Short answer
Inspection image metadata should record lighting type and geometry, optics, sensor and exposure settings, working distance, fixture and part pose, calibration, and the line, lot and time of capture. These conditions decide whether a defect detector transfers to a different line, so request them with the images, and treat missing fields as a risk to quantify, not a formality.
Why imaging conditions decide transferability
A defect is not a property of the image alone. A scratch is visible under low-angle light and invisible under diffuse dome light; a dent shows in reflection but not in a flat overhead image. Two cameras photographing the same part can produce images that look like different domains to a model.
Benchmarks are explicit about this. MVTec AD 2 includes a test_private_mixed split captured under varied lighting conditions, and its scenarios stress small defects in large images, defects at the image border, high normal variability, transparent or reflective objects and overlapping bulk goods (Heckler-Kram et al., 2025). Real-IAD collects 150K high-resolution images across 30 objects of varied materials with multi-view capture (Wang et al., 2024). Both treat capture variation as a first-order factor in how well methods hold up.
A buyer evaluating industrial data therefore needs to know how each image was made. This page is the metadata specification. Context sits in the inspection cluster and the humanoid data section; the downstream effect is covered in domain shift across lines.
What the metadata can and cannot do
Complete metadata lets you stratify results by lighting and camera, decide which subsets match your deployment, build train and test splits that separate setups, and reproduce or approximate conditions in a later collection.
It cannot repair an image. Saturated highlights, motion blur or missing resolution stay lost. Metadata does not make data from one setup valid for another, and it says nothing about robot handling or action. Where a humanoid or mobile manipulator would capture images itself, its camera, pose and motion add variables that archived fixed-station images do not represent; see active inspection by humanoids.
Illustrative metadata specification
The YAML is Illustrative with synthetic values. Fields marked required are those whose absence most often blocks reuse; the list is a starting point for the buyer, not a standard.
# Illustrative image metadata record - synthetic values
image_id: img_000123
part_id: part_77Q
lot_id: lot_2026_041
line_id: line_C
station_id: st_3
timestamp_utc: 2026-03-02T08:14:55Z # required
camera:
model: area_scan_5MP_mono # required
sensor_px: [2448, 2048]
pixel_size_um: 3.45
bit_depth: 12
exposure_us: 800 # required
gain_db: 0
gamma: 1.0
trigger: encoder | hardware | software
optics:
lens_focal_mm: 25 # required
aperture_f: 5.6
working_distance_mm: 300 # required
telecentric: false
filters: [bandpass_630nm]
lighting:
type: ring | bar | dome | coaxial | backlight | dark_field # required
color_nm: 630
angle_deg: 30
intensity_setting: 70
strobed: true
ambient_shielded: true
fixture:
part_presentation: tray_fixed | conveyor | handheld
pose_repeatability_mm: 0.5
background: matte_black
calibration:
last_date: 2026-02-20
scale_um_per_px: 28.4 # required for size claims
flat_field_corrected: true
color_reference: none
processing:
stored_format: png_lossless | jpeg_q90 | raw
resized_from: null
crops_applied: false
Which fields matter most, and why
Lighting type, color and angle. This is the largest single driver of defect visibility. Record the geometry, not just a label such as "ring light."
Resolution and scale. Pixels per millimeter determine the smallest defect visible. If scale is unknown, defect-size labels cannot be tied to a tolerance.
Exposure, gain and bit depth. Saturation and noise limit what training can recover. Eight-bit exports of twelve-bit captures discard range.
Compression. Lossy JPEG at low quality blurs hairline defects. Ask what the original store was and what was exported.
Fixture and pose. Fixed fixtures yield aligned images that make nominal-only methods easy. Hand-held or conveyor-varied presentation yields pose spread that matches robot-collected imagery better.
Calibration status. Calibration date and flat-field correction affect cross-station comparisons.
Line, lot, shift and time. These are the keys for splits. Spread across lots and months reveals drift; a single shift can look consistent and generalize poorly.
Software changes. Auto-exposure or firmware updates mid-collection change the distribution without any visible event.
Three ways missing metadata misleads a buyer
Illustrative scenarios. A supplier delivers 20,000 images of machined brackets with strong results on a held-out split. The split was random within a single week, and every image was taken at one station with a ring light. Your line uses a bar light at a different angle. Without lighting and station fields you cannot tell, and you discover the gap only after integration.
A second supplier's images come from two lines whose cameras differ in pixel size. Defect size labels in pixels look comparable but correspond to different physical sizes. Without scale, a size threshold from the inspection standard cannot be applied.
A third archive was exported as lossy JPEG from a raw store. Fine scratches that appeared in the original are smoothed away, yet the labels still mark them. The label and the image no longer agree, and no model can learn the mark.
In each case the metadata is cheap to ask for and expensive to reconstruct.
Acceptance checks for a delivered set
- Field coverage report. Percentage of images with each required field populated; set a threshold per field.
- Consistency scan. Group images by metadata and compare simple statistics such as mean intensity and sharpness. Unexplained clusters signal undocumented changes.
- Round trip. Pick ten images and have the supplier reproduce the setup description from the line's documentation.
- Scale check. Verify a known feature's size in pixels against the stated scale.
- Split plan. Confirm that metadata supports the train and test separation you need by line, lot or time; see defect detection evaluation sets.
Getting metadata from existing archives
Machine-vision systems record much of this natively, since they must reproduce acquisition. Vision vendors' software often keeps camera and trigger settings in the file header or a configuration store, while lighting and optics may exist only in setup documents or the integrator's drawings. Request the archive together with the station configuration history. See in-line machine vision archives.
Older archives frequently have gaps. A pragmatic approach is a tiered requirement: mandatory fields for the evaluation subset, best-effort for the training pool, and a documented "unknown" value rather than silent defaults.
Commissioned capture
If existing images lack the needed variation, a new capture session can be designed with the metadata template above: multiple lighting setups, deliberate pose variation and logged settings per image. That depends on site access, line downtime and what the supplier agrees. Plan it as a new-collection discussion, and compare against multi-view inspection imagery and golden samples and reference images. For glass and similar parts, see glass and transparent part defects, where lighting design dominates.
Documentation to accompany the images
Metadata fields describe individual images; the set also needs a short document that describes the collection as a whole: the line and product family, the period, selection rules for the images included, known gaps, and changes in setup over time. A datasheet-style summary written by the data holder is often more informative than a spreadsheet of fields, because it records decisions that never reach a column, such as why a lighting change was made in the third month.
Ask also whether the supplier can state which images, if any, were cropped, resized or color-adjusted for storage. Those operations are common in vision software and invisible in the image file.
Limitations and extra preparation
Metadata documents the capture conditions; it does not make images from different conditions interchangeable. Knowing that two archives used different light sources tells you a model may not transfer, but it does not fix the gap. Fields recorded after the fact, reconstructed from memory or from equipment purchase records, can be wrong, and silent defaults (the same exposure value on every file) look complete while hiding missing data. None of these fields describe a humanoid's own sensors, so the viewpoint and calibration mismatch between a fixed inspection camera and a robot's head or wrist camera remains, and robot-specific calibration has to come from the robot itself.
Extra preparation is usually needed: reconciling field names across lines and years, marking unknown values explicitly, checking a sample of files against the physical setup where it still exists, and splitting evaluation data by capture configuration so that a held-out set actually tests a change in conditions rather than repeating the training conditions.
Questions to ask suppliers
- Which camera, lens, light and filter combination produced each subset?
- Were exposure and gain fixed or automatic, and did they change over the collection?
- What is the scale in micrometers per pixel, and when was it last verified?
- Are images stored lossless, and are they resized or cropped from the original?
- How many stations, lines and lots contributed, and are the identifiers preserved?
- Is the part presented in a fixed fixture or at varied poses?
- What documentation exists for the setup, and can it be shared?
- Did any lighting, optics or firmware changes occur between lots?
Discuss your requirements
If you need inspection imagery from US manufacturers with documented capture conditions, describe the parts, views and metadata fields to SourceX. SourceX reviews the requirement and follows up about potential sources; whether a metadata template can be applied depends on each supplier. A request does not guarantee a match.
Tell SourceX what your evaluation set needs to cover
Sources
- The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection — Benchmark with test_private_mixed split under varied lighting, small defects in large images, and transparent or reflective objects.
- Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection — Real-IAD multi-view industrial dataset of 150K high-resolution images across 30 objects.