Estimating the Number of Sources via Deep Learning

This paper presents a deep-learning-based approach to estimating the number of sources in a narrowband angle of arrival (AoA) estimation problem where all signals are coherent. This is an important problem in radar, sonar and communication systems, as many AoA estimators require accurate estimates of the number of sources. Herein, a 15-layer deep learning network with parametric rectified linear units and batch normalization layers is trained using the eigenvalues and the spatially smoothed covariance matrix entries as an input. Our initial results show that the proposed learning based system produces more accurate of estimates of the sources compared to the state-of-the-art techniques.