Sound-field-projection synthesis using latent diffusion model for acousto-optic reconstruction.

Area of Science:

• Acoustics
• Optical Sensing
• Artificial Intelligence

Background:

• Acousto-optic sensing (AOS) offers non-contact sound measurement but provides line-integrated data, hindering 3D sound field interpretation.
• Current volumetric sound-field reconstruction methods require extensive multi-directional data, often necessitating multiple devices or reproducible sound sources.

Purpose of the Study:

• To develop a method for synthesizing sound projection data to overcome limitations in existing 3D sound field reconstruction techniques.
• To improve the accuracy and practicality of volumetric sound-field reconstruction using acousto-optic sensing.

Main Methods:

• Proposed sound projection synthesis using a latent diffusion model conditioned on observed projections and view angles.
• Fine-tuned a pretrained image model on sound-field data, optimizing with pixel-wise loss.
• Integrated synthesized projection data with observed data for 3D reconstruction.

Main Results:

• The latent diffusion model successfully generated realistic sound-field projection data.
• Combining nine observed views with nine synthesized views significantly improved 3D reconstruction accuracy compared to using only observed views.

Conclusions:

• Sound projection synthesis is a viable method to enhance volumetric sound-field reconstruction from acousto-optic sensing data.
• The proposed AI-driven approach offers a more practical and accurate solution for 3D sound field mapping in challenging environments.