Industrial defect image datasets: anomaly detection, annotation and acceptance

A defect definition must match the inspection decision

An industrial anomaly is a deviation from an expected appearance; a defect is a condition that violates a defined requirement. These concepts can overlap without being identical. A harmless surface variation may look unusual, while an unacceptable internal failure may not be visible in an RGB photograph. State which decision the model should support and which evidence establishes the label. Do not force an image annotator to infer manufacturing quality that is not observable or supplied by an appropriate source record.

Define product families, materials, surface finishes, defect mechanisms and severity thresholds. Scratches, dents, contamination, cracks, wrong placement and missing components need different capture and annotation policies. A collection of dramatic defects can be inadequate if real failures are subtle or rare. Include acceptable variation and ambiguous boundary cases. For a process with several camera views, preserve the relationship to the physical part so the buyer can analyze whether the same defect is visible consistently rather than counting each view as an independent part.

What established benchmarks can and cannot show

MVTec AD is a useful industrial-anomaly reference with normal training examples and defect evaluation images across object and texture categories. The original release paper gives 5,354 images. Crucially, the published dataset terms are noncommercial. It should be described as a benchmark with its own restrictions, not as a stock source Fiund can freely relicense for a production system.

A benchmark helps compare methods under a documented protocol, but real inspection depends on a particular part, optics, lighting and quality criterion. A metal scratch photographed under diffuse illumination can become nearly invisible under another setup; a reflective highlight can resemble a defect. Review whether the source captures the buyer’s process and whether labels come from visual annotation, expert judgment or a downstream physical measurement. These forms of truth should remain distinguishable even when they agree on an initial sample.

Likely buyers and process-specific source routes

Potential buyers include machine-vision integrators, manufacturers, inspection-product teams and industrial research groups. They may need a normal-only reference set, labeled defects for supervised learning, a deployment-matched evaluation set or a comparison across sites. These use cases require different balances of ordinary and rare material. A team with an existing model may value carefully documented failure cases more than another large pool of easy examples.

Potential sources include original inspection records, authorized production captures, specialist collection teams and research datasets under their own terms. Existing factory archives may preserve natural prevalence but lack consistent labels. Commissioned captures can control lighting and defect classes but risk producing artifacts or staged examples unlike naturally occurring faults. Ask which defects are authentic process outcomes, deliberately introduced examples or simulated modifications. Keep those origin fields explicit rather than mixing them into one undifferentiated positive class.

Capture specifications determine whether pixels contain the answer

Describe the smallest defect of interest in physical terms and connect it to projected image detail. The relevant requirement is not simply a large file dimension: field of view, working distance, focus, lens distortion and material reflectance determine what the defect looks like. State whether the camera is fixed, whether parts move, whether exposure is controlled and whether multiple illumination states are required. Preserve the original bit depth or color information when it affects inspection.

Document product orientation, surface coverage, line speed and any trigger relationship where available. If the source cannot provide calibration, mark that limitation. A collection intended to compare spatial measurements may need scale information that a general anomaly model does not. Compression, resizing and sharpening can erase fine scratches or create edge artifacts. Review originals and the exact model-input derivatives together, and define which transformations are acceptable before the data is prepared.

• Preserve physical-part, lot, line, machine and capture-session groups when available.
• Record camera view, illumination state and product orientation.
• Separate genuine, deliberately introduced and synthetic defect origins.
• Keep visible labels separate from laboratory or downstream quality measurements.
• Describe normal appearance variation and ambiguous borderline cases.
• Record original image dimensions, processing and minimum observable defect detail.

Annotation and adjudication

A defect taxonomy should include operational examples and an uncertainty policy. For a scratch, define whether the label covers visible discoloration, estimated complete extent or only the region that violates the quality requirement. For contamination, specify how translucent or low-contrast boundaries are handled. For missing components, a conventional foreground mask may be inappropriate because the target is an absence. Different annotation types can be useful, but the export should match the decision and review criteria.

Expert disagreement can reveal an unclear specification rather than a careless annotator. Retain initial labels, adjudicated outcomes and the rule change that resolved a recurring ambiguity. If a quality label comes from a nonvisual test, preserve the relationship and any timing or matching uncertainty. A photograph of one part should not inherit a batch-level failure label without explaining that assignment. Buyers need to know whether the dataset supports image-level prediction or only association with a broader process outcome.

What to request in the record schema

A practical record connects source image, physical object, process context, annotation and permission reference. Machine or supplier identifiers can be pseudonymous when the model task does not require public disclosure. Keep units and missing-value meanings explicit. If a field was reconstructed from folder names, distinguish it from directly supplied process metadata. The purpose is to support reproducible grouping and evidence review, not to imply a factory integration that has not been verified.

Pricing, accepted yield and a worked comparison

No defensible public average commercial license price for industrial defect imagery is established here. Costs can include facility access, collection setup, rare-defect acquisition, source rights, process confidentiality, specialist labeling and validation. A per-image quote is hard to interpret without the number of unique parts and defect instances. A batch with abundant normal frames but few meaningful defects should not be priced or evaluated as though all files provide equivalent supervised evidence.

For a hypothetical annotation workload, suppose 5,000 selected images require two minutes of mask work each and independent review adds another thirty seconds per image. That is about 208 hours. At a chosen loaded rate of $36 per hour, the modeled labor is $7,500 before access, rights, capture, industrial expertise, rework or delivery. These are explicit planning assumptions, not observed rates or a Fiund quote. Complex boundaries, tiny defects and expert adjudication can change both time and yield.

Consider two offers: one has ten views of each of a small number of parts; another has fewer views but more lots and independently occurring defects. Normalize the quotes to unique parts, accepted defect instances, normal-condition coverage and included annotation. Ask whether failed-quality captures and repeated near-identical frames are replaced. Define whether new defect classes trigger additional instructions, collection or expert review. This makes price comparisons useful without inventing a market-wide average.

Sample acceptance and production-relevant splits

Review normal and defect examples across materials, lots, views, lighting and severity. Include acceptable variation that visually resembles defects, and subtle faults close to the decision boundary. Inspect pixel detail at the intended model-input size. A batch can be technically sharp yet fail because the relevant surface is never visible. Separately check image integrity, defect evidence, annotation quality, coverage and source documentation. A single overall pass percentage can conceal a critical missing defect family.

Keep views of one part together and consider holding out lots, time periods or machines to reflect intended deployment. If the product must transfer to a new line, an evaluation split sharing the same background and lighting may answer the wrong question. Preserve evaluation examples that were not used to revise the labeling rules. Report limitations explicitly: absence of a rare failure mechanism is unknown coverage, not evidence that the model handles it.

Observed commercial supplier range: what it does and does not price

As observed on November 12, 2026, LXT’s anomaly-dataset page advertises custom projects from $15,000 to $100,000 or more. This is one vendor’s published project guidance, not a representative market average, an independently verified transaction or a Fiund quote. The actual source, defect coverage, permissions, capture work and review need a selected proposal. It cannot be converted into a reliable per-defect rate by dividing by one of the example counts on the page.

Use the advertised range as a route for asking better questions. Does the proposed source contain naturally occurring faults or controlled examples? How are severity and defect boundaries established? Are accepted unique parts, normal coverage, masks, rework and confidentiality handling included? A price that bundles collection and annotation should not be compared directly with an archive-only license. Obtain a project-specific scope and validate a sample instead of treating a vendor’s general quality or compliance wording as an independent audit.

Commercial context and a Fiund inquiry

Public benchmarks and original industrial captures are complementary routes. A benchmark may be useful for method selection while custom sources address the actual process. No verified industrial-defect-image AI-training deal amount is established by the reviewed sources, and general visual licensing announcements are not reliable process-specific comparables. A source’s inspection-software price is also a different product from a dataset license.

Send Fiund the part family, inspection decision, defect definitions, source styles, capture conditions, minimum visible detail, grouping needs and intended uses. Include whether a representative sample must contain naturally occurring defects, measured labels or multiple views. Fiund can assess source fit and the questions that need resolution before a project proposal. Access, confidentiality handling, rights, preparation, acceptance, delivery and pricing are confirmed for the selected collection; no factory-specific availability is implied by this guide.

Questions buyers ask