TL;DR — A machine vision system is four decisions, not one: what to measure, how to light it, where to infer, and how to act. Most disappointing systems fail at lighting, not at the algorithm. Budget ¥60k–250k for a single inspection station, ¥300k–1M for a multi-station cell with robot guidance, and expect 8–16 months payback where the station replaces manual inspection or manual pick-and-place verification.
Define the measurement before choosing hardware
Write down what “good” and “bad” mean in measurable terms: which feature, what tolerance, how many parts per minute, and what the consequence of a miss is. A system specified as “detect defects” will be argued about forever; a system specified as “detect scratches wider than 0.2 mm on a 40 mm surface, 60 parts per minute, miss rate under 1%” can be designed, tested and accepted.
Lighting decides the image
The single most useful rule: choose lighting that makes the defect change brightness, not one that makes the part look good. Backlight for silhouettes and dimensional edges, low-angle dark field for surface scratches, dome or coaxial diffuse for reflective and curved parts, and structured patterns for 3D height. Then freeze everything — distance, angle, intensity — because lighting drift is the most common cause of field regressions.
Camera choice in one table
| Requirement | Typical choice |
|---|---|
| Defect size ≥0.1 mm, moderate speed | 5–12 MP area scan, global shutter |
| Moving parts, high speed | Global shutter area scan, strobed light |
| Continuous web, long strip | Line scan, 4k–16k pixels |
| Height, volume, flatness | 3D laser profile or structured light |
| Thermal or hidden features | Thermal or X-ray, not visible light |
Edge or cloud inference
Inspection inference belongs at the edge: 10–50 ms latency, no dependency on network availability, and images that never leave the plant. The cloud is useful for training, model versioning and cross-site analytics, not for the reject decision itself. A hybrid pattern is standard: models train centrally, deploy to edge devices per line.
Integration: PLC, robot and MES
The vision decision has to reach an actuator within the takt time. That means a deterministic handshake with the PLC — trigger, exposure, inference, result, reject — and a clear fail-safe rule when inference times out (usually “reject and flag” rather than “pass”). For robot guidance, add calibration that is re-checked on a schedule; a 0.5 mm calibration error becomes a 0.5 mm placement error.
Cost bands
Single-station inspection with camera, lens, lighting, edge PC and software: ¥60k–250k. Multi-station cell with robot guidance and MES integration: ¥300k–1M. Annual maintenance including optics cleaning, calibration and model updates: 5–10% of hardware value.
Selection checklist
Before signing: a written acceptance test with defect samples and pass-rate targets; a shadow-mode period of at least two weeks; documented lighting setup with drift tolerance; a named owner for model updates; and a spare-parts list for cameras and light controllers. Systems that fail acceptance usually fail on the first and third items.
Frequently asked questions
How long does a station take to deploy? Four to ten weeks for a single-station inspection, longer when robot guidance or MES integration is involved.
Can one system handle several part numbers? Yes, with per-family models and a part-number trigger from the PLC.
Do we need a specialist to maintain it? One trained engineer per site is enough for day-to-day operation; model retraining is usually a quarterly task shared with the integrator.
What is the most common specification error? Under-specifying lighting and skipping the written acceptance test — both surface as “accuracy problems” after go-live.
Next step
If you are scoping a first deployment, send us the line layout, defect list and target takt time. We return a feasibility note with expected recall, false-positive budget and a cost band within three working days. See also our AI visual inspection guide for architecture and pricing detail.
