Acoustic source imaging using densely connected convolutional networks

Pengwei Xu

doi:10.1016/j.ymssp.2020.107370

2021-04-01

PRE

DOI：10.1016/j.ymssp.2020.107370

原文链接

原文求助

摘要

En 中文

The localization of acoustic sources using acoustic imaging methods such as acoustic beamforming can be limited by the Rayleigh resolution limit at relatively low frequencies and the output of these methods may also produce spatial aliasing images referred to as sidelobes, particularly at high frequencies. To date, there are very few Deep Neural Network (DNN) applications for acoustic imaging to help alleviate some of these issues. In this study, several DNN models were developed with a training strategy specifically designed for an acoustic imaging task. This proposed DNN-based method is examined for various acoustic source conditions and compared with other classic acoustic imaging methods. The DNN method is able to recognize the pattern behind microphone array signals for different source positions, by using the real-component of the cross-spectral matrix of the received pressure vectors at the microphone positions. This input feature allows the DNN model to accurately locate and quantify the source strengths of complicated distributed acoustic sources, even at low frequencies that typically challenge acoustic beamforming deconvolution methods. The loss function of the DNN model is based on the difference between the estimated and true acoustic source maps, that is used to iteratively improve weighting functions within the DNN hidden layers. DNN models with both a fixed and random number of input sources are simulated for a range of specific frequencies. The DNN model with a random number of input sources is tested against conventional beamforming, CLEAN-SC and DAMAS, revealing a far improved source localization and source strength estimation. These DNN models represent a very promising proof-of-concept for the use of DNN models in the field of acoustic imaging. (C) 2020 Elsevier Ltd. All rights reserved.

Keyword:

Deep learning

Neural networks

Acoustic beamforming