Abstract
Cracks are a ubiquitous and critical form of surface damage in infrastructure; their precise detection and segmentation are paramount for structural safety assessment. To address critical challenges such as feature ambiguity in small-scale cracks, boundary discontinuity, and background interference, this paper proposes Dcmt_Net, a crack segmentation network based on a dual-branch encoder architecture. Specifically, the CNN branch incorporates Fourier frequency-domain enhancement and a Channel-Spatial Group Attention mechanism to bolster edge responses and multi-scale feature modeling. Simultaneously, the Transformer branch integrates Wavelet feature enhancement with a local self-attention structure to effectively capture long-range dependencies and the global continuity of crack images. Furthermore, to achieve efficient complementarity across branches, a Feature Difference Fusion Module is designed to optimize the fusion quality of local and global information. Experimental results on multiple public datasets demonstrate that Dcmt_Net outperforms state-ofthe-art methods. Notably, on the DeepCrack dataset, it achieved 91.46% and 89.13% in key metrics such as RE and F1-score, respectively, validating its robustness and generalization ability in complex crack scenarios.

This publication has 25 references indexed in Scilit: