Research on Constructing Forward-Looking Sonar Datasets Based on Physics-Based Simulation Models

Addressing the scarcity of real annotated data, high acquisition costs, and labeling difficulties in underwater forward-looking sonar image target detection tasks, this paper proposes a synthetic data generation method for forward-looking sonar images based on physical model simulation. By establishing sonar equations, acoustic ray propagation models, and target scattering models, the entire process of acoustic wave emission, propagation, and echo reception is simulated. Coherent speckle noise and reverberation effects are synthesized to generate realistic synthetic sonar images. This method supports automated generation of YOLO-format annotated data, facilitating the construction of large-scale, multi-category forward-looking sonar image target detection datasets. Experimental results demonstrate that the generated synthetic data exhibits high visual similarity to real forward-looking sonar images. The trained YOLOv8 model can accurately identify the positions of targets in simulated sonar images and provide corresponding bounding box localizations, validating the feasibility and potential of physically simulated data for supporting deep learning model training.