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.
- 11
Glass inspection and positioning with AI: transparent and coated panels
How an R&D project located near-invisible transparent and coated glass with laser-line and Raman-shift techniques, feeding a handling robot the position in millimetres.
July 2026 · 8 min read
- 12
Foundry and hot-stamping inspection: high-temperature AI vision on the line
What changes when the part or ambient is hot, and how the AV-H100 (IP54, 0-65 C) runs surface and geometry checks near foundry and hot-stamping stations.
July 2026 · 6 min read
- 13
Chemical and petrochemical inspection: AI vision for hazardous (ATEX) areas
What a hazardous-area (ATEX) inspection unit needs, and how the AV-X100 (IP66) is positioned for chemical and petrochemical lines. Certification to verify.
July 2026 · 6 min read
- 14
Packaging line inspection with AI: label, seal and cap checks at full speed
Inspect caps, labels and seals on a fast conveyor without slowing throughput, anchored to a proven cap cell at 99.65% F1 and 0.69% false-negative rate.
July 2026 · 6 min read
- 15
Metal stamping and machining inspection: burrs, scratches and gauging
How AI surface inspection plus millimetre dimensional gauging catches burrs, deformation and scratches on stamped and machined metal parts.
July 2026 · 6 min read
- 16
The rare-defect problem: why you can never collect enough defect images
Defects are rare by design, so you never collect enough defect images. Good-parts-only anomaly detection trains on about 20 good parts and beats class imbalance.
July 2026 · 7 min read
- 17
Adding a new part variant to a running inspection line
When a customer adds a part, how fast can the line inspect it and who does the work? Capture 20 good images, train in under 48 hours, deploy by USB.
July 2026 · 6 min read
- 18
Unsupervised defect detection: flagging a defect it has never seen
How anomaly detection catches a defect it was never trained on: it learns a good part and scores deviation, so first-seen flaws are flagged, not matched.
July 2026 · 6 min read
- 19
Why training under 48 hours changes the economics of a line changeover
Sub-48-hour training turns a new inspection task into a two-day task, not a multi-month project. Why good-parts-only training changes changeover economics.
July 2026 · 6 min read
- 20
Escapes vs false rejects: the two errors that set inspection cost
One threshold trades escapes against false rejects, and the escape usually costs more. The confusion matrix in shop-floor terms, and 0.69% FNR at 99.65% F1 on a live cap line.
July 2026 · 7 min read