A Dilated Inception Convolutional Neural Network for Gridless DOA Estimation Under Low SNR Scenarios

This paper addresses the direction-of-arrival (DOA) estimation-based source localization problem by using the convolutional neural network (CNN) and root-MUltiple SIgnal Classification (MUSIC) technique. Existing grid-less neural network-based approach employs a LeNet-based CNN, where its network complexity depends on the number of sensors. To overcome this issue, we propose a LeDIM-net CNN that works for a uniform linear array with an arbitrary number of sensors. The proposed LeDIM-net architecture maintains spatial resolution throughout the network while exploiting non-local spatial information. Simulation results demonstrate the effectiveness of the proposed LeDIM-net over the existing grid-less LeNet-based approach and root-MUSIC at low SNRs for arrays with different sensors by maintaining the same network complexity.