TCI-Net: Structural Feature Enhancement and Multi-Level Constrained Network for Reliable Thin Crack Identification on Concrete Surfaces

Accurately identifying thin cracks in concrete surfaces is a key technology for the automated maintenance of structures, with widespread applications in intelligent inspection tasks such as road, bridge, and building façade inspection. Existing crack identification methods, based on encoder–decoder architectures, suffer from inaccuracies and insufficient reliability for diverse concrete surface cracks. This study proposes a reliable crack identification method using the U-Net semantic segmentation network, incorporating image structural feature enhancement and multi-level consistency constraints. This method builds on a pre-trained multi-scale semantic feature encoding network, integrating various crack edge structure information extraction operators, and a lightweight crack spatial detail feature extraction module is constructed. Through an attention mechanism, the extracted multi-scale semantic features and high-resolution structural features are adaptively fused, enabling the reliable and intelligent identification of thin cracks in complex environments. Extensive experiments are conducted on the largest publicly available crack identification dataset, CrackSeg9k, for which the proposed method achieves an F1-score of 74.7%, surpassing state-of-the-art approaches. Furthermore, to assess its practical effectiveness, the trained model is validated on real-world data collected from complex tunnel environments. The results demonstrate that the proposed approach accurately identifies fine cracks and exhibits enhanced robustness against challenging conditions such as illumination variations and environmental interference. The source code is publicly available at https://github.com/SmileLeeCN/ThinCrackIdentification.