Answers are grounded in this paper and cite their sources.
Underwater optical images are often degraded by scattering, uneven illumination, and background clutter, making structural defects difficult to distinguish and localize. This study presents SCA-Net, an RT-DETR-r18-based detector for cracks, spalling, and holes in underwater structures. The network incorporates three modifications to the baseline. A dilated reparameterized residual block (DRRB) replaces the second residual convolution in the backbone. An efficient context transformer (ECT) applies single-head additive attention to the deepest feature map, while a detail–context enhancement fusion (DCEF) module combines parallel atrous convolutions with a projected detail pathway in the bidirectional neck. Channel gates are used in the associated resampling modules. On the image-level split of the underwater defect dataset (UDD), SCA-Net achieved a precision of 0.943, a recall of 0.927, an mAP@0.5 of 0.942, and an mAP@0.5:0.95 of 0.638. Compared with RT-DETR-r18, recall, mAP@0.5, and mAP@0.5:0.95 increased by 11.3, 2.5, and 2.1 percentage points, respectively. Among the tested configurations, the full model achieved the highest recall and both mAP metrics. The evaluation was limited to the image-level UDD split. Because the original videos and acquisition-batch records were not retained, independence across acquisition groups and generalization to other inspection conditions could not be verified.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: