A specialist website by Oxford Industrial Automation
UK robot & vision integration · 01865 416830
Variable defect detection
Apply machine-learning vision where acceptable and defective products cannot be separated reliably by a short set of fixed rules. We connect image capture, model performance, controls, evidence and reject handling into the production process.
Manual inspection shown: defect classes and acceptance boundaries must be defined before automation.
Direct answer
AI vision is most useful when the inspection must recognise complex appearance variation rather than measure one clear edge, dimension or contrast threshold. Rules-based tools remain appropriate for predictable checks; many production systems combine both methods.
Model quality depends on the image set and acceptance rules. The development set should cover normal variation, known defects, borderline cases, product formats and environmental conditions that the system will encounter.
The inspection result must remain associated with the correct product and drive a verified physical action. Thresholds, no-read behaviour, audit records and model-change control are agreed for the application.
Decision to action
Each stage is designed against representative products, real timing and the required physical outcome.
Create stable images of the required feature.
Develop the model from representative examples.
Challenge normal, defect and borderline cases.
Apply the result to the correct product and action.
Application opportunities
Final suitability follows from the product, presentation, rate, environment and acceptance criteria.
Questions manufacturers ask
Use these answers to frame the first application review and the evidence needed next.
No. A deterministic tool is often preferable for clear measurements, code reading or presence checks. AI is considered where the appearance varies too much for a practical fixed-rule approach.
There is no reliable universal number. The important issue is whether the image set represents the real variation, defect range and difficult edge cases. Trials determine the evidence needed.
Systems can retain the image, result, class and confidence information, but the available explanation depends on the chosen method. Production acceptance criteria should not rely on an unexplained score alone.
Yes, when the result is correctly tracked to the product and the reject mechanism, confirmation and exception handling are engineered as part of the system.
Connected guidance
Explore the technologies and buyer guidance most relevant to this application.
Start with the production task
Use the project brief form, or email product photos or video with the target output, current layout and required decision. We will define the next technical step.