Abstract: Acquiring high-quality underwater crack images and achieving efficient identification and quantification are crucial for enhancing dam inspection efficiency. To address the challenges associated with underwater image degradation and crack quantification, this study develops a deep learning and transfer learning-based method for underwater image enhancement and crack analysis. We construct a new platform for underwater imaging and data acquisition, and develop a conditional diffusion model using public marine image datasets as prior knowledge for cross-domain multi-source enhancement. Crack detection is performed using YOLOv12, followed by morphological operations for feature quantification. Experimental results demonstrate our method significantly outperforms conventional approaches in terms of visual quality, no-reference metrics, and pixel allocation. The integrated detection model improves accuracy while reducing missed detections, and the quantification method extracts crack parameters effectively. The enhancement-identification-quantification closed-loop framework developed in this study is an effective technical solution to intelligent underwater inspections.

Key words: underwater structures, underwater cracks, image enhancement, crack quantification, diffusion model