A framework for automatic Real-Time Pixel-Level segmentation of underwater dam concrete cracks utilizing the CRTransU-Net model

Abstract

To address the challenges in detecting concrete cracks in the underwater sections of hydropower station dams, which are prone to interference from water disturbances and light refraction, a real-time pixel-level automatic segmentation framework for underwater dam concrete cracks is proposed. The architecture adopts a symmetric network structure with skip connections across network layers to enhance feature transmission. A combination strategy of ViT and CBAM is employed to extract complex crack effectively features underwater. Lightweight optimization of the network is achieved by integrating channel pruning and Knowledge Distillation techniques. Additionally, Dice Loss is used to optimize the loss function, overcoming the imbalanced foreground and background issues in underwater crack segmentation. The proposed CRTransU-Net model demonstrates the accurate identification of underwater crack regions. Based on an experimental study conducted on an RCC gravity dam project, the method achieved optimal segmentation performance compared to models such as U-Net, U-Net++, FCN, and DeepLabv3+. The model’s mIoU, Recall, Precision, F1-score, PA, and SM values are 0.90127, 0.95867, 0.95449, 0.94676, 0.94218, and 0.92887, respectively. Furthermore, the geometric dimensions of cracks were quantified by combining regional pixel extraction with infrared laser ranging technology. The quantitative results obtained from the predicted masks fit well with those derived from annotated masks.