Sound Source Localization Method Based on Densely Connected Convolutional Neural Network

Abstract

Beamforming and deconvolution techniques play a significant role in identifying sound sources. However, it is commonly known that conventional beamforming (CBF) is difficult to identify sound sources accurately due to its inherent drawbacks, including low spatial resolution and small dynamic ranges, while deconvolution methods are burdened with huge computational costs and fail to provide reliable results. Aiming to overcome the restrictions of conventional beamforming and deconvolution methods, a novel sound source localization method combining conventional beamforming with a deep learning method is proposed. In this paper, the sound source localization task is framed as an image prediction task. Firstly, conventional beamforming (CBF) is utilized for obtaining the initial sound source spatial distribution maps. Secondly, a target map is designed as the ground truth label for training. Then a densely connected convolutional neural network (DCFCN) with an encoder-decoder structure is established for extracting features from CBF maps and predicting the spatial distribution of a single sound source. Finally, the position of a single sound source can be retrieved from the predicted maps generated by DCFCN. Simulations are carried out to verify the effectiveness of the proposed method by comparing it with several traditional sound source localization methods. Results suggest that the proposed method can not only significantly improve the spatial resolution and dynamic range of CBF but also achieve accurate localization with high computational efficiency.