DCDGNet: Dual-frequency cross-feature diffusion GAN for single fusion sonar image generation in exposed subsea...

The exposed subsea pipelines are prone to damage and rupture, increasing the risk of marine pollution; therefore, their safety inspection is imperative. Multibeam Echo Sounder (MBES) and Side-Scan Sonar (SSS) are widely utilized in this task, and their fusion data improves inspection accuracy. However, the limited availability of fusion sonar images with exposed pipelines constrains deep learning performance. Although image generation can ease data scarcity, traditional methods still require large annotated datasets. To address, this paper proposes a Dual-Frequency Cross-feature Diffusion GAN (DCDGNet) for generating fused sonar images of subsea pipelines. The model can generate diverse samples from a single image. To avoid artifacts from spatial modeling, frequency decomposition is applied. Three key modules are designed: the Dual-Frequency Cross-Transformer Fusion Module (DF-CTFM) for global semantic fusion, the Dual-Frequency Collaborative Enhancement Module (DF-CEM) for local detail enhancement, and the Feature Diffusion Module (FDM) for multi-stage feature diffusion. In addition, a label synchronization mechanism enables automatic annotation during image generation process. Evaluations using Fréchet Inception Distance (FID), Kernel Inception Distance (KID) and Inception Score (IS) show that DCDGNet clearly outperforms existing methods in structural fidelity and detail preservation. It achieves a Fréchet Inception Distance of 231.96, a 25.4 percent reduction compared with the second-best method, a Kernel Inception Distance of 0.34, a 21 percent reduction, and an Inception Score of 2.93, a 10.6 percent improvement. Furthermore, UNet++ segmentation shows that generated samples perform comparably to real ones, and data augmentation further enhances exposed pipeline segmentation. Thus, DCDGNet improves detection accuracy and reduces marine pollution risk.