Short answer
Steel and metal surface defect data comes from public benchmarks and from mills, processors and parts makers. Public sets such as NEU (1,800 images, six defect classes) are small and carry research licenses; production archives hold more variety but need rights review, relabeling and metadata. Neither shows robot handling, and benchmark scores rarely transfer to a new surface, finish or camera.
Why metal surfaces have their own data problem
Metal inspection mixes two situations that need different data. Flat products such as hot-rolled strip, sheet and plate are scanned continuously, and defects are texture-like: scale, pits, scratches, inclusions. Machined and fabricated parts are inspected as discrete units with geometry, edges, holes and finish variation, and defects include scratches, dents, burrs and contamination.
Because reflectance and finish dominate appearance, the same defect looks different under different lighting and surface treatment. That is why public benchmarks are valuable for method comparison and limited for deployment. The overview of image dataset sourcing is in sourcing defect image datasets; this page adds the metal-specific route. Cluster context is at inspection and the humanoid data section.
Public benchmarks and what they contain
| Dataset | What its publication supports | Typical use | Watch for |
|---|---|---|---|
| NEU surface defect database | Six hot-rolled steel strip defect types: rolled-in scale, patches, crazing, pitted surface, inclusion, scratches; 1,800 grayscale images; 300 per class; 200x200 pixels (Song) | Classification baselines | Small images, balanced classes unlike real rarity |
| KolektorSDD2 | Over 3,000 images with several defect types; evaluated alongside KolektorSDD, DAGM and Severstal (Božič et al., 2021) | Segmentation and weak-supervision studies | Single product family |
| MPDD | Metal-part images captured under complex production conditions with anomaly-free training images and pixel-precise masks in six categories: bracket_black, bracket_brown, bracket_white, connector, metal_plate, tubes (Jezek et al., 2021) | Anomaly detection on parts | Few categories |
These sets are useful as baselines and for checking a pipeline. They are not a substitute for data from your own material and finish. A model trained on 200x200 grayscale crops of strip surface has limited relevance to polished machined brackets, and a model that scores well on a balanced benchmark may behave differently when defects are rare.
The license matters. Anomalib documentation states that MPDD is released under CC BY-NC-SA 4.0 (Jezek et al., 2021), and the MVTec AD benchmark carries the same license (Bergmann et al., 2021); a non-commercial license restricts commercial training use. For NEU, check the host page terms before relying on it. This is not legal advice, and a commercial buyer should read each license.
What metal defect data can and cannot support
It can support defect detection, classification and segmentation models, anomaly detection on nominal-only data, and evaluation of inspection systems on specific materials. It cannot show handling: how to present a reflective part to a camera, or how a gripper marks a surface. It also does not show downstream consequence; a visible scratch may be cosmetic or may be rejectable depending on the specification, so images need the disposition link.
Surface defects on metal also depend on process stage. A scale mark before pickling is not the same as after. Record the stage.
Preparing metal imagery for use
Before a metal image set is ready for training or evaluation, expect several steps from the data holder or your own team.
Tile or crop with care. Continuous strip images are cut into tiles. Tiling that splits a long defect across tiles creates partial labels. Decide whether labels are per tile or per defect and keep the mapping to the original coil position.
Separate by material identity. Images from the same coil, heat or machining batch share texture and noise. Keep them in one split to avoid leakage; see leakage in defect datasets for the general mechanism.
Relabel to your classes. A producer's codes reflect its customers' specifications. Mapping them to your list may merge or split classes, and the mapping should be documented.
Record surface condition. Oil film, scale, rust bloom and handling marks change appearance without being defects. Ask whether they were annotated as distractors.
Private supply: who holds metal defect imagery
- Steel and aluminum producers and processors hold continuous-line inspection archives from surface inspection systems, often line-scan; see line-scan and web inspection.
- Machine shops and fabricators hold part photos from first-article and final inspection.
- Foundries and forges hold casting and forging images; see casting and molding defects.
- Coaters and finishers hold images of paint and plating defects on metal; see painted and coated surface defects.
- Service firms with 3D scanners hold geometry-based records; see 3D surface inspection data.
Illustrative package for a metal defect request
Illustrative; values are synthetic.
| Element | Example specification |
|---|---|
| Material and finish | Cold-rolled steel, oiled; or machined aluminum, bead-blasted |
| Defect classes | Buyer's list of 6-10, mapped to supplier codes |
| Image type | Line-scan strips or area-scan tiles, 16-bit if available |
| Labels | Image-level class plus boxes; masks on a held-out subset |
| Splits | By coil, lot and week; no shared parts across splits |
| Metadata | Lighting, camera, scale, line speed, process stage |
| Normal data | Defect-free tiles from the same lots |
| Disposition | Accept, downgrade, scrap, with reason |
Evaluating an offer
Distribution realism. Ask for the real class frequencies. A set with equal counts per class has been curated, which is fine for training classifiers but hides the imbalance you will meet in operation; see rare defects and class imbalance.
Label authority. Who judged the defect, with which specification? Surface defect names vary between mills.
Cross-line behavior. Hold out an entire line or coil batch and measure the drop. Performance differences across lines are routine; see domain shift across lines.
Cleanliness. Oil, dust and fingerprints produce false positives. Ask whether cleaning state is recorded.
Limitations and extra preparation
Metal surface imagery is tied closely to the material, finish and optics that produced it. A dataset of hot-rolled strip says little about machined aluminum or brushed stainless steel, and glare, oil films and coolant residue change appearance from one shift to the next. Class definitions vary between plants: what one quality team calls a scratch another calls a handling mark, and severity thresholds are often local customer agreements rather than written rules, so check the applicable specification for each source. Defect classes are usually imbalanced, and the rarest ones are often the ones that matter most.
The data contains no information about how a robot should pick, orient or present a part, so it supports inspection and recognition, not manipulation. Plan for extra preparation: harmonizing the class list across suppliers, recording the finish and lighting per image, re-labeling a sample to check consistency, removing customer part numbers or stamped marks where they are sensitive, and splitting by part, coil or lot rather than by image so that near-identical views do not leak between training and evaluation.
Questions to ask suppliers
- Which surface inspection system or camera setup captured the images, and at what resolution and line speed?
- What material grade, finish and process stage do the images cover?
- How were defects labeled, and by whom, against which specification?
- What fraction of images contain no defect, and how were they selected?
- Can you split by coil, lot or production week?
- Do images show customer part geometry or markings that need approval?
- Are original, uncompressed images available?
- What dispositions and costs were recorded against defective items?
Discuss your requirements
If you need metal surface defect imagery from US manufacturers, describe the material, finish, defect classes and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates licensing and delivery. A request does not guarantee a match.
Tell SourceX what your evaluation set needs to cover
Sources
- NEU surface defect database — NEU database: six hot-rolled steel strip defect types, 1,800 grayscale images, 300 per class, 200x200 pixels.
- Mixed supervision for surface-defect detection: from weakly to fully supervised learning — KolektorSDD2 with over 3,000 images; evaluation on KolektorSDD, DAGM and Severstal.
- Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions (MPDD dataset) — MPDD: metal-part images in complex production conditions with pixel-precise masks, six categories.
- The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — MVTec AD anomaly benchmark license (CC BY-NC-SA 4.0) as a reference for research-license limits.