Under low snapshot conditions, a novel DOA estimation method that integrates covariance matrix reconstruction with deep learning is proposed in this letter. We reconstruct a structured covariance matrix using a reference-auxiliary subarray model combined with diagonal loading. The reconstructed matrix is 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 detection strategy. Simulation results demonstrate that the proposed approach significantly outperforms traditional methods and existing deep learning techniques in terms of accuracy and resolution, particularly under low snapshot and SNR conditions.