Computer Vision for Quality Control on a Packaging Line
A plant packs 60,000 jars per hour. Today 4 operators inspect a sample by eye and defects slip through (badly closed lid, crooked label, crack in the glass), causing returns from large retail chains and contractual penalties. Cameras have been installed on the line that take a photo of every jar.
Where the money comes from
What to do now
- 1Quantify the cost of a defect that reaches the customer.
- 2Install the camera and collect 12,000 labelled images.
- 3Pilot on one line with human inspection in parallel.
For your committee
One page with the decision, the impact, the risks, the steps and the KPI to track.
The print dialog opens: choose "Save as PDF".
Watch out: more volume is not always more profit
A system that is too strict discards good product and slows the line; one that is too lax lets cracks through. The threshold is set using the cost of a recall versus that of a jar.
KPI for the committee
Defects reaching the customer per million and product discarded.
Detail for your data team
Proposed technique: Neural Networks (Deep Learning / CNN)
Why: The information is in the pixels: a crack is a visual pattern that no human can describe as variables in a table. Convolutional networks (CNNs) learn those patterns directly from images. With transfer learning, 12,000 images are enough.
Cost of getting it wrong: A false negative (an undetected crack) can cause a product recall and brand damage; a false positive only costs one jar.
Data needed: High-speed camera images of every unit.; 12,000 images labelled by quality: OK, lid, label, crack, fill level.; Record of returns and penalties by batch.