Underwater dam-crack images captured by remotely operated vehicles (ROVs) often exhibit blurred and unclear features because of the absorption and scattering of water and device shaking. To address these challenges, we propose an underwater crack image-enhancement network (UCE-CycleGAN) that outperforms existing methods on unpaired crack datasets. This network employs multi-feature fusion, incorporating edge maps and dark channels to mitigate blurring and enhance texture details. The blurriness metric and number of matched feature points in the enhanced images improved by 4.12 and 17.11 times, respectively. The mean average precision for detection and segmentation increased by 58.5 % and 39.0 %, respectively, surpassing those of other underwater image-enhancement methods. Furthermore, by integrating YOLOv8 feature recognition and a semantic segmentation network, we developed a real-time automated system (UCADS) for underwater dam-crack detection, thereby expanding the applications of enhanced underwater crack images. The effectiveness of UCADS was validated through field experiments at reservoir dams.