A specialist website by Oxford Industrial Automation
UK machine vision & robot integration · 01865 416830
Inspection method guide
Use the simplest method that proves the quality decision. Deterministic tools remain strong for clear limits; learned models can help where acceptable and defective appearance varies.
Manual measurement shown: deterministic tools remain preferable where a clear tolerance can be proven.
Direct answer
Rule-based tools are often the better choice when the decision can be described with a stable measurement, contrast, position, code or tolerance. AI can help when visual variation is hard to express as fixed rules. Many production systems use both, with controlled imaging and explicit uncertainty handling.
Presence, distance, angle, edge, diameter, code content and known colour limits are usually easier to validate with deterministic tools. The logic remains transparent and recipe changes can be controlled directly.
AI can classify or locate irregular scratches, natural product variation, texture and cosmetic appearance that resist a short rule set. The training and validation dataset becomes part of the machine specification.
Side-by-side comparison
Use the comparison as a screening tool; the final architecture follows from representative products, timing and acceptance evidence.
| Question | Rule-based vision | AI / deep learning | Hybrid approach |
|---|---|---|---|
| Best fit | Stable geometry, codes, known tolerances and repeatable contrast | Variable cosmetic defects, natural materials and complex texture | Applications with both objective checks and learned appearance |
| Evidence needed | Feature visibility, calibration, tolerance and repeatability samples | Representative labelled data across real production variation | Both datasets, plus logic defining how results combine |
| Explainability | Direct measurement or programmed rule | Model score, class, region or anomaly result | Deterministic evidence with learned classification where needed |
| Change control | Recipe, threshold and tool configuration | Dataset, labels, model, threshold and deployment version | Coordinated versioning of recipes, models and PLC response |
Decision pathway
A model demonstration is not a complete inspection system; image control, validation and containment still decide production performance.
Agree defect classes, limits and uncertain-result policy.
Control optics, lighting, trigger and representative variation.
Test unseen products, rare conditions and borderline states.
Version the method and monitor production outcomes.
Questions manufacturers ask
Use these answers to frame the first application review and the evidence needed next.
AI may tolerate some variation, but controlled imaging remains essential. Lighting changes can alter the visual evidence and reduce the reliability of both learned and rule-based methods.
There is no reliable universal number. Coverage of the true production range, defect classes, suppliers, formats and borderline conditions matters more than a headline image count.
Anomaly methods can flag images that differ from a representative normal set, but an unseen defect should not be assumed detectable without challenge testing and a defined review policy.
Yes. A system may use deterministic measurement and code reading alongside learned classification or segmentation, then combine the results into one controlled production decision.
Connected guidance
Explore the technologies and buyer guidance most relevant to this application.
Start with the production evidence
Use the project brief form, or email representative product images or video with the target output and required decision. We will define the next technical step.