- Home
- Capabilities
- Computer Vision
Capability
Inspection and counting that works in the actual lighting.
Vision systems fail in the gap between the lab and the line: the light changes, the camera gets nudged, a new product variant arrives, and accuracy quietly falls off a cliff nobody is watching.
01 / The problem
Where this usually goes wrong.
Vendors demonstrate vision on clean, well-lit, centred images of the defect they brought with them. Your line has vibration, mixed daylight and fluorescent lighting, a camera someone repositioned during maintenance, and defects that occur once in ten thousand units.
That rarity is the real difficulty. A model can score 99.9% accuracy by declaring everything acceptable, and on a line with a 0.1% defect rate that is exactly what an untended training process will produce.
02 / What we build
The parts of the system.
- A capture setup treated as part of the system
- Camera placement, lighting, trigger timing and mounting specified and documented, because the majority of field failures we are called to fix are physical rather than algorithmic.
- Models suited to scarce defects
- Anomaly detection where defect examples are too few to train a classifier, supervised detection where they are not, and targeted augmentation for the variation your environment actually produces.
- Edge deployment
- Inference on hardware at the line, so throughput does not depend on the network and a connectivity outage does not stop production.
- Drift detection on the input
- Continuous monitoring of image statistics, so a shifted camera or a changed lamp raises an alert on the day it happens rather than at the next quality audit.
03 / How it is measured
What we agree to be judged on.
Set before the build starts, against a measured baseline, and reported honestly afterwards — including where the numbers are disappointing.
- Recall at a fixed false-alarm rate
- The operationally meaningful number: what proportion of real defects are caught, given the number of false alarms your line can absorb per shift.
- By defect class
- A system strong on surface scratches and blind to contamination has an average score that hides the thing you needed.
- Under environmental variation
- Tested across shifts, seasons and lighting conditions, because those are the conditions it will run in.
- Against the current inspection
- Human inspection is not perfect either. We measure it, honestly, so the comparison is real.
04 / Honest limits
What this will not do.
A camera can only assess what it can see. Internal defects, subsurface flaws and material properties need a different sensor, and no amount of model work substitutes for that.
Genuinely novel defect types will be missed by definition until examples exist. Systems in this category need a route by which operators report what the model missed, and a schedule for folding that back in.
05 / When it applies
You probably need this if:
- Inspection is manual, tiring and inconsistent between shifts
- Defects are caught downstream, after value has already been added
- Counting or sorting is a bottleneck
- A previous vision pilot worked in a demo and not on the line
Recognising two or more of these is a reasonable trigger for a diagnostic. Recognising none of them is a reasonable trigger for not spending money here yet.
06 / Related
Often scoped together.
These capabilities share data, infrastructure or evaluation approach with computer vision, and are frequently part of the same engagement.
Next step
Is computer vision the right instrument for your problem?
The diagnostic exists to answer exactly that, including the possibility that the answer is no. It is time-boxed, fixed-fee, and the written assessment is yours.