The traditional compressed sensing algorithm achieves high-resolution direction-of-arrival (DOA) estimation through grid partitioning and sparse constraints. However, the grid partitioning typically has a significant impact on the performance of compressed sensing algorithms. Moreover, by minimizing the atomic norm, it is possible to achieve gridless DOA estimation. Nevertheless, the atomic norm minimization (ANM) algorithm has certain limitations, particularly in complex environments where its stability is considerably affected, leading to a decrease in both resolution and accuracy. To enhance the resolution and accuracy of the algorithm, this chapter proposes an atomic norm minimization algorithm based on deep learning. The proposed algorithm preprocesses the received data using an autoencoder and then utilizes a convolutional neural network (CNN) to map the data to the desired Toeplitz matrix. To facilitate network training, we apply weighted summation across multiple subnetworks at the output layer by introducing weighted coefficients. Additionally, we design a corresponding loss function for the algorithm to improve the network’s generalization performance. Experimental results demonstrate that further processing of the atomic norm minimization problem using deep learning can effectively improve DOA estimation resolution and performs well even under low signal-to-noise ratio (SNR) conditions.