Multimodal Detection for Underwater Cultural Heritage: A Physics-Guided Color-Bias-Aware Network and Dataset
Year: 2026
Type: preprint
Abstract: Reliable object detection is essential for underwater cultural heritage research, yet its development remains limited by the scarcity of domain-specific annotations, underwater color distortion, and attenuated object boundaries. To address these challenges, we construct the Wrecks, Ruins, and Fragmentary heritage (WRF) dataset, comprising 4,732 RGB images collected from multiple underwater archaeological sites, corresponding estimated depth maps, 17,059 annotated instances, and 21 categories of artifacts and ship components. We further develop UCH-Net, a dual-stream RGB-depth detector that integrates appearance and geometric cues through physics-guided color-bias-aware fusion, multi-scale feature aggregation, and decoupled localization. On the WRF, UCH-Net achieves an mAP of 77.5% and an mAP75 of 88.0%, surpassing the representative Deformable-DETR baseline by 12.8 and 14.9 percentage points, respectively, while exceeding the strongest competing results by 1.2% in mAP75 and 1.5% in recall. UCH-Net further achieves an mAP of 66.0% on RUOD. These results demonstrate the potential of the proposed dataset and detector for automated screening and object localization in underwater archaeological imagery.
Source: Research Square
Institution: Xiamen University of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223184081
