A guide by SevenLab
A computer vision purchase should begin with the inspection decision and its consequences. Detecting something in an image is only one part of deciding whether a product passes, needs review or should be rejected. SevenLab's AI vision page is the relevant starting point for an enquiry. The inspection designs below are hypothetical scopes to assess, not claims of shipped equipment or integrations.
Define the defect and the cost of each mistake
Before comparing AI inspection software or computer vision providers, write an inspection specification that production and quality teams can both apply. Distinguish classification, reading a label and locating a visible defect. Ask whether conventional machine vision rules could solve a stable, clearly defined check before commissioning a learned model. Record the consequences of accepting a defective item, rejecting a good item and sending an uncertain result to manual review.
Illustrative scenario, hypothetical scope: a packaging line needs automatic visual inspection for missing labels, wrong product labels and damaged print. These are separate decisions, not one general accuracy target. Agree what makes each label unacceptable and which cases need a quality specialist. Acceptance should report missed defects and false rejects separately for each category, with agreed limits based on business consequences. An unreadable image should not silently become a pass.
- Create an agreed defect catalogue with accepted and rejected examples.
- Compare rule-based inspection, AI image recognition and manual checks.
- Define separate acceptance limits and escalation routes for each defect category.
Check whether the images can support the decision
For AI defect detection, ask for an image feasibility review before treating model development as the main task. Sampling should cover relevant product variants, operating shifts, camera positions, lighting and surface conditions. Have qualified reviewers label the images and resolve disagreements. Keep a separate evaluation set from different production batches or capture sessions so that near-identical images do not make the trial look more reliable than it is.
Illustrative scenario, hypothetical scope: a manufacturer wants image recognition software to identify scratches on reflective parts. Some apparent scratches are glare, while shallow defects disappear under the current lighting. Compare revised capture arrangements before choosing a model. Acceptance should require agreement on what is actually visible, evaluation on withheld batches and a declared limit for unsupported finishes or defect types. Synthetic examples may help explore edge cases, but should not replace testing on representative physical products.
- Sample images across product variants, lighting conditions and production shifts.
- Have quality specialists resolve disputed labels before training and evaluation.
- Reserve independent evaluation batches and document unsupported conditions.
Specify how an inspection becomes a line decision
Computer vision integration needs a clear contract between image capture, software output and the action on the line. Ask how an image would be linked to the correct item, how late results would be handled and where uncertain decisions would go. Camera triggers, PLC or MES connections, edge processing and cloud processing are requirements to assess, not confirmed SevenLab integration support. The proposal should explain timing, network dependencies and operator controls.
Illustrative scenario, hypothetical scope: a conveyor inspection returns a result after the item has passed the rejection point. That result must not be applied to the next item. Define a safe operating procedure with the responsible engineering and quality teams, such as manual segregation or a controlled stop where appropriate. Acceptance should include delayed responses, camera resets, network loss and a full review queue, with evidence that item identity and decision history remain traceable.
- Map image identifiers to physical items and the final inspection decision.
- Agree response deadlines and behaviour for late, missing or uncertain results.
- Test interruptions and obtain approval for the operator fallback procedure.
Purchase a staged trial with measurable release gates
When selecting a computer vision company or consulting partner, request separate deliverables for feasibility, offline evaluation, a trial alongside existing inspection and supervised rollout. Require a report by defect type and operating condition, not just an overall accuracy figure. Include manual review effort, false rejects, missed defects and capture failures in the business case. A visually impressive demonstration is not enough evidence to remove an existing quality check.
Illustrative scenario, hypothetical scope: a label supplier changes the material after the inspection system has been accepted. Agree who notices the change, whether the model remains approved and which checks must be repeated. Release criteria should include versioned evaluation data, an operator-approved work instruction and a tested route back to the previous inspection process. A staged rollout lets the buyer assess actual operating burden before extending the system to more products or lines.
- Tie each project stage to evidence and an explicit buyer approval.
- Require results by defect and condition, including manual review workload.
- Assign ownership for material changes, revalidation and fallback to prior inspection.