A cosmetics packaging moulder’s manual inspection could not keep up as volumes grew. An AI vision system trained on 18,000 images lifted detection above 99% on most defects, cut customer escapes from 410 to 32 ppm and freed three inspectors per shift.
Key numbers
A cosmetics packaging moulder produced decorative closures for premium brands, where a single black speck or scratch on a cap could trigger a complaint. Four inspectors per shift watched parts flow past at speed, rejecting defects by eye. As volumes grew, inspection struggled to keep up, and customer escapes, defective parts reaching the customer, were rising.
The situation
Escapes had reached 410 parts per million, and the customer had begun issuing corrective action requests. Inspection was limited to around 60 parts per minute per inspector, and fatigue reduced accuracy towards the end of each shift. A traditional rule-based vision system had been considered years earlier, but natural variation in colour and decoration made it reject too many good parts.
What changed
The company tested AI-based vision, following the principles in computer vision inspection for moulded parts:
- A feasibility trial with sample defects confirmed that each defect type could be detected reliably before any investment was made.
- Cameras and diffuse lighting were installed at a conveyor transfer point, where each closure could be presented consistently. Lighting was designed for the most difficult defects: specks and fine scratches.
- A model was trained on 18,000 labelled images of good and defective parts, covering normal variation across cavities, colours and material lots.
- Rejects were diverted automatically, and images were stored for every part, giving traceability and data for process improvement.
What the data showed
The vision system's detection rate exceeded manual inspection on every defect type: short shots 99.9% versus 94%, black specks 99.4% versus 81%, flash 99.6% versus 88%, colour variation 98.9% versus 72% and scratches 97.8% versus 69%. The largest gains came on the subtle defects that people found hardest to see consistently.
Results
Customer escapes fell from 410 to 32 parts per million, a 92% reduction. Inspection speed rose to around 400 parts per minute. Inspectors per shift fell from four to one, with three redeployed to quality engineering roles, using the stored images to find and fix the causes of defects at the source.
Lessons for your operation
- Run a feasibility trial with real defect samples before investing.
- Lighting matters more than software. Design it for the hardest defect.
- Train on normal variation as well as defects, to avoid false rejects.
- Use stored images and defect data to improve the process, not just to sort parts.
- Validate the system like any measurement system, with known good and bad parts.
