A conditional diffusion-based model for high-resolution acoustic source mapping

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Inverse imaging problems, particularly acoustic source mapping, are challenging due to sparse and peak-shaped source distributions.
• Traditional supervised regression methods struggle to accurately model these complex source structures, often resulting in blurry artifacts.
• Diffusion models offer powerful generative capabilities applicable to inverse problems.

Purpose of the Study:

• To introduce the first diffusion-based framework for acoustic source mapping that directly addresses the deconvolution inverse problem.
• To develop a generative model that explicitly learns the structural prior of source maps, avoiding blurry artifacts.
• To enhance acoustic source mapping accuracy and generalization capabilities.

Main Methods:

• A diffusion-based generative framework conditioned on delay-and-sum beamforming maps and autoencoder-extracted multi-scale point spread function features.
• Utilizing a smoothed target map to guide the model in capturing structural priors.
• Implementing a time-weighted loss function to improve condition exploitation during training.
• Employing an autoencoder to extract frequency-aware features from point spread functions.

Main Results:

• The proposed diffusion model successfully generates high-resolution acoustic source distribution maps with only 20 sampling steps during inference.
• Experimental results demonstrate superior performance compared to traditional and supervised regression-based deep learning methods.
• The framework shows strong generalization across unseen frequencies, varying numbers of sources, and real-world transfer functions.

Conclusions:

• The diffusion-based framework represents a significant advancement in acoustic source mapping, overcoming limitations of previous methods.
• The model's ability to learn structural priors and avoid blurry artifacts leads to more accurate and detailed source maps.
• This approach offers a robust and generalizable solution for acoustic source mapping in complex scenarios.