Short answer
Web inspection data comes from line-scan cameras that image continuous material such as paper, film, metal strip or textile, producing very long images in which defects are rare. It can support defect detection and tiling-aware evaluation. Because it contains no robot actions, it cannot train manipulation, and small public crops do not represent a full production stream.
What is different about continuous material
An area-scan camera takes a picture of a part. A line-scan camera captures one line of pixels at a time while the material moves past it, and software stacks the lines into an image whose length depends on how long the run lasts. A roll of film or a steel strip can therefore produce an image thousands or millions of lines long, with no natural frame boundary.
Three consequences matter to a buyer. First, defects occupy a tiny fraction of the total area, so the data is extremely imbalanced. Second, the scan direction and the material speed fix the geometry: if the speed varies, defects stretch or compress along the length. Third, the relevant context for a defect is often hundreds of meters upstream, such as a roller mark that repeats at the circumference of a roll.
This page covers sourcing such data. For discrete-part image archives, see in-line machine vision archives. The cluster is introduced in the inspection, defect and failure data hub, and the full map is in the humanoid training data guide.
Who may hold it
Producers of paper, plastic film, nonwovens, sheet metal, coated foil and technical textiles often run continuous inspection because manual inspection cannot keep up with the line. The inspection system typically logs defect events with a position along the web, a class from the system's own taxonomy and an image snippet. Whether the full raw scan is kept, or only snippets around detections, varies. Ask early, because snippet-only archives cannot support evaluation of missed defects.
Closely related material pages: metal surface defects and textile and fabric defects.
A public contrast: small crops versus long streams
The NEU surface defect database contains 1,800 grayscale images of six hot-rolled steel strip defect types, with 300 images per class and a size of 200 by 200 pixels (Song, n.d.). Those are small, class-balanced crops. A production stream differs in kind: it is continuous, largely defect-free and not pre-cropped around each defect.
So the dataset is useful as a format and labeling reference, for instance to see how classes can be organized per defect type and how crops are stored. It does not show the scale, imbalance or false-alarm environment of a real web inspection archive, and its balanced class counts should not be read as typical.
What web-inspection data can support
- Defect detection in long scans, with tiling or sliding-window strategies chosen by the buyer.
- Evaluation under realistic imbalance, if the archive includes long clean stretches and a record of true defects. See defect detection evaluation sets.
- Repeat-defect analysis, where the same mark appears at regular intervals.
- Segmentation, if masks exist; see pixel-level defect segmentation.
- Anomaly detection trained on clean material.
Limits: what this data cannot support
It cannot train a robot to handle material. Threading a web, splicing a roll, changing a reel or moving a coil requires robot or teleoperation data with joint states and actions, and often force sensing. Inspection images alone give none of that.
It does not generalize across materials. A model trained on steel strip may not transfer to film, because texture, reflectance, defect geometry and scan geometry differ. Even across two lines of the same material, lighting and camera setup shift appearance; see domain shift across lines.
Defect labels reflect the inspection system's rules. If the archive records only what the system flagged, the buyer cannot know what it missed. Classes may also reflect process vocabulary such as "pinhole" or "streak" without definition.
Images may be compressed heavily or reduced to defect snippets to save storage, which removes context.
Illustrative archive specification
The following is an illustrative specification with synthetic values, not an existing dataset.
| Field | Example | Why it matters |
|---|---|---|
| Raw scan | roll_0192, 8192 px wide, continuous, lossless | Needed for tiling choices |
| Scan geometry | line rate, material speed, pixel pitch | Fixes defect shape and size |
| Defect events | position_m, width_px, class, system confidence | Index into the scan |
| Review label | confirmed / false alarm / missed (by later review) | Separates system output from truth |
| Context | roll ID, product grade, line, shift | Supports splits |
| Disposition | cut out / downgraded / passed | Links defect to outcome |
| Reference | any golden or clean-run segments | Baselines for anomaly models |
Quality and acceptance criteria
Thresholds belong to the buyer. Tests that expose weak archives: confirm raw scans, not only snippets, exist for at least some rolls; check that "false alarm" and "missed" categories are recorded by someone other than the original system; confirm that scan geometry metadata allows physical size recovery; verify that rolls are the unit for splits so adjacent tiles never cross train and test. See rare defects and class imbalance on handling sparse positives.
An evaluation design can use arbitrary numbers for planning; for example, if a roll image were 100 times longer than it is wide, a tile of a given size would give a count of tiles per roll to budget for. Such inputs are arbitrary illustrations, not minimums.
Permissions and provenance
Web images can reveal product grades, coating patterns, security features or customer-specific specifications. Contractual confidentiality with customers is common for coated and printed materials. Ask whether the material is a customer's proprietary product, whether any image shows printed text or marks, and who can authorize licensing. See confidentiality in defect data. This is not legal advice.
Questions to ask suppliers
- Are full raw scans retained, or only snippets around detected defects?
- What are the line rate, pixel pitch and material speed, and do they vary across the archive?
- How were defect classes defined, and has the taxonomy changed?
- Who reviewed detections, and are false alarms and misses recorded?
- How much of the archive is defect-free, and has it been checked?
- Can scans be exported without lossy compression?
- Can position along the web be tied to roll, lot and disposition?
- Do images show proprietary patterns or printed content owned by a customer?
- Can a sample roll with both clean and defective segments be reviewed in advance?
Discuss your requirements
If you need continuous-material inspection data, describe the material, scan format and evaluation needs to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX follows up about potential sources. See also the catalog entry on images and inspection photos.
Tell SourceX what your evaluation set needs to cover
Sources
- NEU surface defect database — hot-rolled steel strip defect crops, used to contrast crops with continuous streams.