Deep Learning-Based Super-Resolution Estimation of DOA at Low SNR

Aiming at the low signal-to-noise ratio (SNR) environment in which the performance of traditional model-based super-resolution direction-of-arrival (DOA) estimation algorithms will be seriously degraded, this paper proposes a deep learning-based super-resolution DOA estimation neural network. To solve the problem of excessive computing power of the classification task, it is built as a regression task, which makes the DOA estimation grid-free, reduces the computing power more significantly, and improves the accuracy. The program first performs data preprocessing of the covariance matrix as an input to the network and builds the corresponding dataset at a low SNR. Then, a convolutional residual attention grid-free DOA estimation network (CRA) is proposed. Finally, a comparison experiment with the traditional algorithm is conducted. The experimental results show that compared with the model-based DOA estimation algorithms, the CRA network can greatly reduce the computational effort and still maintains good performance of super-resolution DOA estimation in low SNR environments.