A Novel DOA Estimation Method for High Resolution Based on Convolutional Neural Network

A method based on convolutional neural network is applied for estimating Direction-of-Arrival (DOA) for high resolution. Firstly, the received 2-D temporal-spatial signals are pre-processed into visualized images and the phase difference induced from various incident angles is transformed into texture difference in the image domain. Then, a convolutional neural network (CNN) is consecutively applied to extract texture features from these images to conduct a DOA estimation. According to the analysis of the relationship between texture features and incident angles of signals, the resolution is defined quantitatively for the proposed method by referring to the traditional definition in peak-searching DOA algorithms. Simulation experiments show that the proposed method can achieve a higher resolution than traditional methods and is more robust in low SNR situations.