Direction-of-Arrival Estimation Using Deep Learning With Covariance Matrix Reconstruction Under Limited Snapshots

Under low-snapshot conditions, traditional direction-of-arrival (DOA) estimation suffers from covariance instability, while existing deep learning methods rely on complex architectures. This letter proposes a hybrid approach that combines model-driven and data-driven theories to strike a balance between estimation performance and computational cost. We reconstruct a structured covariance matrix by applying adaptive diagonal loading. The reconstructed matrix is then transformed into a two-channel input and fed into the proposed squeeze-and-excitation multi-scale deep convolutional network (SE-MSDCN). DOA estimates are obtained via a sub-grid peak interpolation strategy. The experimental results and our analysis validate the efficiency and superiority of the proposed method.