Sound-field image denoising using deep neural network considering physical characteristics of sound

Abstract

This paper presents a method of denoising utilizing deep neural networks (DNN) in sound field images measured by optical methods. Optical sound measurement has attracted much attention because of its capability to visualize sound field with high spatial resolution from a distance, which is difficult with conventional microphones. However, optical methods have the issue of the measured sound-field images often being heavily distorted by noise. This is caused by the weak phase of light changed by sound. Conventionally, noise-reduction filters considering the physical properties of sound have been known to be used for sound-field denoising. In this background, we propose a DNN-based sound-field denoising method that takes into account the physical properties of sound. In the DNN, a network architecture originally used for denoising natural images is employed, and we integrate loss functions based on the Helmholtz equation. Additionally, a 2D sound field dataset obtained from numerical acoustic simulation with random parameters is used during the training. Numerical and experimental data comparison experiments showed that the proposed DNN-based sound-field denoising outperformed the previous non-DNN methods.