Deep Learning-Based Localization of the Cutoff in Automotive Headlamp Images

Abstract

When conventional image processing techniques reach their limits, deep learning models provide a powerful alternative, particularly for objects with high variation in appearance. This thesis addresses the replacement of an existing image processing algorithm used to identify the cutoff (or inflexion point). The cutoff line in light images serves as a reference point and plays a crucial role in the testing of headlamps after manufacturing and assembly. The goal is to replace the first stage (rough determination) in identifying the inflexion point, which is followed by a middle fine and fine determination algorithm. Following a successful feasibility study in cooperation with the University of Applied Sciences Upper Austria in Hagenberg, an optimized and improved architecture is developed under strict requirements. A systematic literature review was conducted to identify state-of-the-art architectures and techniques to maximize performance. Based on these findings, a refined architecture was designed through systematic testing of key components, including different backbones, detection heads (Faster R-CNN, Deta, FCOS), the trade-off between backbone capacity and input resolution, and the effect of Feature Pyramid Networks (FPN). Important hyperparameters were fine-tuned to balance accuracy and inference time. From the experiments, Faster R-CNN with a ConvNext S backbone and an added FPN (C3–C5), as well as Deta with a ConvNext B backbone using only the last feature map (C5), emerged as the best-performing architectures. The refined design achieved a performance improvement of +3.4% accuracy and 97.7% on the extended dataset while reducing outliers. Therefore, the initial model’s limitations in generalization were improved significantly. In the integration tests, the refined model, in combination with the algorithms, successfully identified the inflexion point in 𝟗𝟗. 𝟓𝟖% of cases with an inference time of 𝟒𝟒. 𝟑𝟏 𝒎𝒔, demonstrating that deep learning can reliably replace the previous algorithm in the manufacturing process.

Date of Award	2025
Original language	English
Supervisor	Kurt Niel (Supervisor)