基于侧扫声呐的水下沉船自主探测是水下考古的重点研究方向. 水下目标的稀缺性阻 碍了目标检测模型的训练. 为了解决这一问题, 使用基于稳定扩散模型的人工智能生成内容技 术来补充稀缺的侧扫声呐沉船实例, 通过比较多种生成式技术的效果, 论证了人工智能生成内 容技术在水下沉船探测领域的潜力. 基于该技术, 提出一种无需额外光学数据和人工标注的数 据增强方法, 称作自动扩散生成, 可用于实现高精度的水下沉船探测. 基于YOLOv8n, 运用该 方法训练的检测器在沉船探测任务中达到95.0% 的精度和96.3% 的召回率, 超出仅使用原始 数据训练的检测器1.5% 和1.8%; 基于Faster RCNN, 该方法可以同样促进水下探测的效果, 达到94.8% 的精度和97.3% 的召回率.

Autonomous detection of underwater shipwreck based on side scan sonar (SSS) is the key research direction of underwater archaeology. The limited quantity of underwater targets hinders the training of the detectors. To address this issue, we employ artificial intelligence-generated content (AIGC) technology based on stable diffusion model (SDM) to supplement scarce SSS target instances. By comparing the effects of various generative techniques, we demonstrate the potential of AIGC technology in the field of underwater shipwreck detection. Based on this technique, we propose a data enhancement method without additional optical data and manual annotation, called automatic diffusion generation (ADG), which can be used to achieve high precision underwater shipwreck detection. On the YOLOv8n, the detector trained with this method can achieve 95.0% precision and 96.3% recall for underwater shipwreck detection, exceeding 1.5% and 1.8% of the detector trained with only the original data. On the Faster RCNN, this method can also further improve the detector accuracy, achieving 94.8% precision and 97.3% recall.