End-to-End Regression Neural Network for Coherent DOA Estimation with Dual-Branch Outputs

In this paper, a regression neural network is proposed instead of the commonly used classification neural network to address the direction of arrival (DOA) estimation problem. Unlike the method treating the DOA estimation problem as a classification task, our proposed method utilizes a regression neural network to directly estimate the signal angles by mapping the input data to the high-dimensional features. It is an end-to-end method where the input data is the signal covariance matrix estimated from the received signals by the uniform linear array. The network output consists of two branches: one branch provides the estimated angles of signals, and the other branch provides the confidences corresponding to angles, indicating the probabilities that signals are incident at the estimated angles. Owing to the dual-branch output of our proposed method, it is adept at accommodating a variable number of sources without the requirement to alter the network structure. Compared to classification neural networks, our proposed method circumvents the issue of quantization errors and has the capability to handle the imbalance of positive and negative samples. Simulation results validate the superior performance of our proposed method in terms of estimation success rate, angular resolution, and estimation accuracy, especially in scenarios with a variable and unknown number of coherent signals.