Semantic Segmentation of Condensation in Automotive Headlights Using Deep Learning

Condensation inside automotive headlights can degrade optical performance, complicate durability assessments, and increase maintenance effort. However, comparing condensation behavior across designs or environmental conditions remains difficult due to the lack of a clear, quantifiable metric. One essential step toward objective evaluation is the automatic detection and localization of condensation in image data. In this study, we propose a data-driven approach based on convolutional neural networks to segment fogged regions within headlight assemblies. A U-Net-style architecture is trained on a compact dataset recorded during controlled condensation--decondensation cycles. The model reliably segments condensation patches of varying size and intensity. Performance is evaluated using the standard Intersection over Union (IoU) and a relative IoU variant that normalizes for the size of the ground truth region. Results show stable segmentation quality across the dataset. The approach lays the groundwork for automated condensation quantification and supports future analysis of moisture behavior and design optimizations.