TCDGNet: Crack segmentation network via triple-component dynamic graph channel attention and gated dilated wavelet attention

Cracks are the most common evaluation indicators in structures such as road surfaces and concrete. However, cracks often exhibit complex interference, fuzzy boundaries, and discontinuous topological structures, making accurate crack segmentation tasks extremely difficult. Therefore, a Triple-Component Dynamic adjacency Graph attention network (TCDGNet) was constructed for automatic crack detection in structures such as concrete and pavement. To achieve adaptive modeling of channel dependence, a three component dynamic adjacency graph channel attention (TDA-GCA) mechanism is proposed. In addition, the model integrates two-stage dilated wavelet decomposition with gated feature modulation to enhance the structural edge response while suppressing background interference. Our model was tested on four public benchmarks (DeepCrack, CrackTree260, CrackForest and Crack500), it got mIoU scores of 87.50%, 85.15%, 80.85% and 76.60% on these benchmarks. Respectively, surpassing the ten existing methods. This improvement enhances the structural consistency of cracks and provides a more reliable assessment of the condition of infrastructure and other structures.