Super-Resolution Direction of Arrival Estimation Based on Deep Neural Networks

The problem of estimating the arrival direction of signal source has been a main research area in the field of signal processing in recent decades. By avoiding the matrix inversion operation which is required in traditional subspace-based algorithms, many researches have confirmed that deep neural network (DNN) trained with input output pairs is a powerful tool for achieving DOA estimation in previous works. It has been shown that the DNN approach has lower computational complexity and faster convergence rate than the traditional subspace-based high-resolution DOA estimation method. However, the estimation accuracy of existing DOA estimation techniques based on DNN has limitations due to discretization of the spatial domain. Specifically, the estimation error approaches one half of the grid size when locating source impinges on the boundary between two grids. Therefore, this paper proposes a method of superposing several DNNs whose borders appear at different angles to address this problem. Simulation results corroborate that the appropriate designed DNN architecture can achieve higher DOA estimation accuracy compared with the conventional machine learning methods based on DNNs.