Aiming at the common mixed sources including near-field(NF) and far-field(FF), this paper proposes a mixed sources direction-of-arrival(DOA) estimation method based on Deep Learning(DL). By using the real part and imaginary part of the mixed sources as the input of the neural network, after full separation, the expected FF signals and NF signals can be obtained at the output of the network. Then, the traditional estimation methods for a single source can be used to estimate DOA. The proposed method optimizes some defects existing in the estimation of mixed sources containing near and far fields in the past, such as the inability to separate the mixed sources effectively, and a large loss of array aperture. Simulation results show that the proposed method is effective. Under different SNR or channel error conditions, the proposed method has better DOA estimation accuracy and generalization ability than most super-resolution DOA estimation algorithms.