Machine-vision quality control: how to validate it
Machine vision can inspect visible characteristics and generate alerts or records. It should not promise exhaustive output coverage, a fixed accuracy or operating rate before camera, lighting, cycle, defect, product and integration have been tested.
This guide is for quality and operations leaders seeking a turnkey implementation. If you need a YOLO tutorial or a dataset for your own training, technical resources exist; our service focuses on feasibility, hardware, validation, integration and operation.
Start by defining the defect
An imaging system can only assess what its sensor captures with sufficient signal. Cases that may be studied include:
- surface: visible dents, stains, cracks or finish variation;
- shape and assembly: presence, position, dimension or a missing component;
- labelling: presence, orientation, print or code reading;
- counting and sorting by observable characteristics;
- foreign bodies visible to the chosen imaging modality.
An internal defect with no visual signature, a chemical property or a taste issue needs a different sensor or method. That boundary is documented before the pilot.
Coverage and speed are calculated, not assumed
| Variable | What to validate |
|---|---|
| Coverage | Which faces and areas each camera sees and what remains hidden |
| Cycle | Speed, spacing, exposure, blur and inference time |
| Lighting | Reflections, ambient variation, cleaning and ageing |
| Product | References, batches, tolerances and packaging or finish changes |
| Action | Alert, stop or reject and its safe behaviour |
Inspection may cover a sample, one station or all visible units if design and testing support it. The proposal states the demonstrated scope.
Accuracy needs a protocol
We do not publish a universal percentage. For each defect we agree a validation set, ground truth, metrics and thresholds. In addition to overall accuracy, we measure false negatives, false positives and performance by reference or condition.
A strong result on prepared images does not prove production performance. The pilot must include real variation, boundary pieces and cases not used during tuning. Where risk requires it, the system flags uncertain cases for human review instead of deciding alone.
Manual inspection and vision have complementary roles
| Criterion | Person | Vision system |
|---|---|---|
| Nuance and context | Can apply expert judgement | Limited to validated data and classes |
| Repetition | Depends on workload and conditions | Applies the configured criterion while the environment remains stable |
| Traceability | Needs a recording process | Can log events when designed and integrated to do so |
| Changes | Can adapt with instructions | Needs evaluation and sometimes new data |
The usual aim is to reserve human judgement for exceptions and root-cause analysis, not to claim that a camera always sees more.
Integration with the line, ERP or MES
Detection can connect to actuators, PLCs, ERP, MES or alerts where compatible interfaces exist and the proposal includes them. Before confirming, we review protocols, versions, permissions, latency, safety and failure mode. Not every line can be stopped or reject a part in the same way.
How we deploy it
- Assessment: defect, cost, sensor, line and risks.
- Sample: representative examples and a labelling criterion agreed with quality.
- Pilot: one reference or station with a protocol and predefined thresholds.
- Engineering: camera, lens, lighting, mounting and protection.
- Integration: actions and records included in scope.
- Acceptance test: normal, boundary, failure and recovery cases.
- Operation: monitoring and changes, where contracted.
Schedule depends on plant access, availability of defects, hardware, integrations and shutdown windows. It is agreed after assessment.
ROI: what to measure
- scrap, rework, return and complaint cost;
- inspection time and alert-review time;
- defects caught and defects escaping;
- false rejects of good product;
- hardware, integration, maintenance and change cost.
With low volume, insufficient visual signal or excessive variability, the project may not pay off. The pilot answers that question before scaling.
Frequently asked questions
Do I need to prepare the dataset?
We can include capture and labelling, but your quality owner must define and validate what counts as a defect.
Does it connect to my ERP or MES?
Only after interface, version, permissions and scope have been reviewed. The proposal documents which fields and events are exchanged.
What accuracy will it achieve?
The result demonstrated by the agreed test for each defect and condition. It is not fixed before representative data is available.
Can it cover production at the required throughput?
That can be an objective, but it must be proven against the real cycle, camera coverage and total capture, inference and action time.
Tell us which defect gets through, at what speed and what should happen when it appears. We will design a test with buyable metrics and limits.
Assess my case
