军事知识服务平台
基于分层残差Transformer的声呐图像生成
Hierarchical residual Transformer for sonar image generation
摘要 针对海底底质侧扫声呐图像数据稀缺问题,文章提出一种基于分层残差 Transformer(Hierarchical Residual Transformer,HR-Transformer)的声呐图像生成模型.该模型通过层次化注意力机制与多尺度特征融合策略,旨在生成视觉与结构高保真的声呐图像.在实测侧扫声呐海底底质数据集上的实验表明,HR-Transformer实现了微观细节与宏观结构真实性的平衡,生成图像的峰值信噪比达 33.96 dB,结构相似性达 0.94,较基准模型分别提升了15.4%和 16.0%.频域损失稳定在 0.015以下,感知损失为 0.01,验证了模型对海底底质侧扫声呐图像特征的高效学习能力.该框架在微观纹理(砂纹连续性)与宏观结构(礁石孔隙分布)的物理一致性上均显著优于现有方法.文章通过量化指标验证,为海底底质图像生成提供了一种新范式,并为多模态融合(如引入跨模态注意力机制整合多波束声呐与 LiDAR数据构建几何-纹理复合地图)及动态环境建模(结合 Transformer-LSTM时序模块捕获海底演化规律)等未来研究方向奠定了技术基础.
Abstract:
To address the scarcity of side-scan sonar image data for seabed sediments,this study proposes a sonar image generation framework based on a hierarchical residual Transformer(HR-Transformer)model.The model employs a hierarchical attention mechanism and a multi-scale feature fusion strategy to generate sonar images with high fidelity in both visual appearance and structural detail.Experiments on a self-acquired dataset of side-scan sonar seabed sediment images demonstrate that the HR-Transformer achieves a balance between the authenticity of microscopic details and macroscopic structures.The generated images attain a peak signal-to-noise ratio(PSNR)of 33.96 dB and a structural similarity index(SSIM)of 0.94,representing improvements of 15.4%and 16.0%over benchmark models,respectively.The frequency domain loss remains stable below 0.015,and the perceptual loss is as low as 0.01,validating the model's efficient learning of the characteristics of seabed sediment side-scan sonar images.The framework significantly outperforms existing methods in terms of physical consistency,both in microscopic texture(e.g.,continuity of sand ripples)and macroscopic structure(e.g.,pore distribution of reefs).Validated by quantitative metrics,this study provides a new paradigm for seabed sediment image generation and lays a technical foundation for future research directions such as multimodal fusion(e.g.,integrating multibeam sonar and LiDAR data using cross-modal attention mechanisms to construct geometric-textural composite maps)and dynamic environment modeling(e.g.,combining Transformer-LSTM temporal modules to capture seabed evolution patterns).
刊 名 声学技术 ISTICPKU
年,卷(期): 2026, 45(5)
分类号: O427
机标分类号: TP391.41 P229 TN911.73
在线出版日期: 2026年11月4日
基金项目: 059GJHZ2022022MI:中国科学院海外科教合作中心部署项目BSA202504:中国科学院声学研究所自主部署项目,