Electronics and PCB Inspection Images for AOI Model Training

What public PCB defect datasets cover, and what they miss

Public PCB sets are good for prototyping but rarely match a production SMT line. Widely used academic sets such as DeepPCB pair a defect-free template with a test image of a bare board and label etching-stage classes such as open, short, mousebite, spur, pin hole and spurious copper with bounding boxes. Check each set's paper for how defects were produced; some add artificial defects, which raises defect density far above a real line.

That makes these fabrication (bare-board) benchmarks, not PCBA ones. They have no components, no solder fillets, no 3D height data and no AOI false calls. A review of manufacturing defect benchmarks notes recurring limits such as synthetic generation (DAGM) and images with at most one defect when real products can have several [1]. Benchmarks also saturate: leading anomaly methods already exceed 99% AUROC on MVTec AD, as the 2024 Real-IAD paper reports [3], which says little about your reflow line.

"Semiconductor defect image dataset" usually means something else again. Wafer map benchmarks such as WM-811K encode die pass/fail patterns across a wafer, not optical images of assembled boards. Keep wafer maps, SEM review images and board-level AOI images in separate requirements; they share almost no features. For the general manufacturing case, see industrial defect image datasets.

Define the defect taxonomy before you request images

Write the taxonomy first, because it decides which suppliers and which AOI programs are even relevant. Anchor classes to IPC-A-610, the visual acceptability standard for electronic assemblies, and record which revision the line inspects to. State the product class too: Class 3 acceptance is stricter than Class 2, so the same fillet can be a pass on one line and a defect on another.

A practical SMT and through-hole taxonomy usually includes:

• Solder defects: bridging, insufficient solder, excess solder, solder balls, cold or grainy joints, voids visible at the fillet, head-in-pillow suspects (often needing X-ray, not AOI).
• Placement defects: missing component, misalignment or skew, wrong polarity, wrong part, billboarding, tombstoning, lifted leads.
• Board and process defects: contamination, flux residue, damaged pads, conformal coating gaps.

Record "unknown" and "operator could not decide" as explicit labels. Dropping ambiguous cases inflates offline precision and hides the exact boundary your model has to learn.

Separate AOI false calls from true defects

The most valuable field in AOI data is the operator-verified result, not the machine's call. Inline AOI programs are usually tuned to over-call rather than miss defects, so many flagged locations at a review station turn out to be acceptable. If you train only on AOI calls, you teach the model the old rule set, false calls included.

Ask for three linked fields per flagged location: the AOI call (defect class and machine score), the review-station disposition (true defect, false call, or escalated) and, where available, the downstream outcome such as rework, ICT or functional test failure. False-call images are excellent hard negatives. Also request a sample of unflagged "pass" boards; without them you cannot estimate escapes. For normal-only approaches such as PatchCore [2], clean pass images matter more than defect images; see normal-only image sets for anomaly detection.

Defect classes on a mature line are rare and skewed. Tombstoning or wrong-polarity examples may number in the dozens per product family. Plan pooling across lines or synthetic augmentation early, using the approach in rare defect coverage and class imbalance.

Capture metadata that makes AOI images transferable

AOI images only transfer across lines when the optics are documented. 2D AOI, 3D AOI with phase-shift profilometry or laser height, and solder paste inspection (SPI) produce different signals; mixing them without a modality field breaks training. Lighting matters too: many systems capture multi-angle RGB ring lighting, and a model trained on one lighting recipe can fail on another.

Illustrative example: invented to show structure; it does not describe an available dataset.

Split train and test by board revision, line or date, not by random crop. Random splits leak the same board layout into both sides and overstate accuracy. For delivery, per-sample sharding such as WebDataset tar files, where the image, height map and JSON label share a basename, keeps modalities aligned [4]; see dataset delivery formats.

Licensing board images: whose design is it?

AOI images show a customer's board design, so the factory that captured them often cannot license them alone. Contract manufacturers usually work under customer agreements that treat layouts, BOMs and test programs as the customer's confidential information. A supplier needs to confirm it holds rights in the images and, where needed, the board owner's consent.

Questions to put to any supplier:

1. Who owns the board design shown, and has that owner approved release for AI training?
2. Are silkscreen logos, part markings, serial numbers or barcodes visible, and will they be masked?
3. Do images include operator names, badge IDs or station logins in the metadata?
4. Are CAD files, centroid data or Gerbers included, and under what separate rights? (See CAD and PCB design datasets.)
5. Is the license scoped to training, evaluation, or both, and for which model?

Serial numbers and operator IDs are the main personal or traceable fields; inspection images rarely contain people. Linking images to MES quality records such as rework tickets raises the value of the data, which is covered on manufacturing quality records.

How SourceX sources PCB and AOI inspection data

SourceX sources operational datasets from US companies on request; it does not hold PCB images in stock, and a request does not guarantee a match. You describe the data, such as post-reflow 3D AOI images with operator dispositions for QFN and BGA packages, and SourceX looks for US businesses that hold it. Every release is approved by the supplying company, and nothing is contracted until a supplier agrees.

Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Personal details such as names and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Delivery runs through private, access-controlled workflows after an executed agreement. You can describe your AOI data requirement to SourceX, and browse related categories on the image data hub and inspection photo licensing.

Request PCB and AOI inspection image data

SourceX looks for US electronics businesses that hold the inspection data you describe and manages licensing through Find, Assess, Agree, Transact and Manage. SourceX does not publish prices; terms are agreed per deal. Tell SourceX what PCB inspection images you need.