A closer look at surface quality

The challenge: defects that are hard to see

An automotive manufacturer wanted to automate the inspection of painted parts. Very small surface defects can be difficult to distinguish, especially when lighting does not reveal the feature an inspector needs to examine.

Designing the image and the model together

peoly combined specially developed illumination and camera modules with in-house trained convolutional neural networks. The capture arrangement was an essential part of the solution: it made the relevant surface features visible before classification.

The published work addressed multiple types of paint-coating defects. Its original example shows the lighting arrangement on incorrectly painted car door handles. The portfolio also describes preparation for on-premises and cloud deployment.

What this approach demonstrates

The project brings optical design and machine learning into the same inspection task. A useful image is the starting point; the model then interprets features that relate to the agreed defect categories.

A new inspection project starts with the part, finish, defects of interest and operating conditions. Those details determine the capture setup and the evidence needed to evaluate the model for that production process.