Skip to main navigation Skip to search Skip to main content

A Cross-modal Style Diffusion Sonar Image Synthesis Method Based on Physical Guidance and Structural Constraints

  • Qiaoqiao Zhao
  • , Lichuan Zhang
  • , Xuanfeng Li
  • , Guang Pan
  • Northwestern Polytechnical University Xian

Research output: Contribution to journal › Article › peer-review

Abstract

To address the challenge of scarce high-quality annotated data in underwater sonar perception, a cross-modal style diffusion sonar image synthesis method based on physical guidance and structural constraints is constructed for the synthesis of sonar images from optical remote sensing. The existing diffusion model methods are mostly designed for generic optical images. When dealing with the cross-modal task of converting the remote sensing optical images to the sonar images, the acoustic imaging mechanism is often overlooked, which leads to the degradation of acoustic characteristics and the collapse of geometric structures. To tackle this issue, a frequency-domain noise alignment module and an acoustic prior injection module are constructed. These modules embed the texture distribution of real sonar images and the physical priors of highlight-shadow relationships at the initial stage of diffusion denoising. Meanwhile, an attention mechanism with structural constraints is introduced, which implements a differentiated style transfer strategy for target and background regions. This method ensures the generation of realistic acoustic features while preserving the geometric contours and positional structures of targets, thus enabling the direct transfer of labels from optical remote sensing images. Experimental results demonstrate that the proposed method is superior in terms of both acoustic realism and structural integrity. The mAP_50-95 is significantly improved from 0.724 to 0.771 by using the generated data to assist in training the YOLOv8 detector. This study provides a practical engineering solution for alleviating the problem of data scarcity in underwater sonar perception.

Translated title of the contribution物理引导和结构约束的交叉模态风格扩散声呐图像合成方法
Original languageEnglish
Article number251133
JournalBinggong Xuebao/Acta Armamentarii
Volume47
Issue number9
DOIs
StatePublished - 2026

Keywords

  • cross-modal generation
  • data augmentation
  • diffusion model
  • sonar image synthesis
  • underwater sonar perception

Fingerprint

Dive into the research topics of 'A Cross-modal Style Diffusion Sonar Image Synthesis Method Based on Physical Guidance and Structural Constraints'. Together they form a unique fingerprint.

Cite this