Demo · Engineering & manufacturing
How sensitive should the inspection camera be?
An AI vision system doesn't just pass or fail parts; it gives each one a score, and you choose where to draw the line. Draw it low and defects escape to customers. Draw it high and you scrap good parts or bury inspectors in false alarms. This calculator puts dollars on both sides and compares the camera with and without a person reviewing what it flags.
A day on the line
Where to set the camera
Total daily cost (escapes + scrapped good parts + review labor) as the camera's sensitivity changes. The lowest point is the cost-minimizing setting for these inputs; the dashed line is your current setting.
How it's calculated
The camera's scores follow a standard two-bell-curve model in which the gap between good and defective parts is set by the AUC. Choosing the share of defects to flag fixes the threshold, and that threshold fixes the share of good parts flagged too. Camera only scraps everything flagged. Camera + review sends flags to a person, who confirms real defects (missing the share set above) and returns good parts to the line. Manual has a person inspect every part. DPPM is defective parts per million shipped.
Simplifications: one defect type, independent parts, and a person's review of a flagged part is assumed not to throw away good parts. Real lines see defect rates drift, new defect types the model never saw, and inspector fatigue that rises with volume.