关键词: 裂缝分割, 双分支编码结构, 密集注意模块, 频域特征融合, Mamba架构

Abstract: Cracks represent a common defect in building structures, making their timely and accurate detection crucial for ensuring structural safety. However, due to factors such as the slender and irregular morphology of cracks and interference from complex backgrounds, existing Convolutional Neural Network (CNN)-based methods are often limited by their local receptive fields and struggle to effectively handle cluttered foreground and background interference. To address this, a novel model combining CNN with the Mamba architecture was proposed, featuring a decoupled dual-branch encoder. This design leverages the complementary fusion of a local fine-grained feature extraction branch and a global context modeling branch, balancing detailed preservation with global perception capabilities. Furthermore, to mitigate the semantic discrepancies between the branches, a Frequency-guided Feature Fusion Module (FGFM) was designed. This module facilitates the fusion of high- and low-frequency features within the frequency domain, thereby enhancing the complementarity and expressive efficiency of local and global features. Experiments conducted on three public datasets (Volker, DeepCrack, and CFD) demonstrate that the proposed method outperforms several mainstream models, achieving F1-scores of 83.32%, 87.41%, and 74.86% respectively, validating its superior performance.

Key words: Crack Segmentation, Dual-Branch Encoder Architecture, Dense Attention Module, Frequency-Domain Feature Fusion, Mamba Architecture