Efficient Global–Local Context Fusion with Mobile-Optimized Transformers for Concrete Dam Crack Inspection

To address the difficulties in characterizing fine crack morphology, the limitations of detection accuracy, and the challenge of real-time deployment caused by large model parameter counts in concrete dam crack detection, this paper constructs DamCrackSet-1K, a high-resolution dataset with pixel-level annotations covering multiple crack scenarios; proposes a lightweight semantic segmentation framework, MTC-Net, which integrates a MobileNetV2 encoder with Enhanced Transformer modules to achieve global–local feature fusion and enhance feature extraction; and designs a geometry-sensitive Curvature-Aware loss function to effectively mitigate pixel-level class imbalance for fine cracks. Experiments show that, while significantly reducing the number of model parameters, the method greatly improves crack detection accuracy and inference speed, providing a feasible solution for efficient, real-time crack detection in dams.