UCD-LFE: An Underwater Crack Detection Model for Hydraulic Structures in Complex Environments via Lightweight Design and Feature Enhancement

Underwater crack detection in hydraulic structures is difficult in complex underwater environments, where blur, uneven illumination, low contrast, and rough surface textures can weaken crack cues and introduce background interference, while multiple cracks further increase detection complexity. Practical inspection also requires compact models with low computational demand. To address these coupled challenges, this study develops UCD-LFE, an underwater crack detection model combining lightweight design with feature enhancement. StarNet and a shared-convolution detection head reduce redundant computation, while C2f_Star_CAA enhances crack feature representation during multiscale fusion. The model was evaluated using ROV-acquired images from an operating hydraulic project and a separate held-out set of 93 complex-scene images excluded from model development. UCD-LFE achieved 80.01% mAP@0.5 while reducing computation from 8.1 to 5.5 GFLOPs and model size from 6.0 to 3.8 MB relative to YOLOv8n. On the complex-scene set, mAP@0.5 reached 78.49%, 7.90 percentage points above the baseline, with substantially less performance degradation. Feature-response visualizations covering a clear reference and four challenging scene types—blur and uneven illumination, low contrast, rough texture, and multi-crack scenes—showed more continuous and spatially concentrated responses along crack regions, with fewer dispersed background responses. These results demonstrate that lightweight design and feature enhancement can improve robust crack detection in complex underwater environments while reducing model complexity.