Direction of Arrival (DOA) estimation is a technique that determines the arrival angle of a signal source by processing data received from an antenna array to locate far-field signal sources, with recent research primarily focusing on optimizing algorithms to enhance computational efficiency and estimation accuracy. Classical subspace-based DOA estimation algorithms, which suffer from high computational complexity and reduced prediction accuracy under conditions of low signal-to-noise ratio (SNR). We propose a novel deep learning-based DOA estimation scheme to improve performance in complex environments. In this scheme, the upper triangular elements of the time-domain and frequency-domain covariance matrices are extracted and converted into real-valued vectors, respectively. These real-valued time-domain and frequency-domain vectors are then combined to construct the final input features. Based on this, specialized convolutional neural networks (CNNs) are designed for high-SNR and low-SNR scenarios to optimize the performance under varying SNR conditions. Experimental results demonstrate that the proposed method significantly reduces estimation errors compared to traditional DOA estimation techniques and exhibits robust applicability across various SNR conditions.