Abstract: Cracking, as one of the most common defects in concrete dams, weakens the integrity and durability of dam structures; crack detection has been a crucial task in the operation and maintenance management of concrete dams. Aimed at the drawbacks of traditional image-processing techniques in crack detection－such as substantial manual intervention and limited generalization ability, this paper presents a semantic segmentation model of dam cracks that incorporates multi-scale features and attention mechanisms. This model uses ResNet-50 as its backbone network for integrating the Path Aggregation Network to recycle shallow features, and makes use of the mechanisms of channel attention and spatial attention. These mechanisms enhance the model's ability to identify critical features, thus effectively improving its segmentation accuracy. Then, based on its semantic segmentation results, the digital image technology is adopted to quantify the geometric characteristics of cracks, including area, length, average width, and maximum width. Tests on a crack image dataset show this new model achieves a crack segmentation Intersection over Union of 82.02% and an F1 score of 90.12%; Quantification results of geometric characteristics exhibit an excellent agreement with the real values and a satisfactory accuracy. Thus, our method demonstrates significant potential for application in crack detection and geometric characteristics quantification for concrete dams.

Key words: concrete dam operation and maintenance, crack segmentation, geometric feature quantification, deep learning, attention mechanism