Short answer
Textile and fabric defect data comes from mills, finishers and apparel makers that inspect rolls or finished garments, often with line-scan cameras or on inspection tables. Subtle flaws in repeating texture make labels and roll-level splits critical. Position on the roll, fabric construction and grading outcome should accompany each image. The data supports inspection models, not garment-handling robots.
The buyer's problem
A fabric is a surface made of a repeated structure, and its defects are deviations from that structure: a missing thread, a thick yarn, a stain, a pattern that drifts. Many are only a few pixels wide in an image of a roll several meters across. Others are faint enough that two trained inspectors disagree. That combination of repetition and subtlety is why textile sets behave differently from object-centric defect sets.
A buyer should therefore ask what kind of fabric, what kind of defect and what grade of reader produced the labels, before asking how many images exist.
What defects look like by construction
Defect vocabulary depends on how the cloth is made, so ask the supplier to state its construction types.
- Woven. Broken or missing ends or picks, floats, double threads, reed marks, slubs and knots, uneven density.
- Knit. Dropped stitches, holes, laddering, needle lines, barre (periodic streaks), contamination.
- Nonwoven and technical textiles. Thin spots, holes, lumps, fiber clumps, coating irregularities.
- Printed or dyed goods. Misregistration, color streaks, shade variation across the roll, print skips, stains.
- Finished garments. Seam puckering, skipped stitches, loose threads, stains, misaligned patterns, wrong labels or trims.
Printed patterns raise a further issue: a repeating motif hides anomalies, and a deliberate design element can look like a flaw to a model that has only seen plain cloth. Plain and patterned goods should be separate strata in any split.
A public reference, and its limits
The MVTec Anomaly Detection Dataset has 5,354 high-resolution color images across 15 categories, five of them textures, with pixel-precise ground truth for all anomalies, and a training set that contains only defect-free images (Bergmann et al., 2021). It is licensed under CC BY-NC-SA 4.0.
It is a reference that does not cover textile-mill production: its texture categories are not roll-inspection images of woven or knitted goods. Use it only as a capture, labeling and format reference: it shows how a defect-free training protocol and pixel-level masks are defined for texture-like surfaces. Real fabric sets differ in scale, camera, yarn, weave and defect population, and the non-commercial license is a reason to read terms before any commercial use; this is not legal advice. How the defect-free protocol compares to supervised learning from labeled flaws is covered in anomaly detection vs supervised defect data.
Where the images come from
| Supplier | Typical asset | Gap to expect |
|---|---|---|
| Weaving and knitting mills | Line-scan images from loom or inspection frame | Defect locations logged by meter, not image coordinates |
| Dyeing, printing and finishing houses | Color and print inspection captures | Shade standards held outside the images |
| Technical-fabric and nonwoven producers | Web-inspection archives | Large volumes, mostly clean cloth |
| Apparel makers | Photos from end-of-line garment checks | Inconsistent framing and lighting |
Continuous web captures behave like line-scan and web inspection data: one very long image cut into tiles. Ask how the tiling was done, because a tile boundary can split a defect.
Labels, position and grading
Position on the roll is the key. A mill that records "defect at 143.2 m, 0.6 m from the left selvage" has a label that can be mapped to a pixel region, provided the encoder or timestamp is archived. Without it the image is a picture and the defect log is a separate spreadsheet.
Many mills grade rolls with point-based systems or customer-specific rules, and a defect that earns points on one program may be tolerated on another. Capture the rule set and date. Subtle flaws also produce borderline samples, so ask whether doubtful calls were flagged, and whether more than one grader saw the same cloth; see inter-inspector agreement. The reference condition for a quality level is covered in golden samples and reference images.
Illustrative fabric-defect package schema
# Illustrative only - synthetic structure and values
package: fabric_defect_v0
source: woven_mill_site_007
roll_table:
roll_id: roll_55021
construction: woven_plain
finish: printed # plain | printed | dyed | coated
width_cm: "recorded per roll"
grade_rule: "mill rule set R3, dated"
grade_result: second_quality
captures:
- tile_id: roll_55021_t0418
meter_start: 143.0
meter_end: 144.0
camera_id: linescan_B
tile_overlap_px: recorded
defect_marks:
- {type: missing_pick, box: [x, y, w, h], grader: g_04}
split_key: roll_id # buyer may split by loom or lot instead
excluded: robot actions, joint states, forces
The aim is to be explicit about tile overlap, roll provenance and the grader. None of these values is a requirement; the buyer sets them.
Quality and acceptance pointers
- Defect visibility. Check a sample of labeled flaws at native resolution; a mark that an expert cannot see in the image is a label problem.
- Clean-cloth coverage. If the model is trained on normal examples, the normal set must span shade lots, fabric batches and lighting drift.
- Split by roll. Adjacent tiles of one roll share texture and lighting, so tile-level shuffling inflates scores.
- Class coverage. Count defect types in the delivered set and check whether the types you care about are present at all.
- Resolution. Confirm pixel size against the smallest defect you need to find.
Targets for each of these depend on your model and tolerance, and none has a universal minimum. For common failure modes in small defect classes, see rare defects and class imbalance.
What textile inspection data cannot support
Images of fabric tell you what the cloth looked like at the camera; they do not teach a robot to pick, lay out, feed or fold it. Garment handling is a deformable-object manipulation problem that needs robot or teleoperation data with actions and contact, not archive images or grading logs. Inspection images also say little about wear performance, shrinkage or dye fastness unless those test results are joined separately.
Mill data is mostly clean cloth, so rare flaws are underrepresented, and seasonal products mean the pattern set changes quickly. Models can fail on a new weave or yarn, so hold out by fabric construction and not only by roll when generalization matters.
Questions to ask suppliers
- Which fabric constructions, finishes and pattern types does the archive cover?
- Was imaging done by line-scan, area camera or hand photography, and did it change?
- How are defect positions stored, and can they be mapped to pixels?
- Which grading system or customer rule set defined acceptance, and does it vary by program?
- Who graded the cloth, and were graders independent of production?
- Does the archive include clean rolls, or only flagged ones?
- Were images tiled, compressed or downsampled, and how much do tiles overlap?
- Do any images show customer-owned prints or branded patterns?
- Are second-quality and returned goods linked to their images?
Permissions and provenance
Printed fabrics may embody a brand's artwork, and mills often make goods to customer specifications under contracts that limit the reuse of records. Ask who owns each design, whether any contract restricts use of inspection images, and whether garment labels or brand marks appear. This is not legal advice, and redaction of marks does not on its own resolve permission questions. For the wider rights checklist, see confidentiality in defect data.
Discuss your requirements
If you need textile or fabric defect imagery, describe the construction type, defect classes, labeling level and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources in the US; a request does not guarantee a match. Continue with the inspection data hub, the humanoid training data guide, or the catalog entry for images and inspection photos.
Explore sourcing and licensing options with SourceX
Sources
- The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — public anomaly dataset with texture categories, used as a capture and labeling reference only.