Abstract: To address the challenge of detecting cracks in concrete on underwater dams, this study proposes a method for real-time, pixel-level segmentation and quantification of concrete cracks on underwater dams. The model employs a symmetric architecture with skip connections for feature propagation across layers. By combining ViT and CBAM, it effectively extracts complex underwater crack features. Deep separable convolutions are used to streamline the network, and the Focal Tversky Loss is selected to optimize the loss function, addressing the issue of background imbalance around underwater cracks and enabling accurate identification of crack regions. Underwater crack segmentation experiments were conducted on a gravity dam project. Compared to models such as U-Net, U-Net++, FCN, and DeepLabv3+, the proposed method achieved the best segmentation performance, with mIoU, Recall, Precision, and F1_score values of 0.913, 0.949, 0.957, and 0.953, respectively. Additionally, by combining regional pixel extraction with infrared laser ranging technology, the geometric dimensions of the cracks were quantified, and the resulting quantitative results showed good agreement with those obtained using annotated masks.