RUBIKSCORPAI in factories
Why visual inspection models miss defects
Defects are rare and varied, so models see too few examples of the ones that matter. Changes in material, finish or lighting can also make a trusted model suddenly less reliable.
Define the defect and the cost of missing it
Build a shared defect taxonomy with the quality team. Include examples close to the boundary between acceptable and unacceptable. Record product type, finish, lighting and camera setup so performance can be examined by condition. Keep evaluation examples separate from training examples. If nearly identical images appear in both sets, the reported result may overstate how well the system handles new production.
What to check
- Review label disagreements and confirm how borderline cases should be handled.
- Check rare and severe defects separately from common cosmetic variations.
- Include new batches, material changes, reflections, motion blur and partially obscured surfaces in testing.
What to change
Standardize image capture where possible and collect representative examples with permission. Establish thresholds with the quality team rather than choosing a number from a demonstration. Route uncertain cases for review. Run the system alongside the existing inspection process until acceptance criteria are met and responsibility for missed defects is understood.
How to evaluate the result
Measure defect recall and false rejection rates by defect type. A single accuracy figure can be misleading when most items are good. Review the cost of unnecessary rejects and the consequences of missed severe defects.
What the result does not establish
An inspection pilot covers the tested defects and conditions. New defects and production changes require new review. Critical quality and safety decisions need controls appropriate to the product and process.
Further reading
Want us to fix this for you?
We help you define what to detect and how to test it fairly. Register and we’ll help you plan an inspection pilot that you can measure.