Deep sound-field denoiser: optically-measured sound-field denoising using deep neural network.

Area of Science:

• Acoustics
• Optical Physics
• Machine Learning

Background:

• Optical methods offer high-spatial-resolution sound-field imaging beyond conventional acoustic sensors.
• Optical interferometric measurements suffer from low sensitivity, leading to significant noise in sound-field images.
• Existing denoising techniques struggle to effectively remove noise from these complex optical measurements.

Purpose of the Study:

• To develop a deep neural network (DNN) based method for denoising optically measured sound-field images.
• To enhance the quality and accuracy of sound-field imaging for acoustic phenomena.
• To provide a robust solution for noise reduction in challenging optical acoustic measurements.

Main Methods:

• Time-varying sound-field image sequences were decomposed into harmonic complex-amplitude images using a time-directional Fourier transform.
• Complex images were converted into two-channel real and imaginary parts for processing.
• A nonlinear-activation-free deep neural network was trained on simulated acoustic data for denoising.

Main Results:

• The proposed deep sound-field denoiser demonstrated superior performance over conventional image filters and spatiotemporal filters.
• The DNN-based method significantly outperformed other deep learning architectures in denoising tasks.
• Validation on both numerical simulations and experimental data (parallel phase-shifting interferometry, holographic speckle interferometry) confirmed effectiveness.

Conclusions:

• The proposed deep sound-field denoiser offers a significant advancement in noise reduction for optical sound-field imaging.
• This DNN-based approach overcomes the limitations of conventional methods, enabling clearer acoustic visualizations.
• The study provides a valuable tool for researchers and practitioners in acoustics and optical metrology.