ORAFOL’s AI lens inspection starts with better lighting

Zebra reported on 26 October that ORAFOL Fresnel Optics had automated inspection of more than 1.5 million optical parts and cut inspection time in half. Its account describes existing production systems at the German manufacturer.[1][2]

ORAFOL built an inspection tower using point, transmitted and reflected light. Staff labelled images of scratches, dirt and inclusions to train Aurora’s defect detector; intricate surface structures had made visual inspection difficult.[2]

The surface is doing useful work

ORAFOL’s own product catalogue explains why these surfaces matter. Its microstructured polymer optics produce optical effects in less space than conventional curved optics. Positive Fresnel lenses can focus light or form a parallel beam; other structures distribute or redirect it. The pattern being inspected is therefore part of the component’s function. A successful inspection has to distinguish the intended structure from damage to it.[3]

Teaching the model what counts as a flaw

Aurora’s technical guide describes the work behind that distinction. For its feature-detection tools, users mark the image regions belonging to each class. An unmarked example can become contradictory training material: the software treats it as a negative example. The scale and appearance of training images also need to match production conditions.[4]

The guide even recommends changing the lighting when low contrast prevents learning. Illumination and annotation solve different problems: one makes a difference visible to the camera; the other tells the model which visible differences matter. That is why preparing an inspection system involves both optical engineering and carefully chosen examples.[4]

High-volume lines now inspect and pack parts automatically. Operators replace filled trays; quality specialists adjust settings and retrain models. Manual inspection remains for more variable, lower-volume products.[2]

Report an error or suggest a correction