The model is the easy part
Most failed vision projects fail on data collection and physical setup, not on modelling. If images vary in lighting, angle and focus, no model compensates reliably, and the system that worked in the trial fails on the night shift.
Before anything else, get the camera position, lighting and trigger consistent. That single discipline determines most of the eventual accuracy.
What you need before starting
- Defects visible in an image. If a human cannot see it in the photograph, the model cannot either.
- Hundreds of labelled examples per defect type, including borderline cases.
- Consistent capture — fixed camera, controlled lighting, same trigger point.
- An agreed definition of a defect, because your own inspectors will disagree more than you expect.
- A decision about what happens on detection — reject, divert, flag for human check.
The definition point is the one that derails projects. Ask three inspectors to grade the same fifty parts and compare. If they agree only 80% of the time, that is the ceiling for any automated system, and the first job is agreeing the standard.
Realistic expectations
For well-defined defects with good images, detection rates comparable to or better than human inspection are achievable, with the significant advantage of consistency — a model does not get tired at the end of a shift.
For subtle, variable or novel defects, expect it to catch the known types and miss things it has never seen. This is why most deployments run as augmentation rather than replacement: the system catches the routine majority and humans handle the rest.
Tune the threshold to the cost of being wrong
Every system trades false positives against false negatives. Which error is worse is a business decision, not a technical one.
- Safety-critical parts: tune towards catching everything, accept more false rejects
- High-value parts: false rejects are expensive, so tune towards precision with human review
- High-volume, low-value: optimise for throughput and accept a defined escape rate
Make this decision explicitly and revisit it with real data after a month, because the initial assumption is usually wrong in one direction.
Cost structure
| Component | Typical cost |
|---|---|
| Cameras, lighting, mounting | £3,000–£20,000 per station |
| Image collection and labelling | £5,000–£20,000, often underestimated |
| Model development and validation | £15,000–£40,000 |
| Integration with line control | £10,000–£30,000 |
| Ongoing retraining | Budget annually — products change |
Plan for drift
Products change, suppliers change materials, lighting degrades, cameras drift. A model trained last year will quietly get worse and nobody notices until a customer complains.
Monitor the detection rate distribution and sample outputs for human review on a schedule. Plan for periodic retraining as a normal running cost rather than an unexpected project.