Study on AMSU-Net model for pixel-level concrete cracks detection and its application in dynamic testing of RC structures

Accurate monitoring of the distribution pattern and development trend of surface cracks in concrete structures is of great significance for structural performance analysis and health assessment. To address the problems of low efficiency and insufficient accuracy of manual detection methods under dynamic loads and complex environments, an Attention-based Multi-Scale aggregation U-Net model (AMSU-Net) for pixel-level concrete crack detection is proposed in this paper. Firstly, to alleviate the limitations of traditional U-Net in global semantic feature extraction, an Attention-Enhanced Feature Extraction Module (A-EFEM) combining residual learning and channel attention mechanisms is introduced. Secondly, a dual-branch encoder structure and multi-scale skip connection strategy are integrated to enhance the modelling of global semantic information. Extensive evaluations conducted on several public crack datasets and experimental images of reinforced concrete (RC) structures demonstrate that AMSU-Net significantly outperforms U-Net in segmentation accuracy, exhibiting strong generalisation ability and robustness. Finally, by embedding a crack quantification algorithm, AMSU‑Net has been successfully applied to dynamic tests of reinforced concrete structures, providing strong support for investigating the dynamic propagation and evolution of concrete cracks.