Procházková, JanaMikuláček, PavelŠtarha, PavelGerrits, TimTeschner, Matthias2026-04-212026-04-212026978-3-03868-300-11017-4656https://doi.org/10.2312/egp.20261005https://diglib.eg.org/handle/10.2312/egp20261005Modern vehicles are equipped with a wide range of Advanced Driver Assistance Systems (ADAS) that rely heavily on camera-based perception. Reliable visibility estimation – particularly under fog condition – remains a significant challenge. Accurate fog detection can enable proactive system responses, such as automatic activation of fog lights, and enhance operational safety. We present a contrast-aware anomaly detection framework for image-based fog detection. Our algorithm combines multi-scale Difference of Gaussians responses and Gaussian-weighted local Root Mean Squared contrast with a convolutional autoencoder. The model is trained exclusively on clear-weather imagery to learn the nominal scene distribution, and visibility degradation is detected as a reconstruction deviation from this learned representation. Evaluation on a separate test set containing clear and fog conditions demonstrates an AUC of 0.91, achieved without using fog samples during training. The framework provides a practical basis for camera-based visibility monitoring in automotive environments.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Computer vision; Neural networksCCS ConceptsComputing methodologiesComputer visionNeural networks10.2312/egp.202610052 pages