In this presentation, I will discuss a deep learning architecture developed for reconstructing full 3D acoustic fields from stereo beamforming images. The approach uses a convolutional neural network trained on synthetic data consisting of monopole and line sources to learn the mapping from 2D acoustic input to 3D volumetric output.
I will present results demonstrating the model's ability to accurately reconstruct complex acoustic fields with low computational cost, achieving a localisation error below 50 millimetres and requiring less than 0.5 GFLOPs per inference. The method is validated on both synthetic and measured data and successfully applied to a UAV tracking scenario, where it achieves precise acoustic trajectory estimation.
This work illustrates the potential of deep learning for real-time acoustic diagnostics, source identification, and environmental noise monitoring.