NEU Metal Surface Defects Database

A benchmark produced by the Surface Defects research group at Northeastern University (NEU), China, and the closest thing industrial computer vision has to a shared exam. It contains 1,800 grayscale JPEG images of hot-rolled steel strip surfaces - 300 for each of six typical defect classes: crazing (Cr), inclusion (In), patches (Pa), pitted surface (PS), rolled-in scale (RS) and scratches (Sc). Every image is 200x200 pixels and carries an XML annotation file giving bounding-box locations of the defects.

Two properties explain its persistence. First, dual task support: because both folder-level class labels and per-instance boxes exist, the same record serves classification experiments as-is and object-detection work through the NEU-DET variant, so papers measure at whichever grain their method needs. Second, honest difficulty: these six classes are the ones that resist easy separation under mill conditions - low-contrast crazing, texture-like patches, thin directional scratches - which is precisely why a strong score here means something on a real line.

Inside Datadory's steel pool it plays a different role than every tonnage series around it. The EUROFER European steel data and USGS iron and steel statistics count what the industry produces; this record shows what a defective square inch of that output looks like. Production volume and surface quality are different questions, and the steel data hub holds both.

One row per image, class label attached, every annotated region flattened beside it:

• crazing_1.jpg - class crazing, 200x200 grayscale, three defect instances with boxes such as [12, 45, 88, 130] and [101, 20, 176, 92]
• inclusion_23.jpg - class inclusion, two elongated vertical regions, boxes like [34, 8, 61, 197]
• scratches_10.jpg - class scratches, one thin horizontal band spanning nearly the full width

Upstream, none of this is a table: it is one JPEG plus one PASCAL VOC-style XML file per image, organized into class folders. Datadory delivers the normalized tabular shape above so detection frameworks ingest it directly, or ships the raw paired files untouched if your pipeline expects the original layout. The mono block near the top of this page shows the exact column order; request a sample and real images and annotations arrive cut to the classes you name.

Four fields form the confirmed core, verified during the August 2026 research pass against the archive's own structure. Three details matter before anyone writes a loader. Class identity travels in the filename, not in a separate column, so any reshuffle of folders silently destroys the label unless it is parsed first. Bounding boxes are multiple per image - a single patch can carry several defect regions, which is what separates detection work from classification on this record. And the split field only exists in mirrored distributions: some repackagings add train/validation/test folders and YOLO-format labels, others ship the raw six-class set, so downstream reproducibility depends on knowing which variant arrived.

Everything derived. The base schema stays at four fields so the record remains comparable with the published benchmark; the views below compute from it and confirm alongside your sample when named:

• Per-class summaries - instance counts, mean boxes per image and bounding-area distributions pulled from the XML sidecars.
• YOLO-format labels regenerated from the VOC-style XML for frameworks that prefer normalized coordinates.
• Split manifests materializing train/validation/test assignment as a column, so folds survive reshuffling.
• Balanced sample manifests - n images per class, stratified and reproducible for controlled experiments.
• Confusion-pair flags marking the documented hard separations - crazing versus pitted surface, inclusion versus scratches - for error-analysis setups.

Name what your experiment needs when requesting the sample; nothing speculative is padded into the base schema.

Geography - not geographic. This is laboratory and production-line imagery of hot-rolled steel strip collected by Northeastern University (China); there is no country, region or facility column, and treating the record as a spatial dataset would be a category error. Its provenance is institutional, not territorial.

Temporal - static. The database was published circa 2010s and has not been updated since; that permanence is a feature for benchmarking, because results recorded today remain comparable with results recorded years ago. It is a liability for monitoring, where nothing new ever arrives.

Granularity - image level: 1,800 images, 300 per class, each with per-defect-instance bounding boxes. There is no roll-up layer, no daily aggregate, no hidden hierarchy - just the unit that vision models consume.

Against the wider catalog - mean quality score 7.81 across 1,744 datasets - this record scores 7/10: canonical status, complete annotations and dual-task support, held back by its static nature and the absence of production context beyond the images.

API, files, or your warehouse. Daily, weekly, or hourly.

Cadence matters less for a static benchmark than shape fidelity, so delivery is built around preserving it: images and annotations stay paired, class labels survive any packaging, and the normalized tabular view diffs cleanly against the raw layout if your pipeline switches between them. Take the full 1,800-image set once and version it internally, or keep a warehouse table of image paths, classes and boxes current for team-wide access - the schema holds either way. Name the classes and derived views you want and the sample arrives cut to that scope; it ships first regardless.

• Detection baselines - report mAP against numbers the field already knows, on a set no reviewer will question.
• Cold-start training - stand up a first-pass inspection model before plant labeling begins, then fine-tune on real lines.
• Augmentation research - attack the documented failure modes of crazing contrast and scratch thinness with controlled comparisons.
• Labeling-tool evaluation - measure annotator agreement and auto-label drift against fixed ground truth.
• Teaching - run a complete industrial-vision exercise, data to trained model, inside one course module.
• Domain transfer - pretrain on this benchmark and carry the six-class vocabulary onto your own strip, coil or plate imagery.

Data scientists and ML engineers (relevance 3/3) get the benchmark without mirror archaeology - see data scientists use cases for steel. Journalists, academics and students (3/3) get the citable standard for industrial-inspection scholarship (journalists academics use cases for steel). Market researchers and consultants (2/3) anchor machine-vision market sizing in the dataset vendors benchmark against (market researchers use cases for steel), and competitive-intel product teams (2/3) scope quality-control claims against the field's shared difficulty profile (competitive intel use cases for steel). Investors and quants (1/3) read the benchmark's citation persistence as an adoption signal (investors & quants use cases for steel), and sales and growth teams (1/3) borrow its vocabulary for concrete mill conversations (sales & growth teams use cases for steel).

No other record owns the canonical six-class steel-strip benchmark position, so the honest comparison is against neighbors measuring different things. Within Datadory's steel pool, USGS iron and steel statistics, EUROFER European steel data and monthly crude steel production across 71 countries are tonnage and trade series - they say nothing about what a defect looks like at pixel level. The UNSD commodity stats vs NEU metal defects comparison works through that contrast head-to-head. General industrial-anomaly collections offer broader category coverage but thinner per-class depth, and private mill datasets are richer in context yet unpublished and therefore useless as a shared yardstick.

If the question is whether a model can find and classify the six defect types the industry argues about, this is the yardstick. If it is how many tonnes moved through which market, use the tonnage series.

Three things, stated plainly. First, the record is static: published circa 2010s, never updated since, so it will never tell you anything new about your mill - its value is comparability, and we deliver it as the fixed reference it is. Second, it carries no production context: no line ID, no alloy grade, no process parameters - if your model needs those, pair this benchmark with plant-side data rather than expecting the benchmark to be it. Third, variants differ: classification sets, NEU-DET detection packaging and YOLO-labeled mirrors all circulate, and results are only comparable within one variant - tell us which one your baselines assume and the sample matches it. Name the classes and derived views you want; the sample arrives in exactly the schema shown above.