Abstract: To address the challenge of detecting concrete cracks in underwater dams, this study develops a novel method for real-time pixel-level segmentation and quantification of underwater crack images. It adopts a model of symmetric architecture, with skip connections used for transmitting features across layers, and uses a combination of Vision Transformer (ViT) and Convolutional Block Attention Module (CBAM) to fully extract complicated crack features. To mitigate the issue of foreground-background imbalance in the images, we use the depthwise separable convolution to lighten the network and the Focal Tversky Loss to optimize the loss function, thereby enhancing the accuracy of crack area recognition. In a case study of the Guanyinyan hydropower station, we conduct an ablation experiment of underwater crack segmentation and compare different methods, verifying the performance advantage of our model over other models－with the mean Intersection over Union (mIoU), Recall, Precision, and F1 score of 0.913, 0.949, 0.957, and 0.953, respectively. And, we combine regional pixel extraction with the infrared laser ranging technology to quantify the geometric dimensions of the cracks, and achieve the results that agree well with those derived from the annotated masks.

Key words: underwater dam, crack detection, deep learning, pixel-level segmentation, remotely operated vehicle, infrared laser ranging