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

Diffusion models have recently shown strong generative capabilities in inverse imaging problems. This paper introduces the first diffusion-based framework for acoustic source mapping that directly solves the deconvolution approach for the mapping of acoustic sources inverse problem. Supervised regression-based learning methods cannot well model the sparse and peak-shaped source distributions, but the proposed generative model explicitly learns the structural prior of source maps and thus avoids blurry artifacts in the output map. During training, the diffusion model is conditioned on both the delay-and-sum beamforming map and multi-scale point spread function features extracted by an autoencoder. The beamforming map provides coarse spatial cues on source positions and strengths, while the point spread function features provide frequency-aware information. The target map is a smoothed form of sparse source labels to help the model capture structural priors, and a time-weighted loss is proposed to help model better exploit the conditions. During inference, the model can generate high-resolution source distribution maps in only 20 sampling steps. Experimental results on three generalization tasks, i.e., unseen frequencies, unseen numbers of sources, and real-world transfer functions, demonstrate that the proposed method outperforms existing traditional and supervised regression-based deep learning-based approaches.