Wavelet-Enhanced Crack-Sensitive Mamba UNet for Robust Pixel-Level Crack Segmentation

With the growing prominence of infrastructure aging issues, crack detection plays a critical role in structural health monitoring and road maintenance. However, pixel-level crack segmentation faces significant challenges due to complex backgrounds, noise interference, and the diversity of crack morphologies. To address these issues, this paper proposes a Wavelet-Enhanced Crack-Sensitive Mamba UNet (WECSM-UNet) to improve pixel-level segmentation performance under complex backgrounds and diverse crack patterns. Specifically, we design a Wavelet-Enhanced Crack-Sensitive Scanning Mamba (WE-CSSM) module as encoder, which employs state-space models (SSM) with multi-directional (horizontal, vertical, and dual-diagonal) scanning to capture global dependencies, while utilizing wavelet-domain enhancement to amplify high-frequency edge details and suppress background noise. In the decoder, we integrate a Semantic and Detail Infusion (SDI) module with channel and spatial attention mechanisms to achieve cross-scale feature alignment and adaptive reweighting, alongside progressive dynamic upsampling to restore fine-grained crack structures. Experimental results demonstrate that WECSM-UNet achieves superior segmentation performance on three public benchmarks (Crack500, DeepCrack, and TUT), outperforming existing mainstream methods in key metrics such as F1-score and mIoU, thereby validating its robustness and generalization capability in complex scenarios.