A symmetric dual stream deep learning model with a shared encoder and topological attention gate for crack segmentation

Abstract Semantic crack segmentation is an important task in the automated inspection of civil structures, but preserving crack continuity, connectivity, and structural consistency remains challenging for conventional models. In this work, SkelTAG-Net (Skeleton Topological Attention Gate Network) is proposed as a symmetric dual-stream deep learning model with a shared encoder, composed of a segmentation stream and a supervised skeleton stream, in which structural information is injected into the segmentation decoder through a Topological Attention Gate (TAG) module. The model is evaluated on the CrackVision12K dataset and compared with CNN-based, Transformer-based, and hybrid architectures. The results show that SkelTAG-Net achieves the best overall performance, obtaining an IoU of 0.676, an F1-Score of 0.781, and a clDice of 0.848, outperforming models such as SegFormer-B2 and U-Net. These results indicate that SkelTAG-Net has potential for applications in the automated inspection of civil structures and for supporting Structural Health Monitoring (SHM) systems.