<p>Deep learning, particularly convolutional neural networks (CNNs), has shown great potential for direction of arrival (DOA) estimation. However, directly inputting the complex-valued covariance matrix into CNNs often poses challenges in convergence and performance, as the network must implicitly learn the complex mapping from the abstract matrix space to the physical angle space. This paper proposes a simple yet effective preprocessing technique to accelerate CNN convergence and improve estimation accuracy in DOA tasks. The core of our method involves complex normalization of the received signal’s covariance matrix followed by a two-dimensional Fourier transform (2D-FT). This operation explicitly transforms the intricate inter-element phase difference information into a structured, image-like representation where source directions are manifested as distinct peaks. The real and imaginary components of the resulting spectrum are then concatenated to form a two-channel input. By aligning the input structure with the innate strength of CNNs in processing spatially local features, our method significantly reduces the learning burden. Experimental results demonstrate that CNNs trained with our proposed input exhibit faster convergence speed, lower training loss, and achieve higher estimation accuracy compared to those using the raw covariance matrix. This work provides a simple and efficient pathway for enhancing learning-based DOA estimators.</p>