Machine vision & quality
Machine vision inspects 100% of production at line speed, where the human eye tires. Deep learning has widened its reach to variable, ill-defined defects. But success is decided first on lighting and optics, not on the neural network.
The technical building blocks
Surface & presence inspection
Scratches, stains, presence/absence, code reading and industrial OCR, dimensional measurement with a calibrated camera.
Metrology & profile
Structured light, laser triangulation, stereo vision and point clouds for shape, flatness and volume control.
Classification, segmentation, anomaly
Convolutional networks handle variable defects that rule-based vision cannot describe; unsupervised detection learns "good" and flags the deviation.
Embedded GPU & real time
Inference keeps line speed (a few ms) on an embedded GPU or smart camera, triggers ejection and traces every part.
Lighting first
Half the result is decided before any algorithm: good lighting makes the defect stand out, a bad image cannot be rescued by software. Choosing the lighting is the first setting of a vision cell, not the neural network.
| Lighting | What it reveals |
|---|---|
| Diffuse frontal (dome) | glossy or curved surfaces, without parasitic glare |
| Backlight | silhouette, dimensional measurement, presence/absence |
| Dark-field (grazing) | scratches, engravings, relief defects |
| Coaxial | flat specular surfaces, markings |
| Structured light | 3D relief, height, volume |
| Multispectral / UV-IR | adhesive, contamination, checks invisible to the eye |
Rules or deep learning?
Classic rule-based vision stays unbeatable for crisp metrology tasks: measure, read a code, check presence. Deep learning takes over when the defect is variable and hard to describe — texture, appearance, cosmetic anomaly. Both often coexist in one cell: rules for measurement, network for appearance.
| Criterion | Rule-based vision | Deep learning |
|---|---|---|
| Task type | measurement, code, presence | appearance, texture, anomaly |
| Data required | none (deterministic) | labelled images, or unsupervised |
| Explainability | full | partial (black box) |
| Tolerated variability | low (brittle) | high (generalises) |
| Setup effort | parameter tuning | collection + training + validation |
Metrics to watch: resolution (the smallest detectable defect spans 3 to 5 pixels; resolution = field of view ÷ sensor pixels), throughput (a few milliseconds per part), and the precision/recall pair that trades false rejects against missed defects by the cost of each.
Key challenges
- Lighting & optics first — good lighting (dome, grazing, dark-field, multispectral) solves half the problem before any algorithm. A bad image cannot be rescued by software.
- Data & labelling — real defects are rare; augmentation, synthetic data and unsupervised learning compensate for the lack of labelled examples.
- False rejects vs. escapes — the precision/recall setting arbitrates between needless scrap and an escaping defect; the right point depends on the cost of each.
- Line integration & traceability — handshake with the PLC, ejection trigger, time-stamping and image storage for quality audit (often via OPC UA to the MES).
Standards & references
- EMVA 1288 — Standard for objective camera characterization (noise, sensitivity, linearity) — to compare sensors on a like-for-like basis.
- VDI/VDE 2632 — Machine vision guidelines: specification, acceptance and acceptance testing of systems.
- GenICam / GigE Vision — Camera interface standards (generic transport and API) for cross-vendor interoperability.
Related standard pages
Major players
Cameras & systems
Cognex, Keyence, Basler, Sick, Datalogic, Zebra.
Vision software
MVTec HALCON, Cognex VisionPro, Adaptive Vision, OpenCV.
Deep learning & edge
NVIDIA (Jetson, Metropolis), Landing AI, Intel OpenVINO, Neurala.