Enhancing blind separation of incoherent sources in acoustic imaging with deep learning

Separating multiple incoherent sound sources presents a significant challenge in acoustic imaging. Existing methods, such as principal component analysis (PCA) applied to the cross-spectral matrix, yield "virtual" sources based on decorrelation. However, this approach often fails due to the existence of multiple solutions in the general case. State-of-the-art methods to address this issue involve computing an additional rotation matrix to enforce criteria such as least spatial entropy, or spatial orthogonality, a process that, while effective, significantly increases computational complexity. This work introduces a hybrid approach combining PCA and deep learning for separating source maps from virtual sources. By simulating sound sources in random quantities and locations, a neural network tailored to this task is trained. The permutation problem between PCA-derived virtual sources and pre-simulated labels, as well as the estimation of the number of sources, is addressed by framing source separation as a set prediction problem utilizing the Hungarian loss. This method is array geometry agnostic and frequency agnostic, allowing for robust performance across diverse array configurations and a wide frequency range. Trained solely on simulated data, the model demonstrates effective source separation on real-world datasets, highlighting the potential of integrating deep learning with existing methods.