Direct answer
Start with a defect library and measurable pass/fail rules. Control lighting, camera view, product position, speed and trigger, then prove detection using representative good, bad and borderline packs. Vision can inspect visible features but does not replace hidden seal-strength or contamination tests.
Why this matters
A camera system only performs as well as the inspection definition. “Check the label” must be translated into presence, identity, position, orientation, print, barcode and defect rules.
Decision table
| Inspection | Typical visible feature | Design requirement |
|---|---|---|
| Label | Presence, identity, position and skew | Full label view and controlled product |
| Code | Text, date, batch, barcode and readability | Resolution, contrast and stable print area |
| Closure | Presence, height, tilt and tamper feature | Multiple views where geometry hides defects |
| Fill level | Visible level or headspace | Controlled container and product optics |
| Seal area | Visible contamination or missing material | Lighting and view of complete seal |
Information to collect
- Collect good, bad and borderline samples.
- Define every defect with a measurement or example.
- Record speed, spacing, rotation and presentation.
- Assess reflections, colour and transparency.
- Choose reject and reject-confirmation logic.
- Plan recipe, access, audit images and data retention.
Practical method
- 1
Prove lighting before increasing software complexity.
- 2
Control the product through the imaging zone.
- 3
Separate detection from tracking and rejection.
- 4
Challenge every defect at every relevant position.
- 5
Monitor false reject and missed defect rates.
Common mistakes to avoid
- Training only on perfect samples.
- Using uncontrolled ambient light.
- Inspecting a curved pack from one unsuitable view.
- Setting thresholds from too few samples.
- Assuming a detected defect has been rejected.
Questions people also ask
What can machine vision inspect?
Visible features such as label, code, cap, fill level, colour, presence, position and some seal defects.
Can vision check seal integrity?
It can inspect visible seal contamination or geometry, but hidden leak strength normally requires another method.
What is a false reject?
A conforming pack classified as failed; false-reject rate should be measured alongside missed defects.
Does AI vision remove the need for samples?
No. Representative samples and an agreed defect definition remain essential for training and validation.
Continue from the requirement to the equipment
Use the specialist route that matches the product, pack and process. Final suitability requires representative samples, output targets and site information.