Research on Tunnel Crack Detection Method Based on Multimodal Recognition

ABSTRACT

This paper presents a hypergraph‐enhanced multimodal framework for tunnel crack detection that fixes key limits in standard methods through five linked ideas: (1) a dual‐aligned fusion design that syncs visible light, infrared, and LiDAR for precise crack sizing; (2) lean encoders tuned to keep texture, thermal and geometric cues; (3) bidirectional knowledge distillation in Affine‐Net to enforce cross‐modal agreement; (4) deformable blocks with dynamic sparse attention to handle irregular cracks and noisy scenes; and (5) hypergraph‐based feature alignment that removes spatial mismatch artefacts; we validate on the real tunnels, beat SOTA, and reach 93.8% accuracy, 93.4% recall, and 0.42 h overall.