Blog · AI visual inspection
Notes from the inspection line.
Anomaly detection, few-shot training, edge deployment and the metrics that matter: practical writing for integrators, quality managers and automation engineers.
- 41
Dimensional gauging and tolerance checks in millimetres
How the unit measures feature coordinates, hole positions and dimensions in millimetres with classical CV, then checks tolerance and judges appearance in one pass.
July 2026 · 6 min read
- 42
Discrete I/O or fieldbus: how much integration does your inspection need
A three-tier decision for connecting an inspection unit to a PLC: four 24V bits for pass/fail, a fieldbus for coded results, OPC UA for structured data to MES.
July 2026 · 8 min read
- 43
What is edge AI in visual inspection?
Edge AI runs the model on a device at the line, on fanless Jetson-class compute, deciding in about 30 ms with images staying on the line.
July 2026 · 6 min read
- 44
EtherCAT for high-speed inspection cells: when determinism is the constraint
On fast, motion-synchronized cells, a deterministic bus keeps the pass/fail result aligned with the cycle at 100+ parts per minute. When to reach for EtherCAT and when PROFINET or discrete I/O are enough.
July 2026 · 6 min read
- 45
F1-score vs accuracy: why 99% accuracy can hide a bad inspector
Why accuracy misleads on a rare-defect line, and why F1 with precision and recall is the honest number. 99.65% F1 and 0.69% FNR on a live cap line.
July 2026 · 7 min read
- 46
False positives in inspection: the hidden cost of over-rejection
A false positive is a good part wrongly rejected. Over-rejection scraps good product and erodes trust. On a live cap line, 99.65% F1 kept escapes to 0.69%.
July 2026 · 6 min read
- 47
Retraining in the field: adding a new defect from 20 images
How to add a new defect to an inspection model on the line from about 20 images, retrain on the unit in under 48 hours, and validate before it goes live.
July 2026 · 8 min read
- 48
Fleet analytics without shipping raw frames: metrics up, images stay down
How multi-line inspection dashboards are built from counts and metrics, so you get fleet trends on a web-based dashboard while raw part images stay on each line.
July 2026 · 7 min read
- 49
GDPR and factory imagery: when visual inspection captures people
On-premises edge processing is the simplest data-protection posture when a line camera can catch an operator: frames never leave the site. An architecture choice, not legal advice.
July 2026 · 7 min read
- 50
Global shutter machine vision: why a 12 MP global-shutter sensor matters for moving parts
A rolling shutter skews a part moving at line speed; a global shutter freezes the frame. Why the unit uses an up-to-12-MP global-shutter, C-mount sensor.
July 2026 · 7 min read