Deep learning for beamforming map deconvolution based on u-net architecture

Overview & Synopsis

The beamforming algorithm has established itself over the years as one of the most reliable solutions for sound source localization. It produces acoustic maps that enable the characterization of sound sources for each frequency of interest. However, due to limitations related to microphone array configurations, these maps often contain secondary lobes that can be mistakenly interpreted as nonexistent sources. To address this issue, deconvolution methods have been developed as a complement to acoustic beamforming. More recently, advances in deep learning have demonstrated that neural networks are increasingly capable of effectively processing acoustic images. This article presents a UNet model trained exclusively on synthetic data generated from a shape function designed to replicate acoustic fields. Despite being trained only on numerical data, the model achieved remarkable results when applied to experimental beamforming maps. The proposed approach demonstrated high accuracy in both source localization and sound level estimation, showing promising performance for real-world applications.