Cross-Domain Coupled Convolutional Transformer Network for Concrete Damage Detection

To overcome the challenge where convolutional neural networks (CNNs) struggle to effectively capture the diverse visual features of cracks, spalling, and exposed rebar in concrete structures resulting in inaccurate damage segmentation, a method is proposed, known as the cross-domain coupled convolutional transformer network for concrete damage detection (DamageNet). First, a dual-branch encoder architecture combining CNN and transformer is designed with a hierarchical structure that outputs CNN and transformer features at the same resolution, preserving both local perception and global information. Second, a cross-domain coupling attention module is introduced to integrate the CNN and transformer features effectively, fusing local perception and global modeling information in a complementary manner. Finally, on multiple publicly available multidamage datasets, the proposed network achieves IoU scores of 78.70%, 91.52%, and 73.90% for exposed rebar, cracks, and spalling, respectively, and the mean ± standard deviation across all damage classes obtained from five training repetitions is 82.74% ± 2.46%. Experimental results validate that the proposed network outperforms other mainstream methods, and the feature map visualization demonstrates that the network effectively captures diverse visual features, benefiting concrete multidamage detection tasks.