Deconvoluting acoustic beamforming maps with a deep neural network

Localization and quantification of noise sources is an important scientific and industrial problem, where the use of phased arrays of microphones, known as beamforming, is the standard technique in many applications. However, as non-physical artifacts can appear on output maps, a supplementary step called deconvolution is often performed. While classical deconvolution techniques rely on strong assumptions regarding the environment and the sources, neural network can learn to produce deconvoluted outputs without such explicit assumptions. To do so, information on the acoustic propagation is implicitly extracted from pairs of source-output maps. On this work, a convolutional neural network is trained to deconvolute the beamforming map obtained from synthetic data simulating the response of an array of microphones. Quality of the estimation and the computational cost are compared to those of classical deconvolution methods (DAMAS, CLEAN-SC). Constraints associated with the size of the dataset used for training the neural network are also investigated and presented. Results demonstrate the potential of neural networks in the deconvolution of beamforming maps in simple configurations, paving the way to new robust and effective deconvolution strategies, even in non-ideal acoustic environments