AquaVis-Det: A degradation-aware defect detection framework for reinforced-concrete underwater bridge piers

Underwater bridge pier defect detection is limited not only by poor image quality but also by the spatially non-uniform loss of task-relevant edge, texture, and contrast cues. Existing enhancement-before-detection pipelines optimize visual appearance independently of recognition and therefore do not ensure improved defect observability. This study formulates underwater perception as a detection-driven visibility-compensation problem and proposes AquaVis-Det. A degradation perception unit learns a task-oriented visibility-degradation map under image-prior and detection supervision. This map guides multi-branch compensation of defect-sensitive features, while task-guided cross-attention and visibility-aware query refinement integrate the compensated information into detection. Experiments on a source test set and independent field data from previously unseen remotely operated vehicle inspection locations consistently demonstrate improved defect recognition and direct-transfer performance. Compared with a unified transformer detector baseline, AquaVis-Det improved recall and average precision at 0.50 intersection-over-union by 8.3 and 8.9 percentage points on the source test set, and by 6.4 and 12.4 percentage points on the external field set. It also achieved a model-forward throughput of 31.19 frames per second on the onboard Jetson Orin Nano. These results establish task-oriented visibility compensation as an effective strategy for detection-oriented underwater perception and embedded bridge-pier inspection.