SODA: Structure-Guided Diffusion and Auto-Labeling for Synthetic Side-Scan Sonar Data Generation
IEEE Access · 2026 · 14 · 110671-110682
DOI: 10.1109/access.2026.3714864 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Hyerin Hwang und 2 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: “Development of Maritime Domain Awareness Technology for Sea Power Enhancement”, Korea Institute of Ocean Science and Technology.
Abstract (Originalsprache)
Robust interpretation of side-scan sonar imagery benefits from object information together with surrounding structural cues, but constructing large-scale annotated datasets remains challenging due to the high cost of expert labeling. We present SODA (Structure guided–Object aware Diffusion with Auto-labeling), a synthetic dataset generation pipeline for side-scan sonar imagery. SODA first derives object and shadow masks from real sonar images using SAM3. A Stable Diffusion-based generator is then adapted via LoRA to better capture sonar-specific textures and noise patterns, while ControlNet uses the extracted masks as conditioning inputs for structure-guided image synthesis. Multi-scale cross-attention maps obtained during generation are further analyzed to automatically localize objects and shadow regions, which are converted into COCO-format bounding boxes and segmentation labels. Experiments on object detection show that training with the generated dataset reduces the performance gap relative to training with limited real data while maintaining stable learning behavior. The proposed framework provides a scalable and automated approach for constructing annotated synthetic side-scan sonar datasets.
Förderung
- “Development of Maritime Domain Awareness Technology for Sea Power Enhancement”
- Korea Institute of Ocean Science and Technology
Zitieren
APA
Hwang, H., Im, S., & Oh, J. (2026). SODA: Structure-Guided Diffusion and Auto-Labeling for Synthetic Side-Scan Sonar Data Generation. IEEE Access, 14, 110671-110682. https://doi.org/10.1109/access.2026.3714864
Vancouver
Hwang H, Im S, Oh J. SODA: Structure-Guided Diffusion and Auto-Labeling for Synthetic Side-Scan Sonar Data Generation. IEEE Access. 2026;14:110671-110682. doi:10.1109/access.2026.3714864
BibTeX
@article{Hwang2026,
author = {Hwang, Hyerin and Im, Seung-Uk and Oh, Jung-Hee},
title = {SODA: Structure-Guided Diffusion and Auto-Labeling for Synthetic Side-Scan Sonar Data Generation},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {110671--110682},
doi = {10.1109/access.2026.3714864}
}
EUVIMED