Global Filter Convolutional Network for Near-Field Source Localization

Localizing near-field sources is a challenging topic for next generation wireless communication systems. This paper proposes a novel global filter convolutional network (GFCN) for near-field source localization. The proposed GFCN firstly expands the covariance matrix through a convolutional layer, which effectively captures the spatial characteristics of the received signals. Subsequently, several global filter convolutional blocks are employed to enhance the global features of the received signal vectors by learning the correlation in the frequency domain, which ensure a comprehensive representation of the signal features. Finally, the GFCN outputs the source information through linear classification layers. Particularly, in these blocks we use convolutional layers instead of fully connected layers to strengthen the structure information of the covariance matrix, and put the activation function after the residual connection to improve the ability of representation. The simulation results show that the GFCN achieves excellent performance under challenging conditions, including low signal-to-noise ratio (SNR), limited snapshots, correlated sources and sparse arrays, outperforming existing traditional methods and some deep learning methods.