes using single, double and multiple sampling plans (ISO), so unsampled units may carry no inspection record. See AQL sampling context.
• Late escapes. Parts that passed and failed later appear only if returns or warranty data can be tied back to serials.
• Changed code lists. Disposition codes are redefined after system migrations or standard changes. Ask for the dated history.

What to specify

• Part families, defect types and date range.
• Decision sources required: machine call, station call, final disposition, downstream outcome.
• The disposition vocabulary with definitions, and the reason basis where recorded.
• Join keys and the minimum join rate per link, with unmatched records delivered separately.
• Whether passing images and false rejects are included, and the rule for which passes were kept.
• A mask subset: size, selection rule and label type.
• Split key: serial, then lot, line and date, so all images of a part stay together; see leakage in defect datasets.

Illustrative join audit

A supplier's sample extract, with synthetic figures, shows why join tables should be read before image counts.

Read the last three rows together. A buyer who counts disposition records alone sees under half of the rejects, and one who assumes every reject has a disposition builds a biased label set.

Quality and acceptance criteria

1. Join rates are reported per link and per period, with unmatched records listed and characterized.
2. Every label carries its decision source, and the vocabulary and its change dates are delivered.
3. Image-level and part-level labels are kept apart, and one-to-many relationships are explicit.
4. An independent expert re-reviews a random sample from each disposition class, and the agreement between defect_observed and the disposition is reported per class; see inter-inspector agreement.
5. All images of a serial and lot sit in one split.
6. Passing images and false rejects are included, with the keeping rule documented.
7. Times are in one zone with documented offsets between systems.

What this data cannot support

Joined inspection images support training and testing of defect recognition and, where dispositions are reliable, of acceptance decisions. They do not show how a humanoid should pick up, turn or present a part, and they come from fixed cameras rather than a robot's view; see active inspection by humanoids. A decision rule learned on one line does not carry to another without new data, as domain shift across lines explains. And when the reason for a disposition was never recorded, the join cannot supply it.

Permissions and confidentiality

Disposition and nonconformance records can reveal customers, suppliers, costs, root causes and engineers' or inspectors' identities, in addition to the product design visible in the image. Where parts are built to a customer's design, the customer may need to agree before licensing. Treat it as a diligence item at the start; confidentiality in defect data lists the questions. Pseudonymizing inspector IDs reduces exposure but does not by itself settle permission.

Questions to ask suppliers

1. Which systems hold the images, production records, inspector calls and dispositions, and what key joins them?
2. What share of images has a readable serial or lot, and are the unmatched images random?
3. Is a disposition record created for every machine reject, or only for escalated ones?
4. How are false rejects recorded: deleted, tagged or absent?
5. Do disposition codes separate cosmetic, dimensional and functional reasons?
6. Were code lists or acceptance standards changed during the period, and when?
7. Are passing images retained, and under what sampling rule?
8. Can returns or warranty claims be tied back to serials?

Discuss your requirements

If your inspection model needs images tied to real decisions, describe the parts, decision sources and label types to SourceX. SourceX looks for US businesses that hold suitable inspection records, checks the data and the supplier's licensing permissions, and coordinates pricing, allowed uses and delivery in a license. SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also arrange annotation, scoped case by case. SourceX reviews each request and follows up about potential sources; a request does not guarantee a match.

See also the inspection, defect and failure hub, the humanoid training data guide, and the catalog entries for manufacturing quality and inspection records and inspection reports.

Explore sourcing and licensing options with SourceX

Sources

• The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — curated pixel-level labels as the contrast to production labels.
• Mixed supervision for surface-defect detection: from weakly to fully supervised learning — a few pixel labels mixed with weaker labels.
• ISO 9001:2026 Quality management systems - Requirements — generic requirements, documented information and optional certification.
• ISO 2859-1 Sampling procedures for inspection by attributes - Part 1 — acceptance sampling plans, so unsampled units may have no record.