Extension of Sound Field Image Denoising to High-Frequency Sound Fields by Considering Wavenumber Spectral Loss

Recent development in optical technology enables the capture of sound fields as image sequences. However, these images contain heavy noises derived from measurement devices. Previously, a deep-learning-based denoising method, called deep sound-field denoiser (DSFD), has been proposed to remove noises from sound-field images obtained by optical sound measurement. Although DSFD can remove noises superior to conventional filtering methods, its performance decreases where the frequencies of sound are high. In this paper, we propose a method to improve performance even in high-frequency sound by introducing the wavenumber spectrum obtained by a two-dimensional Fourier transform of the sound-field images. We performed denoising experiments for both simulated and measured data. The results showed that the proposed method was superior to DSFD.