Aimed at the problems thatthe existing Uniform Linear Array (ULA) far-field narrowband non-coherent multi-target estimation algorithms is poor in adaptability to low Signal-to-Noise Ratio (SNR), small snapshots in adaptability, high in computational complexity, and existing Deep Learning (DL) approaches are difficult to effectively extract the complex-valued features of data,a Direction of Arrival (DOA) estimation method based on Deep Convolution Neural Network (DCNN) is proposed . This method is to transform the DOA estimation problem into an inverse mapping problem from the array output covariance matrix to the target DOA, and to utilize the Hermitian characteristic of the array output covariance matrix for extracting the real part, imaginary part, and phase characteristics of an upper triangular array,building input data of a network, and building a deep convolutional neural network with a three-dimensional convolution layer to extract data features, and the labels of the network correspond to the DOAs,realizing the DOA estimation of multiple sources. Theexperimental simulations show that the method can fully extract spatial features, improve DOA estimation accuracy and reduce the complexity of the algorithm.Under condition of low SNR and small snapshots, the estimation accuracy of the proposed method is significantly better than that of the MUSIC, the ESPRIT and the ML algorithms.