TL;DR — AI defect detection replaces threshold-based machine vision with learned models. On real lines it holds 98–99.5% defect recall at 1–3% false positives, runs in 10–50 ms per image on an edge GPU, and typically pays back in 8–18 months. The binding constraint is not the algorithm but the defect sample set: budget 200–500 images per defect class, or use anomaly detection to cold-start with good parts only.
Why rule-based vision fails on real defects
Classical vision compares a region against fixed thresholds: grey-level difference, edge position, blob area. That works for scratches on a uniform surface and fails on textured, curved or reflective parts, where normal variation already exceeds the defect signal. The result is a rule set that either misses low-contrast defects or flags hundreds of false calls per shift, and every new product variant forces the rule set to be rebuilt.
What deep learning changes
A trained detector learns the appearance of defects rather than a threshold. Three consequences matter in production:
- Texture tolerance. Brushed metal, cast surfaces and woven fabric stop being a problem, because the model learns the distribution of normal texture.
- One model, many defect types. Segmentation and classification heads share a backbone, so scratches, stains, missing features and misalignment are handled in a single pass.
- Transfer across variants. Within a product family, fine-tuning on 50–100 new samples usually restores accuracy, instead of rebuilding the rule set.
Accuracy you should expect
| Defect class | Typical recall | False positive rate |
|---|---|---|
| Scratch, stain, contamination | 99–99.5% | 0.5–2% |
| Missing / misplaced component | 99.5%+ | <1% |
| Sub-millimetre dimensional defect | 97–99% | 1–3% |
| Low-contrast internal crack | 94–98% | 2–5% |
Treat these as planning numbers, not guarantees. Any vendor quoting 100% recall without a false-positive figure is quoting half a system.
Data requirements and cold start
Supervised training needs labelled examples. As a rule of thumb, 200–500 images per defect class gives a stable model; rare defects can start at 50–100 and improve as production accumulates cases. When no defect samples exist yet, unsupervised anomaly detection trained only on good parts gives a usable baseline within two to three weeks, and real defects are folded in afterwards.
Takt time and hardware
Inference on an edge GPU takes 10–50 ms per image, covering takt times down to a few hundred milliseconds. Beyond that, the limit moves to optics and material handling: exposure time, part presentation and reject actuation usually dominate before the model does. High-speed lines are solved by multi-camera stations and pre-triggered capture, not by a faster network.
Cost bands
A single-station cell — industrial camera, lens, structured lighting, edge PC and software — typically runs ¥100k–300k (US$14k–42k). A multi-station line system with MES integration runs ¥500k–2M. Ongoing cost is modest: model maintenance is usually a few days per quarter once the defect set stabilises.
Real deployment results
On a metal-parts line running three shifts, a two-camera station replaced four inspectors per shift and cut customer complaints from leaks to zero over six months. On a packaging line, false calls fell from roughly 400 per shift to under 15, which is the number that decides whether operators trust the system.
Frequently asked questions
Can the model handle new product variants without retraining? Within the same family, usually yes after a short fine-tune. A different part category needs new samples, budget two to four weeks.
How do we keep false positives low? Tune the decision threshold against a labelled validation set, not against a single shift of production, and review every false call in the first month.
Do we need a GPU on every station? No. One edge GPU can serve two to four cameras if takt time allows; otherwise one GPU per station is the safe default.
What happens when lighting drifts? Fix the light source and re-validate monthly. Most accuracy regressions in the field trace back to optics, not the model.
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.
