Crack segmentation network based on hybrid-window transformer and dual-branch fusion

Cracks are external manifestations of infrastructure damage, and routine inspections are crucial for assessing their structural safety. However, due to factors such as crack diversity, background noise interference, and information loss, high-precision crack segmentation still faces numerous challenges. To alleviate the influence of these factors, a crack segmentation network based on hybrid-window Transformer and dual-branch fusion (CSHD) is proposed. The CSHD network can effectively capture local texture details and global context modeling to achieve high-precision crack segmentation. First, a hybrid-window attention mechanism(HWA) is designed as the core component, which employs a dual-branch parallel architecture to integrate channel attention and multi-scale depth-wise convolution modules on the value path of window attention, achieving spatial receptive field expansion and cross-window feature interaction. Second, to enhance feature processing capabilities, a locally enhanced gated FeedForward network (LeGN) is proposed, which achieves adaptive feature aggregation through overlapping multi-scale deformable convolution, and a gated unit is designed to optimize the information flow. Thirdly, a dual-branch fusion module (DBF) is introduced in skip-connections of encoder-decoder layers to enhance cross-level feature interaction while effectively mitigating information loss during the downsampling process. Finally, comparative experimental results on three benchmark datasets (CrackLS315, DeepCrack537, and YCD776) with seven advanced networks demonstrate that the proposed network achieves excellent performance, obtaining mean intersection over union (mIoU) scores of 71.26%, 86.58%, and 83.93%, respectively. Code is available at: https://github.com/wjxcsust2024/CSHD.