LGCF Net: A CNN–Transformer Local–Global Fusion Network for Segmenting Deep-Opening Defects at Material Interfaces in Mortared-Stone Levee Revetments

Mortared-stone revetments on the waterside slopes of levees are continuously exposed to flow erosion, environmental variation, and material deterioration during long-term service. Fracture and spalling can therefore occur at the interfaces between stones and cementitious materials, producing anomalously deep openings. Such defects have irregular shapes and marked scale variations and are readily confused with normal joints, shadows, and dark textures. To address these challenges, this study proposes LGCF Net, a semantic segmentation network based on CNN–Transformer local–global feature fusion. ResNet50 and Swin Transformer Tiny are used to extract local texture features and global semantic features, respectively. A Local–Global Context Fusion (LGCF) module adaptively fuses the two feature types, after which a feature pyramid network performs multiscale decoding and defect prediction. LGCF Net achieved a precision of 0.9008, recall of 0.8125, Dice coefficient of 0.8543, and intersection over union (IoU) of 0.7457. Compared with Swin-T+FPN, its Dice and IoU improved by 0.0208 and 0.0311, respectively. Under the current dataset and unified experimental protocol, LGCF Net showed competitive performance relative to CNN, Transformer, and other mainstream semantic segmentation models. The method effectively combines local details with global contextual information and supports two-dimensional candidate-region segmentation of these defects in complex backgrounds.