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.
- 91
Scrap and rework economics: what over-rejection really costs
Over-rejection quietly destroys yield. What a high false-reject rate really costs, why precision is a direct scrap lever, and what to measure before and after AI inspection.
July 2026 · 7 min read
- 92
How to spec an edge-AI inspection unit into an automation quote
Spec inspection as fixed line items: five protocols, four inputs and four outputs at 24V, 90-240 VAC 60 W, and the AV variant chosen by environment.
July 2026 · 8 min read
- 93
Supervised vs unsupervised inspection: which one your line needs
Supervised inspection needs labeled defects you rarely have. Unsupervised learns from about 20 good parts only, so it wins on rare, first-seen defects.
July 2026 · 6 min read
- 94
Synthetic data for defect detection: when it helps and when it misleads
Synthetic defect data helps when real defects are near impossible to capture, and misleads when the model learns the renderer. Good-parts-only often needs neither.
July 2026 · 7 min read
- 95
What "trains on good parts only" actually means
Training on good parts only means the model learns normal from about 20 good images and flags deviation, so you skip defect collection and still reach 99.65% F1.
July 2026 · 7 min read
- 96
Data quality vs data quantity: why clean images beat big datasets
For a good-parts-only inspection model, consistent lighting, framing and a sharp 12 MP capture across about 20 images decide accuracy more than dataset size.
July 2026 · 8 min read
- 97
Troubleshooting a false reject on the inspection line
Most false rejects trace to lighting, part presentation, trigger timing or a tight threshold, not the model. A fixed diagnostic order that finds the cause before you retrain.
July 2026 · 7 min read
- 98
Why vertical integration removes vendor lock-in in machine vision
When one supplier builds the camera, lighting, edge compute and AI model, you avoid multi-vendor blame and data-export traps. Vendor lock-in, explained.
July 2026 · 7 min read
- 99
Vision-guided robot program selection: let recognition pick the program
How a recognition result tells the robot which program to run, shown by a headlight left-vs-right case and bumper variants above 0.9 confidence, over 4 outputs or a fieldbus.
July 2026 · 8 min read
- 100
Why the integrator keeps the customer when the unit ships pre-trained
A pre-trained edge-AI vision unit lets the integrator, not a specialist, own the account: 20 good images, about 30 minutes to install, and the customer stays yours.
July 2026 · 7 min read