Fourier-Mixture of Experts YOLO for concrete crack segmentation with visual interpretability

Abstract

Accurate crack segmentation is essential for structural health monitoring, yet most deep-learning studies treat the task as binary and struggle with varied morphologies. This paper introduces FMOE-YOLO, a Fourier-enhanced Mixture-of-Experts extension of YOLO that integrates a Multi-branched Auxiliary Feature Pyramid Network (MAFPN) and an SPPF_LSKA large-kernel attention head. The Fourier expert captures high-frequency crack cues, while MAFPN with LSKA supplies rich multiscale context. Experiments on three datasets of rising difficulty — Individual, Single-Crack (four classes), and Complex (six classes) — show consistent gains over standard YOLOv8. On the Single-Crack set the model attains 86.2% [email protected], improving performance by 4.7 percentage points. t-SNE and UMAP embeddings reveal tighter, better separated clusters, and Grad-CAM maps confirm sharper crack localization, demonstrating enhanced interpretability. The proposed approach offers strong potential for real-world monitoring, effectively handling diverse crack morphologies and challenging geometric conditions.