Short answer
An in-line machine vision archive holds images captured at inspection stations, either of every part or only of rejects, with the system's pass or fail call, scores and recipe settings. It is often the largest existing source of defect imagery in a plant. Retention is short, labels are the machine's own, reject folders include false alarms, and escapes are missing, so request an archive profile and a verified sample first.
What these archives are
Automated optical inspection and machine-vision stations capture an image at a trigger, run a program (a recipe) and log a verdict. What gets saved is a storage policy, and the policy decides what the archive can teach.
| Storage policy | What is kept | Good for | Blind spot |
|---|---|---|---|
| Store all | Every inspected part | Nominal and defect imagery; escapes can be found later | Volume forces short retention |
| Store rejects only | Parts the system flagged | Defect examples, and false alarms | No nominal images; parts it missed are absent |
| Store on event | Images around alarms, faults or recipe changes | Debugging context | Not representative of production |
| Sampled | Every nth part or a periodic burst | A cheap nominal baseline | Defects are under-represented |
Plants can mix policies across stations and change them over time, so a single archive may hold several regimes.
What a reject folder contains
A reject folder is the reverse of a curated benchmark. MVTec AD's training set, 3,629 of its 5,354 images, is defect-free only, and its test anomalies carry pixel-precise ground truth (Bergmann et al., 2021). A reject archive is defect-heavy by selection, and its labels are the machine's. Expect a mix of:
- true defects of the types the recipe was built to find;
- false rejects caused by dust, flicker, misalignment, a new material lot or a drifted threshold;
- setup, calibration and test images saved under the same folder;
- images from several recipes or product variants with no label saying which;
- duplicates from re-presentation of the same part;
- no examples of defects the system missed.
The last point matters most for training. The archive records what the system saw and flagged, so it inherits that system's blind spots. Reject precision, meaning the share of machine rejects that experts confirm as real defects, is the number to measure. Inspection images linked to outcomes covers attaching final decisions that settle which rejects were real.
A public example of a labeled AOI-style set
DeepPCB consists of 1,500 image pairs, each a template image and a test image, annotated with six common PCB defect types (Tang et al., 2019). Use it as a yardstick for any offer. Does the archive keep the reference image for each recipe? Are defects labeled by type and location, or only by verdict? For electronics specifically, see PCB and solder joint inspection and golden samples and reference images.
Image fidelity: what to check in the files
- Resolution and bit depth. Ask whether files are the sensor's original size or downscaled. Small defects in large images are one of the harder cases added in MVTec AD 2 (Heckler-Kram et al., 2025), and a downscaled archive can erase them.
- Compression. Lossy compression can smear fine defects or add artifacts that look like them. Ask for originals or lossless files where they exist.
- Burned-in overlays. Some systems save the region of interest, verdict text or defect boxes into the pixels. A model can learn the overlay. Ask for a clean image with the result stored separately.
- Crops versus full frames. Region crops drop the context that tells a defect from a feature.
- Multi-view sets. Several cameras may image one part. Ask whether images are grouped by part.
- Lighting and trigger conditions. Record lighting setup, exposure and trigger timing per image; the fields are listed in imaging conditions metadata.
Illustrative archive profile to request before any images
# Illustrative only — synthetic stations and values; the supplier completes one block per station
station: aoi_station_04
product_family: stamped_bracket_family_A
storage_policy: rejects_only # all | rejects_only | on_event | sampled
retention: {oldest_image: 2026-06-14, rule: "overwrite at 120 days"}
counts: {total_images: "supplier-reported", rejects: "supplier-reported", passes_kept: 0}
images: {resolution: "12 MP, original", bit_depth: 8, format: png, overlays_burned_in: false}
recipes: {versions_in_period: 7, changes: ["2026-07-02 threshold", "2026-08-19 lighting"]}
hardware_changes: [{date: 2026-08-19, change: "ring light replaced"}]
identifiers: {serial_read_rate: "supplier-reported", lot_in_filename: true}
result_data: {verdict: yes, score: yes, defect_boxes: yes, reference_image_per_recipe: yes}
export: {native_format: "vendor database + image folders", vendor_support_needed: "to be confirmed"}
ownership: {images: plant, software: vision_vendor_X, customer_design: yes}
Export and format problems
Vision systems often keep metadata in a database and pixels in folders, with filenames encoding parts of the story. Extraction can need the system vendor's help, and the supplier's license for the vision software may govern what can be exported. Volumes can reach terabytes, so transfer and storage planning matter; see delivery and transfer of large datasets. Ask for a pilot export of one day's images with metadata and check it end to end before negotiating the whole archive.
Supplier types
| Supplier | Typical archive | Strength | Watch for |
|---|---|---|---|
| Manufacturers with AOI lines | Reject or all-part images per station | Large volumes, consistent framing | Short retention; recipes tied to the product |
| Packaging and label lines | Print, seal and fill-check images | High image volumes | Brand artwork and customer labels in frame |
| Contract manufacturers | Images of customers' parts | Variety of products | Designs belong to the customers; see contract manufacturers as suppliers |
| Machine-vision integrators | Customer systems they service | Technical knowledge of the archive | Often do not own the images |
Preparation to budget for
Expect expert review to separate real defects from false rejects, de-duplication by part, relabeling into your taxonomy, joins to dispositions, and splits by lot, line and date. A pass over a random sample of rejects before purchase gives you a reject-precision estimate, which sets how much review the full archive needs.
Quality and acceptance criteria
- The archive profile is delivered per station and matches the sample.
- Reject precision is measured on a sample you draw, not one the supplier selects.
- Images are at original resolution and bit depth, with overlays absent or separable.
- Counts per recipe and per camera are given with change dates.
- The serial or lot join rate is stated, and unmatched images are described.
- Nominal images are included, or the sampling rule is documented.
- Timestamps are in one time zone, with clock offsets noted.
- Duplicates and re-presentations are flagged.
What this data cannot support
Archive images do not show how a humanoid would handle, turn or present a part. They come from fixed cameras with lighting tuned for the line, not from a robot's head or wrist view, so plan separate robot-view images; see active inspection by humanoids. They do not show defects the system missed, and the machine's verdict is not ground truth. A model fitted to one line's recipe and lighting often degrades on the next (domain shift across lines).
Rights and confidentiality
Images often show part markings, serial numbers, logos and customer-owned designs, and occasionally people or screens at the edge of the frame. The plant may not own the images outright if a customer, a contract or the vision vendor claims them. Treat this as an early diligence item; confidentiality in defect data lists the questions. Cropping a design out of an image does not by itself settle whether the images may be licensed.
Questions to ask suppliers
- Which stations save images, and under which policy: all parts, rejects, events or samples?
- How long are images retained, and what is the oldest image today?
- How many images exist per recipe and per camera, and when did recipes, cameras or lighting change?
- What share of rejects do your inspectors confirm as real defects?
- Are images saved at original resolution without lossy compression or burned-in overlays?
- Can images be exported in a standard format with metadata, and what vendor support is needed?
- Are passing images kept, and if so, sampled how?
- Who owns the saved images: you, your customer or the vision-system vendor?
Discuss your requirements
If you need machine-vision images for a specific part family, describe the parts, defect types and station details to SourceX. SourceX looks for US businesses that hold suitable inspection data, checks the data and the supplier's licensing permissions, and coordinates pricing, allowed uses and delivery in a license. SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also arrange annotation, scoped case by case. SourceX reviews each request and follows up about potential sources; a request does not guarantee a match.
See also the inspection, defect and failure hub, the humanoid training data guide, the guide to sourcing defect image datasets, and the catalog page on licensing 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 — a curated, defect-free-training benchmark as the contrast to a reject archive.
- Online PCB Defect Detector On A New PCB Defect Dataset — template and test image pairs annotated with six defect types.
- The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection — small defects in large images and varied lighting as hard cases.