@article {WU:2024:0736-2935:7596,
title = "Deep learning-enhanced blind separation of incoherent and spatially disjoint sound sources",
journal = "INTER-NOISE and NOISE-CON Congress and Conference Proceedings",
parent_itemid = "infobike://ince/incecp",
publishercode ="ince",
year = "2024",
volume = "270",
number = "4",
publication date ="2024-10-04T00:00:00",
pages = "7596-7607",
itemtype = "ARTICLE",
issn = "0736-2935",
url = "https://ince.publisher.ingentaconnect.com/content/ince/incecp/2024/00000270/00000004/art00066",
doi = "doi:10.3397/IN_2024_3981",
author = "WU, Xian and ROGER, Boustany and BOULEY, Simon and MINCK, Olivier and DENAYER, Herv{\’e} and ANTONI, J{\’e}r{\^o}me and GRYLLIAS, Konstantinos",
abstract = "Separating incoherent sound sources within complex acoustical fields presents a significant challenge in acoustic imaging. Existing methods, such as Principal Component Analysis (PCA) applied to the Cross-Spectral Matrix (CSM), yield 'virtual' sources based on statistical orthogonality.
However, this approach often fails to accurately identify distinct physical sources, primarily due to its reliance on statistical orthogonality solely. A state-of-the-art method involves computing a rotation matrix to enforce criteria such as least spatial entropy, or spatial orthogonality
among sources, a process that, while effective, significantly increases computational complexity and time. This work introduces a hybrid approach combining PCA and deep learning for predicting spatially disjoint source maps from virtual sources. By simulating sound sources in random quantities
and locations, we train a neural network tailored to this task. We address the order mismatch between PCA-derived virtual sources and pre-simulated labels by framing source separation as a set prediction problem, utilizing the Hungarian loss for efficient mismatch resolution. This method simplifies
the separation process, offering faster post-training computations and eliminating the need for complex optimizations. Validation in simulated environments and real-world datasets has shown the model's effectiveness in source separation for acoustic imaging, indicating the potential of integrating
deep learning with existing methods.",
}