Electronic Components Recognition Data

Where electronics recognition is needed

Three robot workflows depend on recognizing electronic parts. In repair and refurbishment, a robot or an assistant must find the right connector on a board or the right cable in a bin. In e-waste recycling, items must be identified so batteries, boards and housings can be separated. In assembly and kitting, components must be picked and presented by type. In each, the objects are small, similar to one another and often partly covered by labels or cabling.

The core difficulty is fine-grained identity. Two USB connectors, two ribbon cables or two capacitors can be visually close and functionally different. The data must therefore be specified at the level at which the task distinguishes things, which is often a part number, a pin count or a standard, and not just a category.

What published datasets do and do not cover

The public datasets that mention electronics are mostly defect datasets. VisA contains 10,821 high-resolution images, of which 9,621 are normal and 1,200 are anomalous, spanning 12 objects in 3 domains, with image- and pixel-level labels (Zou et al., 2022). DeepPCB supplies 1,500 image pairs, each a template and a test image, annotated with 6 common PCB defect types (Tang et al., 2019).

These are valuable for anomaly detection and board inspection, covered in sourcing defect image datasets. They are not component-identity datasets. A defect label says a trace is open or a pad is missing; it does not say which capacitor sits at a position. A buyer who wants recognition of parts for handling or repair should not read a defect dataset as coverage of component types.

What component recognition data can and cannot support

It can support:

• Part and connector identification from images, including retrieval of a part number from a photograph.
• Counting and sorting in bins and trays when images show realistic piles.
• Orientation estimation for polarized parts, if labeled with pin-1 or key marks.
• Reading of printed markings on chips, housings and cables when resolution allows; see labels, barcodes and text on packages.

It cannot support:

• Insertion and removal. Connector mating force, latch behavior and cable stiffness are physical properties, not image content. Cables are deformable, and their behavior under handling needs deformable product properties.
• Safe handling. Static-sensitive parts, lithium cells and capacitors that hold charge create hazards that images do not capture; add labels as described in hazard properties of objects.
• Diagnosis. Recognizing a part is not knowing why a board failed.

Resolution and optics

Electronic parts push image requirements. The smallest surface-mount passives are tiny, and at typical robot-camera distances they cover only a few pixels. Specify pixels per millimeter at the working distance, not megapixels. Macro lenses, ring lighting and polarization change what is visible; glossy solder and metal shield cans saturate easily. See small object resolution requirements for a way to turn a part size and a required detail into a camera specification.

Supply routes

Part models from manufacturers are covered in manufacturer 3D models and CAD, and distributor data in industrial distributor catalogs. Small hardware that goes with electronics, such as screws and standoffs, is in fasteners and small hardware.

Illustrative label schema

The schema is Illustrative; values are synthetic.

# Illustrative electronics recognition record - synthetic values
image_id: img_5541
capture: {lens: macro, scale_px_per_mm: 24, lighting: ring_diffuse, background: mat_grey}
objects:
  - class: connector
    subtype: usb_c_receptacle
    part_number: PN-000412          # supplier identifier
    bbox: [220, 140, 310, 260]
    orientation_deg: 90
    key_feature_visible: true
    condition: [bent_pin_suspected]
  - class: cable
    subtype: ribbon_20pin
    length_mm_estimate: 150
    state: coiled
board_context: {board_id: brd_088, side: top}
hazard_flags: [static_sensitive]
license_id: lic_000

Acceptance criteria

1. Class definition at task level. Part-number or family-level classes, stated and consistent.
2. Resolution report. Median pixels per millimeter on the smallest class.
3. Near-duplicate classes. A confusion list of similar parts, each with images of both sides of the confusion.
4. Conditions. Counts for new, used, corroded, bent, dirty.
5. Marking legibility. For tasks that read markings, a sample check of legible vs illegible.
6. Hold-outs. Hold out by board design or part lot so that the same unit does not appear in training and evaluation.

Permissions and provenance

Boards and parts can embed proprietary designs, firmware identifiers or customer data. Repair and recycling streams may include devices with personal data in storage; image datasets of exteriors and boards usually do not, but ask. Manufacturer part photos belong to the manufacturer, and a distributor may only hold display rights. Check these before you train on them.

A note on counterfeit and look-alike parts

Repair and recycling streams contain parts that are visually near-identical but electrically different, including remarked chips and cables with the same connector ends and different internal wiring. Images alone cannot separate them, and a classifier trained on labels taken from the printed markings will inherit any false markings. If your application depends on authenticity or electrical rating, treat image recognition as a screening step and add testing or documentation, and tell suppliers so they do not describe visually similar parts as one class.

Limitations and extra preparation

Expect to add dimension and pin-count attributes, a mapping from distributor part numbers to your classes, and orientation labels. If your robot must insert or unplug parts, you also need demonstrations or force data collected on a robot or instrumented fixture, which is outside the scope of recognition imagery; see object recognition data vs manipulation training data.

Evaluating a sample

Ask for fifty images of your hardest confusion pair and judge them yourself. If a human reviewer with a magnifier cannot tell the two apart from the images, a model will not either, and the data needs better optics or a different view. Check that the same physical unit does not recur across many images under different filenames, and that marked parts show their markings facing the camera in a meaningful share of views.

Questions to ask suppliers

1. Which part families are covered, at what level of identity (family, type, part number)?
2. What pixels-per-millimeter do images reach on the smallest parts?
3. Are the images of new parts, used parts, or both, and how many units per part?
4. Do you hold the 3D models or STEP files, and under what license?
5. Are markings legible in the images, and were they transcribed?
6. Which parts, boards or designs belong to third parties?
7. Do any devices in the images contain stored user data, and how was this handled?
8. Can you provide the confusion pairs you know of, such as near-identical connectors?

Discuss your requirements

Define the part families and the resolution you need before you approach suppliers. When you describe the electronics recognition requirement to SourceX, SourceX reviews it and follows up about potential sources, such as US businesses that hold part imagery. A request does not guarantee a match, and nothing is contracted until a supplier agrees. See the humanoid training data guide and the object data hub.

Discuss your humanoid training-data requirements with SourceX

Sources

• SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation — VisA image counts, objects and labels.
• Online PCB Defect Detector On A New PCB Defect Dataset — DeepPCB image pairs and defect types.