An Off-Grid DOA Estimation Method Based on a Frequency-Domain ViT

In this letter, we propose a deep learning-based off-grid Direction of Arrival (DOA) estimation method for low Signal-to-Noise Ratio (SNR) scenarios. Specifically, we develop a dual-branch neural network with residual connections that processes frequency-domain features, consisting of a coarse classification branch and a fine regression branch. The classification branch employs a multi-label approach to obtain on-grid results, while the regression branch predicts the residual between the classification outputs and ground-truth angles. This structural design effectively leverages classification results to avoid convergence difficulties associated with direct off grid angle prediction, thereby enhancing DOA estimation accuracy. Simulation results demonstrate that under low SNR conditions, the proposed method outperforms existing approaches, including both classical model-based and other deep learning-based methods.