Crack segmentation in civil infrastructure using conditional modulation and hierarchical dual-path networks

Yutao Jin, Hui Liu, Xu Cheng, Chunyu Du

Abstract: Highlights•We propose CMHA-Net, an novel method for structural crack segmentation in civil infrastructure. By integrating global context with local feature extraction, it effectively balances edge details and overall image information. Through dynamic feature adjustment and background noise suppression, CMHA-Net improves segmentation accuracy and continuity, significantly enhancing overall performance.•We propose the CMGC module for feature modulation. CMGC uses global conditional variables to refine convolutional features, balancing global context and crack details. With modulation, branch gating, and residuals, it enhances extraction, suppresses noise, and improves accuracy. The HAFM module integrates multi-level features through partitioning, multi-scale processing, upsampling, and gated attention, capturing subtle cracks while reducing noise. The DSCS module fuses Transformer and CNN branches to preserve context and details, further boosting performance.•We assess CMHA-Net using four publicly available datasets, encompassing various scenarios and material structures. The experimental outcomes demonstrate that our model surpasses current SOTA methods, producing high-quality crack segmentation maps.

External IDs:doi:10.1016/j.engstruct.2025.121162