Short answer
PCB and solder joint inspection data comes from public sets and from automated optical inspection (AOI) archives at electronics manufacturers. DeepPCB offers 1,500 template-and-test image pairs with six defect types; VisA adds 10,821 images across 12 objects. Acceptance criteria are codified in standards such as IPC-A-610, so labels need the product class attached. Public sets rarely match a specific board, process or camera.
Why electronics inspection is a distinct sourcing task
Electronics assembly has unusually codified quality criteria. IPC-A-610 provides visual acceptance criteria for electronic assemblies, covering solder joints, placement, cleanliness, coating and marking, with three product classes: Class 1 general, Class 2 dedicated service and Class 3 high performance, the strictest. The current revision is IPC-A-610J (2024), replacing Rev H (2020) (IPC, 2024). Details here come from secondary descriptions rather than the standard's own page, so check the standard itself before relying on them.
The practical effect: the same joint can be acceptable for a Class 1 consumer product and a reject for Class 3. A label of "defect" needs the class attached; the general principle is covered in acceptance standards as label definitions. Beyond that, boards are dense, repetitive and photographed by AOI machines under controlled, programmable lighting, which produces large archives but also strongly line-specific images. The wider framing of reject archives is in in-line machine vision archives. Context is at the inspection cluster and the humanoid data section.
Public data points
DeepPCB provides 1,500 image pairs, each a defect-free template and a test image, annotated with six common PCB defect types; the authors' group pyramid pooling model reported 98.6% mAP at 62 FPS (Tang et al., 2019). The template-and-test pairing suits reference-based comparison methods, which presume precise registration.
VisA contains 10,821 high-resolution images, of which 9,621 are normal and 1,200 anomalous, covering 12 objects in three domains, with image- and pixel-level labels (Zou et al., 2022). It is useful for anomaly detection and segmentation methods, though a buyer should check whether its categories resemble the boards in question.
Treat the reported scores as properties of those datasets and methods. Reference-aligned, uniformly lit, fixed-fixture images make for easier detection than a mix of boards from different customers.
What PCB inspection data can and cannot support
It can support defect detection, anomaly detection against golden boards, solder joint classification, component presence and polarity checks, and evaluation sets for AOI systems. When linked to repair outcomes it supports triage models that reduce false calls; see defect severity and disposition for how to encode the decision.
It cannot support manipulation. Board handling, connector mating, rework soldering and probing are motion tasks that images of finished boards do not describe. It also cannot cover defects invisible from above: voids under ball-grid-array packages and hidden joints need X-ray imaging, covered under X-ray and CT inspection images. Camera-based AOI does not see through components.
AOI archives: what they typically contain
| Item | Description | Value to a buyer |
|---|---|---|
| Board images | Per-panel or per-component crops at stated resolution | Core training and test images |
| Machine verdict | Pass, fail or call by the AOI program | Weak label, with false-call noise |
| Operator review | Human confirmation or override of machine calls | Stronger label |
| Defect code | Component, location (reference designator), defect type | Structured labels |
| Repair record | What was fixed, and the re-test result | Outcome linkage |
| Program and library version | Inspection program, lighting recipe | Domain context |
| Board identity | Product, revision, lot, line | Split keys |
Machine verdicts are not ground truth. They are the output of an existing system, so a model trained on them can inherit its blind spots. The operator review and later test results are better label sources, and are covered in relabeling legacy defect archives.
Preparation before the data is usable
Registration and crop conventions. Reference-based methods need aligned template and test images. If the AOI crops per component, record the crop origin so the full board can be reconstructed.
Recipe grouping. Different board families use different lighting recipes. A pooled set without recipe identifiers hides a major source of variation; see imaging conditions and metadata.
Rebalancing with care. Good joints vastly outnumber bad. Subsampling normals is reasonable for training, but keep evaluation data at the real prevalence so false-call rates are measured honestly.
Panel-level leakage. Several boards are made on one panel and share fixtures and defects from the same print. Keep panels together in one split.
Illustrative request specification
Illustrative; values are synthetic.
| Parameter | Example |
|---|---|
| Product class basis | IPC Class 2 assemblies, revision stated per board |
| Defect types | Insufficient solder, bridging, tombstone, missing part, wrong polarity |
| Image form | Per-component crops with full-board context |
| Labels | Operator-confirmed defect code with reference designator |
| Good data | Operator-confirmed passes from same boards |
| Outcome | Repair and electrical test result |
| Splits | By product revision and line, never by board panel within one lot |
| Redaction | Customer logos, part numbers, silkscreen identifiers |
Evaluation and acceptance
Label authority. Ask whether labels come from the AOI verdict, an operator, or electrical test. Ranking them by authority tells you how much relabeling is needed.
False-call handling. AOI systems flag many good boards. If false calls are not marked, a model learns to reproduce the machine's habits.
Class applicability. Confirm the product class used for acceptance and whether it varied by board. A pooled dataset mixing classes needs the class in each record.
Diversity of designs. Boards from one product family have repeated components; a model may recognize the board instead of the defect. Hold out entire product revisions.
Resolution versus component size. Fine-pitch components need high resolution; check pixel size against the smallest feature.
Rights and sensitivity
Board images expose circuit layout, component choices and markings that belong to the manufacturer's customer. Contract manufacturers usually work under confidentiality terms that restrict disclosure of customer designs. Wiring and connector assemblies are a related but separate asset, described in cable harness and connector inspection. Cropping to solder joints or masking silkscreen reduces exposure, but it does not settle whether the holder may license the images; that comes from the underlying agreement. This is not legal advice.
Limitations and extra preparation
AOI archives reflect one machine's view of one board design. Images from a different AOI vendor, a different camera angle or a different lighting recipe may not transfer, and a new board revision changes component layout enough to break location-based labels. Machine calls are not ground truth: AOI systems are tuned to flag generously, so a large share of flagged items can be false calls, and the true disposition sits in operator review and repair records that may be kept in a separate system or not at all. Labels from a Class 3 product line do not mean the same thing on a Class 2 line, so the product class has to travel with every image.
The data supports inspection and false-call filtering; it does not show how to place, rework or handle a board. Plan for extra preparation: linking images to final disposition, removing customer board names and part numbers where sensitive, normalizing defect codes across machines, and splitting by product revision and line so that boards from one panel or lot do not appear on both sides of an evaluation.
Questions to ask suppliers
- What AOI equipment, lighting recipe and resolution produced the images?
- Are labels machine verdicts, operator confirmations or test results?
- Which IPC class governed acceptance, and which revision of the standard?
- How many boards and designs are represented, and can the images be split by design?
- Are repair outcomes and electrical test results linked?
- What share of calls were false, and are they marked?
- Which customers own the designs, and do their agreements permit licensing images?
- Can you supply X-ray images for hidden joints where relevant?
Discuss your requirements
If you need PCB or solder joint inspection data from US electronics manufacturers, describe the products, defect types and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, and checks the data and the supplier's licensing permissions. A request does not guarantee a match.
Tell SourceX what your evaluation set needs to cover
Sources
- Online PCB Defect Detector On A New PCB Defect Dataset — DeepPCB: 1,500 image pairs annotated with 6 PCB defect types; 98.6% mAP at 62 FPS.
- SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation — VisA: 10,821 images, 9,621 normal and 1,200 anomalous, 12 objects in 3 domains, with image- and pixel-level labels.
- IPC-A-610 Acceptability of Electronic Assemblies — IPC-A-610 visual acceptance criteria for electronic assemblies, three product classes, revision J (2024).