Fourier Mixture-of-Experts for Robust Infrastructure Defect Segmentation

This work addresses the pivotal challenge of robustness in deep learning models for infrastructure crack segmentation. Although these models demonstrate promising accuracy in academic settings, their performance often degrades significantly under the unpredictable conditions encountered in the field. This paper introduces the first comprehensive benchmark to quantify the robustness of commonly used crack segmentation models against a wide spectrum of synthetic environmental corruptions. By adopting a Fourier-based perspective, we diagnose these failures by linking them to the physical processes of image degradation such as optical blur and sensor noise. Based on this insight, we propose the Fourier Mixture-of-Experts, a training framework that systematically enhances robustness with 8.8% to 21.4% improvement across diverse data sets without compromising clean data accuracy.