Machine vision integrators in South Africa: how to choose one

What a machine vision integrator does

A machine vision integrator turns an inspection requirement into a station that works on your line, every shift. The camera is one part. Whether the station works depends just as much on the lighting that makes the defect visible, the mounting and presentation of the product, the trigger, the PLC signals, the tracking to the reject, the reject itself and the tests that prove all of it. An integrator is the party accountable for that whole chain.

Three kinds of supplier

Buyers in South Africa usually meet three kinds of supplier. Each can be the right choice; the table shows when.

Ten questions to ask

Ask every supplier the same questions and compare the answers side by side. Vague answers to the first three are the strongest warning.

1. Samples: Will you test our own good and defective products before proposing hardware, and show us the images?
2. Limits: What can the station not detect on our product, and why?
3. Metrics: How will the false reject rate and the missed defect rate be defined, measured and accepted?
4. Tests: What exactly is tested at FAT and at SAT, with which samples and at what speed?
5. Integration: Who designs the trigger, the PLC signals, the tracking to the reject and the reject confirmation?
6. PLC changes: Who changes our PLC program, and who is responsible if the line stops because of it?
7. Faults: What does the line do if the camera stops, the air drops, a reject is not confirmed or power fails?
8. Changeovers: How are new products and recipes added, and can our team do it after training?
9. Handover: What documentation, backups, licences and spares list do we receive?
10. Support: Who supports the station after handover, from where, how fast, and on what terms?

What to write into the contract

• The defect list, with reference samples of each defect and of normal good variation.
• Acceptance criteria: how the false reject rate and the missed defect rate are measured, on which samples, at what speed, and the levels agreed for FAT and SAT.
• Scope boundaries: PLC changes, electrical work, mechanical changes to the line, the reject mechanism and who does each.
• Fault behaviour: what the line does when the station stops, the reject fails or power is lost.
• Handover set: documentation, recipe backups, licences, training and a recommended spares list.
• Support terms: response times, remote and on-site support, and what counts as a change rather than a fault.

How addaScan runs these steps is described in how we deliver a machine vision integration project.

Warning signs

• A quotation for hardware before anyone has looked at your defective samples.
• A promise of 100 % detection with no false rejects.
• No mention of the reject, its confirmation, or how the failed product is tracked to it.
• Acceptance described as “it works” instead of measured rates on agreed samples.
• No plan for adding products, backing up recipes or supporting the station after handover.

Local support in South Africa

Whichever machine vision integrator South African plants choose, a station that stops the line needs after-sales support from someone who can respond. Ask whether there are local technicians, where the engineers are based, what can be done remotely, how quickly someone can be on site in your province, and which spare parts are held locally or must be imported. addaScan works from Johannesburg for plants across South Africa; projects elsewhere are confirmed case by case. See about addaScan and support after handover.

Compare evidence for the same requirement

Give each shortlisted integrator the same sample identities, product specification and line constraints. Ask them to show the original images, what was classified and what remains unresolved. A demonstration on selected easy samples establishes little about the difficult defect boundary. In a hypothetical bid comparison, one supplier recognises every code in a demonstration while another documents a difficult reflective variant and proposes a lighting trial. The useful comparison is the evidence for that variant and the resulting scope, not the largest headline read-rate claim.

For measured dimensions, ask how repeatability, bias and reference uncertainty are handled. NIST’s gauge R&R material and JCGM 100 sections 4–6 provide useful concepts for reviewing the proposed evidence. (NIST e-Handbook, 2.4 Gauge R&R studies) (JCGM 100:2008, Guide to the expression of uncertainty in measurement) If a supplier cites an accredited calibration provider, check the actual scope relevant to the measurement. South Africa’s accreditation legislation establishes the framework; it does not imply that every calibration or every integrator is accredited. (South Africa: Accreditation for Conformity Assessment, Calibration and Good Laboratory Practice Act 19 of 2006)

For trained inspection models, require an explanation of how development images are separated from the final test. Near-duplicate images of the same physical item should not leak between those groups. (Google Machine Learning Crash Course: dividing datasets) The handover should name who can change the model, how a new version is validated and how to restore the previous approved version. Also agree the ownership of recipes, camera settings, licences, PLC source and retained images. These terms affect maintainability directly and belong in the scope before hardware is ordered.

Record the evidence rating for each requirement

Use a simple evidence rating for each requirement, with the unresolved item and owner written beside it. This is a suggested procurement worksheet, not an industry certification score. A strong aggregate impression must not override a missing critical requirement.