MA-UNet model for crack segmentation in concrete structures utilizing Mamba skip connections in multiple scenarios
Year: 2026
Type: article
Abstract: Concrete structures are prone to cracking during service, which, if left undetected, can lead to severe structural failure. Current deep learning paradigms face significant limitations: Convolutional Neural Networks (CNNs) are constrained by local receptive fields, hindering the capture of global continuity, while Transformer-based models suffer from quadratic computational complexity that impedes real-time deployment. To address these challenges, this paper proposes a concrete surface crack identification method utilizing a hybrid Mamba-attention segmentation model, termed MA-UNet. This model leverages the State Space Model (SSM) to achieve a global receptive field while maintaining linear computational overhead. Built upon the UNet architecture, the proposed framework introduces a novel Multi-head Depthwise-separable Channel Attention (MDCA) module. By integrating lightweight convolutions and a multi-scale feature fusion decoder, the model significantly enhances its capacity to suppress complex background noise. Experimental results demonstrate that MA-UNet outperforms conventional CNN and Transformer-based models across multiple datasets. Specifically, it achieves a maximum improvement of 13.83% in F1-score and 18.41% in mIoU, while reducing parameters by up to 119.58 M. Furthermore, it exhibits substantial advantages in key performance metrics such as FPS and FLOPs, offering a robust new technological trajectory for real-time, high-precision crack monitoring in sophisticated engineering environments.
Authors: Xiaofei Mao, Peng Liu, Lingkun Chen, Hesong Jin
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,165
OpenAlex ID: W7147228119